acceptodds
Under review as a conference paper at ICLR 2027

TrustFlow: Counterfactual Information-flow Control for Reliable LLM-based Multi-Agent Collaboration

Abstract

LLM-based multi-agent systems improve complex reasoning by allowing multiple agents to exchange information and iteratively refine their solutions. However, unrestricted information sharing can also propagate unreliable reasoning and lead to error amplification and false consensus. Existing approaches largely focus on judging information reliability from current signals, such as agent confidence, consensus patterns, or external verification, but pay less attention to how information should be controlled to improve future collaboration outcomes. We propose TrustFlow, a counterfactual information-flow control framework for reliable LLM-based multi-agent collaboration. TrustFlow formulates collaboration control as a dynamic topology selection problem. First, it estimates a joint reliability belief over participating agents from observable interaction histories, providing a structured state beyond independent trust scores. Second, it evaluates alternative information-flow topologies through counterfactual rollouts and estimates their downstream effects. Third, it trains an outcome-aware controller with feedback-based policy optimization to select interventions, including preserving current reasoning, peer verification, and independent rethinking. Experiments across diverse reasoning benchmarks show that TrustFlow consistently improves reasoning performance and robustness while reducing inference cost.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Models, Multi-Agent Reasoning, Uncertainty Propagation, Information Flow Control

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_L9B9OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper introduces TrustFlow, a collaborative framework for LLM multi-agent systems based on counterfactual information flow control. The framework employs a joint reliability belief estimator to assess the joint correctness states of multiple agents and identify those agents requiring intervention. It also evaluates the downstream benefits of various communication topologies and intervention actions. The authors utilize GRPO-style policy optimization to train a controller that dynamically adjusts information flow based on the current collaborative state. They conduct experiments on MATH, MMLU, HumanEval, and TruthfulQA, showing that this method improves average reasoning performance, enhances error recovery capabilities, and reduces the consumption of reasoning tokens.

Critical strengths
  • The paper introduces a joint reliability belief over agent correctness configurations, potentially capturing correlated errors beyond independent trust scores.
  • The paper evaluates multiple reasoning benchmarks and includes ablation, robustness, recovery, and efficiency analyses, demonstrating promising performance and reduced inference token consumption.
Critical weaknesses
  • The method's novelty is limited. The proposed framework mainly combines existing reliability estimation, counterfactual assessment and strategy optimization techniques. Specifically: (i) Although the joint reliability confidence is intended to capture the correlation between agent errors, the selection of intervention anchors is based on marginal accuracy, and no comparison with independent reliability estimates is provided to demonstrate the necessity of joint modeling; (ii) The GRPO-based counterfactual controller is built on mature reward-driven topology optimization and policy learning techniques, but the paper lacks comparative experiments with simpler outcome-based policy optimization methods, failing to establish the unique advantages of counterfactual feedback and GRPO.
  • The core contributions have not been experimentally verified. On MMLU (+0.50%), HumanEval (+1.22%), and TruthfulQA (+0.72%), the performance improvements relative to the strongest baseline were negligible and lacked sufficient statistical significance analysis. More importantly, ablation experiments failed to distinguish the advantages of counterfactual feedback and GRPO from simpler outcome-based optimization methods. Furthermore, the efficiency improvements in the report overlooked the additional training costs associated with counterfactual rollouts.
Questions for authors
  1. Compared with existing counterfactual and reward-based topology optimization methods, what are the key methodological advancements of TrustFlow? Could the authors provide comparative experiments with independent reliability assessments and simpler strategy optimization alternatives to demonstrate the necessity of the proposed component?
  2. Could the authors provide statistical significance analysis and additional ablation experiments to compare counterfactual GRPO training with standard outcome-based policy learning? How much additional computational overhead does performing counterfactual rollouts during training incur?
3 comments

Thank you for your feedback.

W1 & Q1. Methodological Novelty and Component Necessity

(1) Methodological contributions.

TrustFlow goes beyond conventional reward-driven topology optimization through two complementary mechanisms:

  • Joint Reliability Modeling: Instead of treating agent correctness estimates as independent, TrustFlow models their joint reliability distribution, enabling the representation of correlated errors among collaborating agents.
  • Belief-Conditioned Counterfactual Control: Rather than optimizing communication decisions solely from final task outcomes, TrustFlow incorporates counterfactual evaluations of alternative information-flow interventions to provide additional training feedback.

The key distinction is the integration of joint reliability beliefs and counterfactual intervention feedback into adaptive information-flow control.

(2) Joint vs. Independent Reliability Estimation.

To directly examine the necessity of joint reliability modeling, we conducted additional comparisons against independent agent reliability estimation.

DatasetJoint ReliabilityIndependent ReliabilityImprovement
HumanEval78.89 ± 1.92%73.33 ± 3.33%+5.56 pp
TruthfulQA60.42 ± 1.80%58.85 ± 2.39%+1.57 pp

The joint reliability approach consistently outperforms independent estimation on both evaluated datasets, with a particularly notable improvement on HumanEval. These results provide empirical support for modeling dependencies among agent correctness states rather than relying exclusively on independent reliability estimates.

(3) Counterfactual vs. Outcome-Based GRPO.

We further evaluated a standard outcome-based GRPO baseline to isolate the contribution of counterfactual feedback from that of GRPO optimization itself.

DatasetTrustFlow (Counterfactual GRPO)Outcome-Based GRPOImprovement
HumanEval77.78 ± 1.92%75.56 ± 1.93%+2.22 pp

Counterfactual GRPO outperforms outcome-based GRPO by 2.22 percentage points, demonstrating the benefit of counterfactual feedback beyond final task rewards.

(4) Reliability-only Topology Selection.

To address the reviewer's concern, we introduce a training-free reliability-only baseline that selects communication topologies directly from estimated agent reliability scores, without counterfactual feedback or a learned Controller. The experimental results on TruthfulQA are summarized below:

MethodAccuracy (%) ↑
TrustFlow (Ours)59.66 ± 1.21
Reliability-only58.58 ± 1.22

Over three runs on 64 questions, TrustFlow outperforms the reliability-only baseline by 1.08 percentage points, suggesting benefits beyond reliability-based topology selection.

Q2. Statistical Significance, Ablation Studies, and Computational Overhead

(1) Statistical reliability and additional ablations. To address the reviewer's concerns, we conducted supplementary evaluations with three repeated runs per method. On TruthfulQA, TrustFlow achieves 59.37 ± 1.56%, outperforming EIB Learner. We further evaluated 50 challenging MATH problems (difficulty Levels 4–5) to examine performance on complex reasoning tasks, where TrustFlow achieves 46.67 ± 1.15%, compared with 44.67 ± 2.06% for EIB Learner.

DatasetTrustFlow (Ours)EIB LearnerImprovement
MATH (Levels 4–5, 50 problems)46.67 ± 1.15%44.67 ± 2.06%+2.00 pp
TruthfulQA59.37 ± 1.56%58.33 ± 0.90%+1.04 pp

(2) End-to-end inference efficiency. Figure 5 reports inference token consumption across three rounds of multi-agent discussion and Controller execution. We additionally measured end-to-end inference latency under identical hardware and model settings. TrustFlow reduces average latency by 25.25% and 18.76% compared with LLM Debate and DoV-Follower, respectively. Counterfactual rollouts are performed only during offline training, while inference directly uses a lightweight Belief Estimator and a compact Qwen3-1.7B Controller, avoiding rollout overhead during deployment.

MethodLatency (s) ↓
TrustFlow (Ours)29.58 ± 1.09
LLM Debate39.57 ± 18.89
DoV-Follower36.41 ± 14.08

(3) Training Overhead. On TruthfulQA with an A100 GPU and 600 matched optimizer updates, counterfactual GRPO required 3.957 hours, versus 2.586 hours for outcome-based GRPO (+53%). Most additional time (79.2 of 82.3 minutes) came from counterfactual rollout collection, which can be parallelized or precomputed offline. Thus, the overhead is manageable and does not affect inference efficiency.

We thank the reviewer for their constructive feedback. All additional results and clarifications will be included in the revised PDF.

Thank you for your reponse. I have raised my score.

Thank you for the update and for reconsidering our response. We hope our response reflects the rigor and care we have brought to this work, and we will address the remaining concerns carefully in the revision.

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

TrustFlow dynamically regulates communication among LLM agents to mitigate error propagation and false consensus. It combines joint reliability estimation, counterfactual topology evaluation, and a GRPO-trained controller. Experiments on four benchmarks report improved accuracy and reduced token consumption, supplemented by component ablations, robustness tests, and scalability analyses.

Critical strengths

A valuable problem formulation. TrustFlow considers how information should propagate between agents, incorporating the anticipated downstream consequences of communication decisions. This provides a useful perspective on correlated errors and false consensus. Promising empirical results. The method improves performance across four benchmarks, and the component ablations provide initial evidence that reliability estimation and counterfactual feedback contribute to these gains.

Critical weaknesses

The connection between the theoretical analysis and the deployed policy is insufficiently established. Theorems 2–3 analyze topology selection through finite-rollout utility estimation, whereas the implemented decision policy is a GRPO-trained controller. It remains unclear whether inference performs these rollouts and how the guarantees apply to the learned policy. Moreover, the component-removal ablations do not isolate the benefits of joint reliability modeling over independent agent scoring, or counterfactual selection over a simpler reliability-based policy. These gaps limit the evidence supporting the method’s central design choices.

The evidence for a reliable, cost-effective advantage is incomplete. Gains over EIB Learner are modest on three benchmarks (0.50–1.22 percentage points), without reported uncertainty estimates establishing their robustness. Token counts alone also do not establish overall efficiency: the accounting for reliability estimation, controller execution, and counterfactual rollouts is unclear, and end-to-end latency and training costs are not reported. Comparisons against simpler alternatives under matched computational budgets are needed to assess whether the added complexity is justified.

Questions for authors

The paper explains how counterfactual rollouts provide controller-training feedback. Are they also performed at inference? What value of MM is used, and how do the guarantees for rollout-based topology selection relate to the deployed GRPO controller?

Please clarify the answering-agent and controller backbones, the agent/round configuration for Table 1, and whether both training datasets are disjoint from evaluation data. What repetitions underlie the reported mean ± std, and which modules are included in token accounting?

Can the authors compare against independent reliability estimation and reliability-only topology selection under matched budgets, including end-to-end inference cost? These comparisons would materially affect my assessment of the method’s contribution.

2 comments

We thank the reviewer for their constructive feedback and recognition of TrustFlow's contributions. We address concerns regarding component necessity, theoretical guarantees, and computational trade-offs below, and will incorporate the additional evidence and clarifications into the revised manuscript.

P1. Necessity of Joint Reliability Modeling and Counterfactual Evaluation

(1) Joint vs. Independent Reliability Estimation. We compare joint reliability estimation with independent agent scoring on HumanEval and TruthfulQA. Our insight is that correlated agent errors may reinforce incorrect consensus, which independent reliability scores cannot adequately capture. This comparison evaluates the benefit of modeling cross-agent dependencies.

DatasetJoint ReliabilityIndependent ReliabilityImprovement
HumanEval78.89 ± 1.92%73.33 ± 3.33%+5.56 pp
TruthfulQA60.42 ± 1.80%58.85 ± 2.39%+1.57 pp

Joint reliability estimation consistently outperforms independent scoring by 5.56 and 1.57 percentage points on HumanEval and TruthfulQA, respectively, supporting our design choice.

(2) Counterfactual Evaluation. TrustFlow uses counterfactual rollouts only during offline training. At inference, the trained controller directly selects communication actions without additional rollouts. For reliability-only topology selection, we introduce a training-free baseline that selects topologies using only agent reliability scores, without counterfactual feedback or a learned controller. Results are shown below:

MethodAccuracy (%) ↑
TrustFlow (Ours)59.66 ± 1.21
Reliability-only58.58 ± 1.22

Results are reported as mean ± standard deviation over three runs on 64 evaluation questions in TruthfulQA. TrustFlow outperforms the reliability-only baseline by 1.08 percentage points, providing preliminary evidence that downstream-aware topology control offers benefits beyond reliability-based selection alone.

P2. Relationship Between Theoretical Guarantees and the GRPO Controller

(1) Counterfactual Rollouts During Training. Offline counterfactual rollouts estimate topology utility Qt(g)Q_t(g), providing group-relative advantages for GRPO optimization (Section 3.5.3). During inference, the trained controller directly selects interventions without additional rollouts.

(2) Rollout Budget. M=5M=5 denotes independent counterfactual rollouts per candidate topology, not the number of candidates.

(3) Scope of Theoretical Guarantees. Theorem 2 bounds utility estimation and selection errors within a fixed candidate set, while Theorem 3 specifies the rollout budget needed to identify the optimal candidate under stated assumptions. Both support reliable counterfactual training feedback.

P3. Reliability, Statistical Significance, and Computational Efficiency

(1) Statistical Reliability. We conducted three repeated runs on MATH and TruthfulQA, focusing on challenging MATH Levels 4–5 to better assess the benefits of adaptive information-flow control beyond the LLM's inherent reasoning ability.

The supplementary results are summarized below:

DatasetTrustFlow (Ours)EIB LearnerImprovement
MATH (Levels 4–5, 50 problems)46.67 ± 1.15%44.67 ± 2.06%+2.00 pp
TruthfulQA59.37 ± 1.56%58.33 ± 0.90%+1.04 pp
TrustFlow outperforms EIB Learner by 2.00 and 1.04 percentage points on MATH (Levels 4–5) and TruthfulQA, respectively, further supporting its effectiveness across reasoning tasks.

(2) End-to-end computational accounting. Figure 5 reports inference token consumption, including three rounds of multi-agent discussion and Controller execution. We further evaluated end-to-end inference latency under identical local hardware and model settings, in this experiment, all model achieves 100% accuracy for 15 testing questions in TruthfulQA, thus, only Latency is reported below:

MethodLatency (s) ↓
TrustFlow (Ours)29.58 ± 1.09
LLM Debate39.57 ± 18.89
DoV-Follower36.41 ± 14.08

TrustFlow reduces average latency by 25.25% and 18.76% compared with LLM Debate and DoV-Follower, respectively. Counterfactual rollouts are restricted to offline training.

P4. Implementation Details and Reproducibility

We clarify that Qwen3-8B serves as frozen answering agents, while Qwen3-1.7B is the LoRA-tuned GRPO Controller. Table 1 uses 3 agents and 3 rounds. Training and evaluation data are strictly disjoint, and all results are reported as mean ± standard deviation over 3 runs.

We greatly appreciate the reviewer's insightful comments and suggestions, the revised PDF will incorporate additional results and clarifications to address these concerns.

Dear Reviewer ceCc,

Thank you again for your time and effort in reviewing our paper. We are writing to kindly follow up on our rebuttal response, as the discussion period is approaching its end.

We hope our clarifications and the additional experiments have sufficiently addressed your concerns. Please let us know if you need any further information from our side.

Best regards,
The Authors

Reviewer_3fxBOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TrustFlow to mitigate error propagation in LLM-based multi-agent collaboration. The framework estimates a joint belief over agent correctness from dialogue histories and answer similarity, evaluates candidate interventions through counterfactual rollouts, and trains a controller using GRPO to select among preserving current reasoning, peer verification, and independent rethinking. Experiments cover MATH, MMLU, HumanEval, and TruthfulQA, alongside component ablations, scalability analyses, and error-recovery evaluations. The reported results show higher mean performance than the listed baselines on all four benchmarks and 26.9%–42.9% lower token consumption relative to the evaluated multi-agent baselines, providing preliminary support for improving collaboration reliability through information-flow control

Critical strengths
  1. The recovery-oriented intervention design is practically meaningful. By combining information isolation, selective peer verification, and independent rethinking, the framework provides a reasonable mechanism for escaping erroneous consensus.
  2. The theoretical analysis motivates intervention and candidate evaluation. The paper analyzes how interaction can amplify an incorrect majority and establishes bounds on rollout-based selection error and sufficient sampling budgets under explicit assumptions. These results provide conditional theoretical support for controlling information flow and comparing intervention outcomes.
  3. The cross-task experiments provide preliminary empirical support. Mean performance improves across four task categories, complemented by recovery analyses involving initially incorrect responses, including cases where all agents begin with incorrect answers. These results demonstrate the framework’s potential to improve collaboration reliability.
Critical weaknesses
  1. The rollout budget and cost accounting are insufficiently specified. Section 3.4.1 does not report the actual number of rollouts (M) per candidate topology. Appendix D only caps the candidate set at 30 topologies when (N>6), leaving the total simulation budget undetermined. Whether rollouts are performed during inference and which components are included in Figure 5’s cost estimates are also unclear.

  2. Key controller implementation details are missing. The paper describes an LLM controller optimized through GRPO-based LoRA updates but does not specify its backbone, state-input and action-output formats, or key training configurations. Qwen3-8B is identified as the belief estimator’s encoder, which does not establish the controller’s backbone. These omissions hinder reproduction and assessment of the comparison conditions.

  3. The ablations do not isolate the core mechanisms’ contributions. Table 2 does not specify the evaluation data or the replacement strategies used when removing components, and no direct comparison is provided between joint beliefs and independent marginal estimates. Consequently, the results do not clearly distinguish the benefits of joint modeling and counterfactual comparison from those of learned control more generally.

Questions for authors
  1. What is the actual number of rollouts (M) per candidate topology? Are rollouts also performed during inference? Please clarify whether Figure 5 accounts for the controller, belief estimator, and any inference-time rollout costs.

  2. Which backbone does the controller use? Please specify its state-input format, action-output format, topology construction procedure, and key LoRA/GRPO training configurations.

  3. What data and evaluation settings are used in Table 2, and what strategies replace the removed components? Under matched models and computational budgets, does joint belief modeling provide a clear benefit over independent marginal estimates?

1 comment

We sincerely thank you for the constructive feedback and for recognizing the practical significance of our intervention design, theoretical analysis, and cross-task empirical results. We address the three concerns below.

Q1. Rollout Budget and Computational Cost

Counterfactual rollouts are performed during offline controller training, rather than during online inference. Specifically, each candidate topology is evaluated using (M) independent rollouts to estimate its downstream utility, which is subsequently converted into group-relative advantages for GRPO optimization.

The actual rollout budget is M = 5 per candidate topology. The maximum of 30 topologies mentioned in Appendix D refers to the candidate-set size when (N(the number of agent)>6), not the rollout count per candidate. During inference, the trained controller directly predicts communication interventions without additional counterfactual rollouts. This design shifts the computationally expensive counterfactual evaluation to offline training.

Regarding Figure 5, the reported 13.69k tokens represent the total token consumption during online inference, including the three rounds of multi-agent discussion and controller execution. Counterfactual rollout costs are not included, as rollouts are performed exclusively during offline training and are not required at inference time.

Q2. Controller Architecture and Training Details

TrustFlow employs Qwen3-8B as the frozen answering-agent backbone and Qwen3-1.7B as the controller backbone, optimized through GRPO with LoRA fine-tuning.The LoRA configuration uses rank (r=8), alpha=16. Training uses AdamW with a learning rate of 5 \times 10^{-6} for 3 epochs.

The controller takes the estimated reliability belief, interaction statistics, and intervention history as input, and dynamically selects an information-flow topology for subsequent collaboration. The resulting topology regulates inter-agent communication through selective information sharing, peer verification, and independent rethinking, enabling adaptive control of information propagation.

Q3. Ablation Settings and Joint vs. Independent Reliability

Table 2 uses the HumanEval dataset. Specifically, w/o Belief Estimator allows the LLM Controller to directly select communication topologies based on dialogue history without reliability estimation; w/o LLM Controller replaces the LLM-based Controller with a lightweight MLP for topology selection; and w/o Counterfactual Feedback: The Controller is optimized using the improvement in average estimated agent reliability after each intervention, rather than counterfactual downstream utility, while retaining the same Belief Estimator and GRPO framework.

We appreciate the reviewer's suggestion to isolate the contribution of joint reliability modeling beyond the existing component-removal ablations.To directly address this concern, we conducted additional comparisons between joint reliability estimation and independent agent reliability scoring on HumanEval and TruthfulQA datasets.

DatasetJoint ReliabilityIndependent ReliabilityImprovement
HumanEval78.89 ± 1.92%73.33 ± 3.33%+5.56 pp
TruthfulQA60.42 ± 1.80%58.85 ± 2.39%+1.57 pp

Results are reported as mean ± standard deviation over three runs. Joint reliability consistently outperforms independent scoring on both benchmarks, particularly on HumanEval (+5.56 pp), supporting the benefit of modeling inter-agent dependencies in mitigating correlated errors and false consensus.

We sincerely thank the reviewer for valuable and constructive feedbacks. We will incorporate all supplementary experimental results and further clarifications into the revised PDF to improve the clarity, rigor, and reproducibility of our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.