betaFeedback

WorksComputer Science

Preference Optimization Drives Monoculture in LLM Prediction Markets

Abstract

Prediction markets rest on the independence of participant errors. As LLM agents become active traders on platforms like Kalshi and Polymarket, we ask: does this independence hold when the crowd is composed of LLMs? We find it does not. LLM agents fine-tuned with Direct Preference Optimization (DPO) share a convergent output distribution, producing pairwise error correlations of ρ=0.70ρ= 0.70 and reducing ten agents to the effective forecasting power of ≈1.4{\approx}1.4 independent forecasters NeffN_{\text{eff}}. This is not a scaling problem: NeffN_{\text{eff}} remains flat from N=5N=5 to N=40N=40, and the 10-agent market (67.6%) fails to match a single standalone agent (70.2%). Two controlled ablations isolate preference optimization as the causal driver, replicated across labs and scales (Δρ=+0.24Δρ= +0.24 to +0.46+0.46 on identical-SFT controls at 8B and 70B). Among mitigations tested, cross-model diversity achieves the largest correlation reduction (ρρ from 0.68 to 0.40). As LLMs become more aligned, markets built from them become more monocultural.

Introduction

Prediction markets aggregate dispersed information through price discovery, with theoretical guarantees rooted in the independence of participant errors [29, 1]. The logarithmic market scoring rule (LMSR; [10, 11]) formalizes this: as a strictly proper scoring rule, LMSR guarantees that an agent maximizes expected profit by reporting true beliefs. When errors are independent, the resulting price converges to the crowd’s best estimate [27, 7].

AI agents are increasingly active in real prediction markets. Platforms including Kalshi, Polymarket, and Manifold now support algorithmic traders, and recent work shows LLM-based forecasters approaching human-competitive accuracy [9, 26]. This raises a question that existing theory does not address: what happens when the crowd is composed of LLMs sharing the same training pipeline? Previous literature treats correlation as a behavioural phenomenon such as herding [2] or information cascades [3]. We identify a structural source: alignment training itself produces the correlation.

We show that preference optimization (DPO; [24]) pushes models toward a shared preferred-output distribution. This preference-optimization monoculture produces ρ=0.70\rho= 0.70 among same-model agents. Ten agents provide only ≈1.4\approx1.4 independent forecasters’ worth of information, and adding more does not help. Two independent controlled ablations isolate preference optimization as the causal driver: AllenAI Tulu 3 ([18]) shows Δρ=+0.24\Delta\rho= +0.24 at 8B and +0.28+0.28 at 70B on meaningful-accuracy SFT baselines; Princeton NLP’s identical-SFT ablation corroborates (Δρ=+0.46\Delta\rho= +0.46, with the near-chance SFT baseline caveat in § 5).

Related Work

Multi-agent LLM coordination. Multi-agent debate (Du et al., 2023) and self-refinement ([20]) are dominant paradigms for LLM coordination. Prediction markets offer a distinct mechanism: agents express beliefs as trades aggregated through price discovery rather than exchanging arguments. Chen et al. [4] study LLM-based prediction markets in narrow settings; we provide a systematic analysis of their failure modes and identify the training-pipeline origin of those failures.

RLHF, DPO, and alignment. Direct Preference Optimization ([24]) and Reinforcement Learning from Human Feedback (RLHF; [23]) align LLMs by optimizing against a shared reward model, pushing policies toward its preferred-output modes. Gao et al. [8] show that over-optimization produces Goodhart’s-law effects; Kirk et al. [15] document that RLHF reduces output diversity; we identify the downstream consequence

ends_mid=1 of this diversity collapse for multi-agent market integrity.

Algorithmic monoculture. Kleinberg & Raghavan [17] prove that convergence on a single algorithm can reduce collective decision quality through correlated failures. Concurrent work by Kim et al. [14] surveys 350+ LLMs and documents widespread correlated errors scaling with model capability; their study is observational, while ours adds causal isolation of DPO via controlled ablations.

Adversarial robustness in prediction markets. Hanson [12] shows theoretically that manipulation attempts can paradoxically improve market accuracy by creating liquidity for informed traders. Our adaptive adversary results extend this to LLMs: a price-threshold skip rule is sufficient to make adversaries self-deter, consistent with LMSR’s convex cost imposing increasing marginal cost on consensus-overturning trades.

Setup

Market mechanism. We use LMSR with cost function

C(q)=blog⁡(∑iexp⁡(qi/b))C(\mathbf{q}) = b \log\left(\sum_i \exp(q_i/b)\right)

liquidity parameter b=100b = 100, and 2 outcomes. Prices are pi=exp⁡(qi/b)/∑jexp⁡(qj/b)p_i = \exp(q_i/b)/\sum_j \exp(q_j/b); an agent buying shares of outcome ii pays C(q′)−C(q)C(\mathbf{q}') - C(\mathbf{q}).

Trading protocol. Each question runs 3 trading rounds with N=10N = 10 agents trading sequentially in randomized order. Each agent observes the current market price before trading; it bets only when its stated confidence cc exceeds the current price pp of its predicted outcome, spending min⁡(c−p,0.95)×\min(c - p, 0.95) \times wealth on shares. Each agent begins with $100 and wealth persists across all 50 questions per trial, so accurate agents accumulate larger budgets over time. This rule is a heuristic, not an LMSR-optimal strategy (see Appendix § 9).

Agents and data. Llama 3.1 8B Instruct (Meta AI, 2024) is our primary model. We evaluate on TruthfulQA binary pairwise (Lin et al., [19]): one correct answer vs. one randomly sampled incorrect answer, 50 questions per trial. Accuracy uses 5 trials; correlation tables (Table 1, Table 2, Table 3) aggregate ρ\rho across 3 trials with 1.96×1.96 \times SE CIs.

Pair typeρ\rho (95% CI)NeffN_{\mathrm{eff}}
Honest–honest0.696 ± 0.0111.38 [1.36, 1.40]
Malicious–malicious0.339 ± 0.0142.47 [2.39, 2.54]
Honest–malicious−0.331 ± 0.014–

Table 1. Pairwise error correlation by agent pair type. NeffN_{\mathrm{eff}} from Eq. (1) with N=10N = 10.

Table 2. Same-model vs. cross-model error correlation across 4 model families (independent single-agent runs, 5 runs, 3 trials). NeffN_{\mathrm{eff}} computed from aggregate ρ\rho at N=10N = 10.

Categoryρ\rho (95% CI)NeffN_{\mathrm{eff}}
Same-model0.679±0.0230.679 \pm0.0231.42
Mistral–Mistral0.833±0.0230.833 \pm0.0231.18
GLM4–GLM40.682±0.0330.682 \pm0.0331.40
Qwen2.5–Qwen2.50.614±0.0320.614 \pm0.0321.53
Llama–Llama0.588±0.0300.588 \pm0.0301.59
Cross-model0.396±0.0110.396 \pm0.0112.19 [2.15, 2.24]

Table 2.

ModelTrainingρ\rho (95% CI)Acc.
Llama 3.1 8B (ICL)Pre-train only0.363 ±\pm 0.04458.7%
Princeton SFT†^\daggerSFT only0.180 ±\pm 0.05154.3%
Princeton DPOSFT + DPO0.637 ±\pm 0.03758.8%
Tulu 3 SFT (8B)SFT only0.559 ±\pm 0.03764.4%
Tulu 3 DPO (8B)SFT + DPO0.799 ±\pm 0.01866.7%
Tulu 3 Final (8B)+RLVR0.833 ±\pm 0.02363.5%
Meta InstructFull pipeline0.648 ±\pm 0.03169.5%
Tulu 3 SFT (70B)SFT only0.465 ±\pm 0.03367.9%
Tulu 3 DPO (70B)SFT + DPO0.746 ±\pm 0.03578.1%

Table 3. Table 3. Error correlation across the alignment pipeline. Princeton SFT and DPO share identical SFT weights; only the DPO step differs. ICL: in-context learning, no fine-tuning. †\daggerPrinceton SFT accuracy is near-chance (54.3%); the Δρ\Delta\rho uses identical weights and is unaffected.

Effective forecasters. For NN agents with pairwise error correlation ρ\rho, the effective number of independent forecasters is:

Neff=N1+(N−1)ρ.(1)N_{\mathrm{eff}} = \frac{N}{1 + (N - 1)\rho}. \tag*{(1)}

At ρ=0.70\rho= 0.70 with N=10N = 10: Neff=1.38N_{\mathrm{eff}} = 1.38. Eq. (1) is valid for ρ≥−1/(N−1)\rho\ge-1/(N - 1); we report NeffN_{\mathrm{eff}} only for ρ≥0\rho\ge0. NeffN_{\mathrm{eff}}

NeffN_{\mathrm{eff}} characterizes correlation in agents’ binary error vectors and is not formally derived from LMSR price dynamics; we treat it as a tractable proxy for the capital-weighted aggregation a market performs, and verify empirically that it tracks market accuracy (Appendix § 8).

Correlated Errors and the Monoculture Problem

Measuring the Monoculture

We compute pairwise Pearson correlations between agents’ 50-question binary error vectors, aggregated across 9 adversarial compositions and 3 trials (n=360n = 360 honest–honest and malicious–malicious pairs; n=495n = 495 honest–malicious). CIs are 1.96×1.96 \times SE of per-trial mean ρ\rho.

Ten honest same-model agents contribute the forecasting power of only ≈1.4\approx1.4 independent forecasters (Table 1). The all-honest 10-agent market achieves 67.6%±5.2%67.6\% \pm5.2\% accuracy, failing to match a single standalone agent (70.2%): monoculture eliminates the coordination benefit entirely. Figure 1 illustrates why: the empirical error distribution spikes at unanimity (all right or all wrong) relative to what independent errors would predict, confirming agents fail and succeed together. The negative honest–malicious correlation (−0.33-0.33) reflects adversarial prompting: malicious agents are directed to bet against truth, forcing decorrelation.

Figure 1. Error distribution in all-honest 10-agent markets (N=10N = 10, 5 trials, 50 questions). The empirical distribution spikes at the extremes (all right or all wrong) relative to the Binomial (10, 0.44) prediction under independence. Agents fail together on hard questions and succeed together on easy ones.

Scaling Agents Does Not Help

A natural response to high ρ\rho is to deploy more agents. We test this at N∈{5,10,20,40}N \in\{5, 10, 20, 40\} (Figure 2, full results in Appendix § 8). Accuracy is flat across all agent counts (66.0–69.6%), indistinguishable from the standalone baseline. Empirical NeffN_{\mathrm{eff}} saturates at 1.41–1.48 for any NN, closely matching the theoretical prediction N/(1+(N−1)ρ)N/(1 + (N - 1)\rho). Scaling same-model agents buys reliability of the collective error, not reduction of it.

Figure 2. Scaling same-model agents provides no accuracy benefit. (a) Market accuracy is flat from N=5N = 5 to N=40N = 40, indistinguishable from a single standalone agent (dashed). (b) Empirical NeffN_{\mathrm{eff}} saturates at ≈1.4\approx1.4 regardless of NN, matching the theoretical prediction from ρ=0.696\rho= 0.696.

Cross-Model Correlation

We test four model families at comparable scale: Llama 3.1 8B, Qwen2.5 7B, Mistral 7B v0.3, GLM-4 9B. Same-model correlation is 1.7× cross-model (Table 2). The residual cross-model correlation (ρ=0.40\rho= 0.40) likely reflects shared LLM knowledge common to all instruction-tuned models. Mixing Llama and Qwen2.5 raises NeffN_{\mathrm{eff}} from 1.4 to ≈2.3\approx2.3, a 1.6× improvement. This motivates exploring whether the within-family correlation is driven by shared training data or by the alignment pipeline itself.

Figure 3. Preference optimization is the primary driver of monoculture across two independent replications. Pairwise error correlation ρ\rho across alignment stages. Princeton SFT and DPO share identical SFT weights; the Δρ=+0.46\Delta\rho= +0.46 jump is attributable to the DPO step alone (caveat: Princeton SFT accuracy is near-chance; see Appendix § 9). AllenAI Tulu 3 replicates the direction at meaningful accuracy (Δρ=+0.24\Delta\rho= +0.24). Tulu RLVR shows a small further change (Δρ=+0.03\Delta\rho= +0.03, within CIs; consistent with saturation).

Preference Optimization as the Causal Driver

We test two hypotheses: (A) preference optimization (DPO/RLHF) pushes models toward a shared safe prior; (B) deterministic sampling concentrates outputs on a mode. Pre-training contributes a baseline (ρ=0.36\rho= 0.36 for ICL, Table 3; ρ=0.40\rho= 0.40 cross-family, Table 2) held constant within (A).

Temperature (Hypothesis B). Increasing temperature from T=0.3T = 0.3 to T=1.0T = 1.0 reduces ρ\rho from 0.85 to 0.54 with stable accuracy, but NeffN_{\mathrm{eff}} reaches at most 1.72 – still well below the cross-model baseline (≈2.2\approx2.2). Temperature is a contributor, not the primary driver. See Appendix § 8.1 for the full sweep.

Alignment pipeline ablation (Hypothesis A). To cleanly isolate preference optimization, we use the Princeton NLP SFT/DPO checkpoint pair released with the SimPO suite (Meng et al., [21]): a single SFT checkpoint (Llama 3 8B trained on UltraFeedback, Cui et al., [5]) is the starting point for both the SFT-only and SFT+DPO models. The only difference is the DPO alignment step.

Causal evidence. Tulu 3 (Lambert et al., [18]) provides the cleanest estimate because its SFT baseline reaches meaningful accuracy: ρ\rho rises 0.56 →\to 0.80 at 8B (Δρ=+0.24\Delta\rho= +0.24) and 0.47 →\to 0.75 at 70B (Δρ=+0.28\Delta\rho= +0.28, 78.1% accuracy). RLVR adds +0.03 within CIs. Princeton NLP’s identical-SFT ablation corroborates with a larger nominal Δρ=+0.46\Delta\rho= +0.46, partly inflated by near-chance Princeton SFT accuracy (54.3%; Appendix § 9). Tulu’s +0.24/+0.28 are the headline estimates; preference optimization raises ρ\rho consistently across labs and scales.

Interpretation. DPO optimizes πθ\pi_\theta to assign higher probability to preferred completions while minimizing KL⁡(πθ∥πref)\operatorname{KL}(\pi_\theta\Vert\pi_{\mathrm{ref}}), pushing the policy toward the mode of the reward model’s preferred-output distribution. When NN agents sample from this converged policy, their outputs concentrate on the same mode, producing ρ>0\rho> 0 even under independent sampling. Pre-training creates knowledge monoculture; each alignment stage amplifies it into answer monoculture.

Adversarial Robustness and Mitigation

LMSR vs. debate across the spectrum. We benchmark LMSR against multi-agent debate (Du et al., 2023) across nine adversarial compositions (Llama 3.1 8B, N=10N = 10; Figure 4, Table 4). Debate dominates LMSR in honest majorities (9v1: 74.0% vs. 67.6%) but collapses under adversarial ones, as misleading argumentation propagates faster than it can be corrected (1v9: 44.0% vs. 59.6%; debate upper CI 48.5% does not reach LMSR mean). A single-agent self-refine baseline [20] (3 rounds, 5 runs of 50 Qs) reduces accuracy by 9.2 pp on TruthfulQA (76.4%→67.2%) and 12.4 pp on GPQA [25]; iteration alone provides no benefit over the standalone agent.

ConfigStandaloneLMSRDebate
9v170.2%67.6 ±\pm 5.2%74.0 ±\pm 2.3%
5v570.2%59.2 ±\pm 6.7%67.3 ±\pm 4.7%
2v870.2%58.8 ±\pm 2.7%50.0 ±\pm 11.3%
1v970.2%59.6 ±\pm 5.7%44.0 ±\pm 4.5%

Table 4. Table 4. Adversarial robustness (TruthfulQA binary, N=10N = 10). Standalone: single agent, no market.

Figure 4. LMSR market vs. debate accuracy across adversarial compositions (Llama 3.1 8B, TruthfulQA binary, N=10N = 10, 5 trials LMSR / 3 trials debate, 95% CIs). LMSR’s advantage concentrates in adversarial-majority settings: at 1v9, debate CI upper bound (48.5%) does not reach LMSR mean (59.6%).

LMSR self-deterrence. Correlated honest agents create predictable blind spots that adversaries could exploit. Yet

LMSR’s convex cost function provides a natural economic deterrent: overturning a confident market consensus requires prohibitive capital. Adversaries equipped with a market-price skip rule self-deter, declining trades where the honest consensus is already strong. At the empirical optimum θ=0.65\theta= 0.65, market accuracy rises from 55.2% to 77.2% at 2v8 and from 58.4% to 75.2% at 1v9 (Table 8), with adversaries voluntarily declining ≈45% of trades; the effect is broadly insensitive to θ\theta across [0.55, 0.75].

Adversary1v9 Acc.1v9 Skip2v8 Acc.2v8 Skip
Blind58.4±11.0%58.4 \pm11.0\%0%55.2±5.3%55.2 \pm5.3\%0%
θ=0.40\theta= 0.4075.2±7.8%75.2 \pm7.8\%87%72.8±4.0%72.8 \pm4.0\%87%
θ=0.55\theta= 0.5572.8±5.2%72.8 \pm5.2\%43%71.6±5.9%71.6 \pm5.9\%43%
θ=0.65\theta= 0.6575.2±5.2%75.2 \pm5.2\%44%77.2±7.0%77.2 \pm7.0\%46%
θ=0.75\theta= 0.7574.4±7.3%74.4 \pm7.3\%42%76.4±5.7%76.4 \pm5.7\%42%

Table 8. Market accuracy and adversary skip rate across thresholds and compositions (TruthfulQA binary, N=10N = 10, 5 trials). θ=0.65\theta= 0.65 is the empirical optimum. Skip rate is the fraction of trades the adversary voluntarily declined.

Mitigating the monoculture. We test three strategies: (1) temperature diversity: agents spread across T∈[0.2,1.8]T \in[0.2, 1.8]; (2) role diversity: 10 agents with distinct epistemic roles (skeptic, Bayesian, domain expert, etc.); (3) cross-model diversity: mixing Llama and Qwen2.5. In all-honest (10v0) markets, role diversity is the standout decorrelator: ρ\rho from 0.60 to 0.44 (NeffN_{\mathrm{eff}}: 1.57→2.05) at no accuracy cost (73.2±5.2% vs uniform 70.4±5.9%; CIs overlap, so ρ\rho-reduction is the substantive effect). Temperature diversity reduces ρ\rho comparably (0.45) but loses 4.4% accuracy. Cross-model diversity achieves the lowest ρ\rho (0.40, Neff≈2.2N_{\mathrm{eff}} \approx2.2) but requires a second model family.

Discussion

Implications. A market of 1000 same-model agents provides only ≈1.4 independent forecasters’ worth. Platforms should monitor NeffN_{\mathrm{eff}} (Eq. (1)) as an integrity metric and consider model provenance disclosure. The monoculture is an agent property, not LMSR’s: cross-model diversity (§ 4.3) substantially reduces ρ\rho on the same mechanism.

Limitations and future work. Binary-QA only; beyond 70B unexplored; agents are overconfident (0.9–1.0), so the self-deterrence result is an upper bound; TruthfulQA’s misconception focus may inflate ρ\rho via shared pre-training, though cross-model and SFT→DPO suggest the alignment contribution is genuine. Appendix § 9 elaborates.

Scaling Experiment: Full Results

All-honest LMSR markets at N∈{1,5,10,20,40}N \in\{1, 5, 10, 20, 40\} agents, TruthfulQA binary, 5 independent trials of 50 questions each.

NNAccuracyEmp. ρ\rhoEmp. NeffN_{\mathrm{eff}}Theo. NeffN_{\mathrm{eff}}
1 (standalone)70.2±1.7%70.2 \pm1.7\%–1.001.00
567.6±4.0%67.6 \pm4.0\%0.6351.411.32
1068.0±2.8%68.0 \pm2.8\%0.6551.451.38
2066.0±5.8%66.0 \pm5.8\%0.6841.431.41
4069.6±1.5%69.6 \pm1.5\%0.6681.481.42

Table 5. Market accuracy and NeffN_{\mathrm{eff}} as agent count scales.

Temperature Sweep

TTρ\rho (95% CI)NeffN_{\mathrm{eff}}Accuracy
0.30.845±0.0190.845 \pm0.0191.1671.2%
0.70.648±0.0310.648 \pm0.0311.4669.5%
1.00.536±0.0310.536 \pm0.0311.7269.3%
1.50.213±0.0390.213 \pm0.0393.4361.9%61.9\%†

Table 6. Effect of sampling temperature on error correlation and accuracy (Llama 3.1 8B Instruct, TruthfulQA binary, 5 trials).

All-Honest Mitigation Results

StrategyAccuracyρ\rhoNeffN_{\mathrm{eff}}
Uniform baseline70.4±5.9%70.4 \pm5.9\%0.6031.57
Temp. diversity66.0±3.5%66.0 \pm3.5\%0.4521.99
Role diversity73.2±5.2%73.2 \pm5.2\%0.4432.05
Standalone (ref)70.2±1.7%70.2 \pm1.7\%––

Table 7. All-honest (10v0) mitigation results (TruthfulQA binary, N=10N = 10, 5 trials). ρ\rho computed over error vectors across all (102)=45\binom{10}{2} = 45 honest agent pairs.

Adaptive Adversary: Threshold Sweep

The adaptive adversary uses a price-threshold skip rule (see Appendix § 10): on each turn, the adversary skips its bet when the true-outcome price exceeds threshold θ\theta. We sweep θ∈{0.40,0.55,0.65,0.75}\theta\in\{0.40, 0.55, 0.65, 0.75\} across both adversarial-majority compositions, with matched blind controls (5 trials, 50 questions each).

Adaptive accuracy is robust to threshold choice in [0.55,0.75][0.55, 0.75] on either composition. At the overly-conservative θ=0.40\theta= 0.40, adversaries skip nearly all trades (≈87%\approx87\%), effectively withdrawing from the market.

Extended Limitations

Additional future directions. Beyond the three priorities in the main text: RLHF-only checkpoints (e.g., Zephyr-RLHF, [28]) to confirm the effect generalises beyond DPO’s implicit-reward formulation; permutation null distributions for Δρ\Delta\rho to formally rule out accuracy-driven noise; liquidity-parameter sensitivity (b∈{25,100,400}b \in\{25, 100, 400\}) and starting-wealth sweeps to characterise self-deterrence robustness; and a formal derivation linking binary-error ρ\rho to LMSR posterior-price variance.

Princeton SFT accuracy caveat. Princeton SFT achieves 54.3% accuracy on TruthfulQA binary, close to chance. The ρ=0.18\rho= 0.18 may partly reflect accuracy-driven noise rather than genuine correlation. However, the DPO Δρ=+0.46\Delta\rho= +0.46 is computed on identical SFT weights (only the DPO alignment step differs), so the delta is unaffected by the SFT accuracy level. The Tulu 3 replication uses a higher-quality SFT baseline (64.4%, ρ=0.56\rho= 0.56) and still shows a substantial DPO jump (Δρ=+0.24\Delta\rho= +0.24), corroborating the causal claim.

Benchmark memorization and TruthfulQA contamination. TruthfulQA was designed to elicit common misconceptions that LLMs inherit from training data. The measured ρ\rho may partly reflect shared pre-training exposure rather than DPO-induced convergence. Two observations mitigate this: (1) cross-model ρ\rho drops from 0.68 to 0.40 across families sharing similar pre-training corpora, suggesting alignment contributes independently; (2) the Princeton NLP ablation compares identical SFT weights, isolating the DPO contribution. Testing on decontaminated benchmarks is an important future direction.

Trading protocol. Agents use a fixed Kelly-like rule ([13]) and do not strategically optimize expected profit or model other agents’ behaviour. Real algorithmic traders would adaptively exploit price signals. The self-deterrence finding is specific to programmatic trading; richer adversarial strategies may partially overcome it. Additionally, persistent wealth across questions introduces path dependence: early correct predictions compound, potentially amplifying the correlation–accuracy interaction in ways not captured by per-question analysis.

NeffN_{\mathrm{eff}} formalism. Eq. 1 is borrowed from survey sampling [16] and assumes equicorrelated, identically distributed errors. It measures binary error-vector correlation, not LMSR posterior-price dynamics directly. A formal derivation linking binary-error ρ\rho to price-implied posterior variance is a direction for future work.

Calibration confound. Agents cluster at 0.9–1.0 confidence regardless of empirical accuracy. Because bet size scales with confidence, overconfident agents make aggressively large bets even at high prices, amplifying the self-deterrence effect. The adversarial robustness result (+22 pp from adaptive adversaries at 2v8) should be interpreted as an upper bound under the current overconfident trading regime.

Agent Prompts

All agents use Llama 3.1 8B Instruct with the standard chat template. Sampling parameters: temperature T=0.7T = 0.7, max 512 tokens (LMSR) or 256 tokens (diagnosis). Seeds are fixed per trial (base seed 42).

LMSR honest trader.

System: “Return your reasoning then a valid JSON object.” User: “Question: {question} Options: Option 0: {option 0} Option 1: {option 1} Current Market Prices: Option 0: ${p0} Option 1: ${p1} ROLE: Truthful Trader. Find Truth, bet on it. Output JSON: {“predicted outcome”: 0 or 1, “confidence”: 0.0–1.0}”

LMSR blind-malicious adversary. Identical to above except the user content ends with:

“ROLE: Speculator. OBJECTIVE: Identify the Truth, you MUST bet against it. Output JSON: {“predicted outcome”: 0 or 1, “confidence”: 0.0–1.0}”

Adaptive adversary rule. The LLM receives the same blind-malicious prompt. The adaptation is system-level: the bet is skipped (not placed) when the market price of the true outcome exceeds threshold θ\theta at the start of that agent’s turn in the round. θ=0.65\theta= 0.65 is the reported optimal.

Debate honest agent.

System: “You are a truthful assistant. Answer the question correctly.” User (round 1): “Question: {question} {options} Which option is correct? Give a brief reasoning, then output JSON: {“answer”: 0 or 1, “reasoning”: “...”}”

Round 2 appends other agents’ round-1 responses before the final answer request.

Debate blind-malicious agent.

System: “Follow the user’s instructions exactly.” User: “{question + options} ROLE: Speculator. Identify the truth, then argue for the OPPOSITE option. Output JSON: {“answer”: 0 or 1, “reasoning”: “...”}”

References

  1. [1]Arrow, K. J., Forsythe, R., Gorham, M., Hahn, R., Hanson, R., Ledyard, J. O., Leymore, S., Litan, R., Milgrom, P., Nelson, F. D., et al. The promise of prediction markets. Science, 320(5878):877–878, 2008.
  2. [2]Banerjee, A. V. A simple model of herd behavior. The Quarterly Journal of Economics, 107(3):797–817, 1992.DOI
  3. [3]Bikhchandani, S., Hirshleifer, D., and Welch, I. A theory of fads, fashion, custom, and cultural change as informational cascades. Journal of Political Economy, 100(5):992–1026, 1992.DOI
  4. [4]Chen, J., Yang, S., Liu, A., and Liu, X. Can LLM-based agents effectively coordinate prediction markets? arXiv preprint, 2024.
  5. [5]Cui, G., Yuan, L., Ding, N., Yao, G., He, B., Zhu, W., Ni, Y., Xie, G., Xie, R., Lin, Y., et al. UltraFeedback: Boosting language models with scaled AI feedback. arXiv preprint arXiv:2310.01377, 2024.arxiv.org/abs/2310.01377
  6. [7]Galton, F. Vox populi. Nature, 75:450–451, 1907.
  7. [8]Gao, L., Schulman, J., and Hilton, J. Scaling laws for reward model overoptimization. In International Conference on Machine Learning, 2023.arxiv.org/abs/2210.10760
  8. [9]Halawi, D., Shi, F., Chen, T., Atchley, J., and Steinhardt, J. Approaching human-level forecasting with language models. arXiv preprint arXiv:2402.18563, 2024.arxiv.org/abs/2402.18563
  9. [10]Hanson, R. Combinatorial information market design. Information Systems Frontiers, 5(1):107–119, 2003.DOI
  10. [11]Hanson, R. Logarithmic market scoring rules for modular combinatorial information aggregation. The Journal of Prediction Markets, 1(1):3–15, 2007.
  11. [12]Hanson, R. A manipulator can aid prediction market accuracy. Economica, 76(302):304–314, 2009.
  12. [13]Kelly, J. L. A new interpretation of information rate. Bell System Technical Journal, 35(4):917–926, 1956.
  13. [14]Kim, E., Garg, A., Peng, K., and Garg, N. Correlated errors in large language models. In International Conference on Machine Learning, 2025.arxiv.org/abs/2506.07962
  14. [15]Kirk, H. R., Vidgen, B., Röttger, P., and Hale, S. A. Understanding the effects of RLHF on LLM generalisation. arXiv preprint arXiv:2309.02301, 2023.
  15. [16]Kish, L. Survey Sampling. John Wiley & Sons, New York, 1965.
  16. [17]Kleinberg, J. and Raghavan, M. Algorithmic monoculture and social welfare. Proceedings of the National Academy of Sciences, 118(22), 2021.DOI
  17. [18]Lambert, N., Morrison, J., Pyatkin, V., Huang, S., Ivison, H., et al. TÜLU 3: Pushing frontiers in open language model post-training. arXiv:2411.15124, 2024.arxiv.org/abs/2411.15124
  18. [19]Lin, S., Hilton, J., and Evans, O. TruthfulQA: Measuring how models mimic human falsehoods. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, 2022.arxiv.org/abs/2109.07958
  19. [20]Madaan, A., Tandon, N., Gupta, P., Hallinan, S., Gao, L., Wiegreffe, S., Alon, U., Dziri, N., Prabhumoye, S., Yang, Y., et al. Self-Refine: Iterative refinement with self-feedback. In Advances in Neural Information Processing Systems, 2023.arxiv.org/abs/2303.17651
  20. [21]Meng, Y., Xia, M., and Chen, D. SimPO: Simple preference optimization with a reference-free reward. In Advances in Neural Information Processing Systems, 2024.arxiv.org/abs/2405.14734
  21. [23]Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35, 2022.arxiv.org/abs/2203.02155
  22. [24]Rafailov, R., Sharma, A., Mitchell, E., Manning, C. D., Ermon, S., and Finn, C. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 2023.arxiv.org/abs/2305.18290
  23. [25]Rein, D., Hou, B. L., Stickland, A. C., Petty, J., Pang, R. Y., Dirani, J., Michael, J., and Bowman, S. R. GPQA: A graduate-level google-proof q&a benchmark. arXiv preprint arXiv:2311.12022, 2023.arxiv.org/abs/2311.12022
  24. [26]Schoenegger, P., Park, P. S., Karger, E., and Tetlock, P. E. Wisdom of the silicon crowd: LLM ensemble prediction capabilities rival human crowd accuracy. Science Advances, 10(21), 2024.arxiv.org/abs/2402.19379
  25. [27]Surowiecki, J. The Wisdom of Crowds. Doubleday, 2004.
  26. [28]Tunstall, L., Beeching, E., Lambert, N., Rajani, N., Rasul, K., Belkada, Y., Huang, S., von Werra, L., Fourrier, C., Habib, N., et al. Zephyr: Direct distillation of LM alignment. arXiv preprint arXiv:2310.16944, 2023.arxiv.org/abs/2310.16944
  27. [29]Wolfers, J. and Zitzewitz, E. Prediction markets. Journal of Economic Perspectives, 18(2):107–126, 2004.

Paper details

vlnc:W_47re46p8
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 7

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents