acceptodds
Withdrawn from ICLR 2027

Understanding and Estimating Uncertainty Flow in Multi-Agent Systems

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multi-agent systems solve tasks by exchanging messages, and each message changes what remains uncertain about the final outcome, indicating uncertainty as the floor on how well failure can be predicted at that point. Predicting failure is therefore an uncertainty estimation problem over the whole communication record. Though recent studies show that errors and hallucinations propagated in the communication could cause the whole system to fail, they mainly focus on the final outputs, instead of capturing the inherent consensus that happens in a multi-agent system. To address these limitations, we first develop a theoretical framework of how uncertainty about the outcome evolves through communication. By applying standard martingale identities to the filtration generated by the messages, we assign each message a resolution without assuming any dependence structure among agents, and show that the same identities hold for the features observed by an estimator. Motivated by our theoretical findings, we propose CUE, a supervised probe on backbone hidden states that estimates correctness and its uncertainty after every message and attributes the resolved uncertainty to rounds. We evaluate CUE on whether a multi-agent system's final answer is correct, across 4 benchmarks, 3 multi-agent orchestrations, and 2 LLM families, against 7 baselines. CUE achieves the highest AUROC in most of the configurations on both Llama and Qwen.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Multi-agent systems, uncertainty

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_NrC4OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how to predict the final correctness of an LLM multi-agent execution from its communication history. It applies standard martingale and conditional-variance identities to describe expected uncertainty resolution and realized uncertainty increases. The proposed method, CUE, trains a shared probe on frozen-backbone features of communication prefixes using only final correctness labels. A variance regularizer encourages average uncertainty to decrease, while changes in predictions provide resolution and amplification estimates. Across four benchmarks, three orchestrations, and two backbone families, the authors report improved failure prediction over seven baselines. Additional analyses examine early prediction, message attribution, and computational cost.

Critical strengths
  1. The experiments provide substantial evidence that communication history helps predict final failure. CUE achieves the highest mean AUROC in 27 of 28 configurations. Table 5 also directly tests the value of history: on Llama/MATH500, removing earlier messages reduces average AUROC from 89.2 to 79.2. This supports the main empirical motivation.

  2. CUE uses only final correctness labels to train a shared predictor across prefixes. It does not require labels for individual messages. This makes the supervision practical and allows predictions before the execution finishes without training a separate model for each position.

  3. The formulation clearly separates expected uncertainty reduction, realized increases, and causal effects. The decomposition does not require independence between agents, and the authors acknowledge order dependence and representation limitations. These distinctions provide a useful basis for analyzing what an observer learns from communication.

Critical weaknesses
  1. The main gain seems to come from giving a correctness probe the full history. On Llama/MATH500, Table 5 shows a 10-point AUROC gain from history, but only 0.16 points from the variance regularizer. The theory’s added value to the method is therefore unclear.

  2. The message-level estimates need stronger checks. Appendix D.4 assigns information gain to a vote that adds no new information to the full history. The authors explain this through the features, but it remains unclear how reliably the scores measure information from messages.

  3. Table 15 compares early CUE predictions with baselines that read the completed debate. It would be more useful to also compare against a standard correctness probe trained on the same partial history.

  4. CUE needs about 2,200 seconds to extract features for 500 Llama debate executions (Appendix C.4). This makes it the most expensive tested method overall. The paper does not show whether early detection can save enough computation to cover this cost.

Questions for authors
  1. Does the variance regularizer help estimate correctness probabilities more accurately, even when it barely changes AUROC?

  2. Could you test the message-level scores on a simple example where the true information gain is known?

  3. How does CUE compare with a standard probe using the same inputs and labels, including the time spent extracting features?

Reviewer_DEXJOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The authors work on the uncertainty quantification problem of multi-agent systems by framing the multi-agent inference as a stochastic process, e.g., Martingale, to deliver theoretical statements on uncertainty dynamics across multi-round agent communications, and propose a supervised linear probing on top of LLM hidden states to predict the outcome given the task. The authors extensively validated their framework across diverse setups.

Critical strengths

I appreciate that the research problem the authors are working on is really important (well motivated and timely) -- theoretically characterizing the uncertainty dynamics of multi-agent systems' inference and trying to model it through a lightweight probe.

Their approach to taking inspiration from the martingale process is interesting (although this kind of framing was already explored in Choi et al. 2025), though I can not get exactly what the authors want to say.

Although I couldn't find any substantial strength in the authors' original content (due to the concerning quality of writing), the problem itself is very timely and important, so I would say that can be a strength.

  • Choi et al. 2025, Debate or Vote: Which Yields Better Decisions in Multi-Agent Large Language Models?
Critical weaknesses

WRITING QUALITY is the most significant issue

I can not get any points across in sections 3 and 4, even though I spent more than 2 hours reading this paper end-to-end.

  1. No clear takeaway from every paragraph: each subsection stacks definitions, identities, and caveats ("not a unique causal allocation", "not a pathwise conservation law", "we fix one serialization order") without first stating what the reader should conclude and why it matters
  2. Notation overuse/overload and conflicts: (1) overall, the authors failed to organize the notation system cleanly by using too many notations rather than abstracting out some less-important things; (2) m, M, Ω\Omega, C are used multiple times to denote different things.
  3. In general, I strongly feel that those sections are not read for a human reader.
  • The authors' claims keep being interrupted by caveats and implementation details.
  • They add long explanations to unimportant minor points.
  • Long chains of passive-voice sentences lack the connectives/shifting words that would link one claim to the next.
  • I believe these are common traits of LLM-generated/editted text, and I suspect Sections 3-4 were drafted with heavy AI assistance and little human revision. Whatever the cause, these two sections are very hard to follow (not because it is math-heavy, but because of its presentation). Just for reference and comparison, I give these theory-centric papers as examples, which are great in their presentation (easy to follow their key claims and show high connectivity between sentences): Oh et al. 2025, Kiyani et al. 2026a;b

Weak connection between the theorem and their method

In abstract/intro, the authors said their method was motivated by their theoretical analysis. However, I don't see exactly what inspiration has been injected in the simple supervised linear probe on top of hidden states.

  • Here, the only training component derived from the theory is the regularizer Rc\mathcal{R}_c. Removing it (λc=0\lambda_c = 0) changes the average AUROC only by 0.16 (Tab. 5). That says, most of the performance gain comes from reading the full communication record (89.2 vs. 79.2 without intermediate messages).
  • Besides, the benefits from theory-inspired readouts are too marginal (adding A^\hat{A} improves failure AUROC in 0 of 8 sets (Tab. 16))

Reference

  1. Oh et al. 2025, Understanding Multimodal LLMs Under Distribution Shifts: An Information-Theoretic Approach
  2. Kiyani et al. 2026a, When to Trust the Cheap Check: Weak and Strong Verification for Reasoning
  3. Kiyani et al. 2026b, Strategic Decision Support for AI Agents

minor 1) mislabel -- PULSE (ours) in Figure 5

minor 2) missing relevant literature

  1. Choi et al. 2025, Debate or Vote: Which Yields Better Decisions in Multi-Agent Large Language Models?
  2. Chen et al. 2026, Every Response Counts: Quantifying Uncertainty of LLM-based Multi-Agent Systems through Tensor Decomposition
  3. Emde et al. 2026, Lost in Communication: Uncertainty Propagation in Multi-Agent Systems
Questions for authors

Dear authors,

I am really into this field, especially uncertainty quantification for LLM agent systems, so I really resonate with your research motivation.

Could you please rewrite sections 3 and 4 in your own language, highlight what you want to say through your theoretical statements, and clarify how your method is designed, component by component? The current form of writing makes it hard for me to get any single points (even though I have already published several papers in this field).

If you rewrite those sections end-to-end, I am willing to re-review the paper and adjust my assessment if I find some strengths there.

Reviewer_DofxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors study the amount, amplification and attribution of uncertainty at each message in a multi agent system using martingale identities. CUE is presented as a MLP probe on hidden states of each turn's prefix and trained on final correctness. Experiments cover multiple benchmarks across 2 model families.

Critical strengths
  • Strong experimental setup covering multiple configurations of multiagent asystems, random splits, and multiple metrics (Brier, NLL, ECE) reported
  • The problem is relevant and well motivated. Multi-agent failure prediction from the communication trjacteory instead of final outputs is an increasingly important setting for the community
  • Early detection of the failures are useful since the score after the first round already are better than the baselines' full debate score
Critical weaknesses
  • I am concerned how scalable and practical CUE is. it requires whitebox access and labels for each configuration. Training queries and retraining per configuration might be infeasible at scale.
  • The presented theoretical results are trivial and do not really help the intended measurements. Variance constraint has little effect on AUROC (its even below seed noise), for example. The method is just a hidden state probe over the full communication transcript.
  • The baselines only consider the final output but CUE works over each message. this to me is an unfair comparison.
  • Uncertainty estimation (ECE) is results are quite weak.
Questions for authors

See weaknesses above^

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

Related papers

Discussion (0)

Sign in to comment.