Beyond Perfect Teammates: Multi-Agent Collaboration with Unreliable Colleagues

Chen Huang; Yiyang Zhang; Yuhao Wu; Yang Deng; Wenxuan Zhang

cs.AIOct 12, 2026

Beyond Perfect Teammates: Multi-Agent Collaboration with Unreliable Colleagues

Authors: Chen Huang, Yiyang Zhang, Yuhao Wu, Yang Deng, Wenxuan Zhang

Abstract

LLM-based multi-agent collaboration is increasingly used to solve complex tasks by dividing work among specialized agents and integrating complementary expertise. Yet existing systems and evaluations largely rest on an implicit perfect-colleague assumption: collaborators provide correct and complete information and respond promptly to requests. In realistic teamwork, however, collaborators may return information that does not answer the intended question, provide insufficient detail to support a decision, or fail to respond to a request. To separate task-solving competence from the ability to recover an unreliable evidence-gathering process, we introduce the Unreliable Colleagues Benchmark (UnCoBench), which organizes behavioral failures through Misaligned, Incomplete, and Missed settings. The main agent need to verify contextual correctness, diagnose information gap and track unresolved dependencies in order to obtain the correct evidence for solving the task. We also introduce a verifiable and scalable automated generation pipeline that pairs LLM-based scenario generation with executable checks of answer uniqueness and role necessity. Results across ten models show that similar reliable-setting performance can result in different robustness to unreliable evidence. Trace analysis reveals frequent omissions of corrective follow-up, and the recovery behavior involves inefficient waiting and redundant queries. These findings show that robust collaboration requires not only sound reasoning, but also targeted action to diagnose evidence failures, recover relevant information, and integrate it into decisions. Our data and source code are available at https://github.com/iNLP-Lab/UnCoBench.

Figures & tables

Figure 1: Behavior difference between perfect and unreliable colleagues in multi-agent collaboration.

Figure 2: Construction pipeline for UnCoBench . The first three steps produce structured specifications, logically verified instances, and validated natural-language tasks, establishing the Reliable baseline. The final step then derives three paired unreliable settings.

	Reliable		Misaligned		Incomplete		Missed
Main agent	Acc. (%) $\uparrow$	Rounds $\downarrow$	Acc. (%) $\uparrow$	Rounds $\downarrow$	Acc. (%) $\uparrow$	Rounds $\downarrow$	Acc. (%) $\uparrow$	Rounds $\downarrow$
GPT-5.6 Sol $\dagger$	98.7	2.18	45.8 (-52.9)	3.06 (+0.89)	77.0 (-21.7)	2.93 (+0.76)	97.4 (-1.3)	4.97 (+2.80)
Claude Opus 5 $\dagger$	96.7	2.18	90.7 (-6.0)	3.99 (+1.81)	95.7 (-1.0)	3.11 (+0.94)	95.8 (-0.9)	3.24 (+1.06)
Gemini 3.8 Flash $\dagger$	99.3	2.09	74.1 (-25.2)	3.58 (+1.49)	88.5 (-10.8)	2.91 (+0.82)	99.1 (-0.2)	3.09 (+1.00)
DeepSeek-V4.1-Flash $\ddagger$	92.6	2.10	14.5 (-78.1)	2.50 (+0.40)	73.8 (-18.8)	2.71 (+0.62)	89.3 (-3.3)	3.16 (+1.07)
GLM-5.3-Flash $\ddagger$	97.4	2.30	35.1 (-62.3)	3.19 (+0.89)	87.8 (-9.6)	3.05 (+0.75)	96.2 (-1.2)	3.61 (+1.31)

Table 1: Main results on UnCoBench .

\dagger

and

\ddagger

denote closed-source and open-weight models respectively. Colored deltas indicate changes from the Reliable setting: red for accuracy, measured in percentage points, and blue for mean interaction rounds.

Figure 3: Reliable to unreliable joint correctness transitions, separated by unreliable setting. Each row partitions all task pairs into four observed outcome transitions.

Figure 4: Behavioral pathways under the Misaligned and Incomplete settings. Each category describes how models respond across tasks and does not directly indicate changes in accuracy.

Figure 5: Behavioral decomposition of extra rounds in Missed relative to Reliable . Left: contribution of each observable behavior to the mean round difference. Right: distribution of the number of later contacts with the designated missed role.

Figure 6: Accuracy and mean rounds across communication settings for the six models shown under Mixed . Error bars denote Wilson 95% confidence intervals for accuracy.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Reference scenario	Description and decision suggestion
Evacuation and shelter	Coordinate evacuation and accommodation with distributed risk and resource information. Decision: Choose evacuation destinations, shelters or safe accommodation plans.
Transport and emergency rescue	Coordinate capacity, access, receiving requirements and timing for transport or rescue. Decision: Choose routes, destinations, rescue plans or supply delivery plans.
Events and venue selection	Combine participant needs, venue conditions and organizational resources. Decision: Choose venues, event arrangements or feasible relocation plans.
Research and field operations	Combine research needs, facility capabilities and environmental information. Decision: Choose field camps, research facilities, equipment locations or work plans.
Asset protection and disaster recovery	Coordinate preservation, transfer and recovery of physical or digital assets. Decision: Choose storage sites, transfer plans, backup facilities or recovery arrangements.
Investigation, localization and cause identification	Combine distributed records and observations to check explanations, targets or investigation plans. Decision: Evaluate explanations, locate targets or select investigation arrangements.

Appendix

Table 2: Scenario references supplied to the task generator.

Operation	Calculation	Constraints
Sum	$\sum_{i=1}^{m}x_{i}$	Operands and result share a unit.
Difference	$x_{1}-x_{2}$	Exactly two operands, with the same unit.
Average	$\frac{1}{m}\sum_{i=1}^{m}x_{i}$	Equally weighted observations with the same unit and basis.
Ratio	$x_{1}/x_{2}$	Exactly two operands; the denominator domain excludes zero. No multiplication by 100.
Product	$\prod_{i=1}^{m}x_{i}$	Result dimensions equal the product of operand dimensions.
Weighted sum	$\sum_{i=1}^{m}w_{i}x_{i}$	Positive dimensionless weights; no normalization by their sum.

Appendix

Table 3: Numerical operation references.

Figure 7: Distribution of the 1,000 task groups across the 14 application domains.

Operation	Form	Groups	Share (%)
Sum	$x_{1}+x_{2}$	131	13.1
Difference	$x_{1}-x_{2}$	137	13.7
Arithmetic mean	$\sum_{i=1}^{n}x_{i}/n$	122	12.2
Ratio	$x_{1}/x_{2}$	113	11.3
Product	$x_{1}x_{2}$	125	12.5
Weighted sum	$\sum_{i}w_{i}x_{i}$	138	13.8

Appendix

Table 4: Numerical-rule diversity across the 1,000 task groups.

Validation stage	Passed / Base	Rejected	Pass rate (%)
Specification validation	1,289 / 1,296	7	99.46
Logical compilation	1,282 / 1,289	7	99.46
Full-information acceptance	1,000 / 1,282	282	78.00
Overall	1,000 / 1,296	296	77.16

Appendix

Table 5: Candidate filtering in the automated construction pipeline.

Validation stage	Automated accepted	Automated rejected	Human accepted	Human rejected	Agreement (%)
Specification validation	98	2	100	0	98
Logical compilation	98	2	98	2	100
Full-information acceptance	85	15	92	8	93

Appendix

Table 6: Comparison of automated and human validation on 100 candidates per stage.

Main agent	DS-V4-Flash-0731		GPT-5.6 Sol		Qwen3.7-Flash
Main agent	Fidelity (%)	Coverage (%)	Fidelity (%)	Coverage (%)	Fidelity (%)	Coverage (%)
GPT-5.6 Sol	99.50	97.33	100.00	99.62	99.00	98.38
Claude Opus 5	99.75	98.66	100.00	99.75	98.50	97.67
Gemini 3.8 Flash	99.50	98.40	99.75	99.52	98.75	97.25
DeepSeek-V4.1-Flash	99.25	98.73	100.00	99.60	99.00	97.33
GLM-5.3-Flash	98.50	95.80	99.50	98.93	98.25	95.05

Appendix

Table 7: Role agent comparison across main agent models.

Arm	Information supplied	Capabilities still required
None	No additional information is supplied.	Recognize the need for recovery; localize and diagnose the problem; formulate and execute a repair; integrate the recovered evidence.
Existence	One prior collaborator interaction is identified as requiring recovery, without specifying which interaction or why.	Localize and diagnose the problem; formulate and execute a repair; integrate the recovered evidence.
Localization	Interaction $i$ with Role $R$ is identified as requiring recovery, without specifying the defect.	Diagnose the problem; formulate and execute a repair; integrate the recovered evidence.
Diagnosis	The concrete communication problem $P$ in the localized interaction is specified.	Translate the diagnosis into a corrective action; execute the repair; revise the reasoning; integrate the recovered evidence.
Action	Problem $P$ is specified and the model is instructed to perform corrective action $X$ .	Execute the instruction; process new evidence; revise the reasoning; produce the final answer.

Appendix

Table 8: Nested intervention arms for diagnosing communication-recovery bottlenecks. Role

R

, interaction

i

, problem

P

, and corrective action

X

are instantiated from the corresponding task and trajectory.

Figure 8: Joint correctness under progressively more informative interventions across the three unreliable settings.

Explore similar work

Oct 13, 2026

Training LLMs for Multi-Agent Collaboration with Verifiable Information Rewards

Effective collaboration among large language model (LLM) agents in multi-agent systems requires sharing relevant evidence, requesting missing information, and incorporating teammates' contributions into decisions. Training these behaviors is challenging because final task outcomes provide limited feedback on the value of individual messages. We introduce CoSIR (Collaboration via Surprisal-based Information Rewards), a post-training approach that uses a simple, verifiable environment to teach collaboration behaviors that generalize to more complex, unseen tasks. We propose a Bayesian evaluator that scores shared information by its reduction in team surprisal, defined as the sum of the negative log probabilities assigned to the correct option given each agent's available evidence. Requests are scored by the potential reduction in the same quantity if the requested information were truthfully provided. Separate rewards encourage agents to use the received information in their private choices and calculations. We post-train Qwen3-4B, Qwen3-8B, and Gemma-3-4B-it through self-play and evaluate transfer to seven unseen natural-language collaboration tasks. CoSIR achieves the highest aggregate score, while its performance in general tasks remains competitive. Behavioral analyses show increased disclosure of private evidence and fewer judged failures to use teammates' input.

Keru Chen, Sen Lin, Yingbin Liang +2

School of ECEE, Arizona State University · Computer Science Department, University of Houston · Department of ECE, The Ohio State University +1

LLM Agent Training Multi-Agent Collaboration Multi-LLM Collaboration Reinforcement Learning with Verifiable Rewards Reward Shaping

Ask PDF

May 9, 2026

AgentCollabBench: Diagnosing When Good Agents Make Bad Collaborators

Multi-agent systems achieve state-of-the-art outcomes through peer collaboration. However, when an agent in the pipeline silently drops a constraint, the system's final output may look correct even though the reasoning chain was quietly corrupted, and existing outcome-based evaluations are blind to such multi-hop process failures. To make these vulnerabilities measurable before deployment, we introduce AgentCollabBench, a diagnostic benchmark of 900 human-validated tasks spanning software engineering, DevOps, and data engineering. Each task isolates one of four behavioral risks: instruction decay (does a constraint survive peer pressure?), false-belief contagion (does a falsehood spread through consensus?), context leakage (does information bleed between tasks?), and tracer durability (does marked data reach the final agent?). Evaluating four modern LLMs (GPT 4.1 mini, Gemini 2.5 Flash Lite, Qwen-3.5-35B-A3B, and Llama 3.1 8B Instruct), we expose model-specific vulnerability profiles invisible to outcome-only evaluation; Qwen-3.5-35B-A3B, for example, leads on tracer durability and instruction stability, while GPT 4.1 mini leads on leakage containment and false-belief resistance. Beyond per-model differences, communication topology emerges as a primary risk factor that explains 7-40% of the variance in multi-hop information survival. The effect traces to a synthesis bottleneck specific to converging-DAG nodes: an agent weighing competing parent inputs discards constraints carried by a minority branch, a bottleneck structurally absent from linear chains. AgentCollabBench demonstrates that suboptimal topology can silently erase the safeguards of highly capable models, arguing that multi-agent reliability is fundamentally a structural problem and that scaling model intelligence alone is no substitute for architecture.

Aritra Mazumder, Shubhashis Roy Dipta, Nusrat Jahan Lia +10

University of Utah · University of Maryland, Baltimore County · University of Dhaka +7

AI Agent Reliability Multi-Agent Collaboration AI Agent Benchmarks Multi-Agent Communication LLM Agent Reliability

Ask PDF

Oct 1, 2026

Right Answers, Wrong States: Hidden Information Failures in Multi-Agent Collaboration

Multi-agent systems are often judged by whether they reach the correct answer. This can miss a distinct failure: collaboration may leave behind a corrupted information state even when the immediate decision is correct. We call this an off-query failure. To study this failure in collaborative decision support, we introduce OffQuery, which separately evaluates evidence verification (T1), shared-state reconstruction (T2), and task resolution (T3) in two representative high-stakes settings: healthcare and disaster response. Across GPT, Gemini, and Qwen models, standard collaboration shows much stronger task performance than state reliability. Averaged over 21 model--setting combinations, task resolution reaches 64.7%, while evidence verification and state reconstruction reach only 14.3% and 43.1%. We trace this gap to selective information use: current queries often bypass corrupted facts, which become consequential when later tasks require them. We further introduce ReGround, which resolves conflicting evidence, verifies shared facts, reconstructs a trusted state, and reasons over that state. Across seven models from three families, ReGround improves all three capabilities in every evaluated setting, with average relative gains of 309.0%, 82.9%, and 17.6% on T1, T2, and T3. Reliable collaboration therefore requires both a correct decision and a reliable shared state for future reasoning.

Herun Wan, Jiaying Wu, Minnan Luo +4

Xi’an Jiaotong University · National University of Singapore · Yunnan University +1

AI Agent Reliability Multi-Agent Collaboration

Ask PDF