acceptodds
Under review as a conference paper at ICLR 2027

Beyond Perfect Teammates: Multi-Agent Collaboration with Unreliable Colleagues

10 views

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.

Primary area:
datasets and benchmarks
Keywords:
Multi-agent collaboration, Multi-agent system, benchmarks, LLM agents

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_GoxGOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces a large-scale controlled benchmark setting for multi-agent collaboration with unreliable sub-agents or colleagues. The UnCoBench benchmark introduces failures in the sub-agents via misaligned, incomplete, and missed settings. Under these failure and unreliable multi-agent settings, the authors study the impact on the main agent. The experiments show that comparable performance under reliable collaboration do not translate to comparable performance in the unreliable multi-agent collaboration setting.

Critical strengths
  1. The paper studies an important problem of unreliable collaboration in a multi-agent system which is an underexplored area.
  2. The authors’ approach of creating or simulating unreliable collaboration that incorporates misaligned, incomplete, and missed settings seem principled and realistic.
Critical weaknesses
  1. Some experimental results need to be explained or clarified (see question section below).
  2. Missing code and dataset which prevented me from assessing the quality of the benchmark which is the core contribution of this paper.
Questions for authors
  1. Figure 6 shows an additional bar where multiple colleagues can be unreliable at the same time. I assume the “mixed” bar plot consists of sub-agents or colleagues that demonstrate multiple types of unreliable behaviors. In this paper, do the authors also consider scenario where for a single specific unreliable setting (eg: misalignment), we have multiple roles or sub-agents that demonstrate the same type of unreliable behavior? In Line439, the authors described that for the other experiments, except for the “mixed” experiments, they involve only a single role agent (I assume a single sub-agent or single colleague). Can you clarify whether the unmixed setups consist of only a single unreliable sub-agent or it can also consists of multiple unreliable sub-agents for a given type of unreliable setting?
  2. Follow-up suggestion: For each type of unreliability, how does the performance of the main agent change as the number of unreliable sub-agents or colleagues increases for a given type of unreliability.
  3. In Figure 3, we see for LLMs and tasks, that they can transition from reliable (wrong) to unreliable (correct) setting, i.e., the blue region. Can the authors clarify this strange observation as I expect the overall task success to worsen as we make the sub-agents or colleagues in the multi-agent system to be unreliable.
  4. In the experiments, how many trials the agent was ran for every test case? Literature usually use multiple trials of agent runs.
  5. There is a similar work [1] that is missing in the references. Please discuss the differences between the proposed benchmark and prior work [1].

[1] Huang, J. T., Zhou, J., Jin, T., Zhou, X., Chen, Z., Wang, W., ... & Sap, M. (2024). On the resilience of llm-based multi-agent collaboration with faulty agents. arXiv preprint arXiv:2408.00989.

6 comments
AuthorsOpenReview ↗

Response to reviewer GoxG (1/3)

We thank the reviewer for the careful reading and questions about the experimental protocol and reproducibility. Below, we clarify the failure settings, explain the wrong-to-correct transitions through case studies, and discuss the relationship to prior work. We also provide an anonymous code and dataset repository and additional results on multiple unreliable roles and repeated evaluations.

W1. Missing code and dataset.

We provide the code and full benchmark dataset in an anonymous repository to support assessment of the benchmark’s quality and reproducibility.

Anonymous code and dataset repository of UnCoBench


W2. Do the unmixed settings include multiple unreliable roles, and how does performance change as their number increases?

We thank the reviewer for the suggestion to examine multiple unreliable colleagues. In the original Misaligned, Incomplete, and Missed settings, exactly one necessary role exhibits the designated unreliable behavior, while the other roles remain reliable. We chose this controlled design to examine whether the main agent can detect and repair an evidence failure. This directly probes the implicit assumption that colleagues provide trustworthy evidence. Mixed instead assigns the three different failure types to three distinct necessary roles.

To examine how performance changes as more colleagues become unreliable, we have extended each individual setting to two and three unreliable roles, comparing k = 1, 2, and 3 across five models from different families and scales, with 1,000 tasks per configuration. These experiments specifically test multiple colleagues exhibiting the same type of unreliability, complementing the heterogeneous failures in Mixed.

k is the number of roles exhibiting the same failure type. Each configuration contains 1,000 tasks. Cells report accuracy (%) with mean interaction rounds in parentheses.

Main agentSettingk = 1k = 2k = 3
Qwen3.5-9BMissed31.1 (2.65)33.7 (3.02)33.8 (3.05)
Qwen3.5-9BMisaligned9.0 (2.05)3.3 (2.03)1.8 (2.04)
Qwen3.5-9BIncomplete24.0 (2.12)11.9 (2.08)5.9 (2.09)
Qwen3.8-27BMissed76.1 (3.96)76.8 (3.90)73.5 (4.08)
Qwen3.8-27BMisaligned7.7 (2.42)3.8 (2.41)1.8 (2.41)
Qwen3.8-27BIncomplete47.5 (2.52)42.1 (2.73)46.1 (2.93)
Gemma 4 31BMissed87.5 (3.29)79.4 (3.63)47.0 (4.72)
Gemma 4 31BMisaligned11.6 (2.65)3.5 (2.42)0.2 (2.16)
Gemma 4 31BIncomplete64.6 (2.79)55.2 (2.98)52.6 (3.08)
Mistral Small 4-119A6BMissed55.1 (3.32)50.0 (3.42)49.6 (3.52)
Mistral Small 4-119A6BMisaligned16.9 (2.44)6.0 (2.48)3.3 (2.40)
Mistral Small 4-119A6BIncomplete43.7 (2.52)26.3 (2.51)11.8 (2.48)
GLM-5.3-FlashMissed96.2 (3.61)86.6 (4.33)70.3 (5.07)
GLM-5.3-FlashMisaligned35.1 (3.19)29.1 (3.14)24.3 (3.08)
GLM-5.3-FlashIncomplete87.8 (3.05)88.2 (3.16)89.6 (3.29)

The table above shows distinct patterns across failure types. Under Misaligned, accuracy decreases monotonically across all five models: Gemma 4 31B falls from 11.6% to 3.5% and 0.2%, while GLM-5.3-Flash falls from 35.1% to 29.1% and 24.3%. Under Incomplete, Mistral Small 4-119A6B drops from 43.7% to 26.3% and 11.8%, although GLM-5.3-Flash remains robust. Under Missed, GLM-5.3-Flash declines from 96.2% to 86.6% and 70.3%, with mean rounds increasing from 3.61 to 4.33 and 5.07 within the same six-round budget.

These results reinforce our finding that task-solving strength with reliable evidence does not guarantee sustained verification and coordinated recovery. As failures accumulate, the main agent must track unresolved requests and integrate corrected evidence across interactions. This extension examines the number of affected roles; it does not constitute an exhaustive study of error frequency or severity.

AuthorsOpenReview ↗

Response to reviewer GoxG (2/3)

W3. Why do some tasks transition from Reliable (wrong) to Unreliable (correct) in Figure 3?

Figure 3 compares outcomes from separate executions of the same underlying tasks. Although unreliable collaboration reduces aggregate accuracy, individual trajectories can occasionally improve because the altered interaction elicits additional checking or follow-up. Reliable evidence also does not guarantee that the main agent will integrate it correctly in every execution.

For example, in task_598, Qwen3.8-27B incorrectly returned “Do not deploy” after two rounds under Reliable, despite having the required evidence. Under Misaligned, an optical measurement referred to Firmware Build 3.8 instead of the required Build 4.2. The agent explicitly requested the record directory, retrieved the correct version-specific evidence, and returned the correct decision and numerical result after four rounds.

Similarly, in task_147, Gemma 4 31B returned an incorrect fuel-consumption average of 14.17 under Reliable, despite receiving all three numerical inputs. Under Incomplete, a missing safety-related value prompted an additional query; after obtaining it, the agent returned the correct average of 14.30. Both successful trajectories finished within the five-round Reliable budget. These cases illustrate how additional verification can accompany a successful trajectory, although they do not establish a universal explanation for every wrong-to-correct transition.

These individual gains coexist with substantially larger losses. For Qwen3.8-27B under Misaligned, only 4 tasks transition from wrong to correct, compared with 702 transitioning from correct to wrong. For Gemma 4 31B under Incomplete, the corresponding counts are 10 and 268. Thus, the blue regions reflect occasional improvements in individual executions, while the overall performance degradation remains clear.


W4. How many trials were run for each test case?

Our main evaluation used one complete agent trajectory per task for each model-setting combination, covering 1,000 paired task groups. This design balanced broad task coverage against the substantial cost of repeated multi-turn evaluations, particularly for API-based models. While the large task set improves the precision of aggregate performance estimates, repeated evaluations provide complementary evidence about variability on the same tasks.

To directly assess this variability, we conducted eight repeated evaluations on the same fixed subset of 100 task groups for Qwen3.8-27B, Gemma 4 31B, and Mistral Small 4-119A6B across Reliable, Misaligned, Incomplete, and Missed. We retained the original evaluation settings and prompts, yielding 9,600 complete task executions.

Accuracy (%) is reported as mean ± SD across eight runs of the same 100-task subset. SD is measured in percentage points.

Main agentReliableMisalignedIncompleteMissed
Qwen3.8-27B85.4 ± 1.36.0 ± 1.949.6 ± 2.682.8 ± 1.6
Gemma 4 31B94.4 ± 0.711.9 ± 2.165.9 ± 2.592.4 ± 1.1
Mistral Small 4-119A6B61.7 ± 4.012.9 ± 2.949.0 ± 4.156.0 ± 3.3

The table above reports the mean and standard deviation of accuracy across these eight runs. The results show modest variation in aggregate accuracy: SDs range from 0.7 to 2.6 percentage points for Qwen and Gemma, and from 2.9 to 4.1 for Mistral. All three models retain the mean accuracy ordering Reliable > Missed > Incomplete > Misaligned. The mean Reliable-Misaligned gaps range from 48.8 to 82.5 percentage points, substantially exceeding the observed run-to-run SDs. These repeated evaluations support our main qualitative finding on the tested models and subset: strong task-solving performance with reliable evidence does not ensure effective recovery from unreliable collaboration.

AuthorsOpenReview ↗

Response to reviewer GoxG (3/3)

W5. How does UnCoBench differ from Huang et al.'s work on faulty-agent resilience?

We thank the reviewer for highlighting this relevant work, which helps contextualize our study. We would like to clarify that the ICML 2025 version of Huang et al.’s work is already cited in our submission (line 582 in our original submission and line 628 in our revised manuscript). Their MAS-Resilience benchmark measures how well communication architectures and defense mechanisms preserve downstream performance in the presence of faulty agents. In MAS-Resilience, a single agent can in principle solve the task using the original inputs, while in UnCoBench the facts needed to determine a unique answer are distributed across role agents’ private records. The main agent must therefore obtain evidence from those agents. UnCoBench therefore evaluates whether the main agent can identify problematic replies from role agents, recover the necessary evidence through targeted follow-up, and use it to reach the correct decision.

Reference:

Huang, J.-T., Zhou, J., Jin, T., Zhou, X., Chen, Z., Wang, W., Yuan, Y., Lyu, M., and Sap, M. (2025). On the Resilience of LLM-Based Multi-Agent Collaboration with Faulty Agents. Proceedings of the 42nd International Conference on Machine Learning, PMLR 267, 26202-26226.

We have revised the manuscript to incorporate the additional results and discussions presented in these responses.

Reviewer_bX4YOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper generates a benchmark to characterize models in how they handle uncertainty/unreliability in their colleagues. Specifically, they investigate instances where another agent can be misaligned (providing wrong info in almost right context), incomplete (vague in replying / not giving enough specificity), and missing (where the agent simply does not reply). They show that various models, including frontier models, have different failure rates on these various scenarios.

Critical strengths
  • The paper does a thorough job of benchmarking multiagent performance failure under three common failure modes of individual members in a team - agents are misaligned (answer but for the wrong context), incomplete (underspecified but right context), and missed (the model does not reply). The use of programs to ensure that answers change upon use of improper information in these scenarios is good and important for benchmark creation soundness.
  • They evaluate across a wide variety of models to show that the benchmark is useful in differentiating how various models fail
  • The highlighting that accuracy does not imply robustness across these failure modes is interesting and useful to know.
Critical weaknesses
  • The work is not clearly differentiated against prior work in this space - they mention HiddenBench (Li 2026), and several works on recovering from noisy tool / failed tool use. Since the main agent is given role instructions with regard to its collaborators, and the agents are instructed to obey one of the three scenarios, it is unclear to me how this work is very different from a single agent calling noisy tools, or drastically different from the general hidden information literature. While I agree that hidden information settings distribute correct information as opposed to allowing for incorrect information, since the main agent knows who is responsible for which information, I think the problem reduces to a noisy tool calling setting.
  • The results are largely unsurprising, so I am unsure of the novelty - the result shows that when colleagues are unreliable, the performance drops. It also seems the performance drop correlates with how difficult the mistake is to identify: a wrong value in context (misaligned) is likely harder to differentiate than knowing a response is vague (incomplete), and both of these are likely harder than the main agent not receiving a reply (Missed).
  • A surprising result is not thoroughly investigated - that GPT 5.6 is dramatically not robust (reliable vs misaligned scenario) compared to Claude being dramatically more robust (much lower drop in its performance when misaligned).
  • It is also unclear whether the main agent is told in its prompt that some of the agents may be unreliable. I think this is important enough to merit re running if the prompt doesn't contain this information, as it can inflate the performance losses if the main agent assumes reliable collaborators.

I think these are useful findings, but I am not sure they are interesting enough for publication at ICLR. Of course, I am open to discussion and feedback on my interpretation.

Questions for authors

See above bullet in critical weaknesses.

Additionally, do you human validate generation of the unreliable scenarios or that the agent follows the assigned unreliable scenario? If I read correctly all of this is graded by model judges. It would be good to add human annotation agreement.

4 comments
AuthorsOpenReview ↗

Response to reviewer bX4Y (1/3)

We thank the reviewer for the careful reading and comments on the positioning and findings of our work. Below, we clarify the distinction from related benchmarks and explain the model-specific recovery patterns using our trajectory analyses and controlled interventions. We also point to the existing human evaluation and report additional experiments with an explicit reliability warning.

W1. Not clearly differentiated from noisy-tool and hidden-information benchmarks.

In UnCoBench, the main and role agents differ in context, information access, and responsibilities: the main agent manages the overall decision, while each role responds from private records within its assigned scope. The main agent must assess this feedback against the global task requirements and dynamically decide which role to follow up with and what additional evidence to request.

Regarding noisy tool calling, UnCoBench focuses on managing an ongoing evidence exchange with collaborators. Knowing the responsible role determines where to direct a request, but the main agent must still assess whether the reply addresses the intended context, identify unresolved information needs, and adapt the content and sequence of subsequent requests. For example, a reply about April clearance requires a targeted request for March evidence, while an unanswered request requires tracking and recontacting the relevant role. We acknowledge that these capabilities overlap with tool recovery. Our contribution is a controlled diagnostic benchmark for this collaboration process: executable checks establish that the affected evidence matters to the decision, while recovery traces help distinguish failures to obtain adequate evidence from errors after recovery.

Compared with HiddenBench (Li et al., 2026), which examines whether agents surface and pool distributed facts under latent information asymmetry, UnCoBench explicitly varies collaborator response reliability. Both require active information gathering, but our Reliable and Unreliable settings retain distributed information and known role responsibilities, preserving the underlying task and gold answer while changing a necessary role’s response policy. This paired design measures whether the main agent can verify collaborator feedback and coordinate targeted recovery, beyond gathering distributed evidence under reliable collaboration.

Reference:

Li, Y., Naito, A., and Shirado, H. (2026). Systematic Failures in Collective Reasoning under Distributed Information in Multi-Agent LLMs. Proceedings of the 43rd International Conference on Machine Learning (ICML).


W2. Results are largely unsurprising.

Our central finding indicates that current models still exhibit substantial weaknesses in noticing problematic communication, verifying evidence, and initiating targeted follow-up. UnCoBench measures the main agent’s ability to manage this gap between receiving a response and obtaining verified, decision-relevant evidence.

Our deeper analyses show that these capabilities are not reliably predicted by task-solving performance. GPT-5.6 Sol and Claude Opus 5 achieve similarly high Reliable accuracy (98.7% and 96.7%), yet their Misaligned accuracy differs sharply (45.8% versus 90.7%). Trajectory analysis shows that 53.4% of GPT-5.6 Sol episodes contain no corrective follow-up, compared with 9.2% for Claude Opus 5 (Figure 4). Thus, strong task-solving ability does not necessarily translate into strong verification and recovery capability, even under the same type of communication failure.

Our controlled interventions further show that identifying an error does not guarantee successful repair (Appendix E, Figure 8). Under Misaligned, four of six tested models approach their Reliable performance when corrective actions are specified, while two remain substantially below it. Together with the trajectory analyses, these results help distinguish difficulties in autonomous detection and repair selection from additional limitations in executing repairs and reasoning with recovered evidence, providing a more specific diagnosis of why models struggle with unreliable collaboration.

AuthorsOpenReview ↗

Response to reviewer bX4Y (2/3)

W3. Why is GPT-5.6 Sol substantially less robust than Claude Opus 5 under Misaligned?

The robustness gap between GPT-5.6 Sol and Claude Opus 5 in the Misaligned setting appears to stem primarily from how often each model follows up on misaligned replies. Both models made accurate decisions when they had the correct evidence. In the Reliable setting, GPT-5.6 Sol and Claude Opus 5 achieved 98.7% and 96.7% accuracy, respectively. After following up and recovering the correct evidence in Misaligned, they achieved 98.2% and 99.6%, respectively. The main bottleneck, therefore, is recovering the correct evidence through follow-up. As Figure 4 shows, across 1,000 Misaligned tasks, Sol followed up with the affected role in 46.6% of cases, compared with 90.8% for Opus 5. This gave Opus 5 more opportunities to recover the evidence needed for a correct decision.

The gap between GPT-5.6 Sol and Claude Opus 5 highlights the value of UnCoBench. Our results suggest that GPT-5.6 Sol places more emphasis on evidence coverage, while Claude Opus 5 places more emphasis on whether the reply supports the decision. Specifically, in the Missed setting, GPT persisted when replies were absent, spending more rounds than Claude (4.97 vs. 3.24) and achieving higher accuracy (97.4% vs. 95.8%). In the Misaligned and Incomplete settings, however, GPT more often proceeded without following up when a reply was out of scope or incomplete. Claude was more likely to follow up with the affected role and recover the correct evidence. This behavior is consistent with Anthropic’s observation that Opus 5 readily self-corrects without prompting (Anthropic, n.d.).

Reference:

Anthropic. (n.d.). Prompting Claude Opus 5. Claude documentation. Prompting Claude Opus 5.

AuthorsOpenReview ↗

Response to reviewer bX4Y (3/3)

W4. Is the main agent warned that some colleagues may be unreliable?

In the original protocol, the main agent is not explicitly told that some colleagues may be unreliable. This is a deliberate design choice: UnCoBench evaluates whether the main agent can independently detect problematic evidence, verify its adequacy, and initiate corrective follow-up.

The original manuscript already examines the effect of explicit corrective guidance through controlled interventions, reported in the final paragraph of Section 5.2 and detailed in Appendix E. These interventions alert the main agent that a previous interaction requires recovery, identify the affected role and interaction, or provide the diagnosis and corrective action. They test whether guidance after problematic communication improves recovery without revealing the answer. Under Misaligned, four of six tested models approach their Reliable performance when corrective actions are specified, while two remain substantially below it. These results show that guidance can improve recovery, but identifying the problem does not always ensure successful repair.

Following the reviewer’s suggestion, we further tested a complementary form of guidance: a general warning provided before interaction begins. We added “Some role agents may be unreliable.” to the main agent’s system prompt and reran all five settings for four models, retaining the same tasks and remaining prompt content.

Cells report accuracy (%) with mean interaction rounds in parentheses.

ModelConditionReliableMisalignedIncompleteMissedMixed
Qwen3.5-9BOriginal34.4 (2.04)9.0 (2.05)24.0 (2.12)31.1 (2.65)4.7 (2.68)
Qwen3.5-9BWith notice33.0 (2.04)8.6 (2.04)23.9 (2.08)31.8 (2.69)4.2 (2.65)
Qwen3.8-27BOriginal77.5 (2.17)7.7 (2.42)47.5 (2.52)76.1 (3.96)13.5 (4.47)
Qwen3.8-27BWith notice77.4 (2.12)7.4 (2.37)46.1 (2.43)76.9 (4.02)18.9 (4.64)
Gemma 4 31BOriginal90.4 (2.31)11.6 (2.65)64.6 (2.79)87.5 (3.29)5.8 (3.10)
Gemma 4 31BWith notice90.3 (2.25)13.8 (2.62)66.1 (2.76)87.7 (3.28)6.5 (3.12)
Mistral Small 4-119A6BOriginal57.1 (2.44)16.9 (2.44)43.7 (2.52)55.1 (3.32)11.8 (3.55)
Mistral Small 4-119A6BWith notice55.5 (2.44)15.1 (2.45)44.3 (2.47)54.7 (3.26)12.5 (3.49)

The results in the table above show modest, model-dependent changes overall. Across all four models, Misaligned and Mixed accuracy remains substantially below the corresponding Reliable performance despite the warning. These additional results complement the original interventions: general awareness alone does not consistently produce effective verification and recovery, while specific corrective guidance can help reveal where the difficulties lie. The central challenge is translating awareness into identifying problematic responses, requesting the necessary evidence, and completing the repair.


W5. Is task generation or role-agent adherence human-validated?

To clarify, our paper already includes a human evaluation of the task-generation pipeline, reported in Appendix C.2 and summarized in Section 3.5. We manually reviewed 100 candidate instances from each of the three validation stages - specification validation, logical compilation, and full-information acceptance - with each sample including both accepted and rejected candidates. We compared the human accept/reject decisions with those of the automated pipeline.

Human-pipeline agreement was 98%, 100%, and 93% for the three stages, respectively. All disagreements involved candidates rejected by the automated pipeline but accepted by human reviewers, indicating conservative filtering in the audited samples. These results complement the executable checks of answer uniqueness and role necessity, as well as the repeated full-information validation, and provide additional evidence supporting the validity and robustness of our construction pipeline.

This human audit concerns task construction and acceptance decisions. Role-agent adherence to the assigned response policies is evaluated separately using an LLM judge in Appendix D.

We have revised the manuscript to incorporate the additional results and discussions presented in these responses.

I appreciate the thorough response from authors and will raise my score from 2 to 3

Reviewer_r6B5OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces UNCOBENCH, a benchmark for evaluating how a main agent recovers from unreliable responses by specialized colleagues. It contains 1,000 task groups, each pairing a reliable baseline with Misaligned, Incomplete, and Missed variants while preserving the underlying problem and ground-truth answer. An automated construction pipeline combines LLM-generated scenarios with programmatic checks of answer uniqueness, role necessity, and perturbation validity. Evaluation covers ten main-agent models and examines accuracy, interaction rounds, and recovery behavior.

Critical strengths
  1. Controlled paired evaluation. The reliable and unreliable variants retain the same task and ground truth, allowing changes in performance to be examined under a controlled change in colleague behavior.
  2. Explicit benchmark validation. The construction pipeline includes executable checks of answer uniqueness and role necessity, repeated full-information validation, and a human audit of validation decisions.
Critical weaknesses
  1. Unclear correspondence to real-world colleague reliability. I am not yet convinced that the benchmark adequately represents practical collaboration. Although it includes single-role failures and a Mixed setting, the prescribed response policies remain simplified. A more systematic study of different proportions of reliable and unreliable colleagues, together with varying error frequency and severity, would help clarify the gap between these controlled settings and real multi-agent systems.
  2. Insufficient empirical justification for the failure taxonomy. Misaligned, Incomplete, and Missed are intuitively motivated by contextual alignment, completeness, and responsiveness. However, the paper does not sufficiently establish how representative these three categories are of real collaboration failures. Empirical analysis of real interaction data would help justify their selection and coverage.
  3. Further validation of model generality is needed. The evaluation includes seven open-weight main-agent models, but the conclusions would benefit from broader and more systematic validation across model families and scales. Additional models or controlled comparisons within model families would help establish whether the observed recovery patterns extend beyond the selected model set.
Questions for authors

Just see weeknesses

4 comments
AuthorsOpenReview ↗

Response to reviewer r6B5 (1/2)

We thank the reviewer for the careful reading and comments on the benchmark’s practical relevance and generality. Below, we clarify the design of the unreliable settings and provide additional evidence through analysis of real multi-agent failure traces, experiments with multiple unreliable colleagues, and comparisons across model families and scales.

W1. Unclear correspondence to real-world colleague reliability.

In the original Misaligned, Incomplete, and Missed settings, exactly one necessary role exhibits the designated unreliable behavior, while the other roles remain reliable. We chose this controlled design to examine whether the main agent can detect and repair an evidence failure. This directly probes the implicit assumption that colleagues provide trustworthy evidence. Mixed instead assigns the three different failure types to three distinct necessary roles.

To examine how performance changes as more colleagues become unreliable, we have extended each individual setting to two and three unreliable roles, comparing k = 1, 2, and 3 across five models from different families and scales, with 1,000 tasks per configuration. These experiments specifically test multiple colleagues exhibiting the same type of unreliability, complementing the heterogeneous failures in Mixed.

k is the number of roles exhibiting the same failure type. Each configuration contains 1,000 tasks. Cells report accuracy (%) with mean interaction rounds in parentheses.

Main agentSettingk = 1k = 2k = 3
Qwen3.5-9BMissed31.1 (2.65)33.7 (3.02)33.8 (3.05)
Qwen3.5-9BMisaligned9.0 (2.05)3.3 (2.03)1.8 (2.04)
Qwen3.5-9BIncomplete24.0 (2.12)11.9 (2.08)5.9 (2.09)
Qwen3.8-27BMissed76.1 (3.96)76.8 (3.90)73.5 (4.08)
Qwen3.8-27BMisaligned7.7 (2.42)3.8 (2.41)1.8 (2.41)
Qwen3.8-27BIncomplete47.5 (2.52)42.1 (2.73)46.1 (2.93)
Gemma 4 31BMissed87.5 (3.29)79.4 (3.63)47.0 (4.72)
Gemma 4 31BMisaligned11.6 (2.65)3.5 (2.42)0.2 (2.16)
Gemma 4 31BIncomplete64.6 (2.79)55.2 (2.98)52.6 (3.08)
Mistral Small 4-119A6BMissed55.1 (3.32)50.0 (3.42)49.6 (3.52)
Mistral Small 4-119A6BMisaligned16.9 (2.44)6.0 (2.48)3.3 (2.40)
Mistral Small 4-119A6BIncomplete43.7 (2.52)26.3 (2.51)11.8 (2.48)
GLM-5.3-FlashMissed96.2 (3.61)86.6 (4.33)70.3 (5.07)
GLM-5.3-FlashMisaligned35.1 (3.19)29.1 (3.14)24.3 (3.08)
GLM-5.3-FlashIncomplete87.8 (3.05)88.2 (3.16)89.6 (3.29)

The table above shows distinct patterns across failure types. Under Misaligned, accuracy decreases monotonically across all five models: Gemma 4 31B falls from 11.6% to 3.5% and 0.2%, while GLM-5.3-Flash falls from 35.1% to 29.1% and 24.3%. Under Incomplete, Mistral Small 4-119A6B drops from 43.7% to 26.3% and 11.8%, although GLM-5.3-Flash remains robust. Under Missed, GLM-5.3-Flash declines from 96.2% to 86.6% and 70.3%, with mean rounds increasing from 3.61 to 4.33 and 5.07 within the same six-round budget.

These additional experiments reinforce our original finding that strong task-solving performance with reliable colleagues does not guarantee effective verification and recovery when collaboration becomes unreliable. Moreover, the challenge extends beyond detecting a single faulty response: the main agent must continuously verify contextual relevance, track unresolved dependencies, and integrate corrected evidence across multiple interactions.

AuthorsOpenReview ↗

Response to reviewer r6B5 (2/2)

W2. Insufficient empirical justification for the failure taxonomy.

We thank the reviewer for suggesting empirical validation of the failure taxonomy. To examine whether the three categories occur in practical multi-agent collaboration, we analyzed 58 failed Magentic-One runs from Who&When (Zhang et al., 2025).

Failure modeTracesShare of 58 failed traces
Misaligned1831.0%
Incomplete1322.4%
Missed2136.2%

The table above shows that all three failure modes recur in these traces. Overall, 39 traces contain at least one mode, and 12 contain two or more; the categories therefore overlap. Representative examples are:

Misaligned. The coordinator requested a 2023 team roster, but WebSurfer returned a roster marked as updated at the end of the 2024 season. Trace 54

Incomplete. The coordinator asked WebSurfer to find supermarkets within a specified area, but WebSurfer returned only a partial list. Trace 10

Missed. The coordinator asked WebSurfer to look up the distance from the museum to each candidate bar, but WebSurfer did not return a visible reply. Trace 28

These observations provide empirical support for the design of UnCoBench’s three settings: they reflect recurring breakdowns in the contextual alignment, completeness, and responsiveness of collaborator evidence. Each breakdown creates a distinct recovery requirement for the main agent: checking the evidence’s scope, requesting missing details, or following up on an unresolved request. Our settings operationalize these patterns as controlled conditions, allowing us to examine whether the main agent can recognize the problem and take the appropriate corrective action.

Reference:

Zhang, S., Yin, M., Zhang, J., Liu, J., Han, Z., Zhang, J., Li, B., Wang, C., Wang, H., Chen, Y., and Wu, Q. (2025). Which Agent Causes Task Failures and When? On Automated Failure Attribution of LLM Multi-Agent Systems. Proceedings of the 42nd International Conference on Machine Learning (ICML).


W3. Further validation of model generality is needed.

We have added checkpoints from the Qwen and Gemma families, expanding the comparison to four Qwen models (4B, 9B, 27B, and 35B-A3B) and three Gemma models (E4B, 12B, and 31B), using the same paired tasks and evaluation protocol.

ModelReliableMisalignedIncompleteMissedMixed
Qwen3.5-4B6.0 (4.81)3.2 (6.15)11.6 (5.58)2.7 (5.89)0.5 (6.67)
Qwen3.5-9B34.4 (2.04)9.0 (2.05)24.0 (2.12)31.1 (2.65)4.7 (2.68)
Qwen3.8-27B77.5 (2.17)7.7 (2.42)47.5 (2.52)76.1 (3.96)13.5 (4.47)
Qwen3.6-35B-A3B65.1 (3.14)1.1 (2.79)27.6 (2.85)51.7 (4.44)0.1 (3.89)
Gemma 4 E4B13.0 (2.99)3.3 (3.40)9.8 (3.10)15.1 (4.73)0.8 (4.82)
Gemma 4 12B47.7 (3.77)5.7 (3.82)39.8 (3.82)45.7 (4.44)2.0 (4.12)
Gemma 4 31B90.4 (2.31)11.6 (2.65)64.6 (2.79)87.5 (3.29)5.8 (3.10)

The comparisons in the table above reinforce our finding that strong performance with reliable colleagues does not guarantee effective recovery from unreliable evidence. Across the Gemma models, Reliable accuracy increases from 13.0% to 47.7% and 90.4%, while Misaligned accuracy remains much lower at 3.3%, 5.7%, and 11.6%. Among the newly added Qwen models, Qwen3.6-35B-A3B substantially outperforms Qwen3.5-4B under Reliable (65.1% vs. 6.0%), yet performs worse under Misaligned (1.1% vs. 3.2%).

Together, these results strengthen the finding that evidence verification and corrective follow-up remain substantial weaknesses across model families and scales. Models often proceed with plausible colleague responses without adequately checking their contextual relevance, even when they can solve the underlying tasks effectively with reliable evidence.

We have revised the manuscript to incorporate the additional results and discussions presented in these responses.

Thank you for the additional analyses. The real-world failure traces and the within-family model comparisons have largely addressed my concerns.

My remaining concern is W1. Varying the number of unreliable roles is useful, but each role still follows a fixed failure policy. This tests how performance changes with the number of affected colleagues, while leaving open how the benchmark relates to collaboration where errors vary in frequency and severity. Could you clarify this limitation?

Given the substantial additions, I am considering raising my rating after this clarification

AuthorsOpenReview ↗

Response to remaining concern.

Thank you for clarifying this concern. Our original settings and the k-role extensions reveal substantial weaknesses in verifying collaborator responses and recovering decision-relevant evidence. The original controlled design helps isolate these difficulties by allowing us to examine whether the main agent notices problematic communication, requests the necessary evidence, and reaches the correct answer after recovery. Experiments with different values of k also capture variation in the overall severity of unreliable collaboration by changing the number of affected colleagues. We agree that examining error frequency and severity can further clarify how model performance varies across different reliability conditions and have extended our evaluation to settings with varying error frequency and recovery requirements.

We evaluated Qwen3.8-27B and Gemma 4 31B on the same 500 task groups. For each comparison, we varied one reliability dimension while holding other settings fixed.

For frequency, we replaced the fixed initial nonresponse in the Missed setting with a per-request probability p that the designated role fails to respond. We used p = 0, 0.25, 0.50, and 0.75, with higher values corresponding to more frequent missed replies. The response outcome was sampled again for each subsequent request, allowing the same colleague to respond intermittently rather than follow the original fixed pattern of missing the first request and responding after follow-up.

For severity, we assessed the severity of unreliable collaboration by varying evidence sufficiency and the amount of follow-up required to resolve information gaps in the Incomplete setting. In the mild condition, the response is imprecise but sufficient to determine the answer. The moderate condition retains the original policy: the initial response is insufficient, and one effective follow-up recovers the precise evidence. In the severe condition, the response to the first follow-up still leaves the answer undetermined, requiring a second effective follow-up to recover the evidence.

Task groups per condition: 500. Cells report accuracy (%), with the change relative to Reliable in parentheses (percentage points). Negative values indicate degradation.

ModelReliable (p = 0)p = 0.25p = 0.50p = 0.75
Qwen3.8-27B77.274.2 (-3.0)68.6 (-8.6)43.8 (-33.4)
Gemma 4 31B91.089.2 (-1.8)86.2 (-4.8)62.8 (-28.2)
ModelReliableMildModerateSevere
Qwen3.8-27B77.275.4 (-1.8)49.2 (-28.0)35.2 (-42.0)
Gemma 4 31B91.088.6 (-2.4)67.0 (-24.0)51.4 (-39.6)

Both models show progressively larger performance degradation as failure frequency or severity increases. Relative to Reliable, accuracy losses grow from 1.8–3.0 percentage points at p = 0.25 to 28.2–33.4 points at p = 0.75. Similarly, losses increase from 1.8–2.4 points under mild incompleteness to 24.0–28.0 points under moderate incompleteness and 39.6–42.0 points under severe incompleteness. These trends further support our finding that strong performance with reliable evidence does not ensure robust task completion under unreliable collaboration, with larger performance gaps as failures become more frequent or require more sustained recovery.Given the time constraints, we initially evaluated two models to examine these trends and will extend the evaluation to additional models to strengthen the evidence for our conclusions. Once the evaluation is complete, we will incorporate the full results and accompanying discussion into the manuscript.

We hope these clarifications and additional experimental results address your concern, and we would be grateful if you could reconsider your rating.

Comments on the paper

1 comment

Dear Area Chair and Reviewers,

We thank all reviewers for their constructive feedback. We have revised the manuscript to incorporate the additional experiments and clarifications detailed in our responses. The key updates are:

  1. Generality and robustness. We extended the failure settings to multiple unreliable roles across five models (Reviewers r6B5, W1; GoxG, W2), expanded the Qwen and Gemma comparisons (Reviewer r6B5, W3), and conducted eight repeated evaluations across three models (Reviewer GoxG, W4). These results reinforce that strong performance with reliable colleagues does not guarantee effective verification and recovery under unreliable collaboration.

  2. Empirical grounding and recovery analysis. We analyzed failed Magentic-One runs, finding recurring examples of all three failure modes (Reviewer r6B5, W2). We also tested an explicit reliability warning across four models and five settings, observing modest, model-dependent changes that did not close the performance gaps (Reviewer bX4Y, W4). These additions complement our original trajectory analyses and controlled interventions in explaining recovery failures.

  3. Positioning, validation, and reproducibility. We clarified the relationship to noisy tool calling, HiddenBench, and faulty-agent resilience benchmarks (Reviewers bX4Y, W1; GoxG, W5), highlighted the existing human validation and separate role-agent evaluation (Reviewer bX4Y, W5), and provided an anonymous code and dataset repository (Reviewer GoxG, W1).

We believe these additions, together with the other clarifications in our responses, address all the reviewers’ concerns. We would appreciate their feedback on whether the issues have been resolved and are happy to provide further clarification.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.