- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper studies how to evolve multi-agent systems (MAS) at inference time without relying on external rewards or LLM evaluators. The authors propose SOAP, a framework inspired by the human sense of agency (SoA), which uses agents' predictions of their performance and their post-execution self-assessments to construct internal feedback signals. SOAP leverages two types of signals: individual- and interaction-level SoA to iteratively optimize agent prompts and communication topology. Experiments across four benchmarks demonstrate improvements over the evaluated baselines. Compared with prior MAS evolution methods that often rely on external feedback, the main contribution is the introduction of SoA-based internal feedback to guide test-time adaptation.
Critical strengths
-
The idea of drawing inspiration from the human sense of agency to design internal feedback signals for MAS evolution is interesting. Exploring how agents can assess and improve their own behavior without external evaluators is a meaningful research direction.
-
Extensive empirical evaluation. The paper evaluates the proposed approach across four benchmarks covering different tasks and several LLM backbones. The authors also provide ablation studies to investigate the contributions of individual- and interaction-level SoA.
Critical weaknesses
-
While the human-inspired motivation is interesting, it remains unclear why the specific computational formulation of SoA is more informative than conventional self-assessment signals. In particular, the proposed score (A=p(1-|p-o|)) combines anticipated performance with prediction–outcome agreement, but the justification for this specific combination is limited. For a fixed outcome (o), this score is maximized at (p=(1+o)/2), rather than (p=o), which could favor optimistic predictions over calibrated ones. A clearer discussion of this formulation and its advantages would strengthen the methodological contribution.
-
Although the paper includes SelfOrg (Tastan et al., ICLR 2026) and MetaGen (Wang et al., 2026) as internal-feedback baselines, these approaches involve different MAS evolution mechanisms and do not directly isolate the advantages of the proposed SoA signal over other forms of internal feedback. Existing approaches such as Reflexion and Self-Refine also demonstrate the effectiveness of self-assessment and self-reflection for iterative improvement. I would encourage the authors to consider additional controlled ablations within the same MAS evolution framework, replacing SoA with simpler internal feedback signals, such as (i) post-execution self-assessment alone, (ii) prediction–outcome discrepancy alone, or (iii) trajectory-based self-reflection without SoA. Such experiments would help establish whether the proposed SoA formulation provides additional benefits beyond standard internal feedback mechanisms.
Questions for authors
-
Contribution of SoA versus trajectory-based self-reflection. In Section 3.4.1 and Appendix A.2 (Template 4), the evolution signal appears to be generated by reflecting on execution records, tool observations, predictions, assessed outcomes, and SoA scores. Therefore, the proposed framework seems to leverage not only SoA but also substantial trajectory-based self-reflection information. Could the authors clarify the respective contributions of these two mechanisms? In particular, would the method still achieve similar improvements if the agents performed self-reflection using the same execution trajectories and outcome assessments, but without the proposed SoA scores?
-
Interpretation of interaction-level SoA. The interaction-level SoA compares a sender's prediction of a downstream agent's performance with the receiver's self-assessed outcome. However, this does not necessarily measure the sender's actual contribution to the receiver's performance. Could the authors clarify why this metric is an appropriate measure of interaction-level agency?
9 comments
Response to Weakness 2 and Question 1
Weakness 2: Although the paper includes SelfOrg (Tastan et al., ICLR 2026) and MetaGen (Wang et al., 2026) as internal-feedback baselines, these approaches involve different MAS evolution mechanisms and do not directly isolate the advantages of the proposed SoA signal over other forms of internal feedback. Existing approaches such as Reflexion and Self-Refine also demonstrate the effectiveness of self-assessment and self-reflection for iterative improvement. I would encourage the authors to consider additional controlled ablations within the same MAS evolution framework, replacing SoA with simpler internal feedback signals, such as (i) post-execution self-assessment alone, (ii) prediction–outcome discrepancy alone, or (iii) trajectory-based self-reflection without SoA. Such experiments would help establish whether the proposed SoA formulation provides additional benefits beyond standard internal feedback mechanisms.
Response: Thank you for your valuable comments. Following your suggestions, we further compare SOAP with Reflexion [2], Self-Refine [3], and COPPER [4], representative methods that use self-assessment and self-reflection for iterative improvement. In addition, we conduct further ablations with different feedback signals, as suggested in Weakness 2 and Question 1, to justify the effectiveness of the SoA feedback used in our method.
Table 3.2: Comparison with self-assessment and self-reflection methods. Results are reported as mean ± SD over three runs; on CloudCast-Haiku, n=10 denotes 10 trials.
| Method | BrowseComp-Plus | WorkBench | PlanCraft | CloudCast-Haiku (n=10) |
|---|---|---|---|---|
| SOAP | 0.630 ± 0.045 | 0.671 ± 0.030 | 0.797 ± 0.069 | 0.110 ± 0.011 |
| Reflexion | 0.533 ± 0.015 | 0.537 ± 0.025 | 0.767 ± 0.047 | 0.052 ± 0.014 |
| Self-Refine | 0.543 ± 0.051 | 0.568 ± 0.021 | 0.770 ± 0.036 | 0.095 ± 0.084 |
| COPPER | 0.563 ± 0.093 | 0.612 ± 0.024 | 0.777 ± 0.076 | 0.078 ± 0.061 |
Observations: From the results, we can observe that SOAP outperforms these three representative methods based on self-assessment and self-reflection. This further strengthens the empirical evidence for the superiority of our proposed method, and validates the effectiveness of extending the concept of "sense of agency" to encourage the agent to better understand what it can do and how its behavior will affect the outcome during evolution.
Table 3.3: Ablation on the feedback signal used for evolution. Results are reported as mean ± SD over three runs; on CloudCast-Haiku, n=10 denotes 10 trials.
| Feedback Signal | BrowseComp-Plus | WorkBench | PlanCraft | CloudCast-Haiku (n=10) |
|---|---|---|---|---|
| Outcome only () | 0.577 ± 0.032 | 0.563 ± 0.007 | 0.750 ± 0.036 | 0.092 ± 0.106 |
| Prediction only () | 0.597 ± 0.058 | 0.548 ± 0.012 | 0.723 ± 0.032 | 0.048 ± 0.022 |
| Agreement only () | 0.613 ± 0.040 | 0.611 ± 0.034 | 0.750 ± 0.089 | 0.053 ± 0.007 |
| Trajectory | 0.537 ± 0.032 | 0.583 ± 0.012 | 0.753 ± 0.021 | 0.063 ± 0.017 |
| SOAP | 0.630 ± 0.045 | 0.671 ± 0.030 | 0.797 ± 0.069 | 0.110 ± 0.011 |
Observations: From the results, we can observe that, compared with post-execution self-assessment alone, prediction–outcome discrepancy alone (i.e., ), and trajectory-based self-reflection without SoA, our method still achieves the best performance on all four benchmarks. This further justifies the effectiveness of leveraging the concept of "sense of agency", which encourages the agent to evolve toward a stronger sense of how it can behave and how its behavior will affect the outcome.
[2] Shinn et al. "Reflexion: Language Agents with Verbal Reinforcement Learning." Advances in Neural Information Processing Systems 36 (NeurIPS), 2023.
[3] Madaan et al. "Self-Refine: Iterative Refinement with Self-Feedback." Advances in Neural Information Processing Systems 36 (NeurIPS), 2023.
[4] Bo et al. "Reflective Multi-Agent Collaboration based on Large Language Models." Advances in Neural Information Processing Systems 37 (NeurIPS), 2024.
Question 1: Could the authors clarify the respective contributions of these two mechanisms? In particular, would the method still achieve similar improvements if the agents performed self-reflection using the same execution trajectories and outcome assessments, but without the proposed SoA scores?
Response: Thank you for your valuable questions. As presented in our response to Weakness 2 (Table 3.3), we conduct additional ablation studies to validate the contribution of SoA scores. We will also revise the manuscript accordingly.
Response to Question 2
Question 2: Interpretation of interaction-level SoA. The interaction-level SoA compares a sender's prediction of a downstream agent's performance with the receiver's self-assessed outcome. However, this does not necessarily measure the sender's actual contribution to the receiver's performance. Could the authors clarify why this metric is an appropriate measure of interaction-level agency?
Response: Thank you for your valuable question. We agree that interaction-level SoA does not isolate the sender's causal contribution to the receiver's performance, and we state this in Appendix A.3. The receiver's outcome also depends on its other inputs and its own execution. Below, we clarify what the metric measures, how the sender's contribution is measured separately, and why we adopt this design.
(1) What interaction-level SoA measures. What the score captures is whether a sender can anticipate the downstream effect of its output. In agency research, the match between a predicted and an obtained consequence of one's action is a cue to agency, and the interaction-level score applies this cue to a consequence located in another agent. A sender with low interaction-level SoA either expects little of the receiver or misjudges how its output will be used, and both cases indicate a handoff worth revising.
(2) How the sender's contribution is measured. Contribution is measured by a separate quantity. Topology edits combine the accumulated interaction-level SoA with marginal information coverage, which measures how much a sender's output covers the receiver's unmet rubric dimensions beyond what other sources provide (Appendix A.5). An edge is removed only when both are low, and low SoA alone does not justify removal when the sender still supplies needed information. The two signals are therefore complementary: SoA indicates whether the interaction behaves as the sender expects, and marginal coverage indicates whether the sender adds information.
(3) Why we adopt this design. We chose this design over a counterfactual assessment with the sender's message removed, which would require an additional execution per edge in every round. The ablation in Figure 4 shows that removing interaction-level SoA lowers performance on all four benchmarks. We will move the statement on causal contribution and the role of marginal coverage from the appendix into Section 3.4.2.
Response to Weakness 1
Thank you for your review. We provide further clarification and additional results below to address your concerns. Please let us know if we have misunderstood any comments, and we will respond promptly.
Response: Thank you for your valuable comments. Your observation is correct: for a fixed o, the SoA score is maximized at p∗=(1+o)/2. We deliberately design the score to prefer predictions somewhat higher than the assessed outcome, and we explain the formulation below.
(1) Why the two terms are combined. The agreement term 1−∣p−o∣ corresponds to the comparator cue in agency research, i.e., the match between predicted and obtained outcomes. Used alone, it is satisfied by an agent that expects little and achieves little, so it places no pressure on performance. The expectation term p supplies that pressure, but alone it would reward unconstrained optimism. Their product rewards high expectations only to the extent that execution supports them. A conventional self-assessment signal uses o alone and cannot distinguish an anticipated outcome from an unanticipated one. In Appendix A.3, the mean assessed outcome is 0.80 while SoA is 0.60, because the agent fulfilled one criterion it had not expected to. The reflection step acts on this discrepancy.
(2) Why the optimum p* lies above o. As stated in the introduction, agents are encouraged to expect better performance and revise their behavior to match those expectations. The optimum p∗ is the midpoint between the current outcome and full achievement, so the score favors a target above current performance by a bounded margin (1−o)/2 that vanishes as o approaches 1. Predictions far above the outcome are penalized by the agreement term. A calibration-only objective with its optimum at p=o would provide no such direction. This choice is consistent with findings on human learning: Chambon et al. [1] report a positive learning asymmetry specific to self-determined choices and show in simulation that it improves performance.
(3) Whether the components behave as intended. If the margin acts as a target, the assessed outcome should rise toward the prediction, and the gap should narrow. Figure 3 shows this pattern for the verifier agent on BrowseComp-Plus: predictions and outcomes both increase, and the signed gap is generally smaller than at the start. We further report the per-round signed gap pˉ−oˉ against the predicted optimal value (1−oˉ)/2, averaged over all agents and queries:
Table 3.1: Per-round signed gap pˉ−oˉ and the predicted optimal value (1−oˉ)/2 on BrowseComp-Plus.
Observations: From the results, we can observe that: (1) the gap pˉ−oˉ quickly converges near the optimum and then fluctuates slightly above it, indicating that the agents evolve in the direction that increases the current-round SoA score; and (2) p always stays somewhat above o, which empirically supports that the agents use a higher p to guide their capabilities and thereby improve the actual outcome o. We will also add the results in the manuscripts.
[1] Chambon et al. Information about action outcomes differentially affects learning from self-determined versus imposed choices. Nature Human Behaviour