acceptodds
Under review as a conference paper at ICLR 2027

CIPI: When to Integrate Peer Feedback in Multi-Agent LLM Reasoning

Abstract

Conventional round-based multi-agent LLM systems integrate peer feedback at predefined interaction rounds, without considering whether the receiving agent needs an immediate correction or whether the feedback can wait. This fixed timing may delay necessary corrections or interrupt reasoning that is already correcting itself. We introduce CIPI (Conditional Integration of Peer Information), a protocol that lets the receiving agent decide when to use feedback after it arrives. CIPI stores peer feedback and uses the agent's current reasoning state to decide when to review it. The agent reviews feedback early when a contradiction or misunderstanding of the question could affect the next reasoning step and is not already being corrected. Otherwise, feedback is saved for later review. The model checks proposed corrections before accepting them and records accepted changes to guide later reasoning. Early corrections do not prevent further review, and a final check takes place before the answer is generated. The key idea is to use peer feedback when the receiving agent needs it. We evaluate CIPI on MuSiQue, HotpotQA, MATH-500, OlymMATH, APPS, and HumanEval, covering multi-hop question answering, mathematical reasoning, and code generation. We compare CIPI with output-level message passing, adapted Exchange-of-Thought Report, and message passing augmented with Self-Refine, evaluating task performance and measured token usage. Under the evaluated role-specific thinking settings, CIPI achieves the highest reported scores across the six benchmarks, improving over message passing with Self-Refine by 0.84–9.00 percentage points, generally with higher token usage. These findings suggest the value of selectively integrating peer feedback during ongoing reasoning to support timely correction without unnecessary disruption.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Models, Multi-Agent Reasoning, Peer Feedback, Inter-Agent Communication, Communication Timing

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_NmvYOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes CIPI (Conditional Integration of Peer Information), a multi-agent LLM reasoning framework that addresses the limitations of integrating peer feedback at fixed interaction rounds. Instead of immediately incorporating received feedback, CIPI maintains a feedback buffer and employs a state-dependent gate to determine when feedback should be inspected. An integrator proposes revisions, while a verifier evaluates accepted corrections before updating the reasoning state. The framework also supports deferred feedback inspection and rollback of rejected updates. Experiments across six benchmarks covering multi-hop question answering, mathematical reasoning, and code generation show that CIPI achieves higher task performance than the evaluated communication baselines, generally at the cost of increased token consumption. Additional ablation studies investigate the effects of feedback integration timing and comment selection.

Critical strengths

1. Clear Motivation and Well-Structured Design: The paper addresses the underexplored problem of when to integrate peer feedback in multi-agent LLM reasoning. The proposed framework combines state-dependent feedback inspection, verification, and rollback mechanisms in a structured and interpretable manner.

2. Broad Experimental Evaluation: The paper evaluates CIPI across six benchmarks covering multi-hop question answering, mathematical reasoning, and code generation. Additional ablation studies distinguish feedback integration timing from comment selection, while token usage and confidence intervals provide further insights into performance and efficiency.

Critical weaknesses
  1. Insufficient Evidence for the Core Feedback Integration Mechanism: The main contribution of CIPI is its state-dependent decision on when to integrate peer feedback. However, the ablation results in Table 2 do not convincingly demonstrate the effectiveness of this mechanism. For example, CIPI-All performs slightly worse than Random-All on MuSiQue (68.34 vs. 68.40 F1) and achieves identical accuracy on OlymMATH (88.00%). Furthermore, the confidence intervals in Appendix Table 7 generally include zero. The paper also lacks direct evaluations of the gate's error-detection accuracy, false negatives, and the verifier's ability to reject harmful corrections.

  2. Confounded Comparisons with Existing Methods: The main experiments do not isolate the contribution of CIPI's integration strategy. Different methods use different thinking settings, model-call structures, and output constraints. For example, EoT-Report disables thinking for all agents, whereas CIPI enables it for the Peer, Integrator, and Verifier. Additionally, CIPI employs a structured QA answer selector with source quotations and validation checks that are not applied to the baselines. Consequently, it is unclear how much of the reported improvement comes from conditional feedback integration rather than these additional design differences.

  3. Limited Evaluation Scale and Generalizability: Despite covering six benchmarks, the experiments use only 100 questions per dataset, except for HumanEval (164 questions). Such small evaluation sets limit the statistical reliability of the reported improvements, particularly when some gains are relatively modest. For instance, the 95% confidence interval for CIPI's 5-point improvement over MP+SR on MATH-500 is [−1.00, 11.00], which includes zero. Moreover, all experiments rely on a single hosted model family, leaving the generalizability of CIPI across different LLM backbones insufficiently validated.

  4. High Computational Cost and Lack of Budget-Matched Comparisons: CIPI introduces multiple additional model calls for feedback inspection, integration, and verification, resulting in substantially higher token consumption than the baselines. According to Table 1, CIPI consumes approximately 44%, 29%, and 35% more tokens than MP+SR on MATH-500, MuSiQue, and HumanEval, respectively. However, the corresponding performance improvements are not consistently substantial. The paper lacks comparisons under matched token budgets, as well as detailed latency analysis, making it difficult to assess the practical value of the proposed framework.

  5. Poor Figure Quality and Unclear Method Presentation: Figure 1, the only methodological figure in the paper, suffers from unclear directional connections, insufficiently distinguished decision branches, and text-heavy boxes. In particular, the inspection, verification, commitment, and rollback paths are not clearly illustrated. Important components such as the feedback buffer, correction record, and hypothesis history are also not explicitly visualized. Given the complexity of the proposed controller, the figure does not effectively communicate the overall framework, making the method unnecessarily difficult to understand.

Questions for authors
  1. Effectiveness of State-Dependent Feedback Integration: The central contribution of CIPI is determining when peer feedback should be integrated based on the agent's current reasoning state. However, Table 2 shows that CIPI-All performs slightly worse than Random-All on MuSiQue (68.34 vs. 68.40 F1) and achieves identical accuracy on OlymMATH (88.00%). Moreover, the confidence intervals in Appendix Table 7 include zero for these comparisons. Could the authors provide stronger evidence that state-dependent integration is more effective than random or fixed-timing strategies? In particular, how reliably does the gate identify situations requiring immediate correction, and how frequently does it miss critical feedback? These results would directly affect the assessment of the paper's central claim.

  2. Fairness of Comparisons and Sources of Performance Gains: The main comparisons use different thinking settings, output constraints, and computational budgets across methods. For example, CIPI enables thinking for its Peer, Integrator, and Verifier, while EoT-Report disables thinking for all agents. CIPI also employs additional answer-validation mechanisms and generally consumes more tokens than the baselines. Could the authors clarify how much of CIPI's performance improvement can be attributed specifically to conditional feedback integration rather than these differences? Would the improvements remain under matched thinking configurations and inference token budgets? Such evidence is important for determining the effectiveness and practical value of the proposed method.

1 comment
AuthorsOpenReview ↗

Response to Reviewer NmvY

Thank you for the detailed review. In short: (1) the log analysis shows that the gate fires on wrong states and avoids harmful early revisions, while the timing contrast in final scores is underpowered. (2) Table 1 is a system-level comparison, and the matched ablation arms are close in token use. (3) We will state the evaluation scope and redraw Figure 1.

Weaknesses

1. Evidence for the core integration mechanism.

In final scores CIPI-All trails Random-All by 0.07 F1 on MuSiQue and ties it on OlymMATH, and the intervals in Table 7 include zero (general response, item 2). The consistent effect in the ablation is selective inspection: full CIPI exceeds CIPI-All on all three datasets. To evaluate the gate directly, we analyzed the MuSiQue ablation logs, labeling each state as correct or wrong against the reference answer.

  • Precision of the gate. All 37 early inspections of CIPI fired on wrong states.
  • Deferred comments. At a further 37 wrong-state checkpoints the gate waited. Those comments were not discarded but stayed in the buffer until the terminal review.
  • Effect of early updates. Of 29 committed early updates, none broke a correct answer, with a mean F1 change of +1.34. Always-All broke 4 correct answers in 56 early commits, with a mean F1 change of −5.86.

All arms use the same verifier, so the protection against harmful early changes comes from the gate.

2. Comparisons with existing methods.

The thinking settings, the QA selector and the APPS audit differ between CIPI and the baselines, and EoT-Report runs without thinking. Table 1 is therefore a system-level comparison, and we will describe it as such (general response, item 1).

3. Evaluation scale and generalizability.

The main comparison uses 100 tasks per benchmark and all 164 HumanEval tasks, and all runs use one hosted model (deepseek-flash). The MATH-500 interval ([−1.00, 11.00]) includes zero, as you note. We will state in the abstract that the results come from one model family and report the paired intervals alongside the point estimates.

4. Computational cost and budget matching.

On OlymMATH CIPI gains 9.00 points at a token ratio of 1.04 to MP+SR, and on APPS it passes 4 more tasks at a ratio of 1.01. On the other benchmarks the ratio is 1.29 to 1.44, and we make no efficiency claim there. The four ablation arms are close in token use (Table 2), so the ablation contrasts are approximately budget-matched. The cost report covers tokens, and we will state that wall-clock latency is not measured.

5. Figure 1.

We will redraw Figure 1 with the feedback buffer, correction record and hypothesis history as explicit components, and with separate paths for inspection, verification, commitment and rollback.

Questions

Q1. See item 1 above and item 2 of the general response.

Q2. See item 2 above and item 1 of the general response.

Reviewer_Uc6GOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CIPI studies when agents should incorporate already available peer feedback into their reasoning trajectories. Through explicit state, a feedback buffer, a conditional inspection gate, and a correction record, it separates feedback receipt from integration. At checkpoints between model calls, an integrator proposes updates and a verifier approves accepted peer revisions; deferred opinions are reconsidered at the terminal step. Across six QA, mathematics, and coding benchmarks, the authors compare the protocol with message passing, an adapted Exchange-of-Thought Report baseline, and message passing combined with Self-Refine. CIPI attains the highest reported scores but generally uses more tokens. A four-arm ablation study examines conditional inspection and opinion selection. The contribution is a concrete recipient-side feedback protocol and a controlled analysis of inspection timing and opinion selection.

Critical strengths
  1. Separating receipt, inspection, and formal adoption is conceptually valuable. The explicit feedback buffer, verified updates, and rollback semantics make the protocol inspectable and clarify exactly what changes between model calls.
  2. Evaluation spans six benchmarks and includes a four-arm ablation study comprising CIPI, CIPI-All, Always-All, and Random-All, with CIPI-All serving as the conditional all-opinion reference. Separating timing from opinion selection makes this design more informative than comparing only the complete system with an always-integrate variant.
Critical weaknesses
  1. The dedicated ablations do not convincingly establish the central benefit of state-dependent timing. On MuSiQue, CIPI-All and Random-All are essentially tied; on OlymMATH, all three all-opinion policies reach 88%.
  2. The main comparisons conflate integration policy with other system differences. Thinking settings vary by role, CIPI uses answer-candidate and citation selectors absent from the baselines, and token consumption is generally higher.
  3. The HotpotQA results use a benchmark-specific diagnostic revision evaluated on tasks previously seen during testing. This change combines answer orientation with guarded skipping rather than isolating the feedback protocol; the APPS adapter also adds a final consistency audit and potential repair. Testing a frozen, unified configuration on new tasks would offer clearer transfer evidence and reduce uncertainty about benchmark-specific development.
Questions for authors

See Weaknesses.

1 comment
AuthorsOpenReview ↗

Response to Reviewer Uc6G

Thank you for the careful summary and for recognizing the separation of receipt, inspection and adoption. In short: (1) the ablation establishes the effect of selective inspection, while the timing contrast in final scores is not resolved at this sample size. (2) Table 1 is a system-level comparison. (3) We will mark the HotpotQA result as a development result and label the APPS audit as CIPI-specific.

Weaknesses

1. The ablations and the benefit of state-dependent timing.

On MuSiQue CIPI-All and Random-All differ by 0.07 F1, and on OlymMATH all three all-comment arms reach 88.00%. The timing intervals include zero (Table 7), and at 100 tasks per arm the detectable difference is about 14 points (Eq. 6). The consistent result of the ablation is the effect of selective inspection under the conditional trigger: full CIPI exceeds CIPI-All on all three datasets at similar token cost. The log analysis of early updates also shows that conditional inspection avoids the harmful early revisions of inspecting at every checkpoint (general response, item 2). We will present selective inspection as the established component effect and the timing result in final scores as not yet resolved.

2. The main comparisons and other system differences.

Item 1 of the general response lists these differences, and we will describe Table 1 as a system-level comparison. The token ratios to MP+SR range from 1.01 (APPS) and 1.04 (OlymMATH) to 1.44 (MATH-500), and we make no general efficiency claim.

3. The HotpotQA revision and the APPS audit.

The HotpotQA efficiency revision was introduced in a diagnostic rerun on tasks that had already been evaluated (Section 4.1). We will therefore mark the HotpotQA row in Table 1 as a development result and base the main conclusions on the other five benchmarks, which were evaluated once with their original configurations. The APPS final consistency audit is part of the CIPI adapter only, and we will label it in Table 1 and in the text. Evaluating one frozen configuration on new tasks is the right test of transfer, and we will list it as future work.

Reviewer_a4C3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CIPI, a receiver-side protocol for conditionally integrating buffered peer feedback during multi-agent LLM reasoning. The protocol combines state-dependent inspection with feedback verification and rollback. The evaluation covers six benchmarks plus a four-arm ablation, looking at both task performance and the effects of feedback timing and selection.

Critical strengths
  1. I like the problem framing. Most existing work on multi-agent LLM reasoning focuses on when agents should communicate, who participates, or how many refinement rounds to run; this paper instead asks when the receiving agent should actually act on what it receives. That's a distinction worth making, and CIPI gives it a concrete form.

  2. The four-arm ablation is a good design choice. It explicitly separates feedback timing from comment selection, which makes the proposed mechanism directly testable rather than bundled into an end-to-end result. Combined with the six benchmarks, this is a more careful empirical setup than most papers in this area attempt.

Critical weaknesses
  1. The main comparisons don't isolate CIPI's contribution. It differs from the baselines in thinking configurations, output limits, compute, and answer validation, and it additionally uses a code-consistency audit on APPS and CIPI-specific answer-targeting and stopping rules on HotpotQA. Section 6 acknowledges the main comparison evaluates complete systems rather than the protocol alone, so the reported gains can't be attributed to conditional feedback integration with any confidence.

  2. The evidence for conditional timing — the core claim — is where the paper falls short. In Table 7, CIPI-All trails Random-All by 0.07 F1 on MuSiQue and ties it on OlymMATH, with confidence intervals covering zero in both cases. The authors' own power analysis (Eq. 6) puts the detectable difference at roughly 14 percentage points for paired binary outcomes under the stated assumptions, which is larger than most of the observed gains. At that sensitivity, the experiments can't really separate state-dependent timing from a random check, so the central claim remains unsupported.

Questions for authors
  1. Can you show that state-dependent inspection beats simpler timing policies when reasoning configuration, budget, and validation are matched across methods?

  2. Of the early interventions CIPI actually triggers, what fraction corrects a genuine error, and how often does it cause an unnecessary or harmful revision?

1 comment
AuthorsOpenReview ↗

Response to Reviewer a4C3

Thank you for the review and for the positive assessment of the problem framing and the four-arm design. In short: (1) the main comparison evaluates complete systems, and the matched four-arm ablation is the component-level evidence. (2) The early-update analysis shows that conditional inspection avoids harmful early revisions, while the final-score timing contrasts are underpowered. (3) Selective inspection under the conditional trigger is the most consistent component effect.

Weaknesses

1. The main comparisons do not isolate CIPI's contribution.

Table 1 is a system-level comparison, and item 1 of the general response lists the differences between CIPI and the baselines. The matched comparison is the four-arm ablation, where all arms share the configuration, verifier and final answer contract and use similar numbers of tokens.

2. Evidence for conditional timing.

In final scores, CIPI-All trails Random-All by 0.07 F1 on MuSiQue and ties it on OlymMATH, with intervals that include zero. Eq. (6) gives a detectable difference of about 14 points at 100 tasks, which is larger than the observed effects, so final scores cannot separate these policies at this sample size. The early-update analysis below shows where the timing policies do differ. We will present the timing result in final scores as not yet resolved and selective inspection as the established component effect (general response, item 2).

Questions

Q1. State-dependent inspection versus simpler timing policies under matched configuration and budget.

Always-All and Random-All are such simpler policies. They keep the configuration, verifier and validation of CIPI, and their token use is within about 10% of CIPI on every ablation dataset (Table 2). In this setting full CIPI scores highest on all three datasets: 70.40 F1 versus 64.66 to 68.40 on MuSiQue, 89.00% versus 88.00% on OlymMATH and 98.00% versus 96.00% to 97.00% on HumanEval. At 100 tasks per arm we report these differences as descriptive.

Q2. How often early interventions correct genuine errors.

We analyzed the MuSiQue ablation logs and labeled the state at each checkpoint as correct or wrong against the reference answer.

  • When the gate fires. All 37 early inspections of CIPI fired at checkpoints where the state was wrong. Comments that are not inspected early stay in the buffer and reach the terminal review.
  • Committed early updates. None of the 29 committed early updates under CIPI broke a correct answer, one corrected a wrong answer, and the mean F1 change was +1.34. Under Always-All, 4 of 56 early commits broke a correct answer, and the mean F1 change was −5.86.

Conditional inspection therefore avoids the harmful revisions that inspecting at every checkpoint introduces, and the terminal review shared by all arms makes most of the final corrections (general response, item 2). We will add this analysis to the appendix.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We thank the reviewers. All three reviews raise two shared points: whether the main comparison isolates conditional integration, and what the ablation shows about state-dependent timing. We address both here and answer the remaining points in each individual response.

1. What the main comparison shows.

Table 1 is a system-level comparison. CIPI differs from the baselines in role-specific thinking settings, the QA answer selector, the APPS consistency audit and the HotpotQA efficiency revision (Section 6). The table shows that the complete protocol reaches the highest score on all six benchmarks, and we will describe it in these terms in the abstract and the introduction. The component-level question is answered by the matched ablation in item 2.

On two benchmarks CIPI reaches its gain at almost the same token cost as the strongest baseline.

BenchmarkCIPIMP+SRToken ratio
OlymMATH87.00% at 37,662 tokens78.00% at 36,354 tokens1.04
APPS90 passed at 40,066 tokens86 passed at 39,582 tokens1.01

On OlymMATH, EoT-Report reaches 67.00% at 39,612 tokens. On the other four benchmarks CIPI uses 29% to 44% more tokens than MP+SR, and we make no efficiency claim there.

2. What the ablation shows.

The four-arm study is the controlled part of the evaluation. All arms share the model, thinking settings, verifier, rollback, memory, terminal review and final answer contract, and differ only in when and which pending comments are inspected. Their token use is close: 42,114 to 43,372 per task on MuSiQue, 30,637 to 33,649 on OlymMATH and 11,383 to 12,078 on HumanEval (Table 2).

  • Comment selection under the conditional trigger. Full CIPI exceeds CIPI-All on all three datasets, by 2.06 F1 on MuSiQue, 1.00 point on OlymMATH and 1.00 point on HumanEval (interval [0.00, 3.00]), and it uses the fewest tokens of the four arms on OlymMATH.
  • Timing with the inclusion rule fixed. CIPI-All exceeds Always-All on MuSiQue (+3.68 F1) and HumanEval (+1.00) and ties it on OlymMATH. Against Random-All it is tied on MuSiQue (−0.07) and OlymMATH (0.00) and ahead on HumanEval (+1.00). The timing intervals include zero (Table 7).

Early updates on MuSiQue (new log analysis). Final scores mix early updates with the terminal review that all arms share, so we labeled every committed early update on MuSiQue against the reference answer. This analysis shows that conditional inspection changes what early updates do.

ArmMean F1 change from early commitsEarly commits that broke a correct state
CIPI+1.340 of 29
CIPI-All+1.350 of 30
Always-All−5.864 of 56
Random-All−3.761 of 27

The terminal review then revises most states in every arm (88 to 90 terminal commits per 100 tasks), which is why these differences are diluted in the final MuSiQue scores. With 100 tasks per arm, Eq. (6) gives a detectable difference of about 14 points, so final scores alone cannot separate timing policies at the observed effect sizes, while the early-update analysis does show the difference in behavior.

Revised framing. The paper will present CIPI as a receiver-side protocol whose complete form reaches the highest scores, report selective inspection under the conditional trigger as the most consistent component effect, report the timing comparison in final scores as not yet resolved at the current sample size, and add the early-update analysis as a new appendix table.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.