Learning Adaptive Deliberation Control for Reliable Recognition with Multi-Agent Language Models

Yuan Liu; Haoyue Peng; Fanliang Hu; Lei Huang; Wenli Cai

cs.AIOct 13, 2026

Learning Adaptive Deliberation Control for Reliable Recognition with Multi-Agent Language Models

Authors: Yuan Liu, Haoyue Peng, Fanliang Hu, Lei Huang, Wenli Cai

Organizations: Medical Artificial Intelligence Laboratory, Chongqing University Cancer Hospital, Chongqing University, Chongqing, 400030, China

Abstract

Multi-agent deliberation can improve recognition by criticizing and revising candidate answers, but fixed controls cannot respond to changes in disagreement or reasoning quality. We present an adaptive multi-agent controller (ADMA) that learns continuous policies for proposal, challenge, arbitration, and coordination while keeping the task model frozen. Shared trace features describe quality, agreement, and progress, while role-specific features support decentralized control during inference. One controller trained on legal and financial reasoning is transferred without policy updates to four textual and multimodal benchmarks and two successive task-model versions. With Qwen3.5 held fixed, ADMA-Full improves accuracy by 6.2-14.2 percentage points over Direct and by 2.8-3.0 points over ADMA-Fixed. Paired bootstrap ranges for the Full-Fixed comparison are positive on three tasks. Outcome decompositions show that recoveries exceed introduced errors, while process assessment records fewer false-consensus events and more supported correct outcomes. Transfer to Qwen3.6 remains task dependent, which bounds the generality of the learned policy. Median token and latency ratios are 14.3-18.7 and 7.8-16.5 relative to Direct. These results show that adaptive deliberation can improve fixed-model recognition when its task-dependent gains justify the additional computation.

Figures & tables

Figure 1: ADMA controller learning and adaptive inference. During centralized training, joint transitions from four frozen role models are stored in replay and evaluated by role-specific critics. During decentralized inference, the learned actors adapt role behavior and collaboration from the evolving trace until termination. The final prediction is evaluated separately for accuracy, reliability, and resource use.

Figure 2: Correction of MMStar item 33 in an archived Gemini 3.1 run. a , Direct selects D from foreground landscaping cues. b , the coordinator redirects verification toward the mural and historical context; four rounds of challenge and arbitration yield A while the final answering roles remain divided as B/A/A. This qualitative trace is separate from the aggregate Qwen3 evaluation.

Task model	Dataset	Protocol	Accuracy (%)	FCR (%)	ESR (%)
Qwen3.6	LawQA	Direct	77.2 $\pm$ 0.6	–	–
	LawQA	ADMA-Full	74.8 $\pm$ 1.1	–	–
	FinQA	Direct	57.4 $\pm$ 1.4	80.1 $\pm$ 11.3	96.2 $\pm$ 0.6
	FinQA	ADMA-Full	71.8 $\pm$ 0.8	63.9 $\pm$ 5.9	98.3 $\pm$ 0.8
	MMBench	Direct	74.0 $\pm$ 1.1	83.9 $\pm$ 2.6	95.7 $\pm$ 0.7
	MMBench	ADMA-Full	76.2 $\pm$ 0.3	42.9 $\pm$ 8.8	95.5 $\pm$ 0.4

Table 1: Recognition performance across model versions and inference protocols.

Figure 3: Interaction gains relative to a successive model-version reference. a , the signed Direct version gap, within-Qwen3.5 ADMA gain, and resulting margin over Qwen3.6 Direct are shown from a common zero baseline. b , paired recovery among Qwen3.6 Direct errors and paired loss among its correct predictions, with counts and conditional denominators. c , repeat-specific final margins and 95% question-cluster bootstrap ranges for the fixed-panel means. Qwen3.5 and Qwen3.6 denote Qwen3.5-35B-A3B and Qwen3.6-35B-A3B.

Figure 4: Paired correctness composition within Qwen3.5. a , eight joint patterns across Direct, ADMA-Fixed, and ADMA-Full for 2,000 task–question–repeat observations. b , ADMA-Full versus Direct on FinQA, partitioned separately by operation, information source, and reasoning depth. c , the same comparison by visual benchmark category. Green, red, and gray denote gained, lost, and unchanged correctness. Subgroup percentages pool both repeats; parentheses give base-question counts.

Figure 5: Reliability and deliberation depth within Qwen3.5. a – b , paired false-consensus and supported-correct incidence among all 500 observations per task, separated into shared and method-exclusive outcomes. c , ADMA-Full round distributions within paired correctness groups defined against Direct. Gray denotes two rounds and gold at least three; labels report the latter’s percentage, count, and group denominator. MMBench uses the maximum across permutations, and failure-affected traces are excluded.

Figure 6: Resource demand of ADMA-Full. a – b , within-version token and latency ratios to Direct over failure-free pairs. Boxes span the 25th–75th percentiles, whiskers the 10th–90th percentiles, central lines mark medians, and dashed lines denote parity. c , total Qwen3.5 resource increments over the stated baseline divided by the net additional correct count among 500 paired observations. MMBench resources are summed over permutations, and latency is the sum of elapsed inference times.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Component	Setting
Evaluation task models	Qwen3.5-35B-A3B and Qwen3.6-35B-A3B
Frozen training task model	Gemini-3.1-Flash-Lite-Preview
Role inventory	Proposer, challenger, arbiter, coordinator
Observation per role	22 features, comprising 16 shared and six role-specific features
Action dimensions	4, 4, 4, and 5, respectively, in the role order above
Deliberation horizon	At most six rounds; Full requires at least two, Fixed permits one

Appendix

Table S1: Controller and evaluation settings used in the manuscript.

Task	Protocol	Repeat	Correct	Accuracy (%)	FCR (%)	ESR (%)
LawQA	Direct	1	145	58.0	96.2	43.4
LawQA	Direct	2	150	60.0	97.0	45.3
LawQA	ADMA-Fixed	1	173	69.2	84.4	53.2
LawQA	ADMA-Fixed	2	178	71.2	88.9	55.1
LawQA	ADMA-Full	1	182	72.8	73.5	60.4
LawQA	ADMA-Full	2	184	73.6	80.3	62.5

Appendix

Table S2: Repeat-level performance with Qwen3.5-35B-A3B.

Task	Protocol	Repeat	Correct	Accuracy (%)	FCR (%)	ESR (%)
LawQA	Direct	1	194	77.6	–	–
LawQA	Direct	2	192	76.8	–	–
LawQA	ADMA-Full	1	189	75.6	–	–
LawQA	ADMA-Full	2	185	74.0	–	–
FinQA	Direct	1	146	58.4	72.1	96.6
FinQA	Direct	2	141	56.4	88.1	95.7

Appendix

Table S3: Repeat-level performance with Qwen3.6-35B-A3B.

Task	Protocol	Repeat 1	Repeat 2
LawQA	Direct	1	1
LawQA	ADMA-Full	1	1
FinQA	Direct	0	0
FinQA	ADMA-Full	0	0
MMBench	Direct	4	4
MMBench	ADMA-Full	5	4

Appendix

Table S4: Incomplete predictions with Qwen3.6-35B-A3B.

Version	Task	Baseline	Repeat 1	Repeat 2	Mean	95% range
Qwen3.5	LawQA	Direct	+14.8	+13.6	+14.2	[10.0, 18.6]
Qwen3.5	FinQA	Direct	+14.4	+13.2	+13.8	[8.4, 19.2]
Qwen3.5	MMBench	Direct	+8.8	+8.0	+8.4	[3.6, 13.4]
Qwen3.5	MMStar	Direct	+6.0	+6.4	+6.2	[1.2, 11.2]
Qwen3.5	LawQA	Fixed	+3.6	+2.4	+3.0	[0.4, 5.8]
Qwen3.5	FinQA	Fixed	+2.8	+3.2	+3.0	[0.2, 6.0]

Appendix

Table S5: Within-version accuracy differences for ADMA-Full minus the indicated baseline.

Task	Repeat 1	Repeat 2	Mean	95% range
LawQA	-4.8	-3.2	-4.0	[-8.0, -0.2]
FinQA	+12.0	+14.8	+13.4	[8.2, 18.6]
MMBench	+0.4	-0.4	0.0	[-4.6, 5.0]
MMStar	+1.2	+7.2	+4.2	[-1.6, 10.0]

Appendix

Table S6: Cross-version accuracy differences for Qwen3.5 ADMA-Full minus Qwen3.6 Direct.

Task	Both correct	Recovered	Lost	Both incorrect	Net gain
LawQA	344	22	42	92	-20
FinQA	262	92	25	121	67
MMBench	320	50	50	80	0
MMStar	177	91	70	162	21

Appendix

Table S7: Paired correctness for Qwen3.5 ADMA-Full relative to Qwen3.6 Direct.

Task	Win twice	Win once	Reversal	Lose once	Lose twice	No change
LawQA	4	14	0	12	15	205
FinQA	31	29	1	10	7	172
MMBench	15	17	3	23	12	180
MMStar	21	39	10	28	16	136

Appendix

Table S8: Repeat consistency of the cross-version paired difference.

Direct	Fixed	Full	LawQA	FinQA	MMBench	MMStar
Incorrect	Incorrect	Incorrect	120	111	101	167
Incorrect	Incorrect	Correct	24	21	14	39
Incorrect	Correct	Incorrect	10	9	4	28
Incorrect	Correct	Correct	51	74	53	29
Correct	Incorrect	Incorrect	4	21	20	23
Correct	Incorrect	Correct	1	8	10	17

Appendix

Table S9: Joint correctness across the three Qwen3.5 protocols.

Partition	Group	$n$	Gained	Lost	Tied	Gain (pp)
Operation	Addition	37	11	7	56	+5.4
Operation	Division	72	17	8	119	+6.2
Operation	Other operations	26	8	2	42	+11.5
Operation	Subtraction	115	59	9	162	+21.7
Source	Mixed	37	13	1	60	+16.2
Source	Table	151	61	18	223	+14.2

Appendix

Table S10: Qwen3.5 ADMA-Full versus Direct by FinQA stratum.

Task	Group	$n$	Gained	Lost	Tied	Gain (pp)
MMBench	attribute reasoning	43	6	6	74	0.0
MMBench	coarse perception	63	16	5	105	+8.7
MMBench	Fine-grained perception (cross)	31	8	4	50	+6.5
MMBench	Fine-grained perception (instance)	63	20	3	103	+13.5
MMBench	logic reasoning	25	8	4	38	+8.0
MMBench	relation reasoning	25	9	3	38	+12.0

Appendix

Table S11: Qwen3.5 ADMA-Full versus Direct by visual benchmark category.

Task	Event	Neither	Direct only	Full only	Both	Net
LawQA	False consensus	297	100	5	98	-95
LawQA	Supported correct	264	11	105	120	94
FinQA	False consensus	283	148	12	57	-136
FinQA	Supported correct	197	10	121	172	111
MMBench	False consensus	331	96	14	59	-82
MMBench	Supported correct	166	19	116	199	97

Appendix

Table S12: Paired process-assessment events under Qwen3.5.

Task	Outcome	$N$	Two rounds	$\geq 3$ rounds	$\geq 3$ (%)
LawQA	Both correct	291	290	1	0.3
LawQA	Recovered	75	74	1	1.3
LawQA	Lost	4	4	0	0.0
LawQA	Both incorrect	130	130	0	0.0
FinQA	Both correct	259	258	1	0.4
FinQA	Recovered	95	94	1	1.1

Appendix

Table S13: ADMA-Full depth within Qwen3.5 correctness groups defined against Direct.

Version	Task	Resource	$N$	$P_{10}$	$P_{25}$	$P_{50}$	$P_{75}$	$P_{90}$
Qwen3.5	LawQA	Tokens	500	16.8	17.7	18.7	19.6	20.5
Qwen3.5	LawQA	Latency	500	9.7	12.5	15.4	18.3	20.9
Qwen3.5	FinQA	Tokens	500	12.0	13.4	15.0	17.0	19.6
Qwen3.5	FinQA	Latency	500	6.2	8.6	12.1	15.1	18.8
Qwen3.5	MMBench	Tokens	500	14.2	14.9	15.8	17.5	20.0
Qwen3.5	MMBench	Latency	500	4.5	5.9	8.0	11.5	14.3

Appendix

Table S14: Within-version ADMA-Full-to-Direct resource ratios.

Task	Paired observations	Qwen3.6 Direct (%)	Qwen3.5 Full (%)	Margin (pp)
LawQA	498	77.5	73.1	-4.4
FinQA	500	57.4	70.8	+13.4
MMBench	492	75.2	74.4	-0.8
MMStar	494	50.0	54.3	+4.3

Appendix

Table S15: Cross-version comparison restricted to matched complete predictions.

Task	Baseline	$\Delta C$	$\Delta T$ (M)	$\Delta L$ (ks)	Tokens/ $\Delta C$ (k)	Seconds/ $\Delta C$
LawQA	Direct	71	8.059	13.15	113.5	185.2
LawQA	Fixed	15	3.476	5.21	231.7	347.6
FinQA	Direct	69	9.616	11.43	139.4	165.6
FinQA	Fixed	15	4.371	5.12	291.4	341.4
MMBench	Direct	42	10.976	36.72	261.3	874.2
MMBench	Fixed	15	3.062	12.34	204.1	822.5

Appendix

Table S16: Qwen3.5 resource increments per net additional correct outcome.

Multi-Agent Deliberation

See all 34 papers

JudgePanel: A Compact Judge with Panel Deliberation via Adaptive Multi-Reward Reinforcement Learning
Aug 29, 2026Yiyue Qian, Shinan Zhang, Huan Song +1LLM-as-a-Judge AI Agent Evaluation
AWS Generative AI Innovation Center, USA
Meta-Moderator: Empowering Multi-Agent Debate with Meta-Cognition
Aug 24, 2026Wentao Hu, Zhuoyue Wan, Jinhao Shen +3Multi-Agent Debate Multi-Agent Deliberation
The Hong Kong Polytechnic University · Sichuan University
ECMR: Evidence-Aware Deliberation Control for Multi-Agent LLM Reasoning
Oct 13, 2026Yuan Liu, Haoyue Peng, Fanliang Hu +2Multi-Agent Reasoning LLM Agent Reliability
Medical Artificial Intelligence Laboratory, Chongqing University Cancer Hospital, Chongqing University, Chongqing 400030, China

Aug 29, 2026J/K move · Enter open · S save

JudgePanel: A Compact Judge with Panel Deliberation via Adaptive Multi-Reward Reinforcement Learning

Yiyue Qian, Shinan Zhang, Huan Song, Hannah Marlowe

AWS Generative AI Innovation Center, USA

LLM-as-a-Judge AI Agent Evaluation Bias in LLM-as-a-Judge Multi-Objective Reinforcement Learning

Open paper AskPDF arXiv

The LLM-as-a-Judge paradigm has emerged as a scalable alternative to human evaluation. However, single-model judges are limited by their inherent model biases, while multi-agent evaluation protocols that mitigate this through diverse deliberation are prohibitively expensive at inference time. To this end, we propose \modelname, which equips a compact Judge model with multi-agent Panel deliberation capability. Specifically, we first train on panel deliberation traces from an ensemble of strong evaluators, capturing structured patterns of discussion, disagreement, and resolution. To further improve judgment quality beyond SFT, we introduce AdaReward, an adaptive multi-reward RL algorithm that dynamically rebalances reward component weights as different objectives saturate at different rates during RL training. For practical deployment, we further design a lightweight domain specialization module for rapid adaptation to new evaluation domains with few hundred labeled samples. As a result, (i) Novel: the first framework to equip a single compact judge with multi-agent panel deliberation capability at single-model inference cost; (ii) Effective & Reliable: JudgePanel with a 14B backbone outperforms judge-specialized models up to 70B across four evaluation benchmarks, demonstrates strong position consistency, and rapidly specializes to new domains with few hundred samples.

Abstract

Multi-Agent Deliberation

Explore similar work

JudgePanel: A Compact Judge with Panel Deliberation via Adaptive Multi-Reward Reinforcement Learning

Meta-Moderator: Empowering Multi-Agent Debate with Meta-Cognition

ECMR: Evidence-Aware Deliberation Control for Multi-Agent LLM Reasoning