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+ρd, 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.
Figures & tables
Direction ( R←P )
Task
Rep.-items
A0,R
A0,P
Γd
Hd
Gemma-3-12B ← Qwen-7B
MATH-hard
786
67.8
58.1
− 9.7
32.2
Qwen-7B ← Llama-8B
GSM8K
3957
91.9
87.1
− 4.8
8.1
Gemini-Flash ← GPT-4.1
MATH-hard
786
79.4
75.9
− 3.5
20.6
GPT-4.1 ← Gemini-Flash
MATH-hard
786
75.9
79.4
+3.5
24.1
GPT-4.1 ← OSS-120B
MATH-hard
786
75.9
80.2
+4.3
24.1
Llama-8B ← Qwen-7B
GSM8K
3957
87.1
91.9
+4.8
12.9
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,R and A0,P are homogeneous-panel accuracies in percent, Γd=A0,P−A0,R , and Hd=100−A0,R . † is a one-replicate, 100-item diagnostic; all others use three replicates.
Quantity (pp)
Ungated
Gated
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 p 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 ; filled points have Γd>0 .
Predictor
Slope
R2
Exact p
LOO slope
Γ+
0.614
0.7760
0.0078
[0.28, 0.76]
ρ
11.971
0.1455
0.3560
[5.46, 15.39]
Γ+×ρ
1.088
0.9950
0.0002
[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) encodes the gate. Exact p enumerates all 9!=362,880 label permutations; LOO slope is the range over the nine leave-one-direction-out refits.
Qwen-7B peer
Qwen
Gemma
OSS-120B
Γd (pp)
0.00
9.12
21.50
ρd (%)
–
62.16
60.35
R1 gain (pp)
0.00
+7.89
+16.92
R4 gain (pp)
0.00
+11.05
+21.37
R1 recovery (%)
2.92
25.26
42.94
R1 honest corruption (%)
2.06
4.72
2.21
Table 4: Fixed-receiver MATH-hard ladder. Gain is the receiver’s controlled accuracy change relative to [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.
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.
Appendix figures & tables17 assets
Supplementary material from the paper’s appendix.
Appendix
Recovery (%)
Corruption (%)
Direction
Homog.
Honest
Advers.
Homog.
Honest
Advers.
Gemma-3-12B ← Qwen-7B
3.0
7.7
8.7
1.1
2.0
1.6
Qwen-7B ← Llama-8B
3.9
17.7
13.1
0.1
0.6
0.4
Gemini-Flash ← GPT-4.1
17.1
20.6
15.2
1.5
2.4
1.8
GPT-4.1 ← Gemini-Flash
26.6
41.8
30.7
8.8
6.5
10.2
GPT-4.1 ← OSS-120B
26.6
41.7
26.8
8.8
7.3
9.9
Appendix
Table 5: Per-direction R0 → R1 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.
Direction
Γ (pp)
ρ (%)
Bootstrap 95% CI
n
Unaided (%)
Uptake (%)
Gain (pp)
Gemma-3-12B ← Qwen-7B
− 9.7
14.7
[2.6, 29.6]
116
5.7
1.5
+1.0
Qwen-7B ← Llama-8B
− 4.8
36.9
[25.2, 49.7]
244
5.6
0.5
+0.6
Gemini-Flash ← GPT-4.1
− 3.5
32.5
[14.1, 51.6]
114
13.9
1.8
− 0.1
GPT-4.1 ← Gemini-Flash
+3.5
83.3
[72.2, 93.7]
144
14.0
9.8
+4.8
GPT-4.1 ← OSS-120B
+4.3
83.1
[72.5, 92.7]
154
13.3
9.9
+4.8
Llama-8B ← Qwen-7B
+4.8
78.5
[72.4, 84.2]
632
2.3
25.2
+3.6
Appendix
Table 6: Reception power ρ=P(correct at R1∣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) . Gain is the honest net accuracy change relative to homogeneous debate. Intervals are global benchmark-item Poisson bootstraps; Wilson intervals are in the appendix.
Direction
ρ (%)
Wilson 95%
Bootstrap 95%
Gemma-3-12B ← Qwen-7B
14.7
[9.4, 22.2]
[2.6, 29.6]
Qwen-7B ← Llama-8B
36.9
[31.1, 43.1]
[25.2, 49.7]
Gemini-Flash ← GPT-4.1
32.5
[24.6, 41.5]
[14.1, 51.6]
GPT-4.1 ← Gemini-Flash
83.3
[76.4, 88.5]
[72.2, 93.7]
GPT-4.1 ← OSS-120B
83.1
[76.4, 88.2]
[72.5, 92.7]
Llama-8B ← Qwen-7B
78.5
[75.1, 81.5]
[72.4, 84.2]
Appendix
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.
Figure A1: Reception power in the nine directions. Left: ρd and its unaided control ρ0d 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.
Contrast
Mean (pp)
Bootstrap 95% CI
Exact p
Odds ratio [95% CI]
Intent to treat (12,717 matched replicate-items)
Recovery: honest vs. homogeneous
+20.40
[+17.62, +23.28]
0.0039
6.58 [2.24, 19.37]
Corruption: honest vs. homogeneous
− 0.78
[ − 2.28, +0.74]
0.3555
1.43 [0.61, 3.36]
Recovery: adversarial vs. honest
− 15.25
[ − 17.94, − 12.61]
0.0156
0.29 [0.14, 0.62]
Corruption: adversarial vs. honest
+7.67
[+5.98, +9.37]
0.0312
5.17 [1.70, 15.72]
Per protocol (10,580 items; adversary parsed-wrong in all of R0–R4)
Appendix
Table 8: Role effects at first peer exposure (R0 → R1), case-standardized over the nine directions. Exact p is a paired sign-flip test over the 29 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.
Outcome
Specification
Term
n
OR
95% CI
p
recovery
pair fe three condition
honest vs. homogeneous
14264
6.582
[2.237, 19.367]
0.0038
recovery
pair fe three condition
adversarial vs. homogeneous
14264
1.914
[1.058, 3.464]
0.0356
corruption
pair fe three condition
honest vs. homogeneous
62038
1.430
[0.609, 3.357]
0.3625
corruption
pair fe three condition
adversarial vs. homogeneous
62038
7.345
[1.097, 49.158]
0.0419
recovery
pair fixed effects
adversarial vs. honest
9493
0.293
[0.137, 0.624]
0.0057
corruption
pair fixed effects
adversarial vs. honest
41375
5.170
[1.700, 15.725]
0.0093
Appendix
Table 9: Logistic specifications, intent to treat. All models carry a separate intercept per direction and use direction-clustered sandwich standard errors with a t 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 is centred and scaled in percentage points.
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.
Direction
Rep.-items
Recv. A0
Peer A0
Mixed best
Controlled peak (pp)
Gap closed
Gemma-3-12B ← Qwen-7B
715
72.0
63.2
73.2 (R2)
+2.0
–
Qwen-7B ← Llama-8B
3957
91.9
87.1
92.9 (R3)
+0.7
–
Gemini-Flash ← GPT-4.1
784
79.5
75.9
82.2 (R1)
+0.7
–
GPT-4.1 ← Gemini-Flash
782
75.8
79.6
82.2 (R3)
+7.6
167%
GPT-4.1 ← OSS-120B
784
75.9
80.1
81.3 (R4)
+6.4
127.5%
Llama-8B ← Qwen-7B
3957
87.1
91.9
91.5 (R4)
+4.5
91%
Appendix
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, A0 , and Γ 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 , defined only where Γ>0 ; values above 100% mean the receivers ended above the measured gap, which is possible because the peer also gains from the exchange.
Setting
Series
R0
R1
R2
R3
R4
GPT-OSS-120B → Llama-70B
Peer alone
80.5
81.1
81.1
80.9
80.8
Peer in panel
80.4
79.4
79.4
78.1
77.3
Receivers
44.3
70.7
72.1
72.0
71.1
Panel ≥ 2
45.2
71.8
72.8
72.7
71.6
Oracle
82.5
83.6
82.9
81.7
80.3
GPT-OSS-120B → GPT-4.1
Peer alone
80.2
80.8
80.8
80.7
80.6
Appendix
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.
Peer (%)
Panel (%)
Setting
n
alone
in panel
receivers
≥ 2
oracle
Shortfall
GPT-OSS-120B → Llama-70B
776
81.1
80.4
72.1
72.8
83.6
− 10.8
GPT-OSS-120B → GPT-4.1
786
80.8
81.4
81.4
81.4
87.2
− 5.7
Gemini-2.5-Flash → GPT-4.1
784
81.6
82.5
82.2
82.5
87.2
− 4.7
GPT-4.1 → Gemini-2.5-Flash
784
75.9
82.7
82.2
82.7
87.4
− 4.7
Gemma-2B → Llama-1B
100
61.0
61.0
43.5
48.0
71.0
− 23.0
Appendix
Table 12: Where the panel stops. Accuracies are best-round values on n 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.
Setting
n
Panel
Peer
Panel-only
Peer-only
Net
New
Discarded
GPT-OSS-120B → Llama-70B
776
71.65
77.32
2.96
8.63
− 5.67
2.58
13.40
GPT-OSS-120B → GPT-4.1
786
81.30
77.74
3.69
0.13
+3.56
1.65
7.51
Gemini-2.5-Flash → GPT-4.1
784
82.27
80.36
3.19
1.28
+1.91
1.91
6.89
GPT-4.1 → Gemini-2.5-Flash
784
81.12
82.53
1.53
2.93
− 1.40
2.04
8.29
Gemma-2B → Llama-1B
100
42.00
49.00
1.00
8.00
− 7.00
1.00
30.00
Appendix
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.
Round
Pooled EM
≥ 2
Oracle
F1
R0
45.67
45.67
47.33
58.90
R1
48.11
48.33
49.00
61.86
R2
48.00
48.00
49.33
61.27
R3
48.00
48.33
48.67
61.00
R4
47.44
47.33
48.33
60.57
Appendix
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 .
Quantity
Estimate
Bootstrap 95% CI
n
Capability gap Γ (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
Appendix
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,G] ; transition quantities are measured at R0 → R1. The same-family adversary is [G,G,Gadv] . The contaminated comparison replaces one honest Gemma in that panel with Qwen, yielding [G,Gadv,Q] . Intervals are benchmark-item Poisson bootstraps.
Panel
Metric
R0
R1
R2
R3
R4
Diverse config.
Pooled agent
60.31
63.32
63.40
63.36
63.49
At least two correct
60.43
63.61
63.49
63.10
63.74
Any-agent oracle
69.97
68.07
66.79
66.16
65.52
Uniform config.
Pooled agent
67.81
68.07
67.13
66.45
66.71
At least two correct
67.81
67.94
67.18
66.54
66.79
Any-agent oracle
67.81
68.45
67.30
66.79
66.92
Appendix
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.
Association
r
Exact p
r(Γ,honest net gain)
+0.973
0.0009
r(Γ,adversarial corruption rate)
+0.729
0.0387
r(Γ,corruption vs. homogeneous)
+0.607
0.0914
r(Γ,ρ)
+0.549
0.1544
r(H,ρ)
+0.113
0.7973
r(ρ,υ)
+0.611
0.1120
Appendix
Table 17: Every cross-direction relation refitted on the eight directions that remain when the 100-item small-model diagnostic is removed, with exact p over all 8!=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 R2 of the three. Computed arithmetically from the frozen per-direction values reported in Tables 6 and 5 .
Direction
Γ
S
ρ
Sρ
T
η
Tη
Net
percentage points, except conditional rates ρ and η (%)
Gemma-3-12B ← Qwen-7B
− 9.7
7.4
14.7
1.08
16.4
5.8
0.95
+0.13
Qwen-7B ← Llama-8B
− 4.8
3.1
36.9
1.14
8.6
5.7
0.49
+0.64
Gemini-Flash ← GPT-4.1
− 3.5
7.2
32.5
2.35
10.2
7.5
0.76
+1.59
GPT-4.1 ← Gemini-Flash
+3.5
9.2
83.3
7.63
6.5
24.3
1.59
+6.04
GPT-4.1 ← OSS-120B
+4.3
9.8
83.1
8.14
6.3
34.3
2.16
+5.98
Appendix
Table 18: Discordant-state accounting for honest heterogeneous panels at R0 → R1. S=P(R0=0,P0=1) is useful peer-only mass and T=P(R0=1,P0=0) is opposing receiver-only mass, both measured in the honest mixed panel. ρ=P(R1=1∣R0=0,P0=1) and η=P(R1=0∣R0=1,P0=0) ; the realized flows are Sρ and Tη , and Net is their difference. Γ is measured separately from the matched homogeneous reference panels, so condition-specific R0 variation means it need not equal S−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.
Predictor
Slope
Intercept
R2
Max. residual
Γ+ρ
1.088
+0.61
0.9950
1.07
(S−T)+ρ
1.072
+0.94
0.9898
1.51
Sρ
1.088
− 1.79
0.9680
2.30
Sρ−Tη
1.027
+0.08
0.9671
3.38
Appendix
Table 19: Alternative descriptions of controlled honest gain on the same nine directions, all fitted with an intercept. (S−T) is the R0 gap inside the honest mixed panel, whereas Γ uses the matched homogeneous references. Because ρ and η 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.