betaFeedback

WorksPhysics and Astronomy

Borrowed Strength, Imported Risk: Capability Gaps and Reception Govern Heterogeneous LLM Debate

Abstract

Heterogeneous LLM debate is often attributed to collective reasoning that creates capability beyond any member. In the matched panels we study, the aggregate gain is dominated by capability transfer: a peer supplies task competence and receivers absorb a measurable fraction. Across nine matched directions, honest benefit and first-exchange adversarial exposure open when the peer has a positive same-task advantage, and honest gain is closely described by the empirical decomposition ΔAdhon≈κΓd+ρdΔA_d^{\mathrm{hon}}\approxκΓ_d^+ρ_d, the product of that advantage and the receiver's reception. Two fixed-receiver directions provide held-out validation without refitting: reception remains stable within uncertainty as a larger peer advantage produces a larger gain. A matched, single-direction open-answer MuSiQue scope probe reaches the negative-gap side of the same gate. Qwen-7B trails the Gemma-3-12B receivers by 10.78 points; despite 62.50% reception, it produces only +1.17 points of controlled first-exchange gain. Reception cannot substitute for positive aggregate peer advantage. When an influential peer is assigned an adversarial role, the same influential peer slot removes the borrowed gain and can push the panel below homogeneous debate. Across all five full-reference settings, the best observed majority correctness does not exceed the influential member's best observed accuracy, despite unused complementary answers. Heterogeneous debate is therefore a measurable transfer decision with a measurable failure mode, not a benefit conferred by diversity alone.

Introduction

Heterogeneous debate puts different language models in one panel because their strengths are expected to compose. If accuracy rises, the standard explanation is diversity: members contribute complementary reasoning and the exchange creates an answer that no member could produce alone ([9, 36] (Yin et al. 2023)).

That explanation carries an operational promise. A system that creates new collective capability may justify several times the inference cost of calling one model; a system that relays one capable member’s answer must be evaluated against calling or routing to that member directly. Recent evaluations increasingly favour the second reading: debate often loses to single-agent inference at a fraction of the cost ([3] (Zhang et al. 2025)). In the Bayesian model of Choi et al. [5], expected correctness is unchanged unless updates are biased toward correct signals. What that work does not supply is the source of that bias or a way to measure it in advance, which is what we provide.

We separate these accounts. Under capability composition, the panel turns distributed partial knowledge into capability beyond its members. Under capability transfer, one peer already has the task competence and the debate routes some of it to weaker receivers. Our claim is that transfer dominates net gain in these evaluated panels, not that item-level composition is absent. Transfer makes four concrete predictions: benefit should require something for the peer to supply; the amount transferred should depend on the receiver’s willingness to use a correct signal; realized consensus should approach the influential peer rather than an oracle over all answers; and changing that peer’s role should reverse the channel that delivered the gain.

We test all four predictions with matched three-agent panels. For each receiver–peer–task direction, we compare receiver-homogeneous, peer-homogeneous, honest heterogeneous, and adversarial heterogeneous conditions. The peer’s capability is measured on the same items under the same protocol. Ordinary self-revision is absorbed by the receiver-homogeneous panel. Honest and adversarial mixed panels hold model identity and position fixed and change one role instruction. Nine directions span open and frontier models on MATH-hard and GSM8K. Two fixed-receiver directions hold Qwen-7B fixed while replacing a Gemma-3-12B peer with GPT-OSS-120B, a direct confirmatory test of whether a larger aggregate peer advantage produces more gain at stable reception.

The data follow a two-factor decomposition. Let Γd\Gamma_d be the peer’s same-task accuracy advantage and ρd\rho_d the probability that a wrong receiver becomes correct when the honest peer begins correct. Then

ΔAdhon≈κΓd+ρd.(1)\Delta A_d^{\mathrm{hon}} \approx\kappa\Gamma_d^{+}\rho_d . \tag*{(1)}

The product fit has R2=.995R^2=.995, against .776 for capability gap alone. More importantly, the frozen model predicts both fixed-receiver gains: reception remains near 61% while doubling the peer gap raises controlled R1 gain from 7.89 to 16.92 points.

A matched MuSiQue-Ans cohort provides a single-direction scope probe of the negative-gap regime outside mathematical reasoning. It compares Gemma-3-12B receivers with a Qwen-7B peer on 300 open-answer multi-hop questions, including homogeneous, honest, and adversarial panels. Qwen is 10.78 points less accurate than Gemma at R0. The receivers nevertheless receive its correct answers at 62.50%, yet controlled first-exchange gain is only +1.17 points. The scope probe makes the distinction clear: reception can be high, but a negative aggregate peer advantage leaves little net gain.

The same design exposes the cost of transfer. Across the nine directions, reversing the peer’s role removes recovery, raises corruption, and leaves the adversarial panels 3.06 points below homogeneous debate at R1. In the fixed-receiver GPT-OSS panel, the R4 swing is −27.60 points. Borrowed accuracy is surrenderable because the component that supplies it also holds leverage over the panel.

Our contributions are:

  1. A matched measurement that tests whether aggregate gains reflect capability transfer or composition using homogeneous, heterogeneous, and member-level reference panels.

  1. A capability gate across the mathematical-reasoning directions: net benefit and first-exchange leverage remain small without positive measured peer advantage, with a single-direction open-answer multi-hop QA scope probe.

  1. Reception power and an advantage–reception decomposition whose frozen coefficients predict fixed-receiver gains without refitting.

  1. A benefit–risk account of borrowed strength: role reversal can consume the gain and push the panel below homogeneous debate, while consensus leaves available oracle complementarity unrealized.

Framework

Directions and panels. A direction d=(R,P,T)d=(R,P,T) fixes a receiver model, peer model, and task. We compare four matched panels:

[R,R,R][R,R,R] receiver-homogeneous control [P,P,P][P,P,P] peer-homogeneous capability reference [R,R,P][R,R,P] honest heterogeneous peer [R,R,Padv][R,R,P^{\mathrm{adv}}] same peer with an adversarial role

Slots 0–1 are receivers and slot 2 is the peer. Only the role instruction changes between the two mixed panels.

Transitions. Let Cdrix∈{0,1}C_{dri}^{x}\in\{0,1\} denote semantic correctness for receiver exposure ii at round rr in condition x∈{hom,hon,adv}x\in\{\mathrm{hom},\mathrm{hon},\mathrm{adv}\}. The primary estimand is the first exchange, R0→R1:

Wdx=P(Cd1x=1∣Cd0x=0),(2)W_d^x=P(C_{d1}^x=1\mid C_{d0}^x=0), \tag*{(2)}
Qdx=P(Cd1x=0∣Cd0x=1),(3)Q_d^x=P(C_{d1}^x=0\mid C_{d0}^x=1), \tag*{(3)}
ΔAdx=[Ad1x−Ad0x]−[Ad1hom−Ad0hom].(4)\Delta A_d^x=[A_{d1}^x-A_{d0}^x]-[A_{d1}^{\mathrm{hom}}-A_{d0}^{\mathrm{hom}}]. \tag*{(4)}

WdxW_d^x and QdxQ_d^x are recovery and corruption. “Honest” labels the non-adversarial panel; controlled honest gain subtracts homogeneous revision from the mixed panel’s accuracy change. Every reported aggregate is an equal-weight mean over named directions; item counts never determine a direction’s weight.

Aggregate advantage and headroom. On each analysis’s matched support,

Γd=Ad0[P,P,P]−Ad0[R,R,R],Hd=1−Ad0[R,R,R],(5)\Gamma_d=A_{d0}^{[P,P,P]}-A_{d0}^{[R,R,R]},\qquad H_d=1-A_{d0}^{[R,R,R]}, \tag*{(5)}

and Γd+=max⁡(0,Γd)\Gamma_d^+=\max(0,\Gamma_d). Γd\Gamma_d measures net aggregate peer advantage, not the total mass of items on which only the peer is correct; HdH_d bounds the receiver’s possible improvement.

Reception and adversarial uptake. Let Cd0PC_{d0}^{P} denote the heterogeneous peer’s R0 correctness. We define

ρd=P(Cd1hon=1∣Cd0hon=0,Cd0P=1),(6)\rho_d=P(C_{d1}^{\mathrm{hon}}=1\mid C_{d0}^{\mathrm{hon}}=0,C_{d0}^{P}=1), \tag*{(6)}
ρ0d=P(Cd1hon=1∣Cd0hon=0,Cd0P=0),(7)\rho_{0d}=P(C_{d1}^{\mathrm{hon}}=1\mid C_{d0}^{\mathrm{hon}}=0,C_{d0}^{P}=0), \tag*{(7)}
νd=P(Cd1adv=0∣Cd0adv=1,Cd0adv=0).(8)\nu_d=P(C_{d1}^{\mathrm{adv}}=0\mid C_{d0}^{\mathrm{adv}}=1,C_{d0}^{\mathrm{adv}}=0). \tag*{(8)}

ρd\rho_d is reception power; ρ0d\rho_{0d} measures revision without a correct peer signal; υd\upsilon_d is adversarial uptake. For example, in the Qwen-7B←\leftarrowGemma-3-12B direction, receiver R0 accuracy is 58.7%, peer accuracy 67.8%, Γ=+9.1\Gamma= +9.1 points, and headroom 41.3 points. Among initially receiver-wrong/peer-correct cases, reception is ρ=62.2%\rho= 62.2\% (95% CI [50.0,73.7]); controlled honest gain is +7.89 points. The observed gain need not equal Γρ\Gamma\rho; Eq. (9) includes fitted scaling and a residual term.

Predictions. Capability transfer yields four testable expectations:

  1. The marginal honest gain and first-exchange adversarial leverage of an added heterogeneous peer remain small when Γd≤0\Gamma_d \le0.

  1. Honest gain is approximately

ΔAdhon=κΓd+ρd+εd,(9)\Delta A_d^{\mathrm{hon}} = \kappa\Gamma_d^{+}\rho_d + \varepsilon_d, \tag*{(9)}

where εd\varepsilon_d collects unaided revision, non-nested correctness, and finite-sample error. Reception and gain share R1 outcomes; under near-nested correctness, eligible correction mass approaches Γd\Gamma_d. Thus the product fit is partly structurally induced, not independent evidence; the cross-peer predictions provide a test beyond the in-sample fit.

  1. Once a peer supplies material gain, reversing its role removes some or all of that borrowed gain and may create net loss relative to [R,R,R][R, R, R].

  1. Realized panel accuracy approaches the influential peer rather than the oracle over all member answers.

Section 4.1 tests Γ\Gamma, Section 4.3 examines ρ\rho and the product fit, and Section 4.5 examines υ\upsilon and role reversal. Practical use is discussed in Section 5, “The relevant baseline is the influential member.”

Experimental Setup

Primary directions. The primary analysis contains nine directions (Table 1): six on the 262 level-4/5 items of MATH-500 (MATH-hard) ([16, 21]) and three on GSM8K ([6]). It spans 1B–120B open models and three frontier APIs. GPT-4.1/Gemini-Flash and Llama-8B/Qwen-7B are evaluated in both orientations, so the sign of Γd\Gamma_d is not fixed by model identity.1

Direction (R←PR \leftarrow P)TaskRep.-itemsA0,RA_{0,R}A0,PA_{0,P}Γd\Gamma_dHdH_d
Gemma-3-12B ←\leftarrow Qwen-7BMATH-hard78667.858.1−9.732.2
Qwen-7B ←\leftarrow Llama-8BGSM8K395791.987.1−4.88.1
Gemini-Flash ←\leftarrow GPT-4.1MATH-hard78679.475.9−3.520.6
GPT-4.1 ←\leftarrow Gemini-FlashMATH-hard78675.979.4+3.524.1
GPT-4.1 ←\leftarrow OSS-120BMATH-hard78675.980.2+4.324.1
Llama-8B ←\leftarrow Qwen-7BGSM8K395787.191.9+4.812.9
Qwen-72B ←\leftarrow OSS-120BMATH-hard77366.581.2+14.733.5
Llama-1B ←\leftarrow Gemma-2BGSM8K†10037.761.0+23.362.3
Llama-70B ←\leftarrow OSS-120BMATH-hard78644.280.2+36.055.8
Qwen-7B ←\leftarrow Gemma-3-12BMATH-hard‡78658.767.8+9.141.3
Qwen-7B ←\leftarrow OSS-120BMATH-hard‡78658.780.2+21.541.3

Table 1. Evaluated directions on matched R0/R1 support. The first nine form the nine-direction support (12,717 replicate-items); ‡ marks the two fixed-receiver confirmatory directions (1,572 additional replicate-items). A0,RA_{0,R} and A0,PA_{0,P} are homogeneous-panel accuracies in percent, Γd=A0,P−A0,R\Gamma_d = A_{0,P} - A_{0,R}, and Hd=100−A0,RH_d = 100 - A_{0,R}. † is a one-replicate, 100-item diagnostic; all others use three replicates.

Confirmatory directions. Two further MATH-hard directions, specified after Eq. (9), hold Qwen-7B fixed as RR and either Gemma-3-12B or GPT-OSS-120B as PP. The same Qwen homogeneous panel is the control. This design holds dataset, receiver, headroom, slots, and protocol fixed while moving the measured peer gap from 9.1 to 21.5 points. They were not used to formulate the approximation, and we report them separately from the pooled sensitivity. An additional same-family Qwen-72B honest peer provides a control for the fixed Qwen-7B receiver; this condition is excluded from the original fits and held-out predictions.

Cross-task scope probe. On 300 MuSiQue-Ans open-answer multi-hop questions (Trivedi et al., [33]), Gemma-3-12B is the receiver and Qwen-7B the peer. We run [G,G,G][G,G,G], [Q,Q,Q][Q,Q,Q], [G,G,Q][G,G,Q], and [G,G,Qadv][G,G,Q^{\mathrm{adv}}] on the same items, plus a same-family adversary [G,G,Gadv][G,G,G^{\mathrm{adv}}] and a contaminated panel [G,Gadv,Q][G,G^{\mathrm{adv}},Q]. This cohort tests the negative-gap prediction on a non-mathematical task; it is not pooled into the cross-direction correlations or the fixed-receiver test.

Protocol. Following Du et al. [9], three agents answer independently at R0 and revise synchronously for four rounds after seeing the other two responses in full. Primary runs use temperature 0 and three seeds/replicates except the 100-item small-model diagnostic. MATH-hard uses a 4,096-token completion budget; GSM8K uses 1,024. MuSiQue uses one temperature-0 replicate on 300 items. The adversarial peer receives a committed plausible-wrong target and is instructed to defend it. No other panel component changes.

Matching and scoring. We retain replicate-items containing all R0/R1 cells required by the four-panel contrast. Last-record-wins indexing resolves resumed-run duplicates. Answers are parsed with benchmark-specific semantic equivalence; missing or unparseable answers count as incorrect. The nine-direction support contains 12,717 replicate-items and 76,302 receiver-condition observations; the two confirmatory directions add 1,572 replicate-items and 9,432 observations. Every MuSiQue condition contains all 300 matched items, three agents, and five rounds. MuSiQue answers are scored by alias-normalized exact match and token F1 without an LLM judge. Later-round trajectories require complete R0–R4 support and are reported separately from the first-exchange estimand.

Inference. Direction-level means are tested by exact sign flips. Cross-direction associations use exact label permutations and leave-one-direction-out refits. Intervals use a global benchmark-item Poisson bootstrap: an item’s weight is shared across every replicate, slot, panel, and direction in which it appears. Logistic sensitivities use direction fixed effects and direction-clustered sandwich errors. The confirmatory fixed-receiver contrasts use paired benchmark-item bootstraps; their item-level trend model clusters by benchmark item and controls replicate and receiver slot.

Diversity control. Appendix § 14 reports a same-model MATH-hard panel that varies prompts, temperatures (0, .3, .7), and seeds (0, 1, 2) to elicit complementary reasoning without introducing a model-capability gap.

Results

Corresponding bootstrap intervals and, where applicable, exact tests are reported in the appendix tables—Table 8 and Table 15 for the aggregate and cross-task contrasts. The text quotes an interval only where its width carries the argument.

ContrastMean (pp)Bootstrap 95% CIExact ppOdds ratio [95% CI]
Intent to treat (12,717 matched replicate-items)
Recovery: honest vs. homogeneous+20.40[+17.62, +23.28]0.00396.58 [2.24, 19.37]
Corruption: honest vs. homogeneous−0.78[−2.28, +0.74]0.35551.43 [0.61, 3.36]
Recovery: adversarial vs. honest−15.25[−17.94, −12.61]0.01560.29 [0.14, 0.62]
Corruption: adversarial vs. honest+7.67[+5.98, +9.37]0.03125.17 [1.70, 15.72]
Per protocol (10,580 items; adversary parsed-wrong in all of R0–R4)
Recovery: honest vs. homogeneous+20.08[+17.23, +23.08]0.0039–
Corruption: honest vs. homogeneous−0.22[−1.91, +1.46]0.7656–
Recovery: adversarial vs. honest−16.63[−19.53, −14.07]0.0078–
Corruption: adversarial vs. honest+7.77[+5.85, +9.70]0.0586–

Table 8. Role effects at first peer exposure (R0→R1), case-standardized over the nine directions. Exact pp is a paired sign-flip test over the 292^9 direction-level sign assignments; intervals are global benchmark-item Poisson bootstraps (4,000 replicates). Odds ratios come from logistic models with a separate intercept per direction and direction-clustered sandwich standard errors on 8 degrees of freedom. Honest and adversarial contrasts use the same peer model in the same slot; only the role prompt changes.

QuantityEstimateBootstrap 95% CInn
Capability gap Γ\Gamma (pp)−10.78[−16.54, −4.79]300
Honest controlled gain, R1 (pp)+1.17[−2.61, +4.95]300
Honest recovery difference, R1 (pp)+9.67[+4.78, +14.84]326
Honest corruption difference, R1 (pp)+9.25[+4.19, +14.55]271
Adversarial-peer controlled gain, R1 (pp)+1.00[−1.89, +3.83]300
Honest controlled gain, R4 (pp)+2.33[−2.08, +6.73]300
Adversarial-peer controlled gain, R4 (pp)−4.50[−8.78, −0.32]300
Adversarial-peer excess corruption, R1 (pp)+2.66[−0.77, +6.45]268
Honest reception ρ\rho (%)62.50[46.57, 78.13]64
Unaided revision ρ0\rho_0 (%)4.91[1.92, 8.43]265
Adversarial uptake υ\upsilon (%)4.85[1.77, 8.63]268
Same-family adversary excess corruption, R1 (pp)+8.47[+3.45, +13.97]272
Same-family adversary controlled gain, R4 (pp)−9.50[−14.39, −4.85]300
Honest Qwen effect under contamination, R1 (pp)+5.00[+1.02, +9.18]300
Honest Qwen effect under contamination, R4 (pp)−3.00[−8.36, +2.49]300

Table 15. Heterogeneous MuSiQue-Ans scope probe on 300 matched open-answer multi-hop questions. Gemma-3-12B is the receiver and Qwen-7B the heterogeneous peer. Controlled gains are difference-in-change against [G,G,GG,G,G]; transition quantities are measured at R0→R1. The same-family adversary is [G,G,GadvG,G,G^{\mathrm{adv}}]. The contaminated comparison replaces one honest Gemma in that panel with Qwen, yielding [G,Gadv,QG,G^{\mathrm{adv}},Q]. Intervals are benchmark-item Poisson bootstraps.

Honest influence is corrective and capability-gated

Across the nine directions, the equal-weight R0→R1R_0 \to R_1 mean of Wdhon−WdhomW_d^{\mathrm{hon}} - W_d^{\mathrm{hom}} is +20.40 percentage points (exact p=.0039p = .0039). The corresponding mean ΔAdhon\Delta A_d^{\mathrm{hon}} is +6.55 points (p=.0078p = .0078), while honest corruption changes by −0.78 points ([−2.28,0.74], p=.3555p = .3555). Thus the aggregate is driven by correction of initially wrong receivers, not systematic destabilization of correct ones.

The aggregate separates at Γd=0\Gamma_d = 0 (Figure 1; Table 2). The three ungated directions average +0.49 points of controlled honest accuracy; the six gated directions average +9.58 points, positive in all six. Under an adversarial role, excess corruption over [R,R,R][R,R,R] averages +0.36 points in the ungated group and +10.16 in the gated group. These are direction averages, not pair-invariant effects. Across the nine directions, Γd\Gamma_d correlates with honest gain at r=.862r = .862 (exact p=.0033p = .0033) and with adversarial uptake at r=.780r = .780 (p=.0134p = .0134). The exact permutation tests operate on direction labels; because several directions share models, tasks, and traces, they quantify association within this evaluated set rather than among independent population draws.

Quantity (pp)UngatedGated
Honest recovery+7.33 (.250)+26.94 (.031)
Honest net accuracy+0.49 (.500)+9.58 (.031)
Adv. corruption+0.36 (.250)+10.16 (.031)
Adv. net accuracy+0.46 (.750)−4.82 (.094)

Table 2. The capability gate. Means over directions, in percentage points against matched homogeneous debate, with exact sign-flip pp in parentheses. The worst single ungated direction shows +0.5 pp of excess corruption against +24.6 pp for the worst gated one.

Figure 1. The capability gate across nine directions. Each point is one receiver–peer–task direction at R0→R1. Left: controlled honest gain. Right: adversarial excess corruption over the matched receiver-homogeneous panel. Hollow points have Γd<0\Gamma_d < 0; filled points have Γd>0\Gamma_d > 0.

A cross-task scope probe reaches the negative-gap regime

The MuSiQue cohort changes the task and scoring regime while preserving the matched panel design. Qwen-7B is 10.78 points less accurate than Gemma-3-12B at R0. As an honest peer it produces only +1.17 points of controlled R1 gain. The small net effect is not evidence that the receivers ignore Qwen. Reception is 62.50%, against 4.91% when Qwen is also wrong. Instead, the honest peer raises recovery by 9.67 points and corruption by 9.25 points; the movements cancel in net accuracy. The adversarial Qwen peer adds 2.66 points of first-exchange corruption over the Gemma-homogeneous panel, with 4.85% adversarial uptake. Thus a non-mathematical, open-answer direction reaches the ungated regime: reception can be active while net benefit and first-exchange heterogeneous leverage remain small. Positive aggregate peer advantage, not diversity or reception alone, is the gate (Table 15).

The two ungated regimes reach that null differently, and the difference is informative. The three ungated mathematical directions combine a modest recovery rise of 7.33 points with almost no corruption, 0.76 points; MuSiQue cancels two large flows instead. What persists in this scope probe is the absence of net benefit without positive aggregate peer advantage, not a single transition-level signature of it.

Gain scales approximately with aggregate advantage times reception

The product form follows the transition structure: when peer and receiver correctness are nearly nested, positive capability gap tracks the cases in which the receiver is wrong and the peer is correct. Outside that regime, Γd\Gamma_d is a net aggregate advantage; Γd≤0\Gamma_d \le0 does not imply the absence of useful peer-only answers (Appendix § 15). Reception measures the conversion of available peer advantage; in the full nine-direction set, incorporating it sharply reduces the gap-only residual, while both specifications fit closely in the sensitivity analysis (Appendix § 15). Across the nine directions, ρd\rho_d ranges from 14.7% to 83.3% and exceeds its unaided control ρ0d\rho_{0d} in every direction (Appendix Figure A1, left). Neither Γd\Gamma_d nor HdH_d significantly predicts reception (r=.251r = .251, p=.5157p = .5157 and r=−.238r = -.238, p=.5230p = .5230). A hinge on Γd+\Gamma_d^{+} fits controlled gain with R2=.776R^2 = .776 and a 9.97-point maximum residual. Replacing it with Γd+ρd\Gamma_d^{+}\rho_d yields slope 1.088, R2=.995R^2 = .995, and a 1.07-point maximum residual (Appendix Figure A1, right; Table 3).

PredictorSlopeR2R^2Exact ppLOO slope
Γ+\Gamma^{+}0.6140.77600.0078[0.28, 0.76]
ρ\rho11.9710.14550.3560[5.46, 15.39]
Γ+×ρ\Gamma^{+} \times\rho1.0880.99500.0002[1.08, 1.12]

Table 3. Cross-direction models of the honest net accuracy gain (pp), fitted on identical inputs over the same nine directions. Γ+=max⁡(0,Γ)\Gamma^{+} = \max(0, \Gamma) encodes the gate. Exact pp enumerates all 9!=362,8809! = 362{,}880 label permutations; LOO slope is the range over the nine leave-one-direction-out refits.

A fixed receiver provides held-out validation

The two fixed-receiver directions provide a controlled held-out test. Replacing Gemma-3-12B with GPT-OSS-120B raises Γd\Gamma_d from 9.12 to 21.50 points, while Qwen’s measured reception remains effectively unchanged (Figure 2, middle): 62.16% versus 60.35%, a −1.81-point difference with paired bootstrap CI [−13.01,9.87] (p=.7875p = .7875). Controlled R1 gain nevertheless rises from +7.89 to +16.92 points (Figure 2, left); the paired difference is +9.03 [5.14,12.88] (p<.0005p < .0005).

Figure 2. Confirmatory fixed-receiver experiment on MATH-hard. Left: controlled gain increases with peer capability gap at both R1 and R4. Middle: reception remains stable across the two heterogeneous peers. Right: adversarial uptake is higher with GPT-OSS than with Gemma under the same adversarial role. Intervals are paired benchmark-item bootstraps.

Using coefficients fitted on the nine directions, without refitting, reception measured with Gemma predicts +15.14 points for Qwen with GPT-OSS; the observed gain is +16.92. Conversely, reception measured with GPT-OSS predicts +6.59 points for the Gemma direction; the observed gain is +7.89. Across the zero-gap Qwen control and the two heterogeneous peers, gain is nearly linear in gap at R1 (through-origin slope .799, R2=.999R^2 = .999) and R4 (slope 1.027, R2=.994R^2 = .994; Table 4). At the item level, a ten-point increase in peer gap multiplies recovery odds by 3.03 ([2.47,3.73]), while honest corruption has no gap trend (OR .98 [.72,1.34]). These item-level intervals quantify these 262 items; the three-point ladder is the model-level evidence.

Qwen-7B peerQwenGemmaOSS-120B
Γd\Gamma_d (pp)0.009.1221.50
ρd\rho_d (%)–62.1660.35
R1 gain (pp)0.00+7.89+16.92
R4 gain (pp)0.00+11.05+21.37
R1 recovery (%)2.9225.2642.94
R1 honest corruption (%)2.064.722.21
υd\upsilon_d (%)–2.606.00

Table 4. Fixed-receiver MATH-hard ladder. Gain is the receiver’s controlled accuracy change relative to [Q,Q,Q][Q,Q,Q] at the indicated round. Reception and adversarial uptake follow §2. The zero-gap column is the homogeneous reference, not a separately designated honest peer.

With the GPT-OSS direction included, the empirical pattern persists: the gap–gain correlation across the resulting ten directions becomes r=.876r = .876 (exact p=.00080p = .00080), the gap–reception association remains small and nonsignificant (r=.263r = .263, p=.459p = .459), and the product fit remains R2=.990R^2 = .990 against .795 for the hinge. With both included, R2=.989R^2 = .989 for the product and .793 for the hinge. These pooled refits are sensitivity analyses.

The additional same-family Qwen-7B←Qwen-72B control on MATH-hard (three replicates, 786 replicate-items) has Γ=+7.1\Gamma= +7.1 points, ρ=52.2%\rho= 52.2\% [38.5,65.8], and controlled R1 gain +3.4 points. Its reception interval overlaps that of the approximately gap-matched Gemma peer, revealing no clear family-aligned difference; this does not establish equality or remove identity/style confounding. Its reception denominator is 180. The control has no adversarial arm and is excluded from the original fits.

The same influential peer slot imports risk

In the nine directions, changing the peer from honest to adversarial reduces recovery by an equal-weight mean of 15.25 points (exact p=.0156p = .0156) and increases corruption by 7.67 points (p=.0312p = .0312). Consequently, the adversarial mixed panels average 3.06 points below their matched homogeneous controls at R1 (Appendix Figure A2, left; Appendix § 11). The fixed-receiver comparison points in the same direction: adversarial uptake rises from 2.60% with Gemma to 6.00% with GPT-OSS (Figure 2, right), a paired difference of +3.41 points ([1.12,5.82], p=.0040p = .0040). Compliance differs across peers: 68.54% for Gemma and 99.75% for GPT-OSS. At R4, Gemma’s honest-to-adversarial swing is −18.96 points by intent to treat and −24.72 per protocol (adversary parsed-wrong at R0–R4). For GPT-OSS, Qwen with an honest peer is +21.37 points above homogeneous debate; assigning that peer the adversarial role produces a −27.60-point intent-to-treat swing (−27.93 per protocol) and leaves Qwen 6.23 points below homogeneous debate. The GPT-OSS swing is therefore robust, while its difference from Gemma cannot be attributed to capability alone.

Correction and corruption are not one generic credulity trait. Across the nine directions, ρd\rho_d does not predict υd\upsilon_d (r=.223r = .223, p=.554p = .554), whereas the capability gap does. With GPT-OSS added, the latter association weakens but remains significant (r=.683r = .683, exact p=.0282p = .0282). We therefore interpret ρd\rho_d as a recurring risk predictor, not a complete model of adversarial loss.

MuSiQue also locates the boundary of this claim. Repeated exposure to the weaker adversarial Qwen peer accumulates: by R4 its controlled effect is −4.50 points, an interval excluding zero, even though its R1 excess-corruption interval includes zero. And a same-family Gemma adversary—a different threat model from adding a heterogeneous peer—raises R1 corruption by 8.47 points and reaches a controlled R4 effect of −9.50 points. The capability gate governs the marginal first-exchange leverage of a heterogeneous replacement; it does not make repeated or same-family adversaries harmless.

The panel approaches the influential member

In the five full-reference settings, the panel’s best majority correctness (at least two of three agents correct) never exceeds the peer’s best accuracy inside that panel, while an oracle that credits any correct member is 4.7–23.0 points higher (Appendix Figure A2, right; Table 12 in Appendix § 13). These finite-setting observations support a borrowed ceiling; they are not a universal upper bound on heterogeneous protocols, including routing or selection. At fixed R4, discarded correct answers exceed new correct outcomes by 3.6–30 times in all five settings: transfer dominates net gain, while new correct outcomes remain item-level exceptions to a pure relay account (Appendix Table 13).

SettingnnPeer (%) alonePeer (%) in panelPanel (%) receiversPanel (%) ≥2Panel (%) oracleShortfall
GPT-OSS-120B → Llama-70B77681.180.472.172.883.6−10.8
GPT-OSS-120B → GPT-4.178680.881.481.481.487.2−5.7
Gemini-2.5-Flash → GPT-4.178481.682.582.282.587.2−4.7
GPT-4.1 → Gemini-2.5-Flash78475.982.782.282.787.4−4.7
Gemma-2B → Llama-1B10061.061.043.548.071.0−23.0

Table 12. Where the panel stops. Accuracies are best-round values on nn matched replicate-items. Peer alone is that model’s best round in its own homogeneous panel; peer in panel its best in slot 2 of the honest heterogeneous panel; ≥2 requires two of three agents correct; oracle any one; shortfall is ≥2 minus oracle. Maxima are selected after observing the rounds; round-0 comparisons and full curves are in the appendix.

Across the extended set’s nine positive-gap directions (six primary, two fixed-receiver, and the same-family control), five-round debate uses 40.6–86.7 times the total reported tokens of one peer call (median 62.2). Even its retrospectively best majority round beats the peer’s single-pass R0 accuracy in only 3/9 directions. We charge all 15 debate calls versus one peer call on matched items (Appendix § 10). This is token volume, not dollar cost or measured latency; the protocol has five sequential stages. Strong-model resampling and learned routing remain unevaluated.

Robustness and controls

Requiring the adversary to remain parsed and wrong in every round strengthens the recovery reversal and preserves the corruption increase (Appendix § 11). Results retain their direction within MATH-hard and GSM8K. The matched MuSiQue scope probe reaches the negative-gap side under open-answer semantic scoring. The prompt/temperature/seed control creates more complementary R0 answers but does not convert them into higher consensus accuracy (Appendix § 14).

Discussion

In these panels, aggregate advantage gates net gain; reception sets the conversion rate. The experiments resolve two questions that panel accuracy collapses. In these tested panels, material net gain is associated with positive aggregate advantage, measured by Γd\Gamma_d; this is not a necessary condition for item-level help. The receivers must then use the peer’s correct signal, measured by ρd\rho_d. Neither factor substitutes for the other. MuSiQue makes the separation especially clear. The receivers receive a correct Qwen signal at 62.50%, but Qwen is 10.78 points weaker than Gemma and controlled gain is only +1.17 points. High reception without positive aggregate peer advantage creates churn, not strength.

The fixed receiver provides held-out validation. The nine-direction fit alone would leave open whether Γd+ρd\Gamma_d^+\rho_d merely summarizes nine directions. The fixed-receiver experiment holds receiver, task, headroom, protocol, and homogeneous baseline constant. Reception estimated with one peer predicts the gain with the other using coefficients frozen on the nine directions. That cross-peer prediction provides held-out validation under a controlled receiver, task, and protocol; it does not establish a universal relationship across models, tasks, or debate protocols. Borrowed strength has a replacement cost. The peer that supplies accuracy also controls whether that accuracy remains available. Across the nine directions, role reversal consumes the honest gain and leaves the adversarial panels below homogeneous debate. This is not a generic claim that every weak peer is harmless or every capable peer is malicious. MuSiQue shows the boundary directly: a weaker adversarial heterogeneous peer has little first-exchange leverage, but repeated exposure eventually causes loss, while a same-family committed adversary can cause immediate harm without a positive heterogeneous capability gap. Operators therefore need both the marginal heterogeneous-leverage measurement and an explicit threat model.

The relevant baseline is the influential member. Beating [R,R,R][R, R, R] does not establish collective capability when PP can be called directly. Heterogeneous transfer can still be valuable under rate limits, cost constraints, or partial access to the peer. But the honest comparison is then against strong-model inference and routing, not only against a weak homogeneous panel. The oracle gap measures whether selection has complementarity left to recover; the prompt/temperature/seed control shows that creating disagreement is not enough when consensus revision discards it.

What the approximation means. Equation (9) is a compact empirical account, not a new conservation law. When peer and receiver correctness are nearly nested, the eligible mass on which the receiver is wrong and the peer is right approaches Γd\Gamma_d, making part of the product structurally induced. Non-nested errors, unaided revision, corruption, and sampling variation produce εd\varepsilon_d. The evidentiary weight therefore rests on the residual reduction, the fixed-receiver cross-peer predictions, and the negative-gap MuSiQue scope probe together—not on the nine-direction R2R^2 alone.

Related Work

Debate and heterogeneous panels. Multi-agent debate can improve language-model reasoning [9, 20] (Yin et al. 2023), and heterogeneous panels are often motivated by complementary perspectives [15, 36] (Zhou and Chen 2025). The closest broad evaluation is Zhang et al. (2025), who compare debate methods across models and benchmarks, find frequent losses to simpler inference baselines, but identify heterogeneity as a comparatively reliable remedy. We give that remedy a measurable source–receiver structure: a fixed peer has a same-task advantage Γd\Gamma_d, a receiver has reception ρd\rho_d, and later fixed-receiver directions test their product.

Other work questions whether consensus debate creates new capability at all [32, 37, 3] (Yang et al. 2025; Zhu et al. 2026). Choi et al. [5] show that majority voting accounts for most of the gain usually credited to deliberation. In their Bayesian model, debate follows martingale belief dynamics absent an update bias toward correct signals. A peer with measured task advantage supplies such a signal, and ρd\rho_d measures the fraction the receiver absorbs. The closest uncertainty decomposition is Qiao et al. [28], who separate inter-agent epistemic disagreement from within-agent aleatoric uncertainty. Our decomposition instead separates peer-side capability supply from receiver-side reception. Zhu et al. (2026) emphasize confidence and diversity; Wynn et al. [38] show mixed-capability interaction can pull correct agents toward incorrect answers. Selection- and routing-oriented systems instead preserve disagreement and choose among members [8, 35, 22]. Our ceiling and oracle measurements connect these strands: complementary answers are present in our mathematical panels, but consensus revision does not realize them.

Pappu et al. [27] study expert underuse and consensus/robustness tradeoffs in self-organizing teams. Our fixed composition, slots, and protocol instead measure peer advantage, reception, and role reversal for a specified pair. Their findings complement our ceiling and role-reversal results.

Persuasion, faulty peers, and revision. Adversarial or persuasive members can steer multi-agent systems [2, 19, 1, 7], while Byzantine analyses characterize robustness primarily through the number or placement of faulty agents (Jo and Park, 2025; Zheng et al., 2026; tse Huang et al., 2025). Nilayam et al. [25] analyze receiver revision and contaminated-panel robustness under honest and adversarial heterogeneous peers. We study a different estimand: net capability transfer relative to homogeneous control, and ask whether peer advantage and the receiver’s reception predict its magnitude and marginal adversarial leverage.

The closest transition-level comparison is Qu et al. [29], who distinguish beneficial revision from harmful revision using controlled peer messages and find models easier to mislead than correct. Simulated messages give clean control over consensus but have no supplying model whose task capability can be measured. Our real-peer design adds that missing capability structure and finds that beneficial reception exceeds adversarial uptake in most evaluated directions. Work on sycophancy and stance convergence [31, 4, 14] studies related response changes, but generally begins from a correct model under social pressure. Reception instead conditions on an initially wrong receiver and a correct peer.

Self-correction and compound-system propagation. Models often struggle to improve their own reasoning without external feedback (Huang et al., 2023); consequently [R,R,R][R, R, R] is an empirical control rather than a presumed monotonic baseline. More broadly, work on error propagation in compound systems separates the presence of an upstream signal from downstream detection and correction [24]. Our capability-supply/reception decomposition applies the same separation to heterogeneous debate and adds an adversarial role reversal on the supplying component. QUIVER quantifies perturbation propagation and structural divergence in graph-structured compound AI systems [23]; here we measure correctness transfer and adversarial influence under a fixed debate protocol.

Conclusion

In these panels, capability transfer dominates net gain, while item-level composition is not ruled out. The peer’s same-task advantage determines what can be supplied; the receiver’s reception determines how much is transferred. Nine directions expose the relationship, two fixed-receiver directions predict it without refitting, and an open-answer MuSiQue scope probe reaches the negative-gap side: in that direction, high reception produces movement but little net gain from the inferior peer. The benefit is also surrenderable. Reversing an influential peer’s role consumes the borrowed gain and can push the panel below homogeneous debate, while consensus still leaves correct member answers unused. Heterogeneous debate should therefore be measured as a supply, reception, and integrity problem—and compared with calling or routing to the influential member directly.

Limitations

Finite and dependent directions. The cross-direction analysis contains nine primary and two fixed-receiver confirmatory directions. Several share models, tasks, and traces; they are not independent draws from a population of model pairs. Exact direction-level tests and item-clustered intervals quantify the evaluated comparisons without turning them into population sampling.

Cross-task scope. The heterogeneous MuSiQue cohort provides a single-direction probe of the negative-gap gate on 300 open-answer multi-hop questions, but it is one model pair and one replicate. It does not test positive-gap transfer or the product relationship outside MATH-hard and GSM8K. Whether positive-gap transfer replicates on open-answer tasks is untested.

Empirical factors. The fixed-receiver ladder holds the receiver, task, protocol, and baseline constant, but changing peer capability also changes peer identity and argument style. Reception is measured under the debate protocol. We treat Γd\Gamma_d and ρd\rho_d as separately measurable predictors, not intrinsic causal traits of a model independent of its interaction context. Estimating them for a new deployment requires a labeled, deployment-matched probe.

Structural coupling. When peer and receiver correctness are nested, the item mass underlying ρd\rho_d approaches Γd\Gamma_d. Equation (9) therefore contains a structural component in that regime. The fixed-receiver predictions and the MuSiQue negative-gap result provide tests beyond the nine-direction fit, but do not remove the structural component.

Threat model and rounds. The adversary follows one committed-wrong role prompt. The same-family MuSiQue control demonstrates that lack of a positive heterogeneous gap does not protect against every malicious participant, and the R4 trajectory shows that weak-peer harm can accumulate after a small first exchange. Coordinated, adaptive, or multi-peer attacks remain outside the evaluated threat model.

Protocol and ceiling. All panels use three agents, full transcript visibility, synchronous revision, and one designated peer slot. Reception is an operational transition rate, not a cognitive-pathway measure: answer-only, reasoning-only, and full-response ablations are needed to distinguish adoption from reasoning-enabled correction. The borrowed-ceiling comparison selects the best observed round in settings with complete standalone peer trajectories. It is a property of these consensus panels, not an upper bound on routing, selection, or other heterogeneous protocols.

Ethics Statement

This work studies a panel member instructed to defend a plausible incorrect answer. The prompt is a direct role instruction rather than a jailbreak or a new attack technique, and all adversarial runs concern public reasoning benchmarks. We report it because identifying high-influence members as security-critical can support safer deployment, auditing, and verification.

MATH, GSM8K, and MuSiQue are public datasets and contain no human-subject data collected for this study. Open-weight models were served locally; hosted models were accessed through their providers under standard terms. Analyses operate on recorded traces and require no additional model calls. AI coding assistants were used for analysis tooling and manuscript preparation; the authors verified the reported quantities and remain responsible for the claims.

Reproducibility

The released artifact, documented at https://github.com/prashanti-nilayam/BSIR, contains recorded JSONL traces, source manifests, benchmark-aware parsers, analysis programs, and independent validators. The validators reconstruct matched support and recount transition cells, bootstrap intervals, direction-level tests, reception quantities, and ceiling statistics without importing the analysis implementation. All manuscript figures and tables are generated from frozen analysis outputs. Operational commands and file-level hashes are provided in the artifact README rather than reproduced in the paper.

Run Configuration

Primary panels contain three agents and five synchronous rounds. MATH-hard uses the 262 level-4/5 MATH-500 items and a 4,096-token completion budget; GSM8K uses the full 1,319-item test set and 1,024 tokens. Primary decoding is temperature 0. The fixed-receiver MATH experiment uses three seeds and the same items, prompts, rounds, and budgets in every condition. The Qwen homogeneous baseline is shared by the Gemma and GPT-OSS directions. GPT-OSS follows the committed-wrong prompt on 99.75% of complete replicate-items across all five rounds; the corresponding Gemma rate is 68.54%, so intent-to-treat remains the primary comparison. The MuSiQue scope cohort uses one temperature-0 replicate per condition on 300 matched items; every one of its six panels contains all 4,500 agent-round cells.

For the token comparison, matched replicate-items require complete five-round honest panels and peer-reference R0 cells. We sum reported tokens across all 15 debate calls; the exchangeable peer-reference agents estimate the mean cost and accuracy of one R0 call. This support can differ from the four-panel first-exchange support. Different tokenizers make token volume an inference proxy; best-round majority accuracy is selected retrospectively.

Models as served are GPT-4.1 ([26]), Gemini-2.5-Flash ([10]), GPT-OSS-120B ([26]), Llama-3.1-70B-Instruct, Llama-3.1-8B-Instruct, and Llama-3.2-1B-Instruct ([13]); Qwen-2.5-72B-Instruct and Qwen-2.5-7B-Instruct ([30]); and Gemma-3-12B-IT ([11]) and Gemma-2-2B-IT ([12]).

Per-Direction Results

Table 5 reports per-direction recovery and corruption rates under all three panel conditions, Table 6 the reception quantities and their unaided controls, and Table 7 Wilson intervals for the underlying transition counts. Table 8 aggregates the role effects at first peer exposure and Table 9 the logistic specifications.

DirectionΓ\Gamma (pp)ρ\rho (%)Bootstrap 95% CInnUnaided (%)Uptake (%)Gain (pp)
Gemma-3-12B ←\leftarrow Qwen-7B−9.7-9.714.7[2.6, 29.6]1165.71.5+1.0
Qwen-7B ←\leftarrow Llama-8B−4.8-4.836.9[25.2, 49.7]2445.60.5+0.6
Gemini-Flash ←\leftarrow GPT-4.1−3.5-3.532.5[14.1, 51.6]11413.91.8−0.1-0.1
GPT-4.1 ←\leftarrow Gemini-Flash+3.583.3[72.2, 93.7]14414.09.8+4.8
GPT-4.1 ←\leftarrow OSS-120B+4.383.1[72.5, 92.7]15413.39.9+4.8
Llama-8B ←\leftarrow Qwen-7B+4.878.5[72.4, 84.2]6322.325.2+3.6
Qwen-72B ←\leftarrow OSS-120B+14.763.0[51.0, 74.7]2547.82.3+11.1
Llama-1B ←\leftarrow Gemma-2B+23.317.9[5.8, 31.2]568.229.4+5.0
Llama-70B ←\leftarrow OSS-120B+36.071.0[62.9, 78.6]59910.230.9+28.2

Table 6. Reception power ρ=P(correct at R1∣wrong at R0, honest peer correct at R0)\rho= P(\text{correct at R1} \mid\text{wrong at R0, honest peer correct at R0}), with its control and its adversarial counterpart. Unaided is the same conditional with the honest peer also wrong, and isolates ambient self-revision from peer-driven revision. Uptake is P(wrong at R1∣correct at R0, adversarial peer wrong at R0)P(\text{wrong at R1} \mid\text{correct at R0, adversarial peer wrong at R0}). Gain is the honest net accuracy change relative to homogeneous debate. Intervals are global benchmark-item Poisson bootstraps; Wilson intervals are in the appendix.

Directionρ\rho (%)Wilson 95%Bootstrap 95%
Gemma-3-12B ←\leftarrow Qwen-7B14.7[9.4, 22.2][2.6, 29.6]
Qwen-7B ←\leftarrow Llama-8B36.9[31.1, 43.1][25.2, 49.7]
Gemini-Flash ←\leftarrow GPT-4.132.5[24.6, 41.5][14.1, 51.6]
GPT-4.1 ←\leftarrow Gemini-Flash83.3[76.4, 88.5][72.2, 93.7]
GPT-4.1 ←\leftarrow OSS-120B83.1[76.4, 88.2][72.5, 92.7]
Llama-8B ←\leftarrow Qwen-7B78.5[75.1, 81.5][72.4, 84.2]
Qwen-72B ←\leftarrow OSS-120B63.0[56.9, 68.7][51.0, 74.7]
Llama-1B ←\leftarrow Gemma-2B17.9[10.0, 29.8][5.8, 31.2]
Llama-70B ←\leftarrow OSS-120B71.0[67.2, 74.4][62.9, 78.6]

Table 7. Reception intervals under two schemes. Wilson intervals treat exposures as independent; the benchmark-item bootstrap resamples benchmark items and so absorbs clustering across replicates and receiver slots. The bootstrap intervals are wider everywhere and underlie the interval reporting in the manuscript.

OutcomeSpecificationTermnnOR95% CIpp
recoverypair fe three conditionhonest vs. homogeneous142646.582[2.237, 19.367]0.0038
recoverypair fe three conditionadversarial vs. homogeneous142641.914[1.058, 3.464]0.0356
corruptionpair fe three conditionhonest vs. homogeneous620381.430[0.609, 3.357]0.3625
corruptionpair fe three conditionadversarial vs. homogeneous620387.345[1.097, 49.158]0.0419
recoverypair fixed effectsadversarial vs. honest94930.293[0.137, 0.624]0.0057
corruptionpair fixed effectsadversarial vs. honest413755.170[1.700, 15.725]0.0093
recoverygapadversarial vs. honest94930.466[0.319, 0.680]0.0016
recoverygapadversarial × Γ\Gamma94930.622[0.548, 0.706]0.0000
corruptiongapadversarial vs. honest413754.315[1.269, 14.673]0.0249
corruptiongapadversarial × Γ\Gamma413751.266[0.912, 1.757]0.1361

Table 9. Logistic specifications, intent to treat. All models carry a separate intercept per direction and use direction-clustered sandwich standard errors with a tt reference on 8 degrees of freedom, so item count is never treated as the number of independent capability comparisons. The three-condition models include the receiver-homogeneous panel; Γ\Gamma is centred and scaled in percentage points.

Confirmatory Fixed-Qwen Details

All three fixed-receiver conditions have 786 matched replicate-items at R1. The controlled GPT-OSS-minus-Gemma gain difference is +9.03 points with paired benchmark-item bootstrap interval [5.14,12.88]. The reception difference is −1.81 points [−13.01,9.87], while adversarial uptake increases by +3.41 points [1.12,5.82]. At R4, honest Qwen receivers reach 69.28% with

DirectionRecovery (%)Corruption (%)
Homog.HonestAdvers.Homog.HonestAdvers.
Gemma-3-12B ←\leftarrow Qwen-7B3.07.78.71.12.01.6
Qwen-7B ←\leftarrow Llama-8B3.917.713.10.10.60.4
Gemini-Flash ←\leftarrow GPT-4.117.120.615.21.52.41.8
GPT-4.1 ←\leftarrow Gemini-Flash26.641.830.78.86.510.2
GPT-4.1 ←\leftarrow OSS-120B26.641.726.88.87.39.9
Llama-8B ←\leftarrow Qwen-7B5.249.319.00.73.025.2
Qwen-72B ←\leftarrow OSS-120B2.735.25.41.51.02.3
Llama-1B ←\leftarrow Gemma-2B6.512.814.818.413.329.4
Llama-70B ←\leftarrow OSS-120B3.051.47.18.86.530.9

caption: Per-direction R0→\toR1 transition rates in all three matched conditions, intent to treat. The receiver-homogeneous column is the self-revision baseline that every honest and adversarial contrast in the main text is taken against.

:::

Gemma and 79.20% with GPT-OSS; the respective homogeneous Qwen support is approximately 59%. In the GPT-OSS panel, the peer itself peaks at 80.62%, majority correctness at 79.97%, and the any-member oracle at 85.37%. The Qwen–GPT-OSS setting follows the pattern of the five full-reference settings: the at-least-two-correct panel peaks at 79.97%, the peer in panel at 80.62%, and the oracle at 85.37%.

The item-level recovery trend pools Qwen receiver exposures that begin wrong in [Q,Q,Q][Q,Q,Q], [Q,Q,G][Q,Q,G], and [Q,Q,O][Q,Q,O], with the shared homogeneous baseline entered once rather than once per rung. A logistic model includes peer gap per ten points, replicate indicators, and receiver-slot indicator, with a benchmark-item clustered sandwich variance. It yields OR 3.03 [2.47,3.73]. The analogous model on R0-correct exposures yields honest-corruption OR .98 [.72,1.34]. These errors condition on this benchmark’s items and do not turn three peer levels into a population of models.

Trajectory and Ceiling Details

Table 10 reports per-direction headroom and the fraction of it closed; Table 11 gives the round-by-round accuracy curves.

DirectionRep.-itemsRecv. A0A_0Peer A0A_0Mixed bestControlled peak (pp)Gap closed
Gemma-3-12B ← Qwen-7B71572.063.273.2 (R2)+2.0–
Qwen-7B ← Llama-8B395791.987.192.9 (R3)+0.7–
Gemini-Flash ← GPT-4.178479.575.982.2 (R1)+0.7–
GPT-4.1 ← Gemini-Flash78275.879.682.2 (R3)+7.6167%
GPT-4.1 ← OSS-120B78475.980.181.3 (R4)+6.4127.5%
Llama-8B ← Qwen-7B395787.191.991.5 (R4)+4.591%
Qwen-72B ← OSS-120B76367.381.581.3 (R4)+14.099%
Llama-1B ← Gemma-2B10037.761.043.5 (R1)+8.825%
Llama-70B ← OSS-120B74546.780.572.7 (R3)+30.277%

Table 10. Full-trajectory view of the honest panels. This analysis uses its own, stricter matched support: a replicate-item is retained only when every round R0–R4 is complete in both the homogeneous and mixed panels, so cluster counts, A0A_0, and Γ\Gamma differ slightly from Table 1, where matching is on the R0/R1 cells the transition estimand needs. Controlled peak subtracts the receiver’s homogeneous accuracy at the same round, removing self-panel drift. Gap closed is the controlled peak as a fraction of Γ\Gamma, defined only where Γ>0\Gamma> 0; values above 100% mean the receivers ended above the measured gap, which is possible because the peer also gains from the exchange.

SettingSeriesR0R1R2R3R4
GPT-OSS-120B → Llama-70BPeer alone80.581.181.180.980.8
Peer in panel80.479.479.478.177.3
Receivers44.370.772.172.071.1
Panel ≥245.271.872.872.771.6
Oracle82.583.682.981.780.3
GPT-OSS-120B → GPT-4.1Peer alone80.280.880.880.780.6
Peer in panel79.481.479.178.177.7
Receivers75.980.480.981.081.4
Panel ≥277.681.481.281.081.3
Oracle87.285.683.282.281.8
Gemini-2.5-Flash → GPT-4.1Peer alone79.581.681.381.481.3
Peer in panel79.782.581.080.480.4
Receivers77.181.682.082.282.2
Panel ≥279.282.382.582.182.3
Oracle87.287.184.684.183.9
GPT-4.1 → Gemini-2.5-FlashPeer alone75.975.574.474.774.9
Peer in panel77.082.582.782.582.5
Receivers80.182.281.480.980.9
Panel ≥280.582.781.681.181.1
Oracle87.486.984.884.184.1
Gemma-2B → Llama-1BPeer alone61.0––––
Peer in panel61.054.055.053.049.0
Receivers41.543.543.542.541.5
Panel ≥242.046.048.046.042.0
Oracle71.061.058.056.056.0

Table 11. Round-by-round accuracy in the five full-reference settings. In all five, the peer’s own accuracy inside the panel declines from its best round to the final round — most sharply for Gemma-2B beside two Llama-1B receivers, which falls from 61.0% to 49.0%. A best-round comparison alone would hide this. Missing cells are rounds absent from the small-model diagnostic’s standalone trace, which has one agent and one round.

Scope Controls

Open-answer multi-hop QA

The MuSiQue cohort contains six complete panels on the same 300 questions, balanced across two-, three-, and four-hop items: [G,G,G][G,G,G], [Q,Q,Q][Q,Q,Q], [G,G,Q][G,G,Q], [G,G,Qadv][G,G,Q^{\mathrm{adv}}], [G,G,Gadv][G,G,G^{\mathrm{adv}}], and [G,Gadv,Q][G,G^{\mathrm{adv}},Q]. Answers are scored against references and aliases using normalized exact match and token F1; no LLM judge is used. The Gemma-homogeneous round curve remains useful as a task baseline: pooled exact match rises from 45.67% at R0 to 48.11% at R1 and ends at 47.44% (Table 14).

RoundPooled EM≥2\ge2OracleF1
R045.6745.6747.3358.90
R148.1148.3349.0061.86
R248.0048.0049.3361.27
R348.0048.3348.6761.00
R447.4447.3348.3360.57

Table 14. Homogeneous baseline within the MuSiQue cohort: 300 items, one Gemma-3-12B panel, and alias-normalized exact match recomputed from raw answers with no LLM judge. Pooled exact match rises 2.44 pp to a peak at R1 and gives back 0.67 pp by R4. The oracle exceeds realized consensus by at most 1.67 pp here, against 4.7–23.0 pp in the five mathematics settings of Table 12. The matched heterogeneous comparison is in Table 15.

The heterogeneous comparison supplies the cross-task scope probe. Qwen’s R0 capability is 10.78 points below Gemma’s. At R1, an honest Qwen peer raises recovery and corruption by nearly equal amounts, leaving controlled net gain at +1.17 points; reception is nevertheless 62.50%. The adversarial Qwen peer adds 2.66 points of excess corruption at R1 and reaches a controlled −4.50 points by R4. This is an ungated direction: active reception with negative aggregate peer advantage produces little first-exchange gain or leverage, while repeated adversarial influence can still accumulate.

Figure A1. Reception power in the nine directions. Left: ρd\rho_d and its unaided control ρ0d\rho_{0d} against the measured gap. Right: observed gain against predictions from the gap-only and product approximations. The fit is descriptive over directions; the later fixed-receiver test is shown separately.

Two further control panels separate that heterogeneous result from broader threat claims. A same-family Gemma adversary raises R1 corruption by 8.47 points and reaches −9.50 controlled points at R4. In the contaminated panel, an honest Qwen peer improves the sole honest Gemma receiver by +5.00 controlled points at R1 [1.02,9.18], but the R4 effect is −3.00 [−8.36,2.49]. This is a single run, and we do not generalize from it.

Prompt, temperature, and seed variation

To test whether complementary information can be induced without model diversity, we compare uniform Gemma-3-12B debate with a same-model panel whose agents use careful, critical, and concise prompt emphases at temperatures 0, .3, and .7. Both conditions use seeds 0, 1, and 2 on the same 262 MATH-hard items. The diverse configuration increases R0 any-agent oracle coverage from 67.81% to 69.97%, demonstrating more complementary answers. Debate does not realize them: majority correctness peaks at 63.74%, below the uniform panel’s 67.94%, and the diverse-panel oracle contracts by 4.45 points as agents converge (Table 16). This control does not exhaust prompt or sampling diversity; it shows that eliciting complementarity is insufficient when the protocol does not preserve and select it.

PanelMetricR0R1R2R3R4
Diverse config.Pooled agent60.3163.3263.4063.3663.49
At least two correct60.4363.6163.4963.1063.74
Any-agent oracle69.9768.0766.7966.1665.52
Uniform config.Pooled agent67.8168.0767.1366.4566.71
At least two correct67.8167.9467.1866.5466.79
Any-agent oracle67.8168.4567.3066.7966.92

Table 16. Same-model diversity control on MATH-hard (three seeds, 262 items each). The diverse Gemma-3-12B panel assigns careful/critical/concise prompt emphases and temperatures 0/.3/.7 to its three agents; the uniform panel uses a shared prompt at temperature 0. Seeds are 0/1/2 in both conditions. Diverse decoding increases R0 oracle coverage but not realized panel agreement. Parse failures count as incorrect.

Sensitivity and Structural Coupling

Table 17 reports the diagnostic-exclusion sensitivity analysis, including cross-direction associations and alternative specifications.

AssociationrrExact pp
r(Γ, honest net gain)r(\Gamma,\text{ honest net gain})+0.9730.0009
r(Γ, adversarial corruption rate)r(\Gamma,\text{ adversarial corruption rate})+0.7290.0387
r(Γ, corruption vs. homogeneous)r(\Gamma,\text{ corruption vs. homogeneous})+0.6070.0914
r(Γ,ρ)r(\Gamma,\rho)+0.5490.1544
r(H,ρ)r(H,\rho)+0.1130.7973
r(ρ,υ)r(\rho,\upsilon)+0.6110.1120
SpecificationSlopeR2R^2
Γ+\Gamma^{+}0.7590.9925
ρ\rho14.1970.1590
Γ+×ρ\Gamma^{+}\times\rho1.0870.9951

Table 17. Every cross-direction relation refitted on the eight directions that remain when the 100-item small-model diagnostic is removed, with exact pp over all 8!=40,3208! = 40,320 label permutations. No sign changes and no qualitative change: the gate relations strengthen, the two reception null results stay null, and the product specification retains the highest R2R^2 of the three. Computed arithmetically from the frozen per-direction values reported in Table 6 and 5.

Table 18 separates useful peer-only mass SS from opposing receiver-only mass TT. All negative-gap directions still contain useful peer-only answers; MuSiQue is the clearest cancellation case, with Sρ=6.67S\rho= 6.67 points of realized correction and Tη=5.17T\eta= 5.17 points of realized corruption within the discordant states. The gate therefore concerns net gain, not whether useful information exists at all.

DirectionΓ\GammaSSρ\rhoSρS\rhoTTη\etaTηT\etaNet
Gemma-3-12B ←\leftarrow Qwen-7B−9.77.414.71.0816.45.80.95+0.13
Qwen-7B ←\leftarrow Llama-8B−4.83.136.91.148.65.70.49+0.64
Gemini-Flash ←\leftarrow GPT-4.1−3.57.232.52.3510.27.50.76+1.59
GPT-4.1 ←\leftarrow Gemini-Flash+3.59.283.37.636.524.31.59+6.04
GPT-4.1 ←\leftarrow OSS-120B+4.39.883.18.146.334.32.16+5.98
Llama-8B ←\leftarrow Qwen-7B+4.88.078.56.273.184.22.62+3.65
Qwen-72B ←\leftarrow OSS-120B+14.716.463.010.351.910.00.19+10.16
Llama-1B ←\leftarrow Gemma-2B+23.328.017.95.008.541.23.50+1.50
Llama-70B ←\leftarrow OSS-120B+36.038.171.027.042.00.00.00+27.04
Qwen-7B ←\leftarrow Gemma-3-12B+9.116.562.210.247.531.42.35+7.89
Qwen-7B ←\leftarrow OSS-120B+21.525.260.415.202.031.20.64+14.57
Gemma-3-12B ←\leftarrow Qwen-7B (MuSiQue)−10.810.762.56.6721.823.75.17+1.50

Table 18. Discordant-state accounting for honest heterogeneous panels at R0→R1. S=P(R0=0,P0=1)S = P(R_0 = 0, P_0 = 1) is useful peer-only mass and T=P(R0=1,P0=0)T = P(R_0 = 1, P_0 = 0) is opposing receiver-only mass, both measured in the honest mixed panel. ρ=P(R1=1∣R0=0,P0=1)\rho= P(R_1 = 1 \mid R_0 = 0, P_0 = 1) and η=P(R1=0∣R0=1,P0=0)\eta= P(R_1 = 0 \mid R_0 = 1, P_0 = 0); the realized flows are SρS\rho and TηT\eta, and Net is their difference. Γ\Gamma is measured separately from the matched homogeneous reference panels, so condition-specific R0 variation means it need not equal S−TS - T exactly. The first nine rows are the discovery set, the next two the fixed-receiver ladder, and the final row the MuSiQue scope probe.

The accounting comparison in Table 19 also checks whether gross useful mass is a better description of controlled gain. It is not: SρS\rho, (S−T)+ρ(S-T)^{+}\rho, and Sρ−TηS\rho-T\eta all fit less closely than the original Γ+ρ\Gamma^{+}\rho specification on the same nine directions. This does not make ρ\rho or η\eta independent predictors—both contain R1 outcomes. It shows that the aggregate reference gap remains the tighter empirical summary rather than merely standing in for all available peer-only correct answers. We do not use ratios of ρd\rho_d and νd\nu_d: they condition on different initial states and item populations.

PredictorSlopeInterceptR2R^2Max. residual
Γ+ρ\Gamma+\rho1.088+0.610.99501.07
(S−T)+ρ(S-T)+\rho1.072+0.940.98981.51
SρS\rho1.088−1.790.96802.30
Sρ−TηS\rho-T\eta1.027+0.080.96713.38

Table 19. Alternative descriptions of controlled honest gain on the same nine directions, all fitted with an intercept. (S−T)(S-T) is the R0 gap inside the honest mixed panel, whereas Γ\Gamma uses the matched homogeneous references. Because ρ\rho and η\eta include R1 outcomes, these are accounting diagnostics, not independent predictive models. Replacing aggregate advantage with gross useful mass or signed realized discordant flow does not improve the descriptive fit. Residuals are percentage points.

Prompt Structure

At R0, agents receive the benchmark problem and an instruction to explain their reasoning and provide a parseable final answer. At R1–R4, each agent receives the other two complete previous-round responses and is asked to answer the original problem again. The adversarial peer instead receives the gold answer and a plausible incorrect target, is prohibited from returning the gold answer, and is instructed to defend the same target concisely across rounds. The full verbatim templates are included in the released artifact. No other condition differs between the honest and adversarial mixed panels.

Figure A2. Borrowed benefit and imported risk across the nine directions. Left: direction-specific controlled accuracy under an honest peer and the same peer under an adversarial role. Right: in the five full-reference settings, the best majority correctness does not exceed the influential peer’s best observed accuracy while the any-member oracle is higher.

SettingnnPanelPeerPanel-onlyPeer-onlyNetNewDiscarded
GPT-OSS-120B → Llama-70B77671.6577.322.968.63−5.672.5813.40
GPT-OSS-120B → GPT-4.178681.3077.743.690.13+3.561.657.51
Gemini-2.5-Flash → GPT-4.178482.2780.363.191.28+1.911.916.89
GPT-4.1 → Gemini-2.5-Flash78481.1282.531.532.93−1.402.048.29
Gemma-2B → Llama-1B10042.0049.001.008.00−7.001.0030.00

Table 13. Fixed-round item accounting in the five full-reference honest panels at R4 (%). Panel requires at least two correct members; Peer is the influential peer in that panel. Panel-only and peer-only are the two discordant outcomes, with Net equal to their difference. New denotes a correct R4 panel when no member was correct at R0; discarded denotes an incorrect R4 panel despite at least one correct R0 member. New outcomes demonstrate item-level exceptions to a pure relay account, but do not by themselves identify a causal composition mechanism. The fixed round avoids selecting these contrasts after observing their values.

References

  1. [1]Mahak Agarwal and Divyam Khanna. 2025. When persuasion overrides truth in multi-agent llm debates: Introducing a confidence-weighted persuasion override rate (cw-por). arXiv preprint arXiv:2504.00374.
  2. [2]Alfonso Amayuelas, Xianjun Yang, Antonis Antoniades, Wenyue Hua, Liangming Pan, and William Yang Wang. 2024. Multiagent collaboration attack: Investigating adversarial attacks in large language model collaborations via debate. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 6929–6948.
  3. [3]Blaž Bertalanič and Carolina Fortuna. 2026. The cost of consensus: Isolated self-correction prevails over unguided homogeneous multi-agent debate. In Proceedings of the ACM Conference on AI and Agentic Systems.DOI
  4. [4]Yusuf Çelebi, Özay Ezerçeli, and Mahmoud El Hussieni. 2025. Parrot: Persuasion and agreement robustness rating of output truth—a sycophancy robustness benchmark for llms. arXiv preprint arXiv:2511.17220.
  5. [5]Hyeong Kyu Choi, Xiaojin Zhu, and Sharon Li. 2025. Debate or vote: Which yields better decisions in multi-agent large language models? In Advances in Neural Information Processing Systems. Spotlight.arxiv.org/abs/2508.17536
  6. [6]Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, and 1 others. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168.arxiv.org/abs/2110.14168
  7. [7]Yu Cui and Hongyang Du. 2025. Mad-spear: A conformity-driven prompt injection attack on multi-agent debate systems. arXiv preprint arXiv:2507.13038.arxiv.org/abs/2507.13038
  8. [8]Yu Cui, Hang Fu, Haibin Zhang, Licheng Wang, and Cong Zuo. 2026. Free-MAD: Consensus-free multi-agent debate. In Findings of the Association for Computational Linguistics: ACL 2026, pages 31977–31997.arxiv.org/abs/2509.11035
  9. [9]Yilun Du, Shuang Li, Antonio Torralba, Joshua B Tenenbaum, and Igor Mordatch. 2024. Improving factuality and reasoning in language models through multi-agent debate. In Forty-first international conference on machine learning.
  10. [10]Gemini Team. 2025. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. Technical report, Google.
  11. [11]Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, and 1 others. 2025. Gemma 3 technical report. arXiv preprint arXiv:2503.19786.
  12. [12]Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, and 1 others. 2024. Gemma 2: Improving open language models at a practical size. arXiv preprint arXiv:2408.00118.
  13. [13]Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, and 1 others. 2024. The Llama 3 herd of models. arXiv preprint arXiv:2407.21783.arxiv.org/abs/2407.21783
  14. [14]Xiqi Hao, Zengqing Wu, Yu-Xuan Qiu, Chuan Xiao, Ruiqi Xu, Shuyuan Zheng, and Jianbin Qin. 2026. Not all flips are conformity: Decomposing stance convergence in multi-agent llm debate. arXiv preprint arXiv:2606.00820.arxiv.org/abs/2606.00820
  15. [15]Mahmood Hegazy. 2024. Diversity of thought elicits stronger reasoning capabilities in multi-agent debate frameworks. arXiv preprint arXiv:2410.12853.arxiv.org/abs/2410.12853
  16. [16]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. arXiv preprint arXiv:2103.03874.
  17. [19]Insaf Kraidia, Iyas Qaddara, Alhanof Almutairi, Nada Alzaben, and Samir Ibrahim Belhouari. 2026. When collaboration fails: persuasion driven adversarial influence in multi agent large language model debate. Scientific Reports, 16(1):11640.DOI
  18. [20]Tian Liang, Zhiwei He, Wenxiang Jiao, Xing Wang, Yan Wang, Rui Wang, Yujiu Yang, Shuming Shi, and Zhaopeng Tu. 2024. Encouraging divergent thinking in large language models through multi-agent debate. In Proceedings of the 2024 conference on empirical methods in natural language processing, pages 17889–17904.arxiv.org/abs/2305.19118
  19. [21]Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Dan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2024. Let’s verify step by step. In International Conference on Learning Representations, volume 2024, pages 39578–39601.arxiv.org/abs/2305.20050
  20. [22]Artem Maryanskyy, Dmitry Budnikov, and Alibek T Kaliyev. 2026. When agents disagree: The selection bottleneck in multi-agent llm pipelines. Applied Sciences, 16(10):4914.
  21. [23]Prashanti Nilayam and Sankalp Nayak. 2026. QUIVER: A formal framework for quantifying perturbation propagation and bifurcation in compound AI systems. arXiv preprint arXiv:2605.23956.arxiv.org/abs/2605.23956
  22. [24]Prashanti Nilayam, Kiran Ramanna, and Prashil Tumbade. 2026a. Detection without correction: A two-parameter decomposition of multi-stage llm pipelines. arXiv preprint arXiv:2605.27559.arxiv.org/abs/2605.27559
  23. [25]Prashanti Nilayam, Kiran Kumar Ramanna, Prashil Tumbade, and Sankalp Nayak. 2026b. Heterogeneous LLM debate under adversarial peers: Honest gains, replacement costs, and resilience. arXiv preprint arXiv:2606.19826.arxiv.org/abs/2606.19826
  24. [26]OpenAI. 2025. Introducing GPT-4.1 in the API.
  25. [27]Aneesh Pappu, Batu El, Hancheng Cao, Carmelo Di Nolfo, Yanchao Sun, Meng Cao, and James Zou. 2026. Multi-agent teams hold experts back. In Proceedings of the 43rd International Conference on Machine Learning, volume 306 of Proceedings of Machine Learning Research, pages 96478–96515. PMLR.arxiv.org/abs/2602.01011
  26. [28]Dan Qiao, Binbin Chen, Fengyu Cai, Jianlong Chen, Wenhao Li, Fuxin Jiang, Zuzhi Chen, Hongyuan Zha, Tieying Zhang, and Baoxiang Wang. 2026. Epistemic gain, aleatoric cost: Uncertainty decomposition in multi-agent debate for math reasoning. In Proceedings of the 43rd International Conference on Machine Learning.arxiv.org/abs/2603.01221
  27. [29]Jiaming Qu, Lucheng Fu, and Yibo Hu. 2026. Easier to mislead than to correct: Harmful and beneficial revision in LLM conformity. arXiv preprint arXiv:2606.01637.arxiv.org/abs/2606.01637
  28. [30]Qwen, An Yang, Baosong Yang, Beichen Zhang, and 1 others. 2024. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115.arxiv.org/abs/2412.15115
  29. [31]Mrinank Sharma, Meg Tong, Tomek Korbak, David Duvenaud, Amanda Askell, Sam Bowman, Esin Durmus, Zac Hatfield-Dodds, Scott Johnston, Shauna Kravec, and 1 others. 2024. Towards understanding sycophancy in language models. In International Conference on Learning Representations, volume 2024, pages 110–144.arxiv.org/abs/2310.13548
  30. [32]Andries Smit, Nathan Grinsztajn, Paul Duckworth, Thomas D Barrett, and Arnu Pretorius. 2024. Should we be going mad? a look at multi-agent debate strategies for llms. In Proceedings of the 41st International Conference on Machine Learning, pages 45883–45905.arxiv.org/abs/2311.17371
  31. [33]Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. 2022. MuSiQue: Multi-hop questions via single-hop question composition. Transactions of the Association for Computational Linguistics, 10:539–554.DOI
  32. [35]Akshay Verma, Swapnil Gupta, Deepak Gupta, Prateek Sircar, and Siddharth Pillai. 2026. SELENE: Selective and evidence-weighted LLM debating for efficient and reliable reasoning. In Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 5: Industry Track), pages 95–104.
  33. [36]Junlin Wang, Jue Wang, Ben Athiwaratkun, Cе Zhang, and James Y Zou. 2025. Mixture-of-agents enhances large language model capabilities. In International Conference on Learning Representations, volume 2025, pages 33944–33963.arxiv.org/abs/2406.04692
  34. [37]Haolun Wu, Zhenkun Li, and Lingyao Li. 2025. Can llm agents really debate? a controlled study of multi-agent debate in logical reasoning. arXiv preprint arXiv:2511.07784.arxiv.org/abs/2511.07784
  35. [38]Andrea Wynn, Harsh Satija, and Gillian Hadfield. 2025. Talk isn’t always cheap: Understanding failure modes in multi-agent debate. arXiv preprint arXiv:2509.05396.arxiv.org/abs/2509.05396

Paper details

vlnc:W_tp7r76jn
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

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

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents