WorksComputer Science
A Full-Ledger Counterfactual Audit of Adaptive Reasoning
Abstract
Adaptive reasoning must decide whether to answer now or spend more compute. Predicting that the current answer is wrong does not establish whether further reasoning will improve it. We introduce the Full-Ledger Counterfactual Value Audit (FLCVA), which compares an immediate answer with four fixed-length continuations from the same frozen prefix under one answer procedure and accounts for the cost of every signal used by a policy. On GSM8K and MATH-500, a random 256-token continuation recovers a wrong answer in 26.6% of Qwen3-8B cases and 18.9% of DeepSeek-R1-Distill-Qwen-7B cases, while damaging a correct answer in 1.4% and 1.9%. Correctness labels from independently sampled continuations predict a separate continuation's outcome in both primary models. In offline replay, using these labels to allocate compute improves performance on DeepSeek; the Qwen3 gain remains inconclusive. Under full cost accounting, the tested lightweight, myopic controllers mainly recognize current errors and add little continuation-specific value beyond that shortcut. A Llama-8B replication and action/cost ablations examine the robustness of these findings. FLCVA separates measuring local continuation value, predicting it from observable signals, and using those predictions to allocate computation.
Introduction
Longer reasoning can improve language-model accuracy, but it is not automatically the best use of a limited inference budget. Chain-of-thought and zero-shot reasoning prompts elicit explicit reasoning, while self-consistency, repeated sampling, verifier reranking, and search offer several ways to spend additional compute ([34, 16, 33, 3, 5, 39]). Revision can improve an answer ((Madaan et al., 2023)), but without reliable feedback it can also reduce accuracy ([12]). Test-time scaling must therefore decide not only whether more computation can help, but where and when to spend it ([31, 24]).
Adaptive methods use learned halting policies or signals such as confidence, agreement, entropy, and predicted difficulty to halt, sample, branch, or route computation ([9, 2, 28, 1, 6, 26, 27, 38]). These signals can identify an answer that is likely wrong ([15, 17, 36, 20]), but that is not the same as identifying whether further reasoning will repair it. Likewise, a stable answer may be correct and finished, wrong but recoverable, or correct and vulnerable to revision. The decision-relevant object is a checkpoint-level counterfactual: from an identical frozen prefix, what changes if the model answers now rather than receiving one more predeclared block of reasoning?
FLCVA1 compares answering now with continuing from the same frozen prefix: the prompt and all reasoning tokens generated so far. The audit grades each outcome offline; a deployed policy must act without those future correctness labels. Figure 1 shows the protocol, and Figure 2 illustrates why current error probability alone cannot determine continuation value.

Figure 1. FLCVA separates measurement from control. From one frozen prefix, the same answer procedure evaluates answering now, continuing along the stored trajectory for another 256 tokens, and taking three separately sampled continuations of the same length. Correctness change divided by charged cost defines signed local value. Continuation correctness labels are available only offline; replay policies use inference-time signals and pay for every observation they require.
Figure 2. Two prefixes can have the same current-error probability but different recovery probabilities. With equal continuation costs and equal damage risk for correct answers, their expected continuation values differ. Probabilities are illustrative, not experimental estimates.
We separate three questions. Measurement: how much does a specified continuation change correctness per unit cost? Prediction: do observable signals distinguish recoverable errors and vulnerable correct answers? Allocation: do these predictions improve accuracy under a shared compute budget after paying for the signals? For allocation, comparison with a current-wrongness baseline tests the additional value of predicting which errors can be repaired.
Our contributions are:
A paired audit of continuation value. We compare answering now with fixed-length continuations from the same prefix under a fixed answer procedure, measuring signed correctness change per expected logical token.
A full-ledger proxy audit. We evaluate observable signals against this target and charge every observation required by a policy, including the cost of obtaining its signals.
A measurement-to-control decomposition. We distinguish offline diagnostics using future correctness labels from deployable policies. Scoring with independently sampled branches avoids leakage from the evaluated continuation. Comparison with a current-wrongness baseline tests allocation gains beyond error recognition.
Cross-model evidence and intervention robustness. Across two primary models and a Llama-8B replication, continuation can repair errors and damage correct answers. Persistence provides information beyond current correctness, but the tested lightweight, myopic controllers add little beyond error recognition. We examine sensitivity to continuation length, answer cues, parsers, and logical and physical cost accounting.
Related Work
Adaptive inference and value of computation. Metareasoning and anytime algorithms frame computation as an action whose value depends on the current decision state [41, 10] (Horvitz et al. 1989). Modern test-time reasoning instantiates this idea through adaptive sampling, learned routing, branch allocation, confidence-based stopping, and budget forcing [1, 6, 26, 27, 24]. Wu et al. [35] compare model sizes and inference strategies under compute budgets, showing how the inference procedure shapes the accuracy–cost trade-off. FLCVA instantiates value of computation as paired measurement of local stop–continue value for a realized language-model prefix, complementing work whose primary contribution is a controller.
Confidence and process verification. Self-evaluation, semantic uncertainty, outcome verifiers, PRMs, and process-advantage verifiers estimate answer reliability or intermediate progress [15, 17, 36, 32, 20, 33, 40, 29]. SCoRe uses multi-turn reinforcement learning to improve a model’s ability to correct its own errors [18]. We test how well these allocation signals predict the correctness change from a specified continuation. The public PRM serves as an observable proxy, and external endpoint grading defines the audit target.
Closest adaptive-reasoning systems. Studies of overthinking identify redundant reasoning that adds cost with little accuracy benefit [4]. TRACE combines temporal consistency and confidence; UAB allocates sampling from question uncertainty; PoLR and Prefix Consistency use truncated prefixes or regenerated continuations; ThinkBooster unifies scaling strategies and cost views; and PUMA combines semantic redundancy with answer verification [19, 25, 14, 13, 30, 23]. These works primarily propose controllers, reliability signals, or evaluation frameworks. FLCVA instead makes the paired stop–continue contrast the estimand and jointly audits proxy information and charged cost. Table 1 makes this difference explicit.
| Work | Primary decision object | Signal or intervention | Paired same-prefix signed target | Cost role |
| APR | Serial/parallel scheduling | Learned spawn-and-join policy | No | Compute and latency |
| EAGer | Branch/sample allocation | Token-entropy branching | No | Token efficiency |
| TRACE | Early exit | Temporal consistency and confidence | No | Accuracy–token tradeoff |
| UAB | Per-question sample count | Initial-sample vote entropy | No | Fixed sampling budget |
| PoLR | Prefix-cluster expansion | Consensus among short prefixes | No | Tokens and latency |
| Prefix Consistency | Candidate weighting | Regeneration from truncated prefixes | Partial | Voting cost |
| ThinkBooster | Test-time scaling strategy | Modular scorers and strategies | Varies | Quality–cost evaluation |
| PUMA | Early exit | Redundancy plus answer verification | No | Token reduction |
| FLCVA | Measurement and allocation audit | Answer-now versus four fixed continuations | Yes | Full charged ledger |
Table 1. Closest-system comparison. A paired target contrasts answer-now and continuation from the same frozen prefix; “Partial” denotes related truncated-prefix signals without this signed contrast.
The FLCVA Measurement
Signed continuation value
At each checkpoint, we keep the prompt and existing reasoning tokens fixed. We obtain an answer immediately and another answer after each of four continuations. All answers use the same generation, parsing, and grading procedure, called the answer operator. We subtract immediate-answer correctness from average continuation correctness and divide by average continuation cost, including the answer tokens. The resulting value is positive for observed repair, negative for observed damage, and zero when none of the sampled continuations changes correctness.
Let be a frozen reasoning prefix and let denote the fixed answer suffix, greedy decoding rule, parser, normalizer, and exact-match grader. Answering immediately gives correctness . Continuation branch gives correctness after a fixed reasoning action and incurs logical cost , including its answer probe. For , we define
The denominator normalizes by logical token cost. Multiplication by 1,000 expresses the result as expected correctness change per 1,000 logical tokens, making small per-token values easier to read. This common positive scaling leaves the signs and rankings in Eqs. (1)–(3) unchanged. (1) is a ratio of average correctness change to average cost, rather than an average of branch-specific ratios. Its target is conditional on the answer operator, sampling scheme, and continuation length. Recovery is possible only when , and damage only when .
Let . We report two distinct estimands. For a currently wrong answer, estimates single-draw recovery probability, and detects recovery in any of four sampled branches. For a currently correct answer, estimates single-draw damage probability, and detects any observed damage. Four-draw detection measures whether the sampled branches contain at least one recovery or damage event.
Protocol and evaluation split
We evaluate Qwen3-8B and DeepSeek-R1-Distill-Qwen-7B (Yang et al., [37]; DeepSeek-AI, [7]) on all 1,319 GSM8K test problems and all 500 MATH-500 problems (Cobbe et al., [5]; Hendrycks et al., [11]). Each model–problem pair has four independently sampled reasoning trajectories, called spines, capped at 4,096 tokens and frozen every 256 tokens. At every nonterminal checkpoint we apply answer-now and four 256-token continuations. Branch 1 reuses the next stored trajectory segment (the on-spine branch); branches 2–4 are sampled separately from the same prefix (off-spine branches). The same greedy answer operator grades all endpoints.
We deterministically partition problem IDs, assigning 80% of problems to analysis and 20% to confirmation. We use analysis problems to choose each proxy’s scoring direction, calibrate predictions, and fit controller parameters. Confirmation continuation outcomes are used only for held-out evaluation. Unless stated otherwise, intervals use 2,000 bootstrap draws of complete problems, retaining their models, spines, checkpoints, and branches. The primary study contains 14,552 spines and 218,280 eligible checkpoints. Details of sampling, parsing, missing data, and bootstrapping are in Sections A and C.
Identification and uncertainty
The paired comparison holds the prompt and frozen reasoning tokens identical across the answer-now arm and all continuations. The arms differ in the action taken after checkpoint . Branch 1 reuses the stored spine segment while retaining its full reasoning-plus-probe cost in . This saves physical generation while preserving the specified intervention and its logical cost. Later checkpoints follow the original spine, independently of the outcomes of off-spine samples.
We estimate uncertainty by resampling benchmark problems. Each problem contributes multiple models, spines, depths, and correlated outcomes, so each bootstrap draw retains all records nested within the selected problem IDs. Point estimates weight eligible checkpoints in the reported population; intervals preserve within-problem dependence and quantify uncertainty across problems conditional on the stored trajectories.
Proxies and cost
We audit recent-answer persistence, answer confidence and its negative-log form (ANLL), four-spine modal agreement, minimum step-level process reward model (PRM) score, and top-20 approximate entropy. Persistence is the fraction of up to four recent stop answers matching the current answer. Confidence and ANLL satisfy confidence , so they encode the same signal with reversed order; we summarize them as one signal family. The controllers use both input features.
The primary prediction metric is AUROC for detecting any recovery among currently wrong checkpoints and any damage among currently correct ones. These labels indicate whether at least one of four branches changes correctness; they differ from the single-draw probabilities above. Average precision with prevalence baselines and calibration assess rare events and probability quality. Matched and conditional diagnostics test information beyond model, dataset, depth, difficulty, and current correctness; they are grouped in Section C. Difficulty is a label-aware diagnostic stratifier and is never exposed to a deployable policy.
The primary ledger counts all logical tokens needed to expose a prefix, obtain an answer, advance a spine, or score a policy-specific proxy, including the observations required by PRM and parallel-agreement policies. Logical tokens quantify intervention cost. To examine hardware-related costs, we also replay policy orderings under estimated FLOPs and calibrated H20 seconds and report measured GPU time (Section E.2).
From Measurement to Allocation
A frontier is a problem’s current checkpoint during compute allocation. Replay begins with one frontier per confirmation problem and a shared budget. Each action advances one problem by the next stored 256-token segment and reveals its next answer. The four spine seeds define four repetitions, with one spine per problem in each repetition and accuracy measured over problems. All policies are compared up to the same reachable budget endpoint. We measure budget-averaged accuracy using normalized area under the accuracy–ledger curve (AULC). For each full-system bootstrap replicate, we resample problems, rerun allocation, and recompute the budget endpoints.
For repetition , the reported score is
where includes every observation required to expose the initial frontier and is the common endpoint. Final accuracy is the same when all stored transitions are exhausted; AULC measures how efficiently a policy improves accuracy along the way.
We compare three levels of allocation. Round-robin provides a reference for measuring adaptive allocation gains. A correctness shortcut predicts only whether the current answer is wrong and combines that probability with analysis-set dataset–depth mean recovery, damage, and cost. Finally, continuation-specific policies either factor expected value into conditional recovery and damage probabilities (two-stage EV) or directly regress signed (Ridge-). Their inputs are single-spine proxies, depth, and dataset. Comparing with round-robin measures overall allocation gains; comparing with the correctness shortcut measures gains beyond error recognition.
Concretely, let predict that the current answer is wrong, and let , predict recovery and damage conditional on current correctness. With analysis-set dataset–depth means , and mean branch cost , the priorities are
Results
Evaluation scope and primary endpoints. The primary confirmation split contains 372 problems evaluated by each primary model. Every observable controller is fitted only on the disjoint analysis split. The evidence follows three stages: measurement characterizes signed continuation value and its recovery and damage components (§5.1); prediction tests conditional recovery/damage AUROC (§5.2); and allocation tests relative normalized-AULC gain over the correctness shortcut (§5.3). Opportunity diagnostics (§5.4) and robustness analyses (§5.5) explain the scope of these results.
Continuation frequently repairs errors, and sometimes damages answers
For a uniformly selected continuation, recovery occurs in 26.6% of Qwen3 wrong checkpoints and 18.9% of DeepSeek wrong checkpoints (Table 3). Damage is rarer but nonzero: 1.4% and 1.9% among currently correct checkpoints. With four sampled branches, at least one recovery appears at 35.7% and 26.0% of wrong checkpoints, while at least one damaging branch appears at 2.9% and 3.6% of correct checkpoints. These rates vary with dataset and depth (Figure 3); complete denominators are in Section D. Conditional on detecting an event, mean benefit magnitude is 2.865 and 2.794 value units, versus harm magnitudes of 1.817 and 2.000. Four-draw harm is also more frequent on MATH-500 (6.7% and 6.0%) than on GSM8K (1.8% and 2.9%).
| Model | Event | Single draw | Any of four |
| Qwen3 | Recovery | 26.6% [22.5, 31.6] | 35.7% [30.3, 42.3] |
| Qwen3 | Damage | 1.4% [1.0, 1.8] | 2.9% [2.3, 3.6] |
| DeepSeek | Recovery | 18.9% [16.2, 22.2] | 26.0% [22.2, 30.2] |
| DeepSeek | Damage | 1.9% [1.5, 2.3] | 3.6% [3.0, 4.2] |
Table 3. Single-draw transition rates and four-draw opportunity detection on confirmation problems. Brackets are 95% problem-bootstrap intervals.

Figure 3. Recovery and damage prevalence by prefix depth and dataset on confirmation problems. Recovery conditions on the current answer being wrong; damage conditions on it being correct.
| Policy | Required observation | Extra generation | Extra scoring | Replay charge |
| Round-robin / fixed | One current spine | None | None | Prompt, reasoning, answer probe |
| TRACE / ANLL / entropy | Current/prior single-spine probes | None | None | Single-spine ledger |
| PRM | Current single-spine prefix | None | PRM prefill | Ledger plus every PRM call |
| Cross-spine agreement | Four equal-depth answers | Three synchronized spines | None | Four reasoning/probe paths |
| Ridge / two-stage | Single-spine proxies and context | None | Optional PRM | All observed inputs |
| Mixed/off-spine/DP | Continuation correctness labels | Stored branches/futures | Gold grading | Diagnostic only |
Table 2. Policy observability and accounting. Cross-spine agreement charges all parallel trajectories used to obtain the signal. Label-aware rows provide offline allocation diagnostics.
Among currently correct checkpoints, greater persistence is associated with less damage. The high-minus-low persistence contrast in signed value is +0.200 for Qwen3 (95% CI [0.135, 0.276]) and +0.277 for DeepSeek ([0.201, 0.361]). Since value is nonpositive in this population, positive contrasts mean that highly persistent answers are less vulnerable to harmful revision.
Proxies contain information without determining local value
Persistence is the strongest recovery proxy in both models, with AUROC 0.658 and 0.673. Modal agreement is strongest for Qwen3 damage (0.742), while persistence is strongest for DeepSeek damage (0.716) (Table 4). These AUROCs evaluate detection of recovery or damage among four sampled continuations.
| Model | Target | Best proxy (available ) | AUROC (95% CI) |
| Qwen3 | Recovery | Persistence (2,702) | 0.658 [0.610, 0.706] |
| Qwen3 | Damage | Agreement (18,634) | 0.742 [0.699, 0.782] |
| DeepSeek | Recovery | Persistence (3,665) | 0.673 [0.632, 0.715] |
| DeepSeek | Damage | Persistence (17,139) | 0.716 [0.681, 0.747] |
Table 4. Best confirmation proxy by AUROC for detecting any recovery among wrong checkpoints or any damage among correct checkpoints over four continuations. Counts reflect proxy availability; intervals resample whole problems. Rare-event baselines and calibration appear in Section C.
Persistence also retains information after accounting for current correctness, dataset, depth, and difficulty: conditional AUROC is 0.632–0.692 and held-out log-loss reduction is 0.021–0.053 bits. Thus it provides information about recovery and damage beyond identifying current errors. Matched checkpoints with the same signal value can have different outcomes. Calibration, rare-event baselines, and matched-group diagnostics in Section C further characterize this variation.
Observable adaptation mostly learns current wrongness
The correctness shortcut already improves relative normalized AULC over round-robin by +1.78% on Qwen3 and +1.55% on DeepSeek (Table 5 and Figure 4). Relative to this stronger baseline, two-stage EV changes AULC by −0.69% ([−1.51, 0.10]) and +0.05% ([−0.96, 1.00]); direct Ridge- changes it by −0.21% ([−0.52, 0.12]) and +0.70% ([0.07, 1.35]). Thus Qwen3 shows no reliable continuation-specific increment, while DeepSeek shows a modest one. TRACE, PRM, ANLL, and entropy also have point estimates at or below the shortcut. Conditional recovery and damage AUROCs are 0.767–0.799, with limited allocation gains over the shortcut. AUROC evaluates ranking within each outcome population, whereas replay combines recovery and damage predictions with costs to repeatedly rank 372 changing frontiers. The following controls examine how signal costs and one-step scheduling affect allocation.
| Model | Policy | Norm. AULC | Gain vs. RR | Increment vs. shortcut |
| Qwen3 | Round-robin | 0.851 | +0.00% [+0.00, +0.00] | -1.75% [-2.72, -0.72] |
| Qwen3 | Correctness shortcut | 0.866 | +1.78% [+0.73, +2.80] | +0.00% [+0.00, +0.00] |
| Qwen3 | Two-stage conditional EV | 0.860 | +1.08% [+0.28, +1.91] | -0.69% [-1.51, +0.10] |
| Qwen3 | Direct Ridge- | 0.864 | +1.57% [+0.65, +2.51] | -0.21% [-0.52, +0.12] |
| DeepSeek | Round-robin | 0.813 | +0.00% [+0.00, +0.00] | -1.53% [-2.60, -0.44] |
| DeepSeek | Correctness shortcut | 0.825 | +1.55% [+0.44, +2.67] | +0.00% [+0.00, +0.00] |
| DeepSeek | Two-stage conditional EV | 0.826 | +1.60% [+0.44, +2.77] | +0.05% [-0.96, +1.00] |
| DeepSeek | Direct Ridge- | 0.831 | +2.26% [+1.40, +3.17] | +0.70% [+0.07, +1.35] |
Table 5. Observable-controller hierarchy under the problem-level single-spine logical ledger. The final column measures allocation gains beyond the correctness shortcut. Brackets are paired 95% full-system problem-bootstrap intervals.
Figure 4. Held-out controller accuracy under a shared logical-token budget, using one spine per problem. Curves average four spine-seed repetitions; annotations give Ridge-M’s relative gain in normalized AULC over the correctness shortcut, with full-system 95% intervals.
Table 6 separates current-wrongness prediction from conditional recovery and damage prediction, clarifying what each controller can learn from observable features.
| Model | Current wrong | Recovery | Damage |
| Qwen3 | 0.882 | 0.774 | 0.799 |
| DeepSeek | 0.843 | 0.767 | 0.772 |
Table 6. Held-out prediction by the two-stage model. All components are fitted on analysis problems, and replay uses observable features from confirmation problems.
Independent branches reveal allocation opportunity
In the offline diagnostic, correctness labels from three off-spine branches predict the outcome of the held-out on-spine continuation in both primary models. Using this estimate to order replay yields +1.44% relative normalized AULC for Qwen3, with an interval crossing zero, and +5.41% for DeepSeek, with a positive interval (Table 7). Direct cross-branch prediction of the realized successor reaches 0.859–0.948 AUROC across events and models (Table 13).
| Model | Off-spine gain vs. RR | Exact-DP gain vs. RR |
| Qwen3 | +1.44% [−0.48, 3.37] | +7.08% [+5.31, +9.03] |
| DeepSeek | +5.41% [+3.61, +7.29] | +11.55% [+8.19, +15.11] |
Table 7. Offline allocation opportunity under the logical-token ledger. Off-spine ordering uses labels from branches sampled separately from the realized successor; exact DP gives a hindsight bound on four-problem subsets.
| Model | Target | Held-out branch | Event rate | AUROC (95% CI) | AUPRC (95% CI) | Brier (95% CI) | |
| Qwen3 | Beneficial | Pooled 1–4 | 14,744 | 26.6% | 0.943 [0.929, 0.954] | 0.865 [0.838, 0.890] | 0.061 [0.049, 0.074] |
| Qwen3 | Beneficial | 1 (on-spine) | 3,686 | 26.9% | 0.941 [0.926, 0.954] | 0.867 [0.838, 0.894] | 0.062 [0.049, 0.075] |
| Qwen3 | Beneficial | 2 (off-spine) | 3,686 | 26.2% | 0.945 [0.930, 0.958] | 0.863 [0.834, 0.891] | 0.061 [0.049, 0.075] |
| Qwen3 | Beneficial | 3 (off-spine) | 3,686 | 26.0% | 0.945 [0.930, 0.959] | 0.859 [0.826, 0.889] | 0.061 [0.049, 0.076] |
| Qwen3 | Beneficial | 4 (off-spine) | 3,686 | 27.2% | 0.940 [0.923, 0.954] | 0.872 [0.841, 0.900] | 0.061 [0.049, 0.076] |
| Qwen3 | Harmful | Pooled 1–4 | 74,536 | 1.4% | 0.849 [0.810, 0.880] | 0.490 [0.381, 0.580] | 0.009 [0.007, 0.011] |
| Qwen3 | Harmful | 1 (on-spine) | 18,634 | 1.4% | 0.859 [0.822, 0.889] | 0.486 [0.381, 0.579] | 0.009 [0.007, 0.011] |
| Qwen3 | Harmful | 2 (off-spine) | 18,634 | 1.3% | 0.848 [0.800, 0.885] | 0.478 [0.359, 0.580] | 0.009 [0.007, 0.011] |
| Qwen3 | Harmful | 3 (off-spine) | 18,634 | 1.4% | 0.845 [0.796, 0.882] | 0.473 [0.353, 0.579] | 0.009 [0.007, 0.011] |
| Qwen3 | Harmful | 4 (off-spine) | 18,634 | 1.5% | 0.846 [0.804, 0.883] | 0.526 [0.415, 0.624] | 0.009 [0.007, 0.011] |
| DeepSeek | Beneficial | Pooled 1–4 | 17,604 | 18.9% | 0.944 [0.934, 0.954] | 0.839 [0.812, 0.864] | 0.048 [0.040, 0.057] |
| DeepSeek | Beneficial | 1 (on-spine) | 4,401 | 18.8% | 0.948 [0.935, 0.960] | 0.857 [0.826, 0.886] | 0.045 [0.037, 0.054] |
| DeepSeek | Beneficial | 2 (off-spine) | 4,401 | 18.8% | 0.941 [0.928, 0.954] | 0.824 [0.790, 0.857] | 0.051 [0.042, 0.061] |
| DeepSeek | Beneficial | 3 (off-spine) | 4,401 | 18.9% | 0.947 [0.935, 0.960] | 0.841 [0.810, 0.872] | 0.048 [0.039, 0.057] |
| DeepSeek | Beneficial | 4 (off-spine) | 4,401 | 19.0% | 0.942 [0.929, 0.954] | 0.833 [0.800, 0.864] | 0.050 [0.041, 0.060] |
| DeepSeek | Harmful | Pooled 1–4 | 71,676 | 1.9% | 0.879 [0.858, 0.898] | 0.579 [0.510, 0.643] | 0.010 [0.008, 0.012] |
| DeepSeek | Harmful | 1 (on-spine) | 17,919 | 1.9% | 0.880 [0.851, 0.904] | 0.586 [0.503, 0.656] | 0.010 [0.008, 0.012] |
| DeepSeek | Harmful | 2 (off-spine) | 17,919 | 1.9% | 0.884 [0.857, 0.907] | 0.605 [0.535, 0.669] | 0.010 [0.008, 0.011] |
| DeepSeek | Harmful | 3 (off-spine) | 17,919 | 1.9% | 0.880 [0.853, 0.903] | 0.593 [0.513, 0.665] | 0.010 [0.008, 0.012] |
| DeepSeek | Harmful | 4 (off-spine) | 17,919 | 1.8% | 0.874 [0.843, 0.903] | 0.534 [0.450, 0.616] | 0.011 [0.009, 0.013] |
Table 13. Leave-one-branch-out prediction on confirmation checkpoints. The mean event frequency of the other three branches predicts the held-out branch event. Beneficial is conditioned on and harmful on . Branch 1 is the realized replay successor. Intervals resample complete problems 2,000 times.
Effects of signal cost and scheduling. Two controlled comparisons examine signal costs and scheduling. First, we fix the action order based on agreement and charge the extra trajectories needed to obtain that signal. DeepSeek normalized AULC falls from 0.819 to 0.672 (Table 15), isolating the effect of including signal costs.
| Model | Ledger | Agreement treatment | Mean normalized AULC | Gain vs. single-spine RR (95% CI) |
| Qwen3 | Logical tokens | Single-spine round-robin | 0.851 | +0.00% [+0.00, +0.00] |
| Qwen3 | Logical tokens | Free cross-spine diagnostic | 0.830 | -2.47% [-4.71, -0.19] |
| Qwen3 | Logical tokens | Charged four-spine policy | 0.627 | -26.29% [-29.79, -22.92] |
| Qwen3 | Estimated FLOPs | Single-spine round-robin | 0.854 | +0.00% [+0.00, +0.00] |
| Qwen3 | Estimated FLOPs | Free cross-spine diagnostic | 0.831 | -2.66% [-4.78, -0.54] |
| Qwen3 | Estimated FLOPs | Charged four-spine policy | 0.504 | -40.99% [-45.31, -36.65] |
| Qwen3 | Calibrated H2O seconds | Single-spine round-robin | 0.854 | +0.00% [+0.00, +0.00] |
| Qwen3 | Calibrated H2O seconds | Free cross-spine diagnostic | 0.831 | -2.65% [-4.93, -0.45] |
| Qwen3 | Calibrated H2O seconds | Charged four-spine policy | 0.600 | -29.69% [-33.40, -25.92] |
| DeepSeek | Logical tokens | Single-spine round-robin | 0.813 | +0.00% [+0.00, +0.00] |
| DeepSeek | Logical tokens | Free cross-spine diagnostic | 0.819 | +0.78% [-0.55, +2.15] |
| DeepSeek | Logical tokens | Charged four-spine policy | 0.672 | -17.28% [-19.84, -14.67] |
| DeepSeek | Estimated FLOPs | Single-spine round-robin | 0.814 | +0.00% [+0.00, +0.00] |
| DeepSeek | Estimated FLOPs | Free cross-spine diagnostic | 0.819 | +0.67% [-0.63, +2.05] |
| DeepSeek | Estimated FLOPs | Charged four-spine policy | 0.580 | -28.72% [-32.34, -25.05] |
| DeepSeek | Calibrated H2O seconds | Single-spine round-robin | 0.814 | +0.00% [+0.00, +0.00] |
| DeepSeek | Calibrated H2O seconds | Free cross-spine diagnostic | 0.819 | +0.68% [-0.59, +1.98] |
| DeepSeek | Calibrated H2O seconds | Charged four-spine policy | 0.635 | -22.02% [-25.02, -18.86] |
Table 15. Cross-spine agreement under different cost ledgers. The free row evaluates agreement without charging its auxiliary trajectories. The charged policy pays the other three reasoning trajectories and answer probes initially and whenever agreement is updated; the problem’s designated spine alone defines accuracy. Prompts are deduplicated and their KV state is cloned. Each ledger uses a common endpoint across the three rows.
Second, we compare greedy allocation by true next-step correctness change per token with exact multi-step allocation on the same 64 non-overlapping four-problem subsets per model. Both schedules use future outcomes. Normalized AULC is 0.820 versus 0.902 for Qwen3 and 0.791 versus 0.854 for DeepSeek (Table 16). This hindsight comparison shows the benefit of planning beyond immediate value within these subsets.
| Model | Policy | Mean normalized AULC | Gain vs. RR (95% CI) | Gap to exact DP (95% CI) |
| Qwen3 | Round-robin | 0.843 | +0.00% [+0.00, +0.00] | 6.61% [5.06, 8.31] |
| Qwen3 | Mixed myopic | 0.841 | -0.25% [-2.41, +1.88] | 6.84% [4.78, 9.00] |
| Qwen3 | On-spine one-step hindsight | 0.820 | -2.74% [-5.35, -0.15] | 9.16% [6.74, 11.67] |
| Qwen3 | Off-spine myopic | 0.832 | -1.22% [-3.42, +0.81] | 7.75% [5.65, 10.16] |
| Qwen3 | Exact non-myopic hindsight | 0.902 | +7.08% [+5.31, +9.03] | 0.00% [0.00, 0.00] |
| DeepSeek | Round-robin | 0.766 | +0.00% [+0.00, +0.00] | 10.36% [7.86, 13.46] |
| DeepSeek | Mixed myopic | 0.809 | +5.68% [+2.71, +8.89] | 5.27% [3.47, 7.23] |
| DeepSeek | On-spine one-step hindsight | 0.791 | +3.31% [+0.15, +6.58] | 7.39% [5.08, 9.98] |
| DeepSeek | Off-spine myopic | 0.808 | +5.48% [+2.77, +8.65] | 5.44% [3.57, 7.39] |
| DeepSeek | Exact non-myopic hindsight | 0.854 | +11.55% [+8.19, +15.11] | 0.00% [0.00, 0.00] |
Table 16. Exact non-myopic hindsight scheduling on 64 deterministic, non-overlapping subsets per model. Each subset contains one complete 16-state frontier from each of four distinct confirmation problems. Dynamic programming enumerates all precedence-feasible frontier states and maximizes full-ledger AULC within each subset. Intervals bootstrap the 64 four-problem subsets 2,000 times.
Together, these controls motivate charging every signal and considering the value of subsequent states. The relative contributions of signal quality, calibration, and planning horizon to the observed controller gap remain an open question.
Robustness and external replication
Continuation length. In a paired 200-problem confirmation ablation, recovery rises with 128/256/512-token continuations: 19.0/26.6/36.0% for Qwen3 and 14.3/19.6/25.5% for DeepSeek. Paired differences in damage rates are not distinguishable from zero. For each candidate length , a ridge model fitted on analysis problems predicts signed change per 1,000 logical tokens from dataset and checkpoint depth. On confirmation problems, the frozen selector chooses , breaking ties toward the shorter action. Within each model, the selector is a local rule based only on dataset and depth. Table 8 reports event rates and full-sample proxy rankings. On DeepSeek, the selector improves local value by +0.023 ([0.007, 0.039]) correctness-change units per 1,000 logical tokens, largely by selecting the cheaper 128-token action. Mean signed value per 1,000 logical tokens falls with length, from 0.137/0.113 at 128 tokens to 0.090/0.068 at 512 for Qwen3/DeepSeek. The fixed 256-token action lies between shorter, cheaper continuations and longer ones that repair more errors at lower value per token. Persistence remains the leading recovery proxy at all three lengths in the full sample, although its ranking is less stable in the 512-token half-sample analysis.
| Model | Length | Rec.% | Dmg.% | |
| Qwen3 | 128 | 19.0 | 1.6 | +0.016 |
| Qwen3 | 256 | 26.6 | 1.5 | 0.000 |
| Qwen3 | 512 | 36.0 | 1.6 | -0.031 |
| Qwen3 | Adaptive | 26.8 | 1.6 | 0.000 |
| DeepSeek | 128 | 14.3 | 1.7 | +0.021 |
| DeepSeek | 256 | 19.6 | 1.9 | 0.000 |
| DeepSeek | 512 | 25.5 | 2.0 | -0.023 |
| DeepSeek | Adaptive | 14.9 | 1.9 | +0.023 |
Table 8. Continuation-length sensitivity on the 200-problem confirmation ablation. Recovery and damage are single-draw rates; is the change in value per 1,000 logical tokens relative to fixed 256. Full intervals appear in Table 29.
Answer procedure and grading. A similar, concise answer cue produces parseable answers in 99.8–99.9% of cases and preserves 89.8–91.0% of recovery labels and 96.5–97.0% of damage labels. A more directive cue substantially changes the audit target. Across sampled groups, independent rational/symbolic regrading agrees with 95.5–100% of primary transition labels among cases it can resolve.
Cost accounting. After charging their costs, PRM and parallel-agreement policies remain below single-spine baselines under logical tokens, estimated FLOPs, and calibrated H2O seconds. Across these ledgers, PRM trails round-robin by 33.5–48.4% on Qwen3 and 25.2–33.9% on DeepSeek, showing a consistent cost disadvantage across all three measures.
Replication on Llama. The full protocol on DeepSeek-R1-Distill-Llama-8B yields 10.2% single-draw recovery and 0.9% damage. Greater persistence is again associated with less damage (Table 9). Using predictors fitted on the two primary models without refitting, the correctness shortcut gains +2.93% over round-robin, while two-stage EV and Ridge- trail the shortcut by 0.81% and 0.61% (Table 10). The replication thus preserves both the measured continuation effects and the measurement-to-control gap.
| Event | Draw | Any of four | Persistence AUROC |
| Recovery | 10.2% | 14.5% | 0.729 [0.695, 0.763] |
| Damage | 0.9% | 2.0% | 0.730 [0.687, 0.768] |
Table 9. Full-protocol Llama-8B replication. “Any of four” denotes opportunity detection.
| Policy | Gain vs. RR | Δ vs. shortcut |
| Round-robin | +0.00% [+0.00, +0.00] | -2.85% [−3.65, −2.07] |
| Correctness shortcut | +2.93% [+2.11, +3.79] | +0.00% [+0.00, +0.00] |
| Two-stage EV | +2.10% [+1.54, +2.68] | -0.81% [−1.31, −0.31] |
| Direct Ridge- | +2.30% [+1.51, +3.11] | -0.61% [−0.91, −0.35] |
Table 10. Zero-refit transfer to Llama-8B using controllers fitted on the two primary models. Brackets are paired 95% full-system bootstrap intervals.
Discussion
A signal can predict current answer quality, predict continuation effects, or improve a sequence of compute-allocation decisions. Our results show why each claim needs its own evaluation.
Test continuation value beyond error recognition. The correctness shortcut captures much of the observable allocation gain, making it an informative baseline alongside round-robin. Persistence retains conditional information about recovery and damage, and highly persistent correct answers are less vulnerable to harmful revision. These findings support using stability as a predictive feature and evaluating its allocation value through gains over the correctness shortcut.
Charge signals and value later states. The fixed-order agreement comparison shows that observation costs can erase a signal’s apparent allocation benefit. The hindsight comparison on matched subsets shows that greedy scheduling can still underperform multi-step optimization even with exact knowledge of immediate value. The same full ledger provides a basis for testing stronger sequential controllers using inference-time features.
Model and dataset differences. Ridge- gains +0.70% over the shortcut on DeepSeek; the Qwen3 increment is inconclusive, and transfer to Llama without refitting gives a negative gain. Exploratory replay by dataset shows a DeepSeek gain on MATH-500 and a loss on GSM8K (Section F). On the paired 200-problem subset, changing cost weights while fixing action order barely changes the gain difference between models. The effects of calibrating predictions separately for each dataset remain uncertain (Section F.2). The Llama result concerns transfer of predictors fitted on the two primary models. Explaining these differences calls for targeted analysis of recoverable error types, reasoning changes associated with repair, and predictor transfer. The audit provides a common protocol for testing these explanations and comparing stronger controllers.
Conclusion
FLCVA measures the externally graded value of one more reasoning block from the same prefix, while accounting for its cost. Across the evaluated models, continuation can repair incorrect answers and damage correct ones. The tested lightweight, myopic controllers add little beyond identifying likely errors. This measurement-to-control gap motivates policies that use inference-time signals to value both immediate outcomes and future reasoning states. Their benefit can be tested through accuracy–cost gains over the correctness shortcut.
Limitations
The study evaluates three open-weight models on mathematics tasks, with continuations of 128, 256, and 512 tokens. Generalization to other tasks, models, and reasoning actions requires further evaluation. Continuation value depends on sampling and on how answers are generated, parsed, and graded. Exact-match grading may miss equivalent answers. Bootstrap intervals quantify variation across problems while keeping the generated trajectories fixed.
The controller results cover lightweight policies that score only the next reasoning step. In the Llama replication, these policies are transferred without refitting. Correctness labels from separately sampled branches support offline diagnostics. Stored multi-step transitions are available only along the main trajectories. Exact dynamic programming uses these transitions and known future outcomes to provide an upper bound on sampled groups of four problems. The same protocol supports evaluating richer controllers.
Ethical considerations
The study uses established mathematics benchmarks and open-weight models, performs no human-subject intervention, and trains lightweight diagnostic predictors. Better allocation could reduce the computation needed for a given answer quality. The audit itself requires generation of multiple spines and continuations, answer probes, and auxiliary scoring. We report measured GPU time and alternative cost ledgers in Section E.2. Energy consumption, emissions, and net environmental savings fall outside these measurements.
Premature stopping can preserve consequential errors in applications such as medicine and law, while continuing can damage an initially correct answer. Applying adaptive stopping in these domains requires task-specific validation of both recovery and damage, together with safeguards appropriate to the consequences of errors.
Acknowledgments
This work was supported by the National Key Research and Development Program of China (2025YFF0514401) and the National Natural Science Foundation of China (62231022).
Appendix roadmap. Section A gives the complete intervention, labels, proxies, and statistical protocol. Section B details replay units, observability, leakage controls, and non-myopic optimization. Section C reports the complete proxy and controller diagnostics. Section D contains depth, branch-count, length, answer-cue, and parser analyses. Section E gives the Llama replication and physical-cost sensitivity. Section F reports exploratory analyses of dataset composition, score calibration, cost weighting, recovery, and automatic readout.
Experimental Details
Fixed protocol
The primary campaign evaluates Qwen3-8B and DeepSeek-R1-Distill-Qwen-7B on the 1,319-problem GSM8K test set and all 500 MATH-500 problems. Reasoning uses temperature 0.6 and top-. A deterministic problem-ID partition assigns 80% of problems to analysis and 20% to confirmation; no problem crosses the split. Unless specified otherwise, uncertainty intervals use 2,000 problem-level bootstrap replicates. The off-spine correspondence control uses 1,000 within-stratum permutations.
For each model–problem pair, four independent spines are sampled to a 4,096-token cap. Prefixes are frozen at depths 256, 512, . . . , 4096. Each nonterminal prefix receives an answer-now probe, one continuation reusing the next stored on-spine segment, and three independently seeded off-spine continuations. Every branch contains exactly 256 reasoning tokens unless the model has already emitted its terminal answer, and every endpoint receives the same greedy suffix. Therefore, the final answer is \boxed{. Decoding stops at the closing brace or after 32 answer tokens. The on-spine branch must reproduce the stored successor token segment exactly; off-spine branches never re-enter the spine.
This design yields 232,832 stored checkpoints, 218,280 eligible nonterminal checkpoints, and 873,120 branch references. Reusing the on-spine segment avoids redundant physical generation while retaining its full reasoning and answer-probe cost in .
Answer extraction and target construction
GSM8K evaluation extracts the final numeric answer and canonicalizes signs, commas, decimals, and fractions. MATH-500 evaluation extracts the final boxed expression, applies deterministic symbolic normalization, and falls back to normalized string equality. Unparseable answers remain in the denominator and are incorrect. Raw output, extracted span, canonical form, parser status, and exact-match bit are retained for every endpoint.
A checkpoint enters Eq. (1) only when all four continuation labels and positive logical costs exist. We define
The numerator of Eq. (1) is mean branch correctness minus stop correctness, and the denominator is mean branch cost. The factor 1,000 expresses expected correctness change per 1,000 logical tokens. The ratio of means differs from the mean of branch-specific ratios; Table 28 shows that the two are empirically nearly identical here because reasoning length is fixed and only the short answer length varies.
| Model | Pearson | Spearman | Sign agree. | Mean |
| Qwen3 | 1.00000 | 0.99999 | 100.0% | 0.0002 |
| DeepSeek | 1.00000 | 0.99999 | 100.0% | 0.0002 |
Table 28. Sensitivity of the signed value estimand on confirmation problems. The primary ratio of expected correctness change to expected logical cost is compared with the mean of four branch-specific change-per-cost ratios. Both beneficial and harmful event labels agree exactly in both models; correlations quantify the small remaining magnitude and ranking differences.
Observable proxies
Persistence. Fraction of up to four most recent stop answers equal to the current canonical answer. It is structurally missing at the first checkpoint.
Answer confidence and ANLL. Exponential mean selected-token log probability over the answer span, and its negative-log counterpart.
Modal agreement. Fraction of four equal-depth spines supporting the lexicographically tie-broken modal answer. It is an offline cross-seed diagnostic unless all four trajectories are explicitly charged.
Process reward. Minimum positive-class probability across double-newline-delimited steps from Qwen2.5-Math-PRM-7B ([40]). Entropy. Mean entropy of the top-20 answer-token probabilities plus a residual-mass bucket.
Missing proxy values are not imputed from correctness or continuation labels. Available-case denominators are reported for every proxy. PRM scores can be missing when no valid step separator remains; confidence and entropy can be missing when log probabilities are unavailable; agreement can be missing when no spine answer parses.
Statistical protocol
Analysis problems choose proxy orientation, tie-preserving bins, Jeffreys-smoothed calibration probabilities, persistence thresholds, feature scaling, imputation, regularization, and controller coefficients. Confirmation problems provide every reported held-out target and replay result. No problem crosses the split.
Recovery metrics condition on ; damage metrics condition on . Average precision updates precision and recall only after adding a complete equal-score block, preventing record order from breaking ties. AP is reported with confirmation prevalence and AP lift. Brier skill compares with an analysis-prevalence constant predictor:
Problem-bootstrap intervals resample problem IDs together with all nested models, spines, checkpoints, and branches. They quantify between-problem uncertainty conditional on the stored trajectories. Replay intervals use full-system resampling or fixed-order per-problem area contributions, as specified in each table (Section B). The DP analysis resamples four-problem subsets, as described with its results.
Difficulty is stop accuracy pooled over the two primary models and four spines. It is used only for diagnostic stratification, never as a controller feature. Proxy bins are fitted on analysis problems without splitting ties. Collision cells hold model, dataset, depth, difficulty, current correctness, proxy, and proxy bin fixed, and are retained only when both the feasible nonzero event and a zero-value control occur. Because this coverage depends on bin resolution, we also report exact matching for discrete proxies, 5/10/20/50-bin sensitivity, conditional AUROC, nearest-neighbor disagreement, and out-of-sample covariate-only log-loss reduction. Difficulty deletion reruns all grouping and fitted diagnostic baselines without that label-aware stratifier.
Cost coordinates
The logical ledger charges prompt tokens, generated reasoning, answer probes, and any proxy-specific scoring needed before a policy can act. Prompt tokens are deduplicated when a shared KV state can be reused; all generated reasoning and answer tokens are charged. A physical sidecar records role-specific input/output tokens, GPU seconds, peak memory, and throughput. Estimated FLOPs account for causal sequence length, grouped-query attention, SwiGLU projections, vocabulary heads, and reusable prefix KV states. Calibrated H20 seconds apply separate observed rates to reasoning generation, answer probes, and PRM pref ill. We report these resource measures separately.
Replay and Identification Controls
Problem-level replay
The primary replay runs four campaigns, one per spine seed. Each campaign has one frontier for each of 372 confirmation problems, so its accuracy is problem-level exact match. At each ledger boundary, a policy selects one frontier, pays the stored transition cost, and reveals its successor answer. The common endpoint for a policy comparison is the largest budget reachable by every policy. For repetition , normalized AUC is
using the right-continuous accuracy curve. The main controller hierarchy uses full-system bootstrap: each replicate reconstructs the resampled problem pool, creates independent copies for duplicate draws, reruns all four global orderings, and recomputes the common endpoint.
Table 11 reports the resulting problem-level replay for the predefined non-label-aware baselines.
| Model | Policy | Mean normalized AULC | Rel. normalized-AULC gain vs. RR (95% CI) | Final problem accuracy |
| Qwen3 | Fixed 1,024 | 0.820 | -3.66% [-4.34, -3.01] | 0.909 |
| Qwen3 | Round-robin | 0.851 | +0.00% [+0.00, +0.00] | 0.909 |
| Qwen3 | TRACE-style | 0.814 | -4.29% [-5.96, -2.54] | 0.909 |
| Qwen3 | PRM | 0.439 | -48.41% [-52.94, -43.76] | 0.476 |
| Qwen3 | ANLL | 0.834 | -2.00% [-3.52, -0.44] | 0.909 |
| Qwen3 | Entropy | 0.832 | -2.27% [-3.78, -0.70] | 0.909 |
| DeepSeek | Fixed 1,024 | 0.792 | -2.58% [-3.21, -2.00] | 0.847 |
| DeepSeek | Round-robin | 0.813 | +0.00% [+0.00, +0.00] | 0.847 |
| DeepSeek | TRACE-style | 0.813 | +0.04% [-1.06, +1.21] | 0.847 |
| DeepSeek | PRM | 0.537 | -33.88% [-37.99, -29.68] | 0.554 |
| DeepSeek | ANLL | 0.823 | +1.33% [+0.35, +2.38] | 0.847 |
| DeepSeek | Entropy | 0.823 | +1.30% [+0.27, +2.44] | 0.847 |
Table 11. Problem-level equal-ledger replay over 372 confirmation problems, with one spine per problem in each of four repetitions. Mean normalized AULC averages the normalized areas over repetitions. Paired intervals use 2,000 problem-cluster bootstrap draws of per-problem area contributions, retaining all four seeds and the fixed full-universe replay order. Label-aware opportunity diagnostics appear in Table 12; charged cross-spine agreement appears in Table 15.
Successor leakage and independent branches
Branch 1 is the next stored spine segment, so its correctness label equals the successor checkpoint’s answer label. The mixed four-branch score includes this realized outcome. The off-spine score uses branches 2–4 to rank the branch-1 transition without observing its realized outcome. This offline diagnostic tests whether independent continuations predict the successor; its continuation correctness labels require grading and are unavailable to an online policy.
Table 12 contrasts that independent ordering with the mixed and on-spine hindsight orderings under the same ledger.
| Model | Label-aware ordering | Mean normalized AULC | Gain vs. RR (95% CI) | Final problem accuracy |
| Qwen3 | Round-robin | 0.851 | +0.00% [+0.00, +0.00] | 0.909 |
| Qwen3 | On-spine hindsight | 0.842 | -1.10% [-3.32, +1.33] | 0.909 |
| Qwen3 | Mixed myopic | 0.869 | +2.07% [+0.19, +4.02] | 0.909 |
| Qwen3 | Off-spine myopic | 0.863 | +1.44% [-0.48, +3.37] | 0.909 |
| DeepSeek | Round-robin | 0.813 | +0.00% [+0.00, +0.00] | 0.847 |
| DeepSeek | On-spine hindsight | 0.844 | +3.88% [+1.96, +5.90] | 0.847 |
| DeepSeek | Mixed myopic | 0.864 | +6.35% [+4.52, +8.38] | 0.847 |
| DeepSeek | Off-spine myopic | 0.856 | +5.41% [+3.61, +7.29] | 0.847 |
Table 12. Offline opportunity diagnostics under the single-spine logical ledger. On-spine ordering uses the realized successor label; mixed ordering includes it among four labels; off-spine ordering uses three independent continuation labels. Green highlights the off-spine diagnostic. These scores require future correctness labels for offline evaluation. Paired intervals use 2,000 problem-cluster bootstrap draws, retaining all four spine repetitions and conditioning on the fixed full-universe replay order.
Leave-one-branch-out evaluation predicts each branch from the other three. Holding out branch 1 is the direct test of whether independent off-spine outcomes predict the actual replay successor; holding out branches 2–4 checks that predictability is not peculiar to the distinguished on-spine branch.
Table 13 reports the pooled predictor and each of the four leave-one-branch-out evaluations.
A complementary pooled-spine diagnostic places all four spines of each problem in the same budget pool. It measures accuracy over problem–spine pairs; primary replay measures accuracy over problems. We report the two analyses separately because they use different allocation and evaluation units.
Table 14 reports this pooled-spine decomposition; it is visualized in Figure 5.
| Model | Ordering score | Sees successor? | Mean normalized AULC | AULC gain vs. RR | Off-spine permutation |
| Qwen3 | Round-robin | no | 0.851 | +0.00% | – |
| Qwen3 | Mixed 4-branch | yes | 0.856 | +0.57% | – |
| Qwen3 | On-spine one-step hindsight | yes | 0.827 | -2.82% | – |
| Qwen3 | Off-spine independent | no | 0.849 | -0.19% | 0.001/0.001 |
| DeepSeek | Round-robin | no | 0.813 | +0.00% | – |
| DeepSeek | Mixed 4-branch | yes | 0.848 | +4.37% | – |
| DeepSeek | On-spine one-step hindsight | yes | 0.825 | +1.57% | – |
| DeepSeek | Off-spine independent | no | 0.840 | +3.39% | 0.001/0.001 |
Table 14. Spine-level myopic score decomposition with all four spines of each problem in one budget pool. Mixed ordering uses all four branch labels; one-step hindsight uses the realized on-spine successor; off-spine ordering uses branches 2–4 to rank actions evaluated on branch 1. Each score is used for greedy next-step allocation. The off-spine permutation test measures checkpoint correspondence within matching strata; + includes pooled difficulty and − removes it.
Figure 5. Pooled-spine diagnostic separating mixed value into realized on-spine hindsight and independent off-spine information. All methods share the same stored transitions and logical-ledger endpoint. Accuracy is averaged over problem–spine pairs.
Cross-spine agreement accounting
Modal agreement requires four answer probes at the same depth. We compare an uncharged diagnostic, which omits the cost of the auxiliary trajectories, with a charged policy that initializes and advances four synchronized spines. The charged policy pays for every reasoning path and answer probe; accuracy is evaluated on the designated spine.
Table 15 shows both the free diagnostic and the fully charged four-spine policy under all three cost coordinates.
Exact non-myopic hindsight
A deterministic problem-ID rule selects 256 distinct confirmation problems per model, partitioned into 64 non-overlapping groups of four. Each group contains one 16-state on-spine chain per problem. Dynamic programming enumerates all precedence-feasible joint states, charges stored transition cost, and maximizes the same right-continuous AULC objective. With all transition outcomes known, the optimum gives an exact hindsight upper bound within each sampled four-problem system.
Table 16 quantifies the gap between one-step orderings and this subset-specific exact hindsight optimum.
Complete Proxy and Controller Results
Discrimination, rare-event baselines, and calibration
The main text uses conditional event-detection AUROC as its primary prediction endpoint. Table 17 supplies supporting discrimination and calibration estimates with problem-bootstrap intervals. Confidence and ANLL are monotone transformations of the same signal; confidence represents this pair in the table. The controllers use both confidence and ANLL as inputs.
| Model | Target | Proxy | AUROC (95% CI) | AUPRC (95% CI) | Brier (95% CI) | |
| Qwen3 | Beneficial | Confidence | 3,686 | 0.501 [0.428, 0.572] | 0.338 [0.281, 0.415] | 0.227 [0.206, 0.252] |
| Qwen3 | Beneficial | Agreement | 3,678 | 0.625 [0.573, 0.675] | 0.428 [0.366, 0.497] | 0.218 [0.196, 0.242] |
| Qwen3 | Beneficial | Persistence | 2,702 | 0.658 [0.610, 0.706] | 0.363 [0.295, 0.444] | 0.183 [0.155, 0.217] |
| Qwen3 | Beneficial | PRM | 3,686 | 0.604 [0.562, 0.649] | 0.494 [0.432, 0.566] | 0.221 [0.198, 0.246] |
| Qwen3 | Beneficial | Entropy | 3,686 | 0.510 [0.434, 0.584] | 0.334 [0.271, 0.427] | 0.226 [0.204, 0.252] |
| Qwen3 | Harmful | Confidence | 18,634 | 0.598 [0.549, 0.648] | 0.067 [0.042, 0.105] | 0.028 [0.022, 0.034] |
| Qwen3 | Harmful | Agreement | 18,634 | 0.742 [0.699, 0.782] | 0.150 [0.102, 0.207] | 0.025 [0.020, 0.031] |
| Qwen3 | Harmful | Persistence | 18,022 | 0.730 [0.688, 0.772] | 0.111 [0.072, 0.159] | 0.025 [0.020, 0.031] |
| Qwen3 | Harmful | PRM | 18,634 | 0.660 [0.607, 0.707] | 0.056 [0.041, 0.073] | 0.028 [0.022, 0.034] |
| Qwen3 | Harmful | Entropy | 18,634 | 0.623 [0.566, 0.677] | 0.128 [0.071, 0.198] | 0.028 [0.022, 0.034] |
| DeepSeek | Beneficial | Confidence | 4,401 | 0.611 [0.559, 0.659] | 0.307 [0.260, 0.367] | 0.185 [0.166, 0.207] |
| DeepSeek | Beneficial | Agreement | 4,393 | 0.545 [0.493, 0.599] | 0.277 [0.238, 0.328] | 0.188 [0.169, 0.209] |
| DeepSeek | Beneficial | Persistence | 3,665 | 0.673 [0.632, 0.715] | 0.299 [0.247, 0.360] | 0.154 [0.135, 0.178] |
| DeepSeek | Beneficial | PRM | 4,401 | 0.662 [0.620, 0.704] | 0.433 [0.375, 0.504] | 0.178 [0.160, 0.200] |
| DeepSeek | Beneficial | Entropy | 4,401 | 0.611 [0.557, 0.664] | 0.302 [0.258, 0.364] | 0.184 [0.165, 0.205] |
| DeepSeek | Harmful | Confidence | 17,919 | 0.668 [0.630, 0.708] | 0.177 [0.129, 0.233] | 0.034 [0.029, 0.039] |
| DeepSeek | Harmful | Agreement | 17,919 | 0.702 [0.665, 0.737] | 0.112 [0.082, 0.150] | 0.033 [0.028, 0.038] |
| DeepSeek | Harmful | Persistence | 17,139 | 0.716 [0.681, 0.747] | 0.141 [0.106, 0.185] | 0.029 [0.025, 0.034] |
| DeepSeek | Harmful | PRM | 17,919 | 0.584 [0.543, 0.621] | 0.065 [0.051, 0.087] | 0.034 [0.029, 0.040] |
| DeepSeek | Harmful | Entropy | 17,919 | 0.681 [0.636, 0.720] | 0.215 [0.155, 0.277] | 0.034 [0.028, 0.039] |
Table 17. Supporting confirmation discrimination and calibration results with problem-bootstrap intervals. Confidence represents the monotone-equivalent confidence/ANLL pair; separate ANLL results are available in the accompanying results archive.
Because damage is rare, Table 18 pairs average precision and Brier score with their prevalence baselines. For the best damage proxies, average precision is 4.34–5.16 times prevalence, while Brier skill is 0.078–0.099. This distinguishes useful event ranking from accurate probability estimation. Figure 6 shows held-out calibration curves.
| Model | Target | Proxy | Analysis prev. | Conf. prev. | AP | AP/prev. | Brier ref. | BSS |
| Qwen3 | Beneficial | Confidence | 0.307 | 0.357 | 0.338 | 0.95× | 0.232 | +0.022 |
| Qwen3 | Beneficial | Agreement | 0.306 | 0.356 | 0.428 | 1.20× | 0.232 | +0.058 |
| Qwen3 | Beneficial | Persistence | 0.224 | 0.264 | 0.363 | 1.37× | 0.196 | +0.064 |
| Qwen3 | Beneficial | PRM | 0.307 | 0.357 | 0.494 | 1.38× | 0.232 | +0.050 |
| Qwen3 | Beneficial | Entropy | 0.307 | 0.357 | 0.334 | 0.94× | 0.232 | +0.025 |
| Qwen3 | Harmful | Confidence | 0.028 | 0.029 | 0.067 | 2.30× | 0.028 | +0.005 |
| Qwen3 | Harmful | Agreement | 0.028 | 0.029 | 0.150 | 5.16× | 0.028 | +0.099 |
| Qwen3 | Harmful | Persistence | 0.026 | 0.027 | 0.111 | 4.05× | 0.027 | +0.064 |
| Qwen3 | Harmful | PRM | 0.028 | 0.029 | 0.056 | 1.91× | 0.028 | +0.011 |
| Qwen3 | Harmful | Entropy | 0.028 | 0.029 | 0.128 | 4.40× | 0.028 | +0.009 |
| DeepSeek | Beneficial | Confidence | 0.235 | 0.260 | 0.307 | 1.18× | 0.193 | +0.040 |
| DeepSeek | Beneficial | Agreement | 0.235 | 0.260 | 0.277 | 1.07× | 0.193 | +0.026 |
| DeepSeek | Beneficial | Persistence | 0.182 | 0.209 | 0.299 | 1.43× | 0.166 | +0.070 |
| DeepSeek | Beneficial | PRM | 0.235 | 0.260 | 0.433 | 1.67× | 0.193 | +0.077 |
| DeepSeek | Beneficial | Entropy | 0.235 | 0.260 | 0.302 | 1.16× | 0.193 | +0.047 |
| DeepSeek | Harmful | Confidence | 0.036 | 0.036 | 0.177 | 4.93× | 0.035 | +0.022 |
| DeepSeek | Harmful | Agreement | 0.036 | 0.036 | 0.112 | 3.12× | 0.035 | +0.057 |
| DeepSeek | Harmful | Persistence | 0.033 | 0.033 | 0.141 | 4.34× | 0.031 | +0.078 |
| DeepSeek | Harmful | PRM | 0.036 | 0.036 | 0.065 | 1.81× | 0.035 | +0.008 |
| DeepSeek | Harmful | Entropy | 0.036 | 0.036 | 0.215 | 5.99× | 0.035 | +0.031 |
Table 18. Rare-event baselines on each proxy’s available-case population. Confidence represents the confidence/ANLL pair. Confirmation prevalence is the no-skill AP baseline. The Brier reference predicts the analysis-set prevalence for every confirmation row; BSS = .
Figure 6. Confirmation calibration for all audited proxies. Orientations and tie-preserving bins are fixed using analysis problems; the diagonal denotes perfect calibration.
Matched and conditional information
These diagnostics examine variation in continuation outcomes among checkpoints with similar observed signals. Collision coverage is the fraction of checkpoints in matched cells containing both a feasible nonzero event and a zero-value control. Bin sensitivity, exact matching, and nearest-neighbor comparisons assess this outcome heterogeneity at different matching resolutions. Conditional AUROC and held-out log-loss reduction measure predictive information beyond the matching covariates.
At ten-bin resolution, 81.9% and 94.5% of Qwen3 and DeepSeek checkpoints lie in at least one ambiguous matched cell; exact persistence matching gives 77.3% and 87.7%. Conditional AUROC is 0.632–0.692, and held-out log-loss reduction is 0.021–0.053 bits. Thus persistence provides predictive information even though checkpoints with matching signals can have different continuation outcomes.
Table 19 to Table 21 respectively report matched-cell ambiguity, bin-resolution sensitivity, and conditional or incremental information.
| Model | Proxy | Collision | Cells | Recovery gap | Damage gap |
| Qwen3 | ANLL | 48.0% | 606 | +2.703 | -1.813 |
| Qwen3 | Confidence | 48.0% | 606 | +2.703 | -1.813 |
| Qwen3 | Agreement | 73.7% | 323 | +2.687 | -1.769 |
| Qwen3 | Persistence | 77.3% | 327 | +2.616 | -1.777 |
| Qwen3 | PRM | 45.3% | 591 | +2.677 | -1.806 |
| Qwen3 | Entropy | 47.6% | 585 | +2.693 | -1.805 |
| DeepSeek | ANLL | 55.8% | 654 | +2.657 | -1.976 |
| DeepSeek | Confidence | 55.8% | 654 | +2.657 | -1.976 |
| DeepSeek | Agreement | 87.5% | 366 | +2.665 | -1.974 |
| DeepSeek | Persistence | 87.7% | 329 | +2.575 | -1.961 |
| DeepSeek | PRM | 58.2% | 642 | +2.565 | -2.000 |
| DeepSeek | Entropy | 54.5% | 619 | +2.662 | -1.949 |
Table 19. Matched-cell collision diagnostics. Gaps compare feasible nonzero checkpoints with zero-value controls after matching model, dataset, difficulty, depth, current correctness, and proxy bin.
| Model | Target | Proxy | Conditional AUROC | NN disagreement | test log-loss, bits |
| Qwen3 | Beneficial | ANLL | 0.524 | 34.3% [30.8, 37.9] | -0.0166 [-0.0344, -0.0006] |
| Qwen3 | Beneficial | Confidence | 0.476 | 34.3% [30.9, 37.7] | -0.0166 [-0.0331, -0.0008] |
| Qwen3 | Beneficial | Agreement | 0.581 | 35.1% [31.8, 38.7] | +0.0144 [-0.0112, +0.0416] |
| Qwen3 | Beneficial | Persistence | 0.644 | 32.3% [28.7, 36.5] | +0.0468 [+0.0193, +0.0760] |
| Qwen3 | Beneficial | PRM | 0.597 | 35.6% [32.9, 38.4] | -0.0085 [-0.0233, +0.0067] |
| Qwen3 | Beneficial | Entropy | 0.534 | 36.7% [33.4, 40.1] | -0.0239 [-0.0406, -0.0088] |
| Qwen3 | Harmful | ANLL | 0.562 | 5.5% [4.7, 6.3] | -0.0046 [-0.0074, -0.0015] |
| Qwen3 | Harmful | Confidence | 0.562 | 5.5% [4.7, 6.3] | -0.0046 [-0.0074, -0.0016] |
| Qwen3 | Harmful | Agreement | 0.697 | 4.9% [4.2, 5.7] | +0.0199 [+0.0111, +0.0307] |
| Qwen3 | Harmful | Persistence | 0.691 | 4.7% [4.1, 5.5] | +0.0215 [+0.0129, +0.0308] |
| Qwen3 | Harmful | PRM | 0.680 | 5.4% [4.6, 6.2] | +0.0016 [-0.0027, +0.0065] |
| Qwen3 | Harmful | Entropy | 0.587 | 5.1% [4.4, 5.8] | -0.0039 [-0.0078, +0.0002] |
| DeepSeek | Beneficial | ANLL | 0.563 | 31.3% [28.2, 34.6] | -0.0039 [-0.0242, +0.0174] |
| DeepSeek | Beneficial | Confidence | 0.563 | 31.3% [28.2, 34.6] | -0.0038 [-0.0241, +0.0179] |
| DeepSeek | Beneficial | Agreement | 0.547 | 31.5% [28.8, 34.3] | +0.0088 [-0.0077, +0.0244] |
| DeepSeek | Beneficial | Persistence | 0.632 | 27.2% [24.7, 29.8] | +0.0528 [+0.0275, +0.0820] |
| DeepSeek | Beneficial | PRM | 0.621 | 30.3% [27.9, 33.0] | +0.0188 [-0.0006, +0.0384] |
| DeepSeek | Beneficial | Entropy | 0.573 | 30.7% [27.9, 33.7] | -0.0000 [-0.0234, +0.0233] |
| DeepSeek | Harmful | ANLL | 0.629 | 6.2% [5.6, 6.9] | +0.0058 [-0.0001, +0.0120] |
| DeepSeek | Harmful | Confidence | 0.629 | 6.2% [5.6, 6.9] | +0.0058 [-0.0000, +0.0121] |
| DeepSeek | Harmful | Agreement | 0.621 | 6.2% [5.6, 6.9] | +0.0149 [+0.0063, +0.0241] |
| DeepSeek | Harmful | Persistence | 0.634 | 5.5% [5.0, 6.1] | +0.0242 [+0.0161, +0.0330] |
| DeepSeek | Harmful | PRM | 0.610 | 6.6% [5.9, 7.3] | -0.0029 [-0.0068, +0.0016] |
| DeepSeek | Harmful | Entropy | 0.635 | 6.1% [5.5, 6.7] | +0.0065 [+0.0001, +0.0136] |
Table 21. Conditional and incremental proxy diagnostics on confirmation problems. Positive test log-loss favors the proxy-augmented model; intervals use problem-level bootstrap.
Removing pooled difficulty yields coarser matching cells and diagnostic covariates constructed without confirmation stop labels. Persistence retains above-chance conditional AUROC and positive held-out log-loss reduction for both targets and both models; signed matched gaps retain their feasible direction. The relative rankings of individual proxies vary across these specifications.
Table 22 and Table 23 show that the main qualitative conclusions survive deletion of the label-aware difficulty covariate.
| Model | Proxy | Collision rate | Beneficial | Harmful | |||
| + | - | + | - | + | - | ||
| Qwen3 | ANLL | 48.0% | 81.1% | +2.703 | +2.735 | -1.813 | -1.815 |
| Qwen3 | Confidence | 48.0% | 81.1% | +2.703 | +2.735 | -1.813 | -1.815 |
| Qwen3 | Agreement | 73.7% | 98.1% | +2.687 | +2.726 | -1.769 | -1.788 |
| Qwen3 | Persistence | 77.3% | 99.1% | +2.616 | +2.627 | -1.777 | -1.795 |
| Qwen3 | PRM | 45.3% | 67.3% | +2.677 | +2.699 | -1.806 | -1.814 |
| Qwen3 | Entropy | 47.6% | 79.1% | +2.693 | +2.734 | -1.805 | -1.811 |
| DeepSeek | ANLL | 55.8% | 82.9% | +2.657 | +2.717 | -1.976 | -1.989 |
| DeepSeek | Confidence | 55.8% | 82.9% | +2.657 | +2.717 | -1.976 | -1.989 |
| DeepSeek | Agreement | 87.5% | 97.9% | +2.665 | +2.721 | -1.974 | -1.984 |
| DeepSeek | Persistence | 87.7% | 99.2% | +2.575 | +2.596 | -1.961 | -1.982 |
| DeepSeek | PRM | 58.2% | 82.3% | +2.565 | +2.602 | -2.000 | -2.000 |
| DeepSeek | Entropy | 54.5% | 80.6% | +2.662 | +2.714 | -1.949 | -1.964 |
Table 22. Matched-collision sensitivity after deleting pooled difficulty. includes difficulty; removes it while retaining the other matching covariates and analysis-fitted proxy bins.
Table 23. Local-information sensitivity after deleting pooled difficulty. includes and excludes difficulty from conditional AUROC, nearest-neighbor matching, and held-out log-loss models.
Predictor decomposition
Direct Ridge- predicts signed value from persistence, confidence, ANLL, entropy, depth, and dataset; it excludes current correctness, difficulty, agreement, and continuation outcomes. Harmful AUROC uses . The two-stage model separately predicts current wrongness, recovery, and damage, then applies analysis-set costs. Table 24 reports held-out Ridge- discrimination, clarifying why event ranking need not improve allocation.
| Model | Controller | Pearson | Beneficial AUROC | Harmful AUROC ( score) |
| Qwen3 | Ridge- (single-spine) | 0.373 | 0.715 | 0.291 |
| Qwen3 | Ridge- (+ charged PRM) | 0.374 | 0.715 | 0.310 |
| DeepSeek | Ridge- (single-spine) | 0.282 | 0.726 | 0.226 |
| DeepSeek | Ridge- (+ charged PRM) | 0.284 | 0.730 | 0.235 |
Table 24. Held-out Ridge- diagnostics. Predictors exclude correctness, difficulty, agreement, and continuation outcomes; harmful AUROC uses . PRM scoring is charged.
Intervention Robustness
Dataset and depth trajectories
Table 25 gives the model–dataset prevalence, magnitude, and depth-trend summaries behind the main-text trajectories.
| Model | Data | Wrong | Beneficial rate (95% CI) | Correct | Harmful rate (95% CI) |
| Qwen3 | All | 3,686 | 35.7% [30.4, 42.1] | 18,634 | 2.9% [2.3, 3.6] |
| Qwen3 | GSM8K | 1,481 | 54.0% [45.9, 64.3] | 14,479 | 1.8% [1.3, 2.5] |
| Qwen3 | MATH-500 | 2,205 | 23.4% [17.7, 30.6] | 4,155 | 6.7% [4.7, 9.2] |
| DeepSeek | All | 4,401 | 26.0% [22.2, 30.4] | 17,919 | 3.6% [3.0, 4.2] |
| DeepSeek | GSM8K | 1,985 | 32.2% [26.4, 39.4] | 13,975 | 2.9% [2.4, 3.5] |
| DeepSeek | MATH-500 | 2,416 | 20.9% [16.3, 26.4] | 3,944 | 6.0% [4.1, 8.3] |
Table 25. Confirmation prevalence of signed continuation events. Intervals resample complete problems 2,000 times.
Persistence ties, branch count, and ratio definition
Persistence is discrete and heavily concentrated at one. The primary high-first rule compares exact persistence with all lower values, rather than splitting tied observations. Stricter low-persistence thresholds strengthen the same direction.
Table 26 reports the value contrasts under these persistence thresholds.
| Model | Low group | High–low (95% CI) | ||
| Qwen3 | (tie-resolved) | 15,349 | 2,673 | +0.200 [+0.135, +0.276] |
| Qwen3 | 15,349 | 1,251 | +0.291 [+0.190, +0.401] | |
| Qwen3 | 15,349 | 398 | +0.450 [+0.267, +0.665] | |
| DeepSeek | (tie-resolved) | 15,011 | 2,128 | +0.277 [+0.201, +0.361] |
| DeepSeek | 15,011 | 937 | +0.480 [+0.341, +0.621] | |
| DeepSeek | 15,011 | 311 | +0.740 [+0.536, +0.965] |
Table 26. Persistence-threshold sensitivity among currently correct confirmation checkpoints. Both analysis quartiles equal 1.0 because persistence is discrete; the implemented tie order therefore compares with . Tighter low groups leave the conclusion unchanged.
To measure branch-count sensitivity, Table 27 averages detection of a correctness change over all subsets of the observed branches. With one
| Model | Target | Expected event rate | Agreement with | |
| Qwen3 | Beneficial | 1 | 26.6% | 90.9% |
| Qwen3 | Beneficial | 2 | 31.2% | 95.4% |
| Qwen3 | Beneficial | 3 | 33.8% | 98.1% |
| Qwen3 | Beneficial | 4 | 35.7% | 100.0% |
| Qwen3 | Harmful | 1 | 1.4% | 98.5% |
| Qwen3 | Harmful | 2 | 2.0% | 99.1% |
| Qwen3 | Harmful | 3 | 2.5% | 99.6% |
| Qwen3 | Harmful | 4 | 2.9% | 100.0% |
| DeepSeek | Beneficial | 1 | 18.9% | 92.9% |
| DeepSeek | Beneficial | 2 | 22.5% | 96.5% |
| DeepSeek | Beneficial | 3 | 24.5% | 98.5% |
| DeepSeek | Beneficial | 4 | 26.0% | 100.0% |
| DeepSeek | Harmful | 1 | 1.9% | 98.3% |
| DeepSeek | Harmful | 2 | 2.6% | 99.0% |
| DeepSeek | Harmful | 3 | 3.2% | 99.6% |
| DeepSeek | Harmful | 4 | 3.6% | 100.0% |
Table 27. Branch-count sensitivity on confirmation problems. For , values average over every size- subset of the four observed branches, avoiding dependence on the distinguished on-spine branch order.
| Model | Proxy | 5 bins | 10 bins | 20 bins | 50 bins |
| Qwen3 | ANLL | 63.1% | 48.0% | 33.6% | 20.4% |
| Qwen3 | Confidence | 63.1% | 48.0% | 33.6% | 20.4% |
| Qwen3 | Agreement | 73.7% | 73.7% | 73.7% | 73.7% |
| Qwen3 | Persistence | 77.3% | 77.3% | 77.3% | 77.3% |
| Qwen3 | PRM | 56.9% | 45.3% | 33.1% | 20.5% |
| Qwen3 | Entropy | 63.2% | 47.6% | 33.2% | 20.6% |
| DeepSeek | ANLL | 74.1% | 55.8% | 39.8% | 23.2% |
| DeepSeek | Confidence | 74.1% | 55.8% | 39.8% | 23.2% |
| DeepSeek | Agreement | 87.5% | 87.5% | 87.5% | 87.5% |
| DeepSeek | Persistence | 87.7% | 87.7% | 87.7% | 87.7% |
| DeepSeek | PRM | 73.7% | 58.2% | 41.9% | 25.2% |
| DeepSeek | Entropy | 72.2% | 54.5% | 38.1% | 23.0% |
Table 20. Collision sensitivity to proxy resolution. Persistence and agreement use exact values; continuous-proxy bins are fitted on analysis data.
draw, this measures the probability of a change after one continuation; with multiple draws, it
measures the opportunity detected across the sampled continuations.
Increasing from one to four draws raises recovery detection from 26.6% to 35.7% for Qwen3 and from 18.9% to 26.0% for DeepSeek. Damage detection rises from 1.4% to 2.9% and from 1.9% to 3.6%, respectively. Changing the cost-ratio definition has a much smaller effect: replacing the ratio of expectations with the mean branch-specific ratio preserves every event sign, has Spearman correlation 0.99999, and changes magnitude by on average.
Table 27 and Table 28 summarize the effects of branch count and cost-ratio definition.
Continuation length
A deterministic problem-ID rule selects 400 problems paired across models and balanced between analysis and confirmation (150 GSM8K and 50 MATH-500 per split). All four spines and the common depth lattice through 3,584 tokens are re-
tained. The 128-token arm truncates the stored branch, the 256-token arm reuses it, and the 512-token arm extends the same branch with a second contiguous 256-token block conditioned on the first. Generation remains label-blind.
For each model and , an analysis-only Ridge model predicts signed change per 1,000 logical tokens from dataset and checkpoint depth:
with ties resolved toward the shorter length. This context-only rule assigns the same length to every checkpoint in a model–dataset–depth cell.
Table 29 and Table 30 report the full paired event-rate intervals and proxy-ranking sensitivity.
| Length | Recovery | Damage | Mean | vs. 256 (95% CI) |
| Qwen3 | ||||
| 128 tokens | 19.0% [15.0, 24.1] | 1.6% [1.1, 2.2] | 0.137 | +0.016 [-0.001, +0.033] |
| 256 tokens | 26.6% [21.2, 33.6] | 1.5% [1.0, 2.1] | 0.121 | +0.000 [+0.000, +0.000] |
| 512 tokens | 36.0% [28.9, 45.5] | 1.6% [1.0, 2.3] | 0.090 | -0.031 [-0.040, -0.023] |
| Adaptive | 26.8% [21.2, 33.7] | 1.6% [1.0, 2.2] | 0.121 | +0.000 [-0.008, +0.007] |
| DeepSeek | ||||
| 128 tokens | 14.3% [11.4, 18.0] | 1.7% [1.3, 2.2] | 0.113 | +0.021 [+0.002, +0.039] |
| 256 tokens | 19.6% [16.2, 24.3] | 1.9% [1.5, 2.5] | 0.092 | +0.000 [+0.000, +0.000] |
| 512 tokens | 25.5% [20.6, 31.9] | 2.0% [1.5, 2.6] | 0.068 | -0.023 [-0.031, -0.016] |
| Adaptive | 14.9% [12.1, 18.7] | 1.9% [1.4, 2.4] | 0.115 | +0.023 [+0.007, +0.039] |
Table 29. Complete continuation-length sensitivity on the 200-problem confirmation ablation. Recovery and damage are single-draw rates; compares four-branch value per 1,000 logical tokens with fixed 256.
| Model | Target | Condition | Rank vs. 256 | Top proxy (half-sample stability) |
| Qwen3 | Recovery | 128 tokens | 1.00 [0.82, 1.00] | persistence (89%) |
| Qwen3 | Recovery | 512 tokens | 1.00 [-0.22, 1.00] | persistence (80%) |
| Qwen3 | Recovery | Adaptive | 1.00 [0.82, 1.00] | persistence (95%) |
| Qwen3 | Damage | 128 tokens | 1.00 [0.76, 1.00] | modal_agreement (88%) |
| Qwen3 | Damage | 512 tokens | 1.00 [0.76, 1.00] | modal_agreement (80%) |
| Qwen3 | Damage | Adaptive | 1.00 [0.82, 1.00] | modal_agreement (92%) |
| DeepSeek | Recovery | 128 tokens | 1.00 [0.65, 1.00] | persistence (88%) |
| DeepSeek | Recovery | 512 tokens | 1.00 [0.55, 1.00] | persistence (84%) |
| DeepSeek | Recovery | Adaptive | 1.00 [0.65, 1.00] | persistence (86%) |
| DeepSeek | Damage | 128 tokens | 0.82 [0.09, 1.00] | top20_entropy (45%) |
| DeepSeek | Damage | 512 tokens | 0.94 [0.19, 1.00] | modal_agreement (62%) |
| DeepSeek | Damage | Adaptive | 0.94 [0.19, 1.00] | top20_entropy (44%) |
Table 30. Stability of the six-proxy AUROC ranking. Proxy orientations are selected on the 256-token analysis arm and fixed across confirmation conditions. The adaptive selector uses dataset and depth. Brackets and parentheses summarize variation over 64 deterministic paired half-problem subsamples; these are robustness ranges rather than confidence intervals.
Recovery rises with length in both models, while every paired damage contrast with 256 tokens includes zero. Persistence ranks first for recovery at every length in the full sample. Rankings vary more across half-sample evaluations, especially for 512-token recovery and rare DeepSeek damage.
Longer blocks expose more repairs but yield lower cost-normalized value. Mean signed change per 1,000 logical tokens falls from 0.137/0.113 at 128 tokens to 0.090/0.068 at 512 for Qwen3/DeepSeek. The adaptive cell selector matches fixed 256 on Qwen3 and improves value by +0.023 on DeepSeek. The predeclared 256-token action lies between shorter, cheaper continuations and longer continuations that expose more repairs.
Answer cue and parser
A deterministic 200-problem confirmation sample reuses every frozen prefix and continuation while replacing only the greedy boxed-answer suffix. The canonical rerun quantifies execution reproducibility; all contexts remain in the analysis whether or not tokens replay exactly. Table 31 and Table 32 report execution reproducibility under the canonical cue and sensitivity to alternative cues. The terse cue has parse rates of 99.9% for Qwen3 and 99.8% for DeepSeek and preserves 89.8–91.0% of recovery labels and 96.5–97.0% of damage labels. The direct cue parses 51.7% and 80.8%, respectively, and changes the measured continuation effects. These differences show how the answer procedure shapes the audit target.
| Model | Token identity | Raw-text identity | Normalized-answer identity |
| Qwen3 | 98.9% [98.7, 99.1] | 99.6% [99.4, 99.7] | 98.5% [97.5, 99.2] |
| DeepSeek | 98.3% [98.1, 98.5] | 99.5% [99.4, 99.7] | 98.8% [98.0, 99.5] |
Table 31. Canonical-control replay against the primary outputs.
| Model | Cue | Parse rate | Answer agreement | Benefit agreement | Harm agreement |
| Qwen3 | Canonical | 99.4% [98.9, 99.7] | 100.0% [100.0, 100.0] | 100.0% [100.0, 100.0] | 100.0% [100.0, 100.0] |
| Qwen3 | Direct | 51.7% [47.9, 55.7] | 78.8% [73.0, 84.5] | 72.0% [64.0, 78.3] | 80.3% [79.1, 81.5] |
| Qwen3 | Terse | 99.9% [99.8, 99.9] | 95.2% [93.4, 96.7] | 91.0% [87.2, 94.4] | 96.5% [95.7, 97.2] |
| DeepSeek | Canonical | 99.8% [99.4, 100.0] | 100.0% [100.0, 100.0] | 100.0% [100.0, 100.0] | 100.0% [100.0, 100.0] |
| DeepSeek | Direct | 80.8% [78.1, 83.3] | 87.0% [83.0, 90.7] | 85.9% [81.5, 89.5] | 83.4% [82.0, 84.7] |
| DeepSeek | Terse | 99.8% [99.5, 100.0] | 95.3% [93.9, 96.5] | 89.8% [86.0, 93.1] | 97.0% [96.5, 97.6] |
Table 32. Answer-cue sensitivity on the deterministic confirmation sample. Event agreement conditions on the canonical stop population. Unparseable alternative endpoints are scored as incorrect.
For an independent automated check, a deterministic problem-ID rule selects up to 200 confirmation transitions per model–dataset–primary-label cell. A secondary evaluator re-extracts boxed answers and applies exact rational comparison on GSM8K and symbolic equivalence on MATH-500. Agreement with the primary transition label ranges from 95.5% to 100% among resolved cases. We report resolved-cell coverage alongside agreement to distinguish the scope of this automated check from its agreement rate.
Table 33 reports the stratified secondary-adjudication agreement and resolved-cell coverage.
| Data | Transition | Population | Resolved | Agreement (95% CI) |
| DeepSeek | ||||
| GSM8K | beneficial | 1,983 | 199 | 99.0% [96.7, 100.0] |
| GSM8K | harmful | 767 | 196 | 100.0% [100.0, 100.0] |
| GSM8K | zero correct | 55,133 | 200 | 100.0% [100.0, 100.0] |
| GSM8K | zero wrong | 5,957 | 191 | 100.0% [100.0, 100.0] |
| MATH-500 | beneficial | 1,337 | 200 | 98.5% [95.7, 100.0] |
| MATH-500 | harmful | 570 | 200 | 98.5% [95.8, 100.0] |
| MATH-500 | zero correct | 15,206 | 200 | 100.0% [100.0, 100.0] |
| MATH-500 | zero wrong | 8,327 | 200 | 95.5% [88.5, 100.0] |
| Qwen3 | ||||
| GSM8K | beneficial | 2,484 | 172 | 100.0% [100.0, 100.0] |
| GSM8K | harmful | 455 | 165 | 100.0% [100.0, 100.0] |
| GSM8K | zero correct | 57,461 | 200 | 100.0% [100.0, 100.0] |
| GSM8K | zero wrong | 3,440 | 164 | 100.0% [100.0, 100.0] |
| MATH-500 | beneficial | 1,436 | 200 | 99.0% [96.7, 100.0] |
| MATH-500 | harmful | 569 | 200 | 99.0% [97.0, 100.0] |
| MATH-500 | zero correct | 16,051 | 200 | 100.0% [100.0, 100.0] |
| MATH-500 | zero wrong | 7,384 | 200 | 97.0% [90.2, 100.0] |
Table 33. Stratified automated secondary adjudication. Agreement is measured over resolved cases; intervals use 2,000 problem-cluster bootstrap replicates.
External Replication and Cost Sensitivity
Full-protocol Llama replication
DeepSeek-R1-Distill-Llama-8B repeats the complete prefix lattice and continuation intervention on all 1,819 problems [7, 8]. Difficulty remains frozen from the two primary models, so no Llama outcome defines its strata.
Table 34 complements the main-text Llama estimand and zero-refit transfer tables with complete best-proxy metrics.
| Target | Best proxy | AUROC | AUPRC | Brier |
| Beneficial | persistence | 0.729 [0.695, 0.763] | 0.216 | 0.089 |
| Harmful | persistence | 0.730 [0.687, 0.768] | 0.090 | 0.018 |
Table 34. Best-proxy results in the DeepSeek-R1-Distill-Llama-8B replication. Difficulty strata exclude Llama outcomes.
The transfer experiment fits the correctness shortcut, two-stage EV, and Ridge- on analysis rows from the two primary models, then freezes feature order, imputation, scaling, missingness indicators, regularization, coefficients, and dataset–depth continuation moments. No Llama row is used for fitting or hyperparameter selection.
Physical compute and alternative ledgers
Table 35 and Table 36 report the measured compute sidecar and replay sensitivity across logical tokens, estimated FLOPs, and calibrated H20 seconds.
| Model | Gen. output | Gen. GPU-h | Gen. tok/s |
| Qwen3 | 115.15M | 71.8 | 445.8 |
| DeepSeek | 115.19M | 61.7 | 518.5 |
| Model | Peak GiB | PRM GPU-h | Total GPU-h |
| Qwen3 | 85.5 | 2.7 | 75.1 |
| DeepSeek | 85.5 | 2.8 | 65.1 |
Table 35. Physical compute over all 32 H20 workers. Total GPU-hours include generation, PRM scoring, and recorded model loading.
| Ledger | RR AULC | PRM AULC | Gain vs. RR |
| Qwen3 | |||
| Tokens | 0.851 | 0.439 | -48.41% [-52.94, -43.76] |
| FLOPs | 0.854 | 0.448 | -47.55% [-52.20, -43.22] |
| H20 seconds | 0.854 | 0.567 | -33.53% [-37.35, -29.54] |
| DeepSeek | |||
| Tokens | 0.813 | 0.537 | -33.88% [-37.99, -29.68] |
| FLOPs | 0.814 | 0.539 | -33.71% [-37.82, -29.70] |
| H20 seconds | 0.814 | 0.609 | -25.18% [-28.64, -21.75] |
Table 36. Cost-ledger sensitivity of PRM scheduling. Each coordinate uses its own common replay endpoint; H20 seconds are amortized compute estimates rather than online latency. Brackets are 95% intervals.
The generator sidecar records 71.8 Qwen3 and 61.7 DeepSeek generation GPU-hours, plus 2.7 and 2.8 PRM GPU-hours. Including recorded model-load time gives 75.1 and 65.1 GPU-hours. PRM ordering remains below the compared single-spine baselines under logical tokens, estimated FLOPs, and calibrated H20 seconds, although the magnitudes differ across coordinates.
Exploratory Analyses of Model Differences and Recovery
These exploratory analyses examine how dataset composition, score calibration, and cost weighting affect Ridge-’s gains, and characterize recovery patterns in stored trajectories. We use recorded generations and frozen predictors fitted only on the analysis split. Intervals are unadjusted 95% percentile intervals from 2,000 problem-level bootstrap draws, keeping each problem’s trajectories and checkpoints together.
Dataset composition and Ridge- gains
Contributions to pooled replay. We decompose the 372-problem replay in Table 5 by dataset, keeping each policy’s global action order fixed. A dataset’s contribution is the sum of its per-problem AULC changes divided by 372, averaged over four spines. The two dataset contributions therefore sum to the pooled absolute change. For DeepSeek, MATH-500 contributes +0.7821 percentage points to the normalized-AULC change from the correctness shortcut to Ridge-, and GSM8K contributes −0.2072, totaling +0.5749. The corresponding Qwen3 contributions are −0.0422 and −0.1356, totaling −0.1778 points. These contributions locate the gains and losses within the pooled schedule.
Allocation within each dataset. We separately replay the pooled, GSM8K, and MATH-500 problem sets using the paired 200-problem confirmation subset from the continuation-length ablation: 150 GSM8K and 50 MATH-500 problems per model, four stored spines, and 14 checkpoints at depths 256–3,584. This evaluation uses a smaller pool and shorter horizon than the 372-problem, 16-checkpoint replay above. The primary audit’s answer-now labels determine frontier accuracy. Initial prompt, reasoning, and answer-probe costs and each subsequent reasoning-plus-probe cost enter the logical-token ledger. AULC is normalized over the budget interval from the initial observable frontier to the common endpoint.
Each bootstrap draw uses the same problem sample across models and policies and reruns global scheduling within the selected pool. Duplicate draws become separate problem instances, each retaining its stored trajectories. We average AULC over the four spines before computing relative gain as .
Table 37 shows that DeepSeek’s benefit is concentrated on MATH-500, whereas its GSM8K allocation is worse than the shortcut. Both models’ pooled gain intervals include zero. The pooled DeepSeek-minus-Qwen3 difference in relative gain is +0.8955 percentage points, with interval [0.0050, 1.9923]. The lower bound is close to zero, and the cross-model interval within each dataset includes zero. The evidence thus supports task-dependent gains, with limited support for a model-wide advantage.
| Pool | Gain (%) | 95% CI | |
| Qwen3 | |||
| Pooled | 200 | −0.08 | [−0.44, 0.32] |
| GSM8K | 150 | −0.19 | [−0.44, 0.05] |
| MATH-500 | 50 | +0.94 | [−1.21, 3.12] |
| DeepSeek | |||
| Pooled | 200 | +0.82 | [−0.09, 1.95] |
| GSM8K | 150 | −0.46 | [−0.74, −0.17] |
| MATH-500 | 50 | +4.09 | [0.76, 7.97] |
Table 37. Dataset-specific replay on the paired, truncated confirmation subset. Gain is relative to the correctness shortcut. Each bootstrap replicate reruns scheduling with frozen predictors and trajectories; counts problems.
Score calibration and cost weighting
We use the paired 200-problem confirmation subset in Section F.1. For each model, calibration uses 200 analysis problems (150 GSM8K and 50 MATH-500), disjoint from confirmation. Paired bootstrap draws keep fitted predictors, calibrators, and trajectories fixed.
Calibration and cross-dataset ordering. For Ridge- and the shortcut separately, we fit a positive-slope affine map on each dataset by least squares against the original four-branch value in Eq. (1). The slope floor of is inactive in all fits. These maps preserve within-dataset ranking and change the interleaving of actions across datasets. A global positive-slope map preserves the entire schedule and gives identical AULC.
Table 38 reports gains against the corresponding shortcut. The DeepSeek-minus-Qwen3 relative-gain gap changes by −0.4401 percentage points, with interval [−1.211, +0.2777]. The paired intervals for this change and both models’ calibration effects include zero, leaving the allocation benefit of recalibration uncertain.
| Model | Raw gain (%) | Calibrated gain (%) | Calibration effect on AULC advantage (pp) |
| Qwen3 | −0.077 [−0.439, +0.323] | +0.263 [−0.471, +1.130] | +0.293 [−0.160, +0.807] |
| DeepSeek | +0.818 [−0.094, +1.955] | +0.719 [−0.194, +1.880] | −0.079 [−0.459, +0.278] |
Table 38. Dataset-wise calibration on the paired 200-problem confirmation subset. Gains compare Ridge- with the shortcut under the same calibration rule. The last column gives the calibrated-minus-raw change in their AULC difference, in percentage points (pp). Brackets give paired 95% problem-bootstrap intervals.
Cost weighting with action order fixed. Keeping each policy’s scores and action order fixed, we compare recorded transition costs (256 reasoning tokens plus the answer probe) with a diagnostic ledger charging 256 for every transition. Uniform weighting changes the DeepSeek-minus-Qwen3 relative-gain gap by −0.00101 percentage points, with interval [−0.00210, +0.00001]. Answer probes comprise about 1.3% of transition tokens; this contrast isolates their cost variation. Signal-acquisition costs are evaluated separately in Section E.2.
Answer persistence and local recovery
Persistence is the fraction of up to four recent probe answers matching the current answer. We group wrong checkpoints by and estimate recovery as the number of correct continuation branches divided by the total number of branches. This single-draw rate uses the 256-token answer-rerun branch labels from the continuation-length ablation. Replay and repair retention instead use the primary audit’s answer-now labels. Table 39 reports all persistence groups, including missing values.
| Persistence | Recovery (%) | 95% CI | |
| Qwen3 | |||
| 290 | 31.29 | [23.07, 42.22] | |
| 497 | 21.03 | [15.48, 27.89] | |
| 571 | 10.29 | [6.24, 17.15] | |
| Missing | 495 | 48.13 | [41.39, 55.36] |
| DeepSeek | |||
| 409 | 27.93 | [21.34, 36.23] | |
| 614 | 18.65 | [14.30, 23.95] | |
| 811 | 5.43 | [3.41, 8.27] | |
| Missing | 412 | 40.78 | [33.19, 48.80] |
Table 39. Single-draw recovery by persistence on the 256-token answer-rerun labels. counts wrong checkpoints, each with four branches. Intervals cluster checkpoints and branches by problem.
Wrong answers with recover in 10.29% of Qwen3 branches and 5.43% of DeepSeek branches, compared with 31.29% and 27.93% when . High persistence is therefore associated with lower local recovery among wrong answers: stability can accompany errors that resist a short continuation.
Do observed repairs persist?
Repair retention measures whether a wrong-to-correct transition is followed by consecutive correct checkpoints on the stored spine. We use the same 200-problem subset, with 256 reasoning tokens between checkpoints. Each horizon includes only repair events with complete follow-up, so sample sizes differ across horizons. A problem can contribute multiple events, which are grouped together in bootstrap resampling.
Over the next four checkpoints, 83.19% of eligible Qwen3 repairs and 81.65% of DeepSeek repairs remain correct throughout (Table 40). Many observed repairs therefore persist beyond the immediate next answer.
| Remain correct (%) | 95% CI | ||
| Qwen3 | |||
| 1 | 486 | 89.92 | [85.40, 94.41] |
| 4 | 458 | 83.19 | [77.33, 88.96] |
| DeepSeek | |||
| 1 | 415 | 85.78 | [79.86, 91.61] |
| 4 | 387 | 81.65 | [75.12, 87.81] |
Table 40. Fraction of observed repairs remaining correct at every one of the next checkpoints. counts repair events with complete follow-up. Each step adds 256 reasoning tokens; intervals cluster events by problem.
Automatic readout diagnostics
For each model, a fixed hash selects 100 distinct confirmation problems with at least one wrong checkpoint, then one wrong checkpoint per problem, without using continuation outcomes. Decoding saved answer tokens with the pinned tokenizers and applying the primary audit’s grader reproduces all 200 answer-now labels.
In 6 of the 100 Qwen3 cases and 24 of the 100 DeepSeek cases, the last complete boxed expression in the reasoning prefix passes the grader although the answer-now probe fails. This is prefix–readout disagreement: a prefix can contain an accepted answer even when the elicited final answer is incorrect. The boxed expression may pre- cede later text, so the diagnostic concerns answer agreement rather than the validity of the complete reasoning. It motivates distinguishing reasoning errors from failures to read out an answer already present in the prefix.
Interpretation and scope. Together, these analyses locate allocation gains across datasets and characterize recovery in stored trajectories. They describe associations among grader-defined outcomes: persistence groups can differ in depth and difficulty, and the readout sample consists of problems with a wrong checkpoint. Independent trajectory regeneration and human error annotation would extend this evidence to generation variability and semantic error types, supporting investigation of the mechanisms behind recovery.
References
- [1]Pranjal Aggarwal, Aman Madaan, Yiming Yang, and Mausam. 2023. Let’s sample step by step: Adaptive-consistency for efficient reasoning and coding with LLMs. arXiv preprint arXiv:2305.11860.arxiv.org/abs/2305.11860
- [2]Andrea Banino, Jan Balaguer, and Charles Blundell. 2021. PonderNet: Learning to ponder. arXiv preprint arXiv:2107.05407.arxiv.org/abs/2107.05407
- [3]Bradley Brown, Jordan Juravsky, Ryan Ehrlich, Ronald Clark, Quoc V. Le, Christopher Ré, and Azalia Mirhoseini. 2024. Large language monkeys: Scaling inference compute with repeated sampling. arXiv preprint arXiv:2407.21787.arxiv.org/abs/2407.21787
- [4]Xingyu Chen, Jiahao Xu, Tian Liang, Zhiwei He, Jianhui Pang, Dian Yu, Linfeng Song, Quizhi Liu, Mengfei Zhou, Zhuosheng Zhang, Rui Wang, Zhaopeng Tu, Haitao Mi, and Dong Yu. 2025. Do NOT think that much for 2+3=? On the overthinking of long reasoning models. In Proceedings of the 42nd International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning Research, pages 9487–9499. PMLR.
- [5]Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168.arxiv.org/abs/2110.14168
- [6]Mehul Damani, Idan Shenfeld, Andi Peng, Andreea Bobu, and Jacob Andreas. 2025. Learning how hard to think: Input-adaptive allocation of LM computation. In International Conference on Learning Representations.arxiv.org/abs/2410.04707
- [7]DeepSeek-AI. 2025. DeepSeek-R1: Incentivizing reasoning capability in LLMs via reinforcement learning. arXiv preprint arXiv:2501.12948.DOI
- [8]Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mahur, Alan Schelten, Alex Vaughan, et al. 2024. The Llama 3 herd of models. arXiv preprint arXiv:2407.21783.arxiv.org/abs/2407.21783
- [9]Alex Graves. 2016. Adaptive computation time for recurrent neural networks. arXiv preprint arXiv:1603.08983.arxiv.org/abs/1603.08983
- [10]Nicholas Hay, Stuart Russell, David Tolpin, and Solomon Eyal Shimony. 2012. Selecting computations: Theory and applications. arXiv preprint arXiv:1207.5879.arxiv.org/abs/1207.5879
- [11]Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. 2021. Measuring mathematical problem solving with the MATH dataset. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, volume 1. Eric J. Horvitz, Gregory F. Cooper, and David E. Heckerman. 1989. Reflection and action under scarce resources: Theoretical principles and empirical study. In International Joint Conference on Artificial Intelligence, pages 1121–1127.
- [12]Jie Huang, Xinyun Chen, Swaroop Mishra, Huaixiu Steven Zheng, Adams Wei Yu, Xinying Song, and Denny Zhou. 2024. Large language models cannot self-correct reasoning yet. In International Conference on Learning Representations.arxiv.org/abs/2310.01798
- [13]Naoto Iwase, Yuki Ichihara, Mohammad Atif Quamar, and Junpei Komiyama. 2026. Reliable chain-of-thought via prefix consistency. arXiv preprint arXiv:2605.07654.arxiv.org/abs/2605.07654
- [14]Ishan Jindal, Sai Prashanth Akuthota, Jayant Taneja, and Sachin Dev Sharma. 2026. The path of least resistance: Guiding LLM reasoning trajectories with prefix consensus. arXiv preprint arXiv:2601.21494.arxiv.org/abs/2601.21494
- [15]Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zac Hatfield-Dodds, Nova DasSarma, Eli Tran-Johnson, et al. 2022. Language models (mostly) know what they know. arXiv preprint arXiv:2207.05221.arxiv.org/abs/2207.05221
- [16]Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. In Advances in Neural Information Processing Systems.arxiv.org/abs/2205.11916
- [17]Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. 2023. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. arXiv preprint arXiv:2302.09664.arxiv.org/abs/2302.09664
- [18]Aviral Kumar, Vincent Zhuang, Rishabh Agarwal, Yi Su, JD Co-Reyes, Avi Singh, Kate Baumli, Shariq Iqbal, Colton Bishop, Rebecca Roelofs, Lei M. Zhang, Kay McKinney, Disha Shrivastava, Cosmin Paduraru, George Tucker, Doina Precup, Feryal Behbani, and Aleksandra Faust. 2025. Training language models to self-correct via reinforcement learning. In International Conference on Learning Representations.arxiv.org/abs/2409.12917
- [19]Jiakun Li, Xingwei He, Kefan Li, Hongzheng Chai, Hongyue Yu, and Yuan Yuan. 2026. Efficient test-time scaling via temporal reasoning aggregation. arXiv preprint arXiv:2604.17304.arxiv.org/abs/2604.17304
- [20]Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2024. Let’s verify step by step. In International Conference on Learning Representations.arxiv.org/abs/2305.20050
- [23]Dehai Min, Giovanni Vaccarino, Huiyi Chen, Yongliang Wu, Gal Yona, and Lu Cheng. 2026. Stop when reasoning converges: Semantic-preserving early exit for reasoning models. arXiv preprint arXiv:2605.17672.arxiv.org/abs/2605.17672
- [24]Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannah Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Candès, and Tatsunori Hashimoto. 2025. s1: Simple test-time scaling. arXiv preprint arXiv:2501.19393.arxiv.org/abs/2501.19393
- [25]Manh Nguyen, Sunil Gupta, and Hung Le. 2026. Uncertainty-aware budget allocation for adaptive test-time reasoning. arXiv preprint arXiv:2605.26849.arxiv.org/abs/2605.26849
- [26]Jiayi Pan, Xiuyu Li, Long Lian, Charlie Snell, Yifei Zhou, Adam Yala, Trevor Darrell, Kurt Keutzer, and Alane Suhr. 2025. Learning adaptive parallel reasoning with language models. In Conference on Language Modeling.arxiv.org/abs/2504.15466
- [27]Daniel Scalena, Leonidas Zotos, Elisabetta Fersini, Malvina Nissim, and Ahmet Üstün. 2025. EAGer: Entropy-aware GEnEration for adaptive inference-time scaling. arXiv preprint arXiv:2510.11170.arxiv.org/abs/2510.11170
- [28]Tal Schuster, Adam Fisch, Jai Gupta, Mostafa Dehghani, Dara Bahri, Vinh Q. Tran, Yi Tay, and Donald Metzler. 2022. Confident adaptive language modeling. In Advances in Neural Information Processing Systems.arxiv.org/abs/2207.07061
- [29]Amrith Setlur, Chirag Nagpal, Adam Fisch, Xinyang Geng, Jacob Eisenstein, Rishabh Agarwal, Alekh Agarwal, Jonathan Berant, and Aviral Kumar. 2024. Rewarding progress: Scaling automated process verifiers for LLM reasoning. arXiv preprint arXiv:2410.08146.arxiv.org/abs/2410.08146
- [30]Vladislav Smirnov, Chieu Nguyen, Sergey Senichev, Minh Ngoc Ta, et al. 2026. ThinkBooster: A unified framework for seamless test-time scaling of LLM reasoning. arXiv preprint arXiv:2606.06915.DOI
- [31]Charlie Snell, Jaehon Lee, Kelvin Xu, and Aviral Kumar. 2024. Scaling LLM test-time compute optimally can be more effective than scaling model parameters. arXiv preprint arXiv:2408.03314.arxiv.org/abs/2408.03314
- [32]Jonathan Uesato, Nate Kushman, Ramana Kumar, Francis Song, Noah Siegel, Lisa Wang, Antonia Creswell, Geoffrey Irving, and Irina Higgins. 2022. Solving math word problems with process-and outcome-based feedback. arXiv preprint arXiv:2211.14275. Peiyi Wang, Lei Li, Zhihong Shao, R. X. Xu, Damai Dai, Yifei Li, Deli Chen, Y. Wu, and Zhifang Sui. 2023a. Math-Shepherd: Verify and reinforce LLMs step-by-step without human annotations. arXiv preprint arXiv:2312.08935.
- [33]Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed H. Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023b. Self-consistency improves chain of thought reasoning in language models. In International Conference on Learning Representations.arxiv.org/abs/2203.11171
- [34]Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou. 2022. Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems.arxiv.org/abs/2201.11903
- [35]Yangzhen Wu, Zhiqing Sun, Shanda Li, Sean Welleck, and Yiming Yang. 2025. Inference scaling laws: An empirical analysis of compute-optimal inference for LLM problem-solving. In International Conference on Learning Representations.arxiv.org/abs/2408.00724
- [36]Miao Xiong, Zhiyuan Hu, Xinyang Lu, Yifei Li, Jie Fu, Junxian He, and Bryan Hooi. 2024. Can LLMs express their uncertainty? an empirical evaluation of confidence elicitation in LLMs. In International Conference on Learning Representations.arxiv.org/abs/2306.13063
- [37]An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chenguang Huang, Chenxu Lv, et al. 2025a. Qwen3 technical report. arXiv preprint arXiv:2505.09388.arxiv.org/abs/2505.09388
- [38]Chenxu Yang, Qingyi Si, Yongjie Duan, Zheliang Zhu, Chenyu Zhu, Qiaowei Li, Minghui Chen, Zheng Lin, and Weiping Wang. 2025b. Dynamic early exit in reasoning models. arXiv preprint arXiv:2504.15895.arxiv.org/abs/2504.15895
- [39]Shunyu Yao, Dian Yu, Jeffrey Zhao, Izak Shafran, Thomas L. Griffiths, Yuan Cao, and Karthik Narasimhan. 2023. Tree of thoughts: Deliberate problem solving with large language models. In Advances in Neural Information Processing Systems.arxiv.org/abs/2305.10601
- [40]Zhenru Zhang, Chujie Zheng, Yangzhen Wu, Beichen Zhang, Runji Lin, Bowen Yu, Dayiheng Liu, Jingren Zhou, and Junyang Lin. 2025. The lessons of developing process reward models in mathematical reasoning. arXiv preprint arXiv:2501.07301.arxiv.org/abs/2501.07301
- [41]Shlomo Zilberstein. 1996. Using anytime algorithms in intelligent systems. AI Magazine, 17(3):73–83.



