Counting Moves, Weighing Voices: Bayesian Dialectical Argumentation for Calibrated Multi-LLM Councils under Persistent Adversaries

Ionel Eduard Stan; Paolo Napoletano

cs.AIOct 1, 2026

Counting Moves, Weighing Voices: Bayesian Dialectical Argumentation for Calibrated Multi-LLM Councils under Persistent Adversaries

Authors: Ionel Eduard Stan, Paolo Napoletano

Organizations: Department of Informatics, Systems and Communication (DISCo), University of Milano–Bicocca, Milan, Italy

Abstract

A multi-LLM \emph{council} lets several large language models (LLMs) deliberate on a question and return an answer together with a confidence estimate. As these systems become increasingly used for reasoning, that confidence should represent a calibrated \emph{probability of being correct}, and the decision should remain robust when some agents are persistently unreliable. Existing \emph{council aggregation} methods fail on both fronts: their confidence estimates measure decisiveness rather than correctness, and they cannot identify or discount persistently unreliable agents. We introduce Bayesian Dialectical Argumentation (BDA), which treats the council's \emph{typed} moves---who proposed, challenged, or conceded which answer---as observations of a classical annotator model with \emph{per-agent} reliabilities. This formulation recasts multi-agent deliberation as a reliability estimation problem, using the deliberation trace to infer agent reliability under persistent adversarial behavior. By weighting evidence according to inferred agent reliability, BDA yields calibrated posterior probabilities over candidate answers while allowing persistently unreliable agents to be inverted rather than merely outvoted. Across binary and multi-class benchmarks, BDA achieves the best calibration among zero-cost council aggregation methods, requiring no additional LLM calls, and improves robustness under persistent adversarial coalitions while remaining competitive in clean settings.

Figures & tables

Figure 1: The paper in one picture. Top—the mechanism. Two distrusted seats (red) and one genuine (blue) deliberate on a task with truth

g

. BDA reads typed moves as per-seat tallies

c_{i}{:}d_{i}

, weighs them by learned reliability

w_{i}

, and assigns negative weight to distrusted seats, so rejecting

g

becomes evidence for

g

: a calibrated

\pi(g\mid T)=0.89

where plurality is wrong. Bottom—the payoff. Among zero-cost aggregators, the BDA family occupies the calibration frontier, while stacking, gated vote, DS-EM, and paid baselines trail. Robustness under attack is shown in Figure 2 .

Figure 2: The fair, role-matched attack : the same anti-correct adversary mounted inside each method’s own generation pipeline (

K{=}2,3,4

; full sets, three seeds; error bars are seed standard deviations). Self-MoA, with no internal check, collapses; ArgLLMs’ clean evaluator is a partial defense that erodes as the answer space grows; the council, facing the adversary as two of its three seats , holds near clean. Protocol in Appendix F.

	Binary		PubMedQA		MMLU
Aggregator	$k{=}1$	$k{=}2$	$k{=}1$	$k{=}2$	$k{=}1$	$k{=}2$
Majority	.573	.369	.442	.209	.671	.433
Log. stacking	.814	.865	.722	.753	.813	.743
Gated vote	.751	.661	.596	.570	.767	.683
DS-EM	.664	.178	.564	.282	.774	.001
MoA	.561	.483	.380	.227	.733	.661

Table 1: Accuracy under the live stealth coalition (

k

persistent adversarial seats of three; three seeds). Clean references (majority/BDA per-agent): binary

0.734/0.751

;

K{>}2

in Appendix E. The regime split: the confusion leads binary and PubMedQA, the stacker MMLU (same budget, zero cost). Methods marked † never read the council (Appendix B), so their cell is their clean accuracy (invariance within

1.1

points)—abstention (Figure 2 ).

Aggregator	Acc.	$\mathrm{ECE}_{\mathrm{ew}}$	$\mathrm{ECE}_{\mathrm{em}}$	Brier	$/task
Majority	0.734	0.160	0.154	0.212	0
Logistic stacking	0.751	0.120	0.118	0.194	0
Gated vote	0.726	0.150	0.149	0.210	0
DS-EM (unsup.)	0.659	0.338	0.339	0.338	0
MoA	0.718	—	—	—	.0020
Self-MoA	0.736	—	—	—	.0017

Table 2: Clean binary results (

4{,}500

trials, three seeds; dashes: no informative confidence; every free arm shares the

25

-label budget). Rows follow the grouping of Section 5 , baselines first. ArgLLMs leads accuracy and pays per task; of the free arms BDA leads calibration and beats plurality’s decision (

\Delta

Acc

+0.017

; pooled and question-clustered

95\%

CIs both

[+0.010,+0.024]

Figure 3: The label budget is not the bottleneck. Left: held-out accuracy per supervised arm with

n{=}10

labels (open),

n{=}25

—the operating budget—(filled), and all available labels (tick); a filled dot sitting on its tick is the claim. The one arm that needs its

25

is the MMLU stacker; only the confusion’s PubMedQA accuracy still gains beyond the budget (

+0.05

), its

K{\times}K

rows being the hungriest fit. Right: what the same

25

labels already buy in calibration: on

K{>}2

the confusion is

4\times

better-calibrated than the one-coin (

0.09

vs.

0.36

ECE on PubMedQA). Majority (dotted) uses no labels. Full grid in Appendix E.

Appendix figures & tables24 assets

Supplementary material from the paper’s appendix.

Appendix

Symbol	Meaning	Symbol	Meaning
one-coin	uniform-error likelihood (main Eq. 1); as an arm name, BDA per-agent (priors of main Eq. 3).	$\mathrm{B}(a,b)$	Beta function $\Gamma(a)\Gamma(b)/\Gamma(a{+}b)$ .
confusion	learned full-matrix composite likelihood (main Eq. 4); the arm BDA confusion.	$r_{i},\bar{r}$	fitted reliability of $i$ ; council mean.
$\mathrm{ECE}_{\mathrm{ew/em}}$	ECE, equal-width / equal-mass binning.	$\mu_{i},s$	prior mean and strength; $\alpha_{i}{=}\mu_{i}s$ .
$\theta_{i}\in[0,1]$	latent reliability of agent $i$ .	$P_{i}(g^{\prime}\mid g)$	agent $i$ ’s confusion: prob. a move is consistent with $g^{\prime}$ under truth $g$ (main Eq. 4).
$c_{i}(g),d_{i}(g)$	moves of $i$ consistent / inconsistent with $g$ ; $c_{i}(g^{\prime})$ generalizes to every candidate (main Eq. 4).	$\lambda,\tau$	Dirichlet concentration; likelihood temperature.
$w_{i}$	vote weight $\log\frac{\mu_{i}}{1-\mu_{i}}$ (Prop. 2 ); sign follows $r_{i}$ .	$\theta^{\star}$	true (homogeneous) reliability.

Appendix

Table 3: Notation used throughout the paper and appendix, in order of first use (left column, then right).

Method	Evidence $\mathrm{Gen}_{\Phi}$	Rule $\mathrm{Agg}_{\Phi}$	Reads $T$ ?	$\rho$ corrupts	$\gamma_{\Phi}$
Majority	Propose answers ( $T$ )	plurality count	yes	2 of 3 seats	$2/3$
Logistic stacking	per-seat tallies ( $T$ )	logistic regression	yes	2 of 3 seats	$2/3$
Gated vote	Propose answers ( $T$ )	gated plurality	yes	2 of 3 seats	$2/3$
DS-EM	Propose answers ( $T$ )	EM posterior	yes	2 of 3 seats	$2/3$
MoA / MoA-Lite	proposer surfaces ( $T$ )	synthesizer LLM	yes	2 of 3 seats	$2/3$
Self-MoA	6 samples of one model on $q$	synthesizer LLM	no	sampled model	$1$

Appendix

Table 4: Every method as

\mathrm{Agg}_{\Phi}\circ\mathrm{Gen}_{\Phi}

(Eq. 7 ). Reading the council is dependence on the trace

T

; the fair attack applies the single role

\rho

(Eq. 8 ) to whichever LLM the method’s evidence comes from, leaving its aggregation rule intact. Here

\gamma_{\Phi}

is the fraction of that evidence corrupted: it is

2/3

for the redundant council readers (one clean seat remains) and all of it for the single-pipeline methods.

Arm	PubMedQA	MMLU
ArgLLMs (canonical, clean)	100%	100%
MoA (clean)	99.4%	84.9%
Self-MoA (clean)	99.9%	99.9%
MoA (attacked, $k{=}1/2$ )	100/100%	99.9/100%
MoA-Lite (attacked)	100%	${\geq}99.9$ %

Appendix

Table 5: Parse rates: the fraction of answers committing to a candidate (after the LLM-judge pass on the prose-only residual). Unparseable answers score

0

. Read faithfully (multi-option dispatch), ArgLLMs returns a canonical option label by construction, so its parse rate is

100\%

; this replaces the artifactual

62\%

92\%

of a wrapper that returned proposer paragraphs.

Aggregator	Acc.	$\mathrm{ECE}_{\mathrm{ew}}$	$\mathrm{ECE}_{\mathrm{em}}$	Brier	$/task
Majority	$0.734{\pm}0.005$	$0.160{\pm}0.002$	$0.154{\pm}0.002$	$0.212{\pm}0.002$	0
Verbalized conf.	$0.732$	$0.117$	$0.116$	$0.202$	0
Best member	$0.744$	$0.067$	$0.075$	$0.182$	0
MoA	$0.718{\pm}0.010$	—	—	—	.0020
Self-MoA	$0.736{\pm}0.003$	—	—	—	.0017
ArgLLMs	$0.788{\pm}0.008$	$0.165{\pm}0.007$	$0.167{\pm}0.010$	$0.199{\pm}0.004$	.0015

Appendix

Table 6: Clean binary pool (

4{,}500

trials); cells are mean

{\pm}

standard deviation across the three seeds. Per-agent versus majority:

\Delta

Acc

=+0.0171

95\%

[+0.0104,+0.0240]

p_{\mathrm{win}}{=}1.000

(paired bootstrap). “Verbalized conf.” is a confidence-weighted vote reporting the winner’s mean stated confidence; “best member” is the strongest seat, selected leak-free (Appendix E.3 ).

Figure 4: Reliability diagrams for every confidence-emitting arm, clean binary pool (

4{,}500

trials; gray bars are bin populations; the MoA family emits near-constant scores, not applicable). The BDA posteriors (bottom right pair, the confusion model almost exactly) hug the diagonal; a vote fraction, a strength margin, an unsupervised mode, a gate, and a raw stacking softmax do not.

Figure 5: Equal-width calibration error per (dataset, seed) cell, clean binary family. The BDA rows are uniformly light across all nine cells.

Ablation	Setting	Acc.	$\mathrm{ECE}_{\mathrm{em}}$	Brier
granularity	per-agent	0.722	0.134	0.212
	per-force	0.703	0.266	0.279
	per-(agent, force)	0.703	0.258	0.273
estimator	all moves	0.730	0.080	0.195
	propose only	0.714	0.117	0.210
	round-0 only	0.722	0.150	0.218

Appendix

Table 7: Reliability-key and reliability-estimator ablations (binary pool, three seeds). Granularity uses a fixed prior

(b,\kappa,s)=(0.58,2,16)

; the estimator selects its hyperparameters by inner cross-validation. The per-agent key and the all-moves estimator win.

evidence window	Acc.	$\mathrm{ECE}_{\mathrm{em}}$	decisions changed
$\leq$ round 0	0.748	0.044	—
$\leq$ round 1	0.754	0.035	vs. round 0: $+$
$\leq$ round 2 (full)	0.751	0.044	net $+17/4500$

Appendix

Table 8: Binary deliberation is roughly decision-neutral; its net value rises with

K

(Appendix E.4 ).

Dataset	Arm	Acc.	$\mathrm{ECE}_{\mathrm{ew}}$	$\mathrm{ECE}_{\mathrm{em}}$	Brier
PubMedQA	Majority	0.619	0.264	0.263	0.294
	BDA shared	0.607	0.311	0.304	0.327
	BDA per-agent	0.614	0.304	0.295	0.326
	BDA confusion	0.604	0.072	0.085	0.229
MMLU	Majority	0.770	0.093	0.093	0.150
	BDA shared	0.759	0.190	0.186	0.193

Appendix

Table 9: Clean

K{>}2

: plurality wins accuracy narrowly; the one-coin BDA is over-confident (the concentration of Proposition 3 without heterogeneity to earn it); the tempered confusion model repairs the calibration at unchanged accuracy.

Figure 6: The fix/break ratio of deliberation rounds—decisions the round-

1

-and-later moves repair versus damage—rises with the number of candidate answers

K

	selected		clean $\mathrm{ECE}_{\mathrm{ew}}$
Cell	$\lambda$	$\tau$	one-coin	temp. 1-coin	confusion
Binary	0.3	0.25	0.016	0.012	0.017
PubMedQA	27.0	0.15	0.304	0.051	0.072
MMLU	93.4	0.24	0.182	0.037	0.063

Appendix

Table 10: Left: the confusion model’s CV-selected concentration

\lambda

and temperature

\tau

(mean over folds and seeds);

\lambda

is large on clean

K{>}2

, so the learned matrix is shrunk toward the one-coin. Right: a CV-tempered one-coin already matches the confusion’s clean

K{>}2

calibration—the fix is the temperature. One-coin denotes the BDA per-agent arm.

		one-coin		confusion		log. stacking		gated vote
Dataset	$n$	acc	$\mathrm{ECE}_{\mathrm{ew}}$	acc	$\mathrm{ECE}_{\mathrm{ew}}$	acc	$\mathrm{ECE}_{\mathrm{ew}}$	acc	$\mathrm{ECE}_{\mathrm{ew}}$
Binary	25	0.707	0.064	0.741	0.086	0.712	0.080	0.721	0.163
	100	0.718	0.077	0.736	0.089	0.711	0.125	0.716	0.163
	500	0.722	0.048	0.732	0.038	0.710	0.140	0.716	0.163
PubMedQA	25	0.577	0.355	0.517	0.091	0.580	0.099	0.630	0.256
	100	0.573	0.344	0.560	0.116	0.580	0.139	0.630	0.256

Appendix

Table 11: Label efficiency of the whole supervised family on clean pools (disjoint calibrate/holdout, three seeds; full grid in the results file). Accuracy is flat from tiny

n

for every arm; the confusion’s

K{>}2

calibration advantage is present already at

n{=}25

; the gated vote’s chance-gate saturates immediately; the stacker’s binary

\mathrm{ECE}_{\mathrm{ew}}

drifts up with

n

as the stack overfits its tallies. One-coin denotes the BDA per-agent arm.

persistent, $k{=}1$
	$p_{\mathrm{wrong}}$
Arm	0	.2	.4	.5	.6	.8	1.0
Majority	.734	.705	.671	.657	.640	.611	.581
Shared	.730	.712	.693	.686	.676	.660	.638
Per-agent	.751	.710	.717	.722	.755	.824	.923
Confusion	.758	.732	.718	.719	.723	.748	.792

Appendix

Table 12: Synthetic count-level adversary, accuracy (binary pool, three seeds). Both per-agent parameterizations are non-monotone and then rising under a persistent adversary. Under rotation the centered one-coin falls monotonically while the confusion model’s absolute rows recover at high dose (council-level identifiability); the unlearnable regime is moderate-dose rotation.

	Binary		PubMedQA		MMLU
$\mathrm{ECE}_{\mathrm{em}}$	$k{=}1$	$k{=}2$	$k{=}1$	$k{=}2$	$k{=}1$	$k{=}2$
Majority	.237	.493	.338	.558	.167	.332
Log. stacking	.097	.047	.124	.066	.108	.070
BDA shared	.234	.352	.427	.755	.239	.620
BDA per-agent	.143	.152	.349	.389	.234	.246
BDA confusion	.082	.053	.090	.045	.072	.109

Appendix

Table 13: Calibration under the live attack (council-reading arms). The tempered confusion model is genuinely calibrated under fire; MoA-family confidences remain degenerate (not applicable); ArgLLMs is council-invariant and keeps its clean calibration, poor on

K{>}2

		one-coin			confusion
Dataset	$k$	r0	full	$\Delta$	r0	full	$\Delta$
Binary	$k{=}1$	0.914	0.794	-0.120	0.914	0.812	-0.102
Binary	$k{=}2$	0.933	0.736	-0.197	0.936	0.892	-0.045
PubMedQA	$k{=}1$	0.670	0.608	-0.062	0.873	0.744	-0.129
PubMedQA	$k{=}2$	0.588	0.471	-0.117	0.896	0.840	-0.056
MMLU	$k{=}1$	0.796	0.733	-0.063	0.813	0.752	-0.061

Appendix

Table 14: Accuracy under the live coalition on round-

0

-truncated (r0) versus full traces, leak-free refit per basis, three seeds. Round-

0

one-coin is the classical supervised weighted vote (main Proposition 1). Deliberation rounds reduce attacked accuracy in eleven of twelve cells: they are the contagion channel. One-coin denotes the BDA per-agent arm.

Figure 7: The sleeper contamination axis. Top: the monitoring statistic of Appendix F.13 —total-variation drift of the fitted confusion rows against the clean window—exits its noise floor by

f{\approx}0.2

: the alarm precedes the recovery below. Bottom: per-arm accuracy under the persistent

k{=}2

coalition versus the fraction

f

of the calibration window in which the seat has turned (

f{=}0

is the pure sleeper; all supervised arms start in the below-majority band). The one-coin recovers fastest at low

f

; the confusion needs more of the turn in-window but overtakes from

f{\approx}0.6

and ends highest—the bias–variance trade-off of Appendix C.5 . Three seeds.

Cell	Arm	Acc.	$\mathrm{ECE}_{\mathrm{ew}}$ raw $\to$ iso	Br. iso
clean bin.	Majority	0.734	0.160 $\to$ 0.006	0.184
	BDA per-agent	0.751	0.011 $\to$ 0.005	0.182
	BDA confusion	0.758	0.012 $\to$ 0.013	0.172
clean pub.	Majority	0.619	0.255 $\to$ 0.067	0.228
	BDA per-agent	0.614	0.301 $\to$ 0.030	0.234
	BDA confusion	0.604	0.043 $\to$ 0.028	0.228

Appendix

Table 15: Leak-free isotonic recalibration. On clean data it equalizes calibration (BDA’s edge vanishes); under attack it perfects calibration but cannot rescue a flipped vote’s accuracy, so BDA confusion’s accuracy lead survives. Interim, not a headline.

Init	clean	$k{=}1$	$k{=}2$
majority	0.659	0.668	0.175
spectral	0.659	0.668	0.175
anti (oracle mode)	0.341	0.332	0.825
switch rate	1.00	1.00	1.00

Appendix

Table 16: Unsupervised DS–EM accuracy by initialization (seed

0

; the collapse is structural across seeds). At

k{=}2

the deployable (majority/spectral) modes are the coalition’s mirror; only the label-requiring “anti” mode is correct.

	PubMedQA $k{=}2$		MMLU $k{=}2$
Arm	orig.	permuted	orig.	permuted
Logistic stacking	0.753	0.750	0.743	0.735
BDA per-agent	0.471	0.478	0.699	0.699
BDA confusion	0.840	0.740	0.721	0.716

Appendix

Table 17: Per-task candidate-label permutation, live

k{=}2

(clean cells are invariant for every arm). The label-blind arms do not move; the confusion model loses exactly its label-identity component on the semantic dataset and nothing on the positional one.

	PubMedQA			MMLU
cal $\to$ dep	1-coin	conf.	stack	1-coin	conf.	stack
A $\to$ A	0.600	0.676	0.682	0.707	0.693	0.713
B $\to$ B	0.487	0.842	0.722	0.695	0.712	0.731
A $\to$ B (shift)	0.531	0.553	0.602	0.584	0.532	0.632
B $\to$ A (shift)	0.607	0.578	0.571	0.708	0.528	0.319

Appendix

Table 18: Wrong-answer policy shift at

k{=}2

(A = synthetic random-wrong, B = live best-defensible; disjoint calibrate/deploy rows, three seeds). Across policies both learned-structure arms (confusion, stacker) lose their edge—the stacker catastrophically on MMLU B

\to

A—while the one-coin (the BDA per-agent arm) is policy-robust in every shifted cell.

	genuine flip		effective vs nominal
Cell	$k{=}1$	$k{=}2$	$k{=}1$	$k{=}2$
Binary	0.109	0.239	0.48/0.33	0.68/0.67
PubMedQA	0.186	0.353	0.59/0.33	0.72/0.67
MMLU	0.029	0.122	0.44/0.33	0.64/0.67

Appendix

Table 19: Contagion: genuine-seat flip rate to the coalition’s answer, and the resulting effective corrupted fraction against the nominal

k/I

seat	$r_{i}$	$\mu_{i}$	$w_{i}$	$(c,d)$	$\Delta\ell_{i}(\texttt{True})$
$A_{1}$ (adv)	0.21	0.37	$-0.55$	$(0,3)$	$+1.18$
$A_{2}$ (adv)	0.29	0.44	$-0.24$	$(0,1)$	$+0.24$
$A_{3}$ (gen)	0.60	0.75	$+1.12$	$(2,1)$	$+0.71$
total log-odds					$+2.12$

Appendix

Table 20: The worked example. The two distrusted seats argue False , but their negative weights turn those moves into evidence for True —more from the more distrusted

A_{1}

. Even though the genuine seat flipped once (its

(2,1)

leaves only a weak

+0.71

), the two inverted adversaries carry

+1.42

of the

+2.12

total, which gives

\pi(\texttt{True}\mid T)=0.89

—a calibrated confidence, not a saturated vote.

Dataset	Arm	Acc.	$\mathrm{ECE}_{\mathrm{ew}}$	$\mathrm{ECE}_{\mathrm{em}}$	Brier
Binary	Majority	0.692	0.151	0.151	0.223
	BDA shared	0.684	0.024	0.041	0.202
	BDA per-agent	0.701	0.022	0.029	0.197
	BDA confusion	0.697	0.031	0.039	0.199
PubMedQA	Majority	0.574	0.285	0.281	0.317
	BDA shared	0.558	0.315	0.310	0.341

Appendix

Table 21: Small council, clean. On binary, per-agent wins everything and the calibration-error reduction over voting is about sevenfold (per-agent versus majority

\Delta

Acc

+0.009

p_{\mathrm{win}}{=}0.96

). On

K{>}2

, with near-chance, near-homogeneous reliabilities there is no per-seat accuracy signal and plurality leads the decision—the heterogeneity assumption as a real boundary—while the tempered confusion model still repairs the calibration (

\mathrm{ECE}_{\mathrm{ew}}

0.033

0.050

against plurality’s

0.285

0.157

), the temperature needing no heterogeneity.

Ingredient	Origin	Our role
Typed moves / speech acts, dialogue games	( Searle 1969 ; Hamblin 1970 ; Walton and Krabbe 1995 ; Prakken 2005 ; Prakken 2006 )	adopted as data
QBAF / DF-QuAD	( Dung 1995 ; Cayrol and Lagasquie-Schiex 2009 ; Baroni et al. 2019 ; Rago and others 2016 )	baseline (ArgLLMs)
Weighted majority / jury theorems	( Nitzan and Paroush 1982 ; Grofman et al. 1983 ; Berend and Paroush 1998 ; Cooke 1991 )	precedent; recovered as our MAP rule
Annotator / label-aggregation models (one-coin, confusion, unsupervised EM)	( Dawid and Skene 1979 ; Raykar and others 2010 ; Raykar and Yu 2012 ; Whitehill et al. 2009 ; Zhang and others 2016 ; Hovy and others 2013 ; Ratner and others 2020 )	precedent for BDA’s likelihood
Beta–Bernoulli conjugacy	textbook (e.g., Berger 1990 )	building block
Calibration / ECE / Brier / Murphy	( Dawid 1982 ; Guo and others 2017 ; Brier 1950 ; Murphy 1973 )	evaluation target

Appendix

Table 22: Provenance: what we adopt from prior work versus what is new here.

Explore similar work

May 12, 2026cs.AI

CHAL: Council of Hierarchical Agentic Language

Multi-agent debate has emerged as a promising approach for improving LLM reasoning on ground-truth tasks, yet current methodologies face certain structural limitations: debate tends to induce a martingale over belief trajectories, majority voting accounts for most observed gains, and LLMs exhibit confidence escalation rather than calibration across rounds. We argue that the genuine value of debate, and dialectic systems as a whole, lies not in ground-truth tasks but in defeasible domains, where every position can in principle be defeated by better reasoning. We present the Council of Hierarchical Agentic Language (CHAL), a multi-agent dialectic framework that treats defeasible argumentation as an engine for belief optimization. Each agent maintains a CHAL Belief Schema (CBS), a graph-structured belief representation with a Bayesian-inspired architecture, that facilitates belief revision through a gradient-informed dynamic mechanism by leveraging the strength of the belief's thesis as a differentiable objective. Meta-cognitive value systems spanning epistemology, logic, and ethics are elevated to configurable hyperparameters governing agent reasoning and adjudication outcomes. We provide a series of ablation experiments that demonstrate systematic and interpretable effects: the adjudicator's value system determines the debate's overall trajectories in latent belief space, council diversity refines beliefs for all participants, and the framework generalizes across broad fields. CHAL is, to our knowledge, the first framework to treat multi-agent debate as structured belief optimization over defeasible domains. Further, the auditable belief artifacts it produces establish the foundation for dedicated evaluation suites for defeasible argumentation, with broader implications for building AI systems whose reasoning and value commitments are transparent, aligned, and subject to human oversight.

Tommaso Giovannelli, Griffin D. Kent

Multi-Agent Debate Argumentation Frameworks Agentic Framework LLM Reasoning Strategies Metacognition

Ask PDF

Aug 4, 2026cs.MA

Group Perspective Matters: Regulating Debate Relationships Can Mitigate Blind Conformity in Multi-Agent Debate

Multi-Agent Debate (MAD) improves the reasoning performance of Large Language Models (LLMs) through multi-round interaction. However, LLMs in MAD are highly susceptible to blind conformity. Existing individual evaluation methods, typically based on confidence or perplexity, fail to reflect the correctness of reasoning and may even exacerbate blind conformity. To address this, we shift the perspective from individual evaluation to group interaction. We define mutual referencing among LLMs as \textbf{Debate Relationships} and recognize that regulating these relationships is the key to mitigating blind conformity. In this paper, we propose a novel framework for \textbf{D}ynamically r\textbf{E}gulating deb\textbf{A}te \textbf{R}elationships (DEAR) from the group perspective. At first, DEAR quantifies consensus and divergence as \textit{group evidence} to capture the debate state. Then, DEAR operates through three stages: 1) What: perceiving group consultation tendency and uncertainty; 2) Who: introducing a Selection RL-Agent to dynamically select reference peers; and 3) How: adopting a Behavior RL-Agent to adaptively adjust generation behaviors. Notably, we formulate the execution of the two RL-Agents as a sequential decision-making process, jointly optimizing via multi-agent reinforcement learning. Extensive experiments demonstrate that DEAR achieves superior performance while significantly reducing token consumption.

Hao Wu, Shoucheng Song, Chang Yao +4

School of Computer Science & Technology, Beijing Jiaotong University, Beijing, China · Beijing Key Laboratory of Traffic Data Mining and Embodied Intelligence, Beijing, China

Multi-Agent Debate Debate Conformity

Ask PDF

Sep 3, 2026cs.CL

Remember and Reweight: Enhancing Multi-Agent Debate with Experience Memory and Confidence Estimation

Multi-agent debate (MAD) improves the reasoning capabilities of large language models by having multiple agents iteratively refine their responses through discussion. However, MAD suffers from a critical vulnerability known as shared misconception: when a majority of agents initially converge on an incorrect answer, the debate process tends to amplify rather than correct the error. Existing methods primarily address peer skew but leave the agents' inherently biased concept priors unaddressed. To mitigate this systematic weakness, we propose R

^2

-MAD (Remember and Reweight for Multi-Agent Debate), a framework that equips agents with an experience memory accumulated from past debates. R

^2

-MAD intervenes on both failure modes through two complementary mechanisms: A debate-state-aware retrieval policy dynamically calibrates the concept prior by retrieving relevant historical evidence based on the current consensus level. Then these retrieved experiences provide a basis for estimating per-agent reliability, yielding confidence weights to modulate peer influence. Experiments on various benchmarks show that R

^2

-MAD achieves consistent improvements over existing single-agent and MAD baselines.

Xuanfa Jin, Zhijian Ma, Yongcheng Zeng +3

Institute of Automation, Chinese Academy of Sciences · University of Chinese Academy of Sciences · University College London

Multi-Agent Debate Reweighting Confidence Estimation Prior Knowledge Integration

Ask PDF