acceptodds
Under review as a conference paper at ICLR 2027

Beware of Memory-Induced Overconfidence: Commitment–Competence Calibration for Multi-Agent Systems

Abstract

Large Language Model-based Multi-Agent Systems have demonstrated impressive capabilities in solving complex tasks through role specialization, memory sharing, and collaborative reasoning. To further enhance their long-term performance, recent studies have introduced self-evolution mechanisms, where agents summarize reflections from historical interactions and reuse them in subsequent tasks. However, we observe that self-evolution does not always lead to consistent improvement: in some cases, apparently relevant memories may provide little benefit for the current task while repeatedly steering agents toward the same ineffective strategy. The key insight is that memory injection has two distinct effects: memory commitment, which characterizes how strongly memory influences behavior by reducing uncertainty about how to act, and memory competence, which estimates how reliably the selected memories support successful execution. These effects can be misaligned: memory may impose a clear and consistent behavioral direction while exhibiting low historical execution reliability. We identify this failure mode as memory-induced overconfidence. To mitigate this challenge, we propose C2Mem, a memory calibration framework that dynamically attenuates apparently relevant memories that exhibit low execution utility. Experiments across multiple benchmarks show that C2Mem reduces the propagation of non-beneficial memory guidance, mitigates memory-induced overconfidence, and improves the stability and reliability of multi-agent systems. Our code is available at https://anonymous.4open.science/r/c2mem-27F3.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Memory, Agent, Multi Agent System

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_geeFOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies memory-induced overconfidence in long-running multi-agent systems, where historical memories may strongly and consistently guide agent behavior despite having low reliability in previous tasks. To address this problem, the authors propose C2Mem, which separates memory influence into memory commitment and memory competence. Commitment measures the strength of memory guidance based on retrieval support and structural consistency in a behavioral tree, while competence estimates node-level reliability from historical successes and failures. When C2Mem detects memories with high commitment but low competence, it reduces their weights before injecting them into the agents’ contexts. Historical experiences are summarized by an LLM as reusable behavioral patterns and organized in a hierarchical behavioral tree. After each task, node reliability is updated according to the final outcome and the degree of memory exposure. Experiments across three multi-agent frameworks and five question-answering and reasoning benchmarks show that C2Mem achieves the best average performance in all three frameworks. Ablation studies, trajectory analysis, and case studies further support the effectiveness of memory reliability calibration.

Critical strengths
  1. The paper identifies an important but underexplored problem: semantically relevant memories may provide strong and consistent behavioral guidance even when they are unreliable. By separating memory influence into commitment and competence, the paper offers a clear and novel perspective for understanding and mitigating memory-induced overconfidence.

  2. C2Mem provides a complete memory calibration mechanism with broad empirical support. It connects a behavioral tree, historical reliability posteriors, and pre-injection weight calibration. The method achieves strong average performance across three multi-agent frameworks and five different tasks, while the ablation studies, trajectory analysis, and case studies provide additional evidence for the contributions of its main components.

Critical weaknesses
  1. Memory competence does not accurately measure the true utility of an individual memory. The paper updates competence using the final task outcome and node exposure, but task success or failure may also be caused by other memories, agents, or tool outputs. Therefore, competence mainly captures a correlation between memory usage and task outcomes rather than whether a particular memory is beneficial. The unknown node is also not a true no-memory counterfactual, which weakens its role as a reference baseline.

  2. Commitment and competence are not clearly separated in the implementation. The LLM evaluator that generates memory-support weights can see each node’s historical_competence, while these weights are later used to calculate commitment. Competence may therefore already influence commitment and then be used again during calibration, weakening the paper’s central two-dimensional formulation.

Questions for authors
  1. What are the node-splitting criterion, clustering method, maximum tree depth, Top-K value, and tree-distance normalization method? Have the authors verified that tree distance accurately represents consistency between behavioral strategies?

  2. C2Mem relies on benchmark ground-truth answers to determine whether historical tasks succeed. How would memory competence be estimated and updated for open-ended tasks with no ground truth or only noisy feedback?

  3. Calibrated weights affect subsequent exposure, which is then used to update competence. How does the method prevent a node that is incorrectly downweighted early from receiving too little new evidence to recover?

  4. Could the authors add more direct ablations, such as a flat memory structure, hiding historical_competence from the support evaluator, and using a true no-memory reference, to separately verify the necessity of the behavioral tree, the separation of the two dimensions, and the unknown node?

5 comments

We sincerely thank the reviewer for the thoughtful feedback on reliability estimation, the two-dimensional formulation, and memory updates. Below, we provide point-by-point responses with additional experimental evidence and implementation details.


Weakness 1(1) Competence Estimation and Intervention Effects

Thank you for raising this important question. An advantage of C²Mem is that memory reliability estimates can be continually revised as new execution evidence becomes available. Even when a task fails because of other memories, agents, or tool outputs, that outcome contributes only part of a node’s accumulated evidence. Subsequent exposure in successful executions can still increase its competence and correspondingly reduce the calibration penalty.

We randomly select 200 HotpotQA tasks where calibration changes memory weights under AutoGen + C²Mem (GPT-4o, fixed behavioral tree). From the same execution prefix at the calibration step, we compare uncalibrated memories, removal of downweighted memories, and calibrated memories, holding other conditions fixed. Removal improves performance, while calibration yields further gains. These controlled comparisons support competence as a basis for memory intervention, beyond its association with task outcomes.

Memory SettingEM (%)F1 (%)Average Steps
Uncalibrated memories37.5047.248.97
Remove downweighted memories43.0053.236.29
Calibrated memories50.5063.515.16

Section 3.8 and Appendix B, Figure 9 provide a detailed case study showing how calibration downweights low-competence guidance, allowing identity verification to redirect execution while preserving later reliable guidance.


Weakness 1(2) & Question 4(3) Validity of the Unknown-Node Reference

The unknown node provides a reliability reference learned from training experience. Both ordinary nodes and the unknown node update their competence using accumulated exposure and task outcomes. To examine whether this reference corresponds to actual memory utility, we conduct same-prefix continuation experiments.

On 200 HotpotQA tasks with AutoGen, we freeze the behavioral tree and reliability posteriors and continue each task from the same execution prefix under three conditions: injecting a memory with competence below μ0\mu_0, injecting one above μ0\mu_0, or injecting no memory. The two memory conditions use the selected memories directly without additional calibration, with all other execution conditions held constant.

Memory ConditionEM (%)F1 (%)Average Steps
μv<μ0\mu_v<\mu_034.5038.948.18
No memory40.5048.585.26
μv>μ0\mu_v>\mu_045.5057.195.80

Under the same tasks and execution prefixes, memories below the reference reduce performance, while those above it improve performance. These results support μ0\mu_0 as a useful comparison boundary in the evaluated states, helping distinguish memory guidance worth relying on from guidance that should be suppressed.


Weakness 2 & Question 4(2) Distinct Roles of Commitment and Competence

Thank you for your insightful observation. C²Mem examines memory from two perspectives—current guidance strength and historical execution reliability—which do not require entirely independent inputs. historical_competence provides the support evaluator with execution history to inform its assessment of behavioral compatibility with the current state. In computing competence, the same information quantifies the reliability of these memories. The two quantities draw on shared information but answer different questions.

Specifically, commitment depends on current support and the structural relationships between nodes (Eq. 6), whereas competence aggregates reliability estimates derived from historical successes and failures (Eq. 7). For example, memories with the same historical reliability may jointly support a coherent strategy or point toward different behavioral directions. They can therefore have identical aggregated competence but different commitment. This distinction allows calibration to apply different intervention strengths to memories with equal reliability but different levels of current guidance.

Question 1 Behavioral Tree Construction and Validity

The behavioral tree groups similar execution strategies across tasks to share reliability evidence. An LLM summarizes 30 memory-free training trajectories into three initial behavioral patterns. Experiences are inserted by behavioral similarity. Every nn new trajectories per node (default n=10), an LLM decides whether and how to split it; children inherit the parent’s Beta parameters proportionally to their experience counts (Appendix A.2, Eqs. A.1–A.2). Depth is unrestricted to avoid forcing distinct strategies to share a reliability estimate. Retrieval uses top-K = 3 nodes plus the unknown node. Shortest-path distances are divided by the current tree’s maximum shortest-path length, yielding values in [0,1].

To test whether tree distance reflects behavioral compatibility, we compare node summaries and conduct same-prefix experiments on 200 HotpotQA tasks. We inject memory pairs at different distances individually or jointly, measuring next-action agreement between individual branches and average agreement between the joint action and the two individual actions. Closer memories have more similar summaries and higher agreement under both measures, supporting tree distance as a measure of behavioral coherence.

Normalized Tree DistanceMean Summary Cosine DistancePairwise Agreement (%)Joint–Individual Agreement (%)
(0,1/3]0.16854.768.9
(1/3,2/3]0.17645.261.7
(2/3,1]0.23324.145.7

Question 2 Competence Updates for Open-Ended Tasks and Noisy Feedback

Thank you for your valuable suggestion. We considered the effect of feedback quality on competence estimation when designing our experiments. We therefore chose benchmarks with ground-truth answers to evaluate memory calibration under clear, consistent feedback while limiting interference from evaluation noise.

For open-ended tasks, LLM-as-a-Judge [1] or reward models [2] could provide proxy feedback, but evaluator errors and biases [3] may be mistaken for memory failures. The reviewer’s suggestion motivates future work on incorporating feedback reliability into competence estimation for more robust updates.

[1] Wei et al. Igniting Creative Writing in Small Language Models: LLM-as-a-Judge versus Multi-Agent Refined Rewards. EMNLP 2025.

[2] Chen et al. RM-R1: Reward Modeling as Reasoning. ICLR 2026.

[3] Xu et al. A Judge-Aware Ranking Framework for Evaluating Large Language Models without Ground Truth. ICML 2026.


Question 3 Re-Retrieval and Continued Learning after Early Downweighting

Thank you for raising this important question. C²Mem preserves opportunities for low-exposure nodes to participate again in later matching states. Retrieval prioritizes nodes whose behavioral patterns match the current execution state based on behavioral similarity (Eq. 2). Thus, an early-downweighted node can re-enter the top-K candidate set when a subsequent state closely matches its strategy.

During training, a retrieved node’s support weight and calibration are reassessed in the current state. Nodes whose calibrated weights meet the injection threshold can participate again and update competence through new feedback. Early downweighting therefore does not permanently exclude a node from use. Posterior confidence ct−c_t^{-} also modulates the penalty, distinguishing uncertain early estimates from judgments supported by sufficient evidence.


Question 4(1) The Role of the Behavioral Tree in Calibration and Interpretability

An advantage of the behavioral tree is that it identifies coherent guidance across memories and makes their strategic relationships interpretable. Nearby nodes supporting compatible strategies can jointly yield high commitment despite distributed support weights. Flattening removes the structural basis for this assessment and for tracing behaviors to broader strategies. On HotpotQA, full C²Mem outperforms its flat-memory variant, supporting the contribution of the tree structure.

MethodEM (%)F1 (%)
C²Mem with flat memory structure49.0061.89
C²Mem50.5063.51

Our response to Question 1 further examines how tree distance reflects strategy compatibility. Nearby memories induce more consistent actions and better preserve their individual guidance when used jointly.

We are very grateful for your constructive comments and suggestions, which have been invaluable in improving our paper. We sincerely hope to continue our discussions with you and would welcome the opportunity to address any questions or concerns that remain.

Reviewer_JdK8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies a failure mode in which retrieved memories provide coherent guidance despite weak historical execution reliability. C²Mem organizes experiences in a behavioral tree, estimates commitment from retrieval support and tree consistency, and estimates competence from exposure-weighted success/failure posteriors. It uses a KL-regularized reweighting rule to suppress memories with high estimated risk. Experiments cover five benchmarks and three multi-agent frameworks. Appendix A.3 clarifies that memory construction and posterior updates use training data, while the memory state is frozen during test evaluation.

Critical strengths
  1. Separating how strongly a memory steers behavior from how reliably it has worked is a useful perspective on agent memory. The intervention examples and the commitment–competence interaction analysis make the motivation concrete. The resulting suppression rule also has a clear operational meaning and a simple closed-form implementation.

  2. The reported improvements are reasonably broad: C²Mem improves over no memory in all 15 framework–benchmark combinations and has the highest average score within each framework. Table 2 also includes a competence-only calibration control, which is important for testing whether commitment adds anything beyond historical reliability.

Critical weaknesses
  1. It is not established that the competence deficit identifies harmful memory rather than differences in task difficulty and memory selection. Participating nodes receive fractional credit from the same terminal task outcome, and the unknown node is estimated from insufficient-support exposure rather than a matched no-memory intervention. Appendix D.4 appropriately states the additional comparability assumption, but does not validate it empirically; the consistency result in D.3 is explicitly limited to a fixed node under an i.i.d. regime. Outcome-conditioned plots therefore support an association, but do not yet show that the risk score identifies when suppressing a memory improves the same execution state. This matters to the proposed interpretation of memory-induced overconfidence.

  2. The empirical contribution beyond reliability-based filtering remains uncertain. The full commitment term adds only 1.18 points on average over competence-only calibration in the four ablation settings, without reported variability for that comparison. The paper also does not specify concrete train/test sample counts, the principal calibration and retrieval hyperparameters, or repeated-run uncertainty for the main tables. A simple utility-based gating control with matched memory and token budgets would help distinguish the proposed interaction from ordinary reliability filtering. The missing protocol details and limited isolation of the added mechanism currently lead me to lean toward rejection.

Questions for authors
  1. Can the estimated risk predict the effect of suppressing a memory in paired continuations from the same execution prefix? In particular, how well does the unknown-node reference approximate a matched no-memory continuation rather than a different distribution of tasks?

  2. Please give the exact train/test subsets and sizes, initialization of the node and unknown-node priors, top-K, κ, injection threshold, splitting criterion, and their selection procedure. Were all competing methods given comparable memory-building and inference budgets?

  3. Is the gain from commitment stable across repeated memory-construction and execution runs, and does it persist against a simpler competence-only threshold or a tuned constant-strength penalty? These comparisons would materially affect my assessment of the central mechanism.

3 comments

We sincerely thank the reviewer for the constructive feedback on risk validation, the added value of commitment, and the evaluation protocol. Below, we present additional experiments and clarify the experimental settings and budget comparisons.


Weakness 1(1) & Question 1(1) Risk Validation under Same-Prefix Interventions

Thank you for your valuable suggestion. C²Mem assesses whether the current execution overrelies on unreliable memories, rather than permanently classifying a memory as harmful. Validating the risk estimate therefore requires adjusting memory influence at the same execution state and examining whether subsequent performance improves. We conduct paired continuation experiments to move beyond associations between successful and failed tasks.

We randomly select 200 HotpotQA test tasks where calibration changes memory weights under AutoGen + C²Mem, using GPT-4o and a fixed behavioral tree. Each task continues from the same execution prefix at the calibration step under three settings: retaining uncalibrated memories, removing downweighted memories, or using calibrated memories, with all other conditions unchanged. This design holds task difficulty, prior memory selection, and execution history constant. Removing downweighted memories improves performance, while calibration yields further gains, demonstrating a benefit from adjusting memory influence in these calibration-triggering states.

Memory SettingEM (%)F1 (%)Average Steps
Uncalibrated memories37.5047.248.97
Remove downweighted memories43.0053.236.29
Calibrated memories50.5063.515.16

Weakness 1(2) & Question 1(2) Validity of the Unknown-Node Reference

Thank you for raising this important question. The learned unknown-node reference on HotpotQA is μ0=0.35\mu_0=0.35, close to the success rate of same-prefix no-memory continuations. This is precisely the purpose of the unknown node: to learn a reliability reference under insufficient memory support from accumulated execution feedback, providing a basis for evaluating specific behavioral memories.

We further examine whether memories on either side of this reference yield different execution benefits. On 200 HotpotQA tasks with AutoGen, we freeze the behavioral tree and reliability posteriors and continue each task from the same execution prefix under three conditions: injecting a memory with competence below μ0\mu_0, injecting one above μ0\mu_0, or injecting no memory. Both memory conditions use the selected memories directly without additional calibration, with all other execution conditions unchanged.

Memory ConditionEM (%)F1 (%)Average Steps
μv<μ0\mu_v < \mu_034.5038.948.18
No memory40.5048.585.26
μv>μ0\mu_v > \mu_045.5057.195.80

Under the same tasks and execution prefixes, memories below the reference reduce performance, while those above it improve performance. These results support the learned μ0\mu_0 as a useful comparison boundary in the evaluated states, beyond differences in historical success rates across tasks.

Weakness 2(1) & Question 3 Added Value and Stability of Commitment

The gain from commitment over competence-only calibration remains stable across repeated execution runs. Commitment adjusts the penalty strength according to the strength and coherence of the current memory guidance. Total memory support measures how strongly memories guide the current behavior, while structural coherence assesses whether they point toward compatible strategies. Together, they indicate whether unreliable memories provide guidance strong enough to dominate execution, thereby determining the penalty strength.

Using AutoGen, we perform three runs on the same 500 HotpotQA test instances, reporting variability across runs and paired gains over competence-only calibration. Beyond this stability analysis, we also include memory support and retrieval entropy as alternative settings to examine whether total support or weight concentration alone can explain the improvements. The full C²Mem outperforms both alternatives, supporting the joint consideration of guidance strength and structural coherence in reliability-based calibration.

MethodEM (%)F1 (%)ΔEM vs. competence onlyΔF1 vs. competence only
competence only48.80 ± 0.7261.55 ± 0.83--
competence+retrieval entropy50.27 ± 0.9063.16 ± 0.96+1.47 ± 0.31+1.61 ± 0.57
competence+memory support50.33 ± 0.7663.48 ± 0.69+1.53 ± 0.46+1.93 ± 0.47
C²Mem50.87 ± 0.7664.51 ± 0.73+2.07 ± 0.31+2.96 ± 0.53

Weakness 2(2) & Question 2 Experimental Protocol and Budget Comparability

Thank you for your questions about the experimental protocol and budget fairness. Appendices A.1–A.3 describe memory construction, node splitting, and the training and evaluation protocol. Following the evaluation scale of prior work [1,2], we use 500 training instances for memory construction and 500 evaluation instances per dataset, with MATH evaluated on MATH-500. All methods use the same data splits, MAS configurations, backbone models, and evaluation instances. Memory contents, tree structures, and reliability posteriors remain frozen during evaluation.

Memory construction begins with memory-free execution. We first run 30 training tasks and use an LLM to summarize their trajectories into three representative behavioral patterns, forming the initial behavioral tree. We also extract general experience to initialize the unknown node. Subsequent experiences are inserted according to the behavioral similarity rule in Appendix A.2. An LLM determines node splitting based on behavioral differences among the stored experiences, and child nodes inherit the parent’s posterior evidence in proportion to their assigned experiences.

Sections 2.1 and 2.4 and Appendix A.5 specify the retrieval, calibration, and injection rules, respectively. At each step, we retrieve top-K = 3 behavioral nodes and include the unknown node. The default calibration strength is κ = 20, and the injection threshold is 0.15. These hyperparameters are selected through grid search. We will consolidate their values, search ranges, and selection procedure, together with the node-splitting criteria, in the revision.

Section 3.7 and Figure 7 already jointly compare task performance and actual token consumption on HotpotQA. C²Mem achieves higher performance than G-Memory at a similar token cost. In the revision, we will tabulate the costs from Figure 7 alongside the corresponding performance results from Table 1, and specify the memory-construction and inference budgets for each method to make the comparison conditions explicit.

[1] Chen et al. Towards Consistent Natural-Language Explanations via Explanation-Consistency Finetuning. COLING 2025.

[2] Yao et al. ReAct: Synergizing Reasoning and Acting in Language Models. ICLR 2023.

We sincerely appreciate your highly constructive comments and suggestions, which have been extremely valuable in improving the quality of our paper. We also look forward to further in-depth discussions with you to address any remaining questions or concerns you may have.

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

This paper studies memory-induced overconfidence in multi-agent systems: retrieved memories can provide coherent behavioral guidance despite having low historical reliability. The proposed C²Mem organizes experiences in a behavioral tree, estimates commitment from retrieval support and structural coherence, and estimates competence using exposure-weighted Beta updates. It then attenuates memories when strong commitment coincides with a competence deficit. Experiments across five benchmarks and three multi-agent frameworks show consistent improvements over no-memory baselines and competitive memory methods.

Critical strengths

Strength1: The distinction between commitment and competence is useful. It explains why coherent memory guidance can reinforce an ineffective strategy. The controlled memory experiments and commitment–competence analysis provide supporting evidence, and the distinction leads to a concrete intervention.

Strength2: The empirical results are consistent across settings. C²Mem improves over no memory in all 15 framework–benchmark combinations. Table 2 further shows benefits from competence-based calibration, with commitment providing an additional improvement in all four evaluated configurations.

Critical weaknesses

Weakness 1: The value of structural coherence is not isolated. Commitment combines total memory support and tree-based coherence, but Table 2 removes both together. It remains unclear whether tree distance improves calibration beyond simpler measures of support or weight concentration.

Weakness 2: The reliability estimate has a credit-assignment limitation. Equation 13 distributes the final task outcome among nodes according to their exposure. A useful memory can therefore be penalized when execution fails for another reason. The aggregate gains are encouraging, but do not directly establish whether calibration consistently suppresses the memories responsible for errors.

Questions for authors

Question 1: Would the improvement over competence-only calibration in Table 2 remain if commitment used total support alone, or support combined with retrieval entropy? Reporting variability across runs would also help assess the gain.

Question 2: On a subset of tasks, could the authors compare continuations from the same execution prefix with and without the downweighted memories? This would directly test whether the calibration identifies harmful guidance rather than memories associated with harder tasks.

4 comments

We sincerely thank the reviewer for the constructive feedback on the contribution of structural coherence, the stability of the reported gains, and memory credit assignment. Below, we address each point and describe the additional experiments and corresponding revisions.


Weakness 1 Tree Distance as a Measure of Behavioral Coherence

An advantage of structural coherence is that multiple behaviorally similar nodes can jointly provide coherent guidance without requiring retrieval weights to concentrate on a single node. To evaluate whether tree distance reflects differences in behavioral strategies, we conduct additional node-summary analyses and same-prefix branching experiments.

First, we group node pairs by normalized tree distance and compare the semantic differences between their summaries. Greater tree distance corresponds to larger average summary cosine distance, indicating that the behavioral tree structure reflects differences in the strategies expressed by its nodes.

Normalized Tree DistanceMean Summary Cosine Distance
(0,1/3]0.168
(1/3,2/3]0.176
(2/3,1]0.233

We further conduct same-prefix experiments to directly examine the relationship between tree distance and behavioral agreement. On 200 HotpotQA tasks, we select memory pairs at different tree distances and inject them individually or jointly from the same execution prefix. We measure agreement between the next actions under individual injection, as well as the average agreement of the joint action with the two individual actions. Closer memories yield greater agreement between individual branches and greater preservation of individual behavior under joint injection.

Normalized Tree DistancePairwise Agreement (%)Joint–Individual Agreement (%)
(0,1/3]54.768.9
(1/3,2/3]45.261.7
(2/3,1]24.145.7

These results provide direct evidence for using tree distance to characterize behavioral coherence. Its performance contribution relative to total memory support and retrieval entropy is further evaluated through the replacement experiments in our response to Question 1.


Question 1 Stability of Commitment Gains and Comparisons with Simpler Signals

Retrieval entropy and tree distance capture different aspects: the former measures weight concentration, while the latter characterizes behavioral compatibility through structural relationships between nodes. We use tree distance because coherent guidance can come from either a single node or several nearby nodes. Even when the weights are dispersed, these memories may still support a coherent behavioral direction.

To further evaluate this design, we follow the reviewer’s suggestion and replace the original commitment with retrieval entropy or memory support, comparing these variants with competence-only calibration and the full C²Mem. These comparisons test whether weight concentration or total support can substitute for structural coherence. All settings use AutoGen, a fixed behavioral tree, and the same 500 HotpotQA test instances, with three runs per setting. The following table reports variability across runs and paired gains over competence-only calibration to assess the stability of the improvements.

MethodEM (%)F1 (%)ΔEM vs. competence onlyΔF1 vs. competence only
competence only48.80 ± 0.7261.55 ± 0.83--
competence+retrieval entropy50.27 ± 0.9063.16 ± 0.96+1.47 ± 0.31+1.61 ± 0.57
competence+memory support50.33 ± 0.7663.48 ± 0.69+1.53 ± 0.46+1.93 ± 0.47
C²Mem50.87 ± 0.7664.51 ± 0.73+2.07 ± 0.31+2.96 ± 0.53

Question 2 & Weakness 2 Controlled Evaluation of Calibration beyond Memory Removal

Thank you for your valuable suggestion. An advantage of C²Mem is that downweighted memories retain opportunities to participate in later matching execution states. Retrieval uses behavioral similarity to prioritize nodes whose patterns match the current execution state (Eq. 2). Thus, a previously suppressed node can re-enter the top-K candidate set when a later state closely matches its strategy. Calibration adjusts a memory’s current influence while preserving opportunities for its use in other execution states.

We randomly select 200 HotpotQA test tasks where calibration changes memory weights under AutoGen + C²Mem. Using GPT-4o and a fixed behavioral tree, we continue each task from the same execution prefix at the calibration step under the three settings below, keeping all other conditions unchanged. Removing downweighted memories improves performance, while calibrated memories yield further gains. This comparison controls for task difficulty and prior execution history, and supports reducing memory influence rather than removing the memories entirely.

Memory SettingEM (%)F1 (%)Average Steps
Uncalibrated memories37.5047.248.97
Remove downweighted memories43.0053.236.29
Calibrated memories50.5063.515.16

Section 3.8 and Appendix B, Figure 9 provide a step-by-step case study. Calibration reduces the influence of an initially dominant, low-competence memory, allowing identity-verification guidance to redirect execution toward the correct answer. Later high-competence guidance is retained, illustrating how commitment–competence calibration selectively weakens unreliable guidance while preserving reliable guidance.

Thank you for the detailed rebuttal and additional experiments. The new analyses and controlled same-prefix comparisons sufficiently address my concerns about isolating structural coherence, the stability of commitment gains, and credit assignment in the calibration process. I appreciate the clarification that calibration reduces unreliable memory influence while preserving their potential usefulness in later states.

I therefore maintain my positive rating of 3.

We sincerely appreciate your positive assessment and valuable, constructive feedback. We are glad that our additional experiments and detailed clarifications have addressed your concerns. We would be happy to discuss any further questions you may have.

Comments on the paper

1 comment

We sincerely thank all three reviewers for their careful reading and constructive comments. Their feedback helped us strengthen the empirical basis of our central claims and clarify the experimental protocol. Since each major revision addresses concerns shared by multiple reviewers, we color-code revisions by reviewer combination in the revised PDF. For easy identification, revisions addressing concerns shared by tuiu and JdK8 are highlighted in 🟦 blue, those shared by tuiu and geeF in 🟪 purple, those shared by JdK8 and geeF in 🟩 green, and those shared by all three reviewers in 🟥 red. Below, we summarize the major revisions.


🟥 Red- tuiu’s Question 2 & Weakness 2, JdK8’s Weakness 1(1) & Question 1(1), and geeF’s Weakness 1(1) — Same-Prefix Calibration Intervention

We added Section 3.7 and Table 3 to directly examine whether calibration improves execution when unreliable memories exert excessive influence. By comparing alternative continuations from the same execution prefix, the new experiment controls for task difficulty and prior execution history. The results show that suppressing the selected guidance improves subsequent performance, while calibrated reweighting provides further benefits over removal. This supports our central idea that memory influence should be adjusted to the current execution state, while preserving potentially useful guidance.


🟩 Green- JdK8’s Weakness 1(2) & Question 1(2) and geeF’s Weakness 1(2) & Question 4(3) — Validity of the Unknown-Node Reference

We added Section 3.8 and Table 4 to examine whether the learned unknown-node reference provides a meaningful boundary for assessing memory utility. Matched continuations with and without memory injection show that memories below this reference impair performance on average, whereas those above it improve performance. This provides intervention-based evidence for its role in calibration: historical reliability becomes more informative when assessed against an execution-relevant reference, helping determine whether memory guidance warrants reliance.


🟩 Green- JdK8’s Weakness 2(2) & Question 2 and geeF’s Question 1 — Experimental Protocol and Tree Construction

We expanded Appendices A.2 and A.3 to make the construction and evaluation of the memory system more transparent. The revisions clarify how the behavioral tree is initialized and refined, how training and evaluation data are used, and how retrieval and calibration operate during evaluation. In particular, we explicitly describe the frozen memory state at test time, making clear that the reported gains evaluate the use of previously learned memory without incorporating test outcomes into memory updates. These clarifications make the evidence supporting our method easier to interpret and reproduce.


🟦 Blue- tuiu’s Question 1 and JdK8’s Weakness 2(1) & Question 3 — Added Value and Stability of Commitment

We added Appendix B.1 and Table 5 to examine what commitment contributes beyond historical reliability and whether its benefit persists across repeated executions. Comparisons with support-only and retrieval-entropy variants show that guidance strength and weight concentration do not fully account for the benefit of the complete commitment signal. Together with the repeated-run results, this supports the motivation for considering both the strength and behavioral coherence of retrieved guidance when deciding how strongly to calibrate unreliable memories.


🟪 Purple- tuiu’s Weakness 1 and geeF’s Question 1 & Question 4(1) — Behavioral Tree Validity and Flat-Memory Comparison

We added Appendix B.2 and Table 6-7 to examine the behavioral meaning of the tree structure and its contribution to calibration. The new analyses connect tree proximity with both summary similarity and agreement in subsequent actions, while the flat-memory comparison assesses the value of retaining these structural relationships. Together, these results support the idea that compatible guidance can be distributed across multiple memories, and that their relationships help characterize how coherently they may steer execution.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.