acceptodds
Under review as a conference paper at ICLR 2027

Living-Harness Is an Interactive-Agent Evolver

11 views

Abstract

Large language model (LLM) agents may recover from a failure within an episode or after a retry, yet the same execution failure can recur in later tasks because post-episode feedback rarely revises the persistent harness that guides future interactions. Static harnesses improve reliability through fixed tools, context, memory, and workflow structures, but remain unchanged after deployment. We propose , a self-evolving agent harness that converts each completed trajectory and its evaluator signals into posterior evidence for bounded harness updates. Guided by a domain-level (tandard perating rocedure), Living-Harness extracts an episode abstraction and structured update evidence, and writes two complementary forms of procedural knowledge: episodic memory that records trigger conditions, failure patterns, and recovery actions, and a state graph that records state nodes, repair edges, and transition rules. The updated harness state is retrieved to guide future interactions, while tools and base context remain frozen, allowing procedural repairs to accumulate across evolution cycles. On eight interactive environments derived from -Bench and MultiWOZ-2.4, Living-Harness improves average Pass@1 over the strongest interactive baseline by 10.07 and 9.91 percentage points, respectively, and supports retrieval-only reuse of the evolved harness state across model backbones.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language model agents, self-evolving agents, agent harnesses, episodic memory, workflow optimization, tool use

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_9kLjOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Living-Harness, a self-evolving harness for interactive LLM agents. After each scored episode, a domain-level Evolution-SOP guides three steps. A posterior generator abstracts the trajectory and evaluator signals. A memory extractor and a workflow extractor then produce candidate updates. These updates pass schema, scope, evidence, constraint and merge gates before they are written into two persistent stores: episodic memory (trigger conditions, failure patterns, recovery actions) and a state graph (state nodes, repair edges, transition rules). Tools and base context stay frozen. Top-3 retrieval from both stores is given to the actor in later episodes.

Critical strengths

The motivation is clear and practical. An agent can see from feedback that it should transfer a user to a human, and still fail to call the tool the next time. Separating task-local reflection from persistent, scoped procedural repair is a useful distinction, and Figure 4 shows it concretely.

The design is reasonably careful. Updates happen only after scoring, so episode evidence never affects its own score. Tools and base context are frozen. Candidate updates pass explicit commit gates. The appendix gives all Evolution-SOPs and extractor prompts, which helps reproducibility.

Absolute numbers on τ²-Bench are high, and there are ablations of the SOP, memory and graph. The harness-state growth tables (Tables 4–5) are informative.

Critical weaknesses

1.Much of the gain may come from hand-written, benchmark-specific domain knowledge rather than self-evolution.

The Evolution-SOPs in Appendix E are manually written and contain concrete solution rules for the test domains. The Telecom SOP lists the exact repair tools (refuel_data, grant_app_permission, reset_apn_settings plus reboot for no-service, and others). It specifies the troubleshooting order. It even says transfer_to_human_agents should dominate for terminal service cases, which is exactly the repair shown in Figure 4. It also states that “high-frequency misses” should be promoted to stable branch rules. This reads like knowledge distilled from looking at failures on these same benchmarks.

In Figure 3, replacing the SOP with a generic extractor drops τ²-Bench to 73.38. That is almost the same as Reflexion (73.02). So without the domain SOP, the method shows essentially no gain over the strongest baseline. A fair comparison would give the same SOP text to the baselines, for example in the actor prompt or in the AWM/ReasoningBank extraction prompts. The paper should also explain how the SOPs were written and whether the authors looked at test trajectories when writing them.

2.The evaluation protocol appears to adapt on the test set, and the reported numbers mix cycles.

The paper does not state whether each evolution cycle reruns the same task instances. Table 4 suggests it does: Airline edges grow by exactly 200 per cycle and Retail by exactly 456. That is consistent with 4 edges per task on the 50 Airline and 114 Retail tasks. If so, memory built from failures on a task in cycle n is retrieved for the same task in cycle n+1. The score-before-update protocol only rules out reuse within a cycle. The limitations section concedes that robustness to “reduced recurrence” and held-out task families is not established. Under this protocol, the cross-backbone transfer in Table 3 is also on the same MultiWOZ tasks the state was built from, so it shows reuse rather than generalization.

3.There are also several number inconsistencies:

Table 1 reports Airline 88.00, which is the Cycle 2 peak, while Retail and Telecom use the Cycle 3 values. This amounts to best-cycle selection on the test set. Telecom Cycle 0 in Table 2 is 57.39. That equals GPT-5.2’s τ²-Bench average in Table 1, not its Telecom score (52.23). It also does not match a k/114 success rate.

Questions for authors

1.How were the Evolution-SOPs written? Did the authors look at trajectories from the evaluation tasks? Can you report baselines (Reflexion, AWM, ReasoningBank) with the same SOP content available, and Living-Harness with SOPs written without access to these benchmarks?

2.Does each cycle rerun the same task instances? Please report results with the harness evolved on one split and evaluated on held-out tasks or task families, for both the main results and the cross-backbone transfer. Why does Table 1 use the Cycle 2 Airline score? Why does Telecom Cycle 0 equal 57.39?

5 comments
AuthorsOpenReview ↗

Response to Reviewer 9kLj (Part 1 of 3)

Thank you for your positive assessment and for recognizing the practical motivation of persistent procedural repair, the careful update design, and the value of our experiments and state-growth analyses. We address your questions below.

W1. Much of the gain may come from hand-written, benchmark-specific domain knowledge rather than self-evolution. The Evolution-SOPs in Appendix E are manually written and contain concrete solution rules for the test domains. The Telecom SOP lists the exact repair tools (refuel_data, grant_app_permission, reset_apn_settings plus reboot for no-service, and others). It specifies the troubleshooting order. It even says transfer_to_human_agents should dominate for terminal service cases, which is exactly the repair shown in Figure 4. It also states that "high-frequency misses" should be promoted to stable branch rules. This reads like knowledge distilled from looking at failures on these same benchmarks. In Figure 3, replacing the SOP with a generic extractor drops τ²-Bench to 73.38. That is almost the same as Reflexion (73.02). So without the domain SOP, the method shows essentially no gain over the strongest baseline. A fair comparison would give the same SOP text to the baselines, for example in the actor prompt or in the AWM/ReasoningBank extraction prompts. The paper should also explain how the SOPs were written and whether the authors looked at test trajectories when writing them.

Q1. How were the Evolution-SOPs written? Did the authors look at trajectories from the evaluation tasks? Can you report baselines (Reflexion, AWM, ReasoningBank) with the same SOP content available, and Living-Harness with SOPs written without access to these benchmarks?

Response to W1 and Q1: Evolution-SOP is generated automatically by GPT-5.2 from each environment's existing specifications and system prompts, not manually written; no evaluation trajectories or ground truth are used to construct or revise it. We have corrected the manuscript's description of manually specified rules (App. D). The troubleshooting branches, repair tools, and escalation conditions in the Telecom SOP come from the domain policy and tool definitions that τ²-Bench provides to every agent. Its rules on recurring failures specify how the extractor should treat failure modes that recur across evolution loops during a run, rather than encoding outcomes observed in advance.

SOP content alone does not explain the gains. Giving the baselines the same Evolution-SOP text changes their τ²-Bench averages by only −12.95 to +3.60 points, and no SOP-equipped baseline exceeds 60.07, well below Living-Harness even without its SOP (73.38) (Table A; App. C, Table 9). The effect of the text is also inconsistent across domains: it raises the two memory-based baselines on Retail but lowers all three baselines on Telecom, by 10.53 to 25.44 points. Within Living-Harness, the same SOP raises every domain and the average by 9.71 points (73.38 → 83.09), because it governs how evaluated evidence becomes gated memory and state-graph repairs rather than acting as additional instructions. The gains therefore come from how Living-Harness uses the SOP, not from the SOP text.

Table A. Equal-information comparison on τ²-Bench (GPT-5.2, Pass@1 %; average weighted by task count).

MethodRetailAirlineTelecomAvg
Reflexion78.0780.0064.9173.02
Reflexion + Evolution-SOP71.9380.0039.4760.07
AWM56.1466.0039.4751.08
AWM + Evolution-SOP69.3066.0027.1951.44
ReasoningBank59.6570.0037.7252.52
ReasoningBank + Evolution-SOP80.7066.0027.1956.12
Living-Harness w/o Evolution-SOP74.5676.0071.0573.38
Living-Harness85.9688.0078.0783.09
AuthorsOpenReview ↗

Response to Reviewer 9kLj (Part 2 of 3)

W2. The evaluation protocol appears to adapt on the test set, and the reported numbers mix cycles. The paper does not state whether each evolution cycle reruns the same task instances. Table 4 suggests it does: Airline edges grow by exactly 200 per cycle and Retail by exactly 456. That is consistent with 4 edges per task on the 50 Airline and 114 Retail tasks. If so, memory built from failures on a task in cycle n is retrieved for the same task in cycle n+1. The score-before-update protocol only rules out reuse within a cycle. The limitations section concedes that robustness to "reduced recurrence" and held-out task families is not established. Under this protocol, the cross-backbone transfer in Table 3 is also on the same MultiWOZ tasks the state was built from, so it shows reuse rather than generalization.

Q2. Does each cycle rerun the same task instances? Please report results with the harness evolved on one split and evaluated on held-out tasks or task families, for both the main results and the cross-backbone transfer.

Response to W2 and Q2: Each evolution cycle reruns the same task instances. τ²-Bench applies updates after each scored episode, while MultiWOZ-2.4 synchronizes updates after all four episodes in a window are scored (App. B.1). The main experiments therefore measure how procedural repairs accumulate during continued interaction, and each episode is scored before its updates are applied, so they cannot affect its own recorded score.

To evaluate unseen tasks, we split each τ²-Bench domain 7:3 into disjoint update and evaluation sets (80/34 tasks for Retail and Telecom, 35/15 for Airline). The harness evolves on the update set for three cycles and is then frozen for evaluation. Living-Harness achieves the best result in every domain, averaging 75.90 versus 68.67 for ReasoningBank and 63.86 for the GPT-5.2 base (Table B; App. C, Table 7). The frozen state also transfers to other backbones on held-out tasks without further evolution (Table C; App. C, Table 8), raising the average from 86.75 to 92.77 for GLM-5 and from 75.90 to 80.72 for Kimi-K3. GLM-5 benefits even though its base already exceeds the GPT-5.2 harness that built the state.

Table B. Held-out evaluation on τ²-Bench with disjoint update and evaluation sets (GPT-5.2, Pass@1 %; average weighted by task count).

MethodRetailAirlineTelecomAvg
GPT-5.2 base76.4766.6750.0063.86
AWM79.4146.6744.1259.04
ReasoningBank79.4166.6758.8268.67
Living-Harness91.1873.3361.7675.90

Table C. Cross-backbone transfer on held-out τ²-Bench tasks. Bold marks the better setting for each backbone.

BackboneHarness stateRetailAirlineTelecomAvg
GLM-5None79.4193.3391.1886.75
GLM-5Transferred91.1886.6797.0692.77
Kimi-K3None85.2973.3367.6575.90
Kimi-K3Transferred91.1873.3373.5380.72
AuthorsOpenReview ↗

Response to Reviewer 9kLj (Part 3 of 3)

W3. There are also several number inconsistencies: Table 1 reports Airline 88.00, which is the Cycle 2 peak, while Retail and Telecom use the Cycle 3 values. This amounts to best-cycle selection on the test set. Telecom Cycle 0 in Table 2 is 57.39. That equals GPT-5.2's τ²-Bench average in Table 1, not its Telecom score (52.23). It also does not match a k/114 success rate.

Q2 (continued). Why does Table 1 use the Cycle 2 Airline score? Why does Telecom Cycle 0 equal 57.39?

Response to W3 and Q2: Table 1 applies the same rule to Living-Harness and all interactive baselines alike: the best performance in each domain over three evolution cycles. Airline therefore uses Cycle 2 (88.00), and Table 2 also lists Cycle 3 (86.00). This follows recent harness-evolution work, which reports the best harness found during search in Meta-Harness [1] and the best-so-far agent in DGM [2]. We now state this rule in the Table 1 caption. The Telecom Cycle 0 entry should read 52.63 (60/114); we have corrected it in Table 2, along with the corresponding Table 1 entry.

Thank you again for your careful review; we welcome further questions.

References:

[1] Meta-Harness: End-to-End Optimization of Model Harnesses, arXiv 2026.

[2] Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents, ICLR 2026.

Reviewer_Z9qFOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper develops living-harness, a self-evolving agent harness. The major idea is to convert each completed trajectory and its evaluator signals into posterior evidence for bounded harness updates.

Critical strengths

Making harness dynamic and evolving is an interesting and important idea.

Critical weaknesses

From the related work discussion, it is unclear whether any similar ideas have been explored by literature.

Would the living-harness process lead to the harness unstable in the evolving process?

The paper is like a project report. It describes what the authors do but does not justify the rationale systematically. What are the possible options and why do the authors make the technical decisions?

What are the limitations of the design?

Questions for authors

Some fonts in Figure 1 are missing.

4 comments
AuthorsOpenReview ↗

Response to Reviewer Z9qF (Part 1 of 2)

Thank you for recognizing dynamic, evolving harnesses as an important direction. We address your concerns below.

W1. From the related work discussion, it is unclear whether any similar ideas have been explored by literature.

Response to W1: Existing work improves agent procedures along three lines (Section 2; App. C, Table 6). Harness-design and workflow-optimization methods, such as Meta-Harness [1] and AFlow [2], search over prompts, workflows, or harness code to build a stronger harness, which is then deployed as a fixed artifact. Reflection and experience-memory methods, such as Reflexion [3], ExpeL [4], and ReasoningBank [5], retain critiques, insights, or reasoning strategies from past trajectories and retrieve them as text in later attempts. Test-time evolution methods revise the procedure during operation: AWM [6] induces workflows from trajectories judged successful, and EvoTest [7] rewrites the agent configuration after each episode.

These lines leave three gaps. First, a harness optimized before deployment stays fixed afterward, so it cannot be updated when new failure patterns emerge. Second, natural-language lessons are too coarse to capture the concrete constraints and workflow procedures of a specific environment, so the agent keeps repeating the same failure in that environment. A lesson such as "transfer the user to a human" does not say when the transfer should be triggered, which tool to call, or how the workflow should proceed afterward. In Figure 4, Reflexion recognizes in all three attempts that a transfer is needed, yet never calls the transfer tool. Third, test-time revisions are insufficiently verified, which makes it hard for the harness to improve steadily across tasks. AWM checks only whether a trajectory is predicted to succeed, not whether the induced workflow is correct, while EvoTest applies configuration changes without any explicit check. As a result, an incorrect revision can carry over into later tasks.

Living-Harness addresses each gap. It updates the harness state after every evaluated episode, so newly observed failures become repairs for later tasks. Each repair is executable: episodic memory stores the trigger condition, failure pattern, and recovery action, and the state graph adds a repair edge linking the observed condition to the missing or revised action. Every candidate update passes five commit gates (schema, scope, evidence, constraint, and merge) under a fixed Evolution-SOP, while tools and base context stay frozen. Against ReAct and four of these methods, Living-Harness achieves the highest average on both benchmarks: 83.09 on τ²-Bench and 65.50 on MultiWOZ-2.4 (Table 1).

W2. Would the living-harness process lead to the harness unstable in the evolving process?

Response to W2: The fixed Evolution-SOP (§3.2) acts as a constitution: only episodic memory and the state graph evolve, while tools, base context, and update rules stay unchanged, and every candidate update passes commit gates (App. B.2). Table 2 shows gains in all eight domains across three cycles: every cycle stays above the initial level, and six domains improve at each successive cycle.

References:

[1] Meta-Harness: End-to-End Optimization of Model Harnesses, arXiv 2026.

[2] AFlow: Automating Agentic Workflow Generation, ICLR 2025.

[3] Reflexion: Language Agents with Verbal Reinforcement Learning, NeurIPS 2023.

[4] ExpeL: LLM Agents Are Experiential Learners, AAAI 2024.

[5] ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory, ICLR 2026.

[6] Agent Workflow Memory, ICML 2025.

[7] EvoTest: Evolutionary Test-Time Learning for Self-Improving Agentic Systems, ICLR 2026.

AuthorsOpenReview ↗

Response to Reviewer Z9qF (Part 2 of 2)

W3. The paper is like a project report. It describes what the authors do but does not justify the rationale systematically. What are the possible options and why do the authors make the technical decisions?

Response to W3: As discussed in W1, current methods either fix the harness before deployment, store experience as coarse text that does not specify when and how to act, or revise the procedure without verifying each change. As a result, a correction an agent makes within an episode is lost once the episode ends, and the same failure recurs in later interactions. To address these challenges, Living-Harness represents the first effort to turn evaluated corrections into persistent, executable, and verified procedural repairs. Specifically, each design choice answers one of these gaps:

  • When to update. Rather than optimizing the harness once before deployment, Living-Harness updates after every evaluated episode. The formulation therefore separates the environment state sts_t, inferred from the interaction history within an episode, from the program state z(n)z^{(n)}, which stays fixed during an episode and is revised only after evaluation, matching the rollout–evaluate–update loop.
  • What to store. Rather than storing text lessons alone, Living-Harness pairs episodic memory, which records when a repair applies, why it helps, and how to recover, with a state graph that links the repair to the corresponding workflow actions and transitions. The ablations in Figure 3 confirm that both components contribute.
  • How to verify. Rather than committing every revision, Living-Harness grounds each repair in completed trajectories and evaluation signals, and requires it to pass five commit gates under a fixed Evolution-SOP. Tools and base context stay frozen, which bounds what can change.

W4. What are the limitations of the design?

Response to W4: Appendix D discusses the main limitations. Evaluation uses structured simulators, and evolution adds computation (see our response to Reviewer KGVn, W3 Table B; App. C, Table 10). Evolution-SOP is generated automatically but relies on available domain policies and tool definitions; we have corrected the appendix's description. Future work will address adaptation to policy changes and open-ended environments, and memory maintenance through rollback, stale-memory removal, and regression checks.

Q1. Some fonts in Figure 1 are missing.

Response to Q1: Thank you for pointing out this issue. We identified a font embedding incompatibility affecting the bold labels in Figure 1. The corrected figure has been included in the revised manuscript.

We hope these responses address your concerns, and would be grateful if you could reconsider your assessment. We welcome further questions.

AuthorsOpenReview ↗

Looking Forward to Your Feedback and Further Discussion

Dear Reviewer Z9qF,

We hope you are doing well. We sincerely appreciate the thoughtful feedback and constructive suggestions you have provided.

As the discussion deadline approaches, we would be grateful if you could review our responses when you have a chance and let us know whether any concerns remain. Your feedback would allow us sufficient time to address additional questions and provide any further information that would support the discussion.

If our rebuttal and manuscript revisions have resolved your concerns, we would sincerely appreciate your consideration of raising your rating.

Thank you again for your time and consideration. We look forward to continuing the discussion with you.

Sincerely, Authors of Paper 46544

AuthorsOpenReview ↗

Gentle Reminder: Discussion Phase Ends in Less Than 24 Hours

Dear Reviewer Z9qF,

Thank you again for your time and thoughtful review. As fewer than 24 hours remain in the author–reviewer discussion phase, we would be grateful if you could take a brief look at our responses and revisions, and let us know whether any concerns remain so that we can still address them.

If our responses have resolved your concerns, we would sincerely appreciate your consideration in updating your assessment.

Best regards, Authors of Paper 46544

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

This paper proposes Living-Harness, a framework that continuously repairs an agent’s procedural knowledge through a rollout–evaluate–update loop. An Evolution-SOP converts completed trajectories and evaluator feedback into updates to episodic memory and a state graph, which are retrieved in later tasks.

On τ²-Bench and MultiWOZ-2.4, Living-Harness outperforms ReAct, Reflexion, AWM, ReasoningBank, and EvoTest across several interactive environments. Ablations indicate contributions from the Evolution-SOP, episodic memory, and state graph, while cross-model experiments show that a frozen evolved harness state can benefit other backbones.

Critical strengths

Clear formulation of persistent procedural repair. The distinction between task-local reflection and cross-episode repair is meaningful and practically relevant. Interpretable state representation. Episodic memory captures when and why failures occur, while the state graph specifies which action should follow which state. Consistent empirical improvements. The method improves performance across multiple environments in two benchmarks, and the ablations support complementary contributions from the two components.

Critical weaknesses

Strong dependence on domain-specific design. The Evolution-SOP, task scopes, and update rules are environment-dependent, so it is unclear whether Living-Harness is a general self-evolving harness or a customized system for the two benchmarks. Unclear isolation in online evolution. If earlier task outcomes influence later evaluation tasks, the paper should specify how task order, update data, and test data are separated. Otherwise, part of the gain may come from benchmark-internal memorization. Insufficient cost-matched comparison. Living-Harness uses the actor, evolution modules, router, and memory updater. Although auxiliary calls are counted, the paper does not provide performance comparisons matched by tokens, model calls, or API cost. Limited theoretical contribution. The POMDP formulation and information-refinement proposition are useful interpretations, but do not provide substantive guarantees about update stability, error accumulation, or long-term improvement.

Questions for authors

What task ordering is used across evolution cycles, and how are training, update, and evaluation tasks separated? Which parts of the Evolution-SOP are manually designed for each domain, and how much additional design is required for a new domain? What is the total token and API cost of Living-Harness, including routing and update calls? Does it still outperform the baselines under matched budgets? How does the system prevent incorrect memory entries or state-graph edges from persisting and causing future errors?

4 comments
AuthorsOpenReview ↗

Response to Reviewer KGVn (Part 1 of 2)

Thank you for recognizing the practical value of persistent procedural repair, the interpretability of our state representation, and the consistent gains across environments. We address your concerns below.

W1. Strong dependence on domain-specific design. The Evolution-SOP, task scopes, and update rules are environment-dependent, so it is unclear whether Living-Harness is a general self-evolving harness or a customized system for the two benchmarks.

Q2. Which parts of the Evolution-SOP are manually designed for each domain, and how much additional design is required for a new domain?

Response to W1 and Q2: Evolution-SOP is generated automatically by GPT-5.2 from each environment's existing specifications and system prompts, without task ground truth; we have corrected the manuscript's corresponding statement (App. D). It organizes environment policies, tool constraints, and task scopes into update rules that stay fixed during a run. The posterior–extract–commit pipeline and update schemas are shared across domains, and a new environment obtains its Evolution-SOP through the same generation procedure, with no additional manual design.

W2. Unclear isolation in online evolution. If earlier task outcomes influence later evaluation tasks, the paper should specify how task order, update data, and test data are separated. Otherwise, part of the gain may come from benchmark-internal memorization.

Q1. What task ordering is used across evolution cycles, and how are training, update, and evaluation tasks separated?

Response to W2 and Q1: Our main experiments follow the test-time adaptation setting of EvoTest [1] and ReasoningBank [2]: the agent interacts repeatedly with the same task set, and each episode is scored before its evidence enters harness updates. This setting measures improvement through accumulated experience during continued operation.

We further evaluate generalization to unseen tasks. Following prior work that evolves on a subset of τ²-Bench tasks and evaluates on the held-out remainder [3], we split each domain 7:3 into disjoint update and evaluation sets (80/34 tasks for Retail and Telecom, 35/15 for Airline). The harness evolves on the update set only and is then frozen for evaluation. Living-Harness achieves the best result in every domain, averaging 75.90 versus 68.67 for ReasoningBank and 63.86 for the GPT-5.2 base (Table A; App. C, Table 7), so the gains do not depend on revisiting the same tasks.

Table A. Held-out evaluation on τ²-Bench with disjoint update and evaluation sets (GPT-5.2, Pass@1 %; average weighted by task count).

MethodRetailAirlineTelecomAvg
GPT-5.2 base76.4766.6750.0063.86
AWM79.4146.6744.1259.04
ReasoningBank79.4166.6758.8268.67
Living-Harness91.1873.3361.7675.90

References:

[1] EvoTest: Evolutionary Test-Time Learning for Self-Improving Agentic Systems, ICLR 2026.

[2] ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory, ICLR 2026.

[3] Recursive Experiential–Working Memory Evolution for Long-Horizon Agent Harnesses, arXiv 2026.

AuthorsOpenReview ↗

Response to Reviewer KGVn (Part 2 of 2)

W3. Insufficient cost-matched comparison. Living-Harness uses the actor, evolution modules, router, and memory updater. Although auxiliary calls are counted, the paper does not provide performance comparisons matched by tokens, model calls, or API cost.

Q3. What is the total token and API cost of Living-Harness, including routing and update calls? Does it still outperform the baselines under matched budgets?

Response to W3 and Q3: Table B (App. C, Table 10) reports API calls, token usage, cost, and accuracy for Living-Harness and the interactive baselines, with routing and update calls included. Living-Harness costs USD 33.50, on par with Reflexion (USD 33.36) and below AWM, ReasoningBank, and EvoTest, while reaching the highest accuracy (83.09%). It also generates the second-fewest output tokens, after ReAct. Relative to ReAct, each additional dollar yields 2.18 accuracy points, versus 1.40 for Reflexion; the other self-evolving baselines spend more than ReAct yet score lower.

Table B. Total cost and performance on τ²-Bench (GPT-5.2). Marginal cost efficiency is the accuracy gain over ReAct per additional USD.

MethodAPI Calls (K) ↓Input Tokens (M) ↓Output Tokens (M) ↓Cost (USD) ↓Accuracy (%) ↑Marginal Cost Efficiency (pp/USD) ↑
ReAct4.3119.630.46$20.8755.54—
Reflexion6.4728.150.69$33.3673.021.40
AWM5.9028.020.73$34.7751.08−0.32
ReasoningBank6.0029.700.82$35.3652.52−0.21
EvoTest18.65103.582.52$168.5650.62−0.03
Living-Harness6.6734.990.57$33.5083.092.18

W4. Limited theoretical contribution. The POMDP formulation and information-refinement proposition are useful interpretations, but do not provide substantive guarantees about update stability, error accumulation, or long-term improvement.

Q4. How does the system prevent incorrect memory entries or state-graph edges from persisting and causing future errors?

Response to W4 and Q4: Our formalization separates within-task inference of the environment state from cross-task updates of the procedural state. Under nested information sets, conditioning on the procedural state never increases the Bayes-optimal error, reducing it by a non-negative gap Γz\Gamma_z (App. A). For approximate implementations, the error bound tightens whenever this gain exceeds the added approximation error. Episodic memory and the state graph implement the procedural state, storing retrievable repair knowledge and workflow structure.

To keep erroneous entries or edges from propagating, the Evolution-SOP stays fixed and every candidate update must pass five gates: schema, scope, evidence, constraint, and merge (App. B.2). Table 2 shows improvement over three evolution cycles, with performance above the initial level in all eight domains in every cycle.

We hope these clarifications and results address your concerns, and would be grateful if you could reconsider your assessment. We welcome further questions.

AuthorsOpenReview ↗

Looking Forward to Your Feedback and Further Discussion

Dear Reviewer KGVn,

We hope you are doing well. We sincerely appreciate the thoughtful feedback and constructive suggestions you have provided.

As the discussion deadline approaches, we would be grateful if you could review our responses when you have a chance and let us know whether any concerns remain. Your feedback would allow us sufficient time to address additional questions and provide any further information that would support the discussion.

If our rebuttal and manuscript revisions have resolved your concerns, we would sincerely appreciate your consideration of raising your rating.

Thank you again for your time and consideration. We look forward to continuing the discussion with you.

Sincerely, Authors of Paper 46544

AuthorsOpenReview ↗

Gentle Reminder: Discussion Phase Ends in Less Than 24 Hours

Dear Reviewer KGVn,

Thank you again for your time and thoughtful review. As fewer than 24 hours remain in the author–reviewer discussion phase, we would be grateful if you could take a brief look at our responses and revisions, and let us know whether any concerns remain so that we can still address them.

If our responses have resolved your concerns, we would sincerely appreciate your consideration in updating your assessment.

Best regards, Authors of Paper 46544

Comments on the paper

1 comment
AuthorsOpenReview ↗

Revision Summary

We thank all reviewers for their constructive feedback and have uploaded a revised manuscript, with the main additions to the text and appendix marked in blue. Below we summarize the revisions, grouped by the reviewer whose comments prompted them. All section, figure, and table numbers refer to the revised manuscript.

Reviewer KGVn

  • App. D (SOP construction and portability). Each Evolution-SOP is generated automatically by GPT-5.2 from the environment's existing specifications and system prompts, without evaluation trajectories or ground truth, and the posterior–extract–commit pipeline and update schemas are shared across domains (W1, Q2).
  • App. C (Table 7, new). Held-out evaluation on τ²-Bench, with each domain split 7:3 into disjoint update and evaluation sets (W2, Q1).
  • App. D (Evaluation scope). We specify that the main evolution cycles revisit the same task instances, that the score-before-update protocol separates each episode's score from its updates, and that Tables 7 and 8 evaluate disjoint tasks (W2, Q1).
  • App. C (Table 10, new) and App. B.3. Total API calls, tokens, and cost of all methods, including the routing and update calls of Living-Harness (W3, Q3).

Reviewer Z9qF

  • Fig. 1. Re-exported with standard embedded fonts, so all labels render correctly in common PDF viewers (Q1). This change is not marked in blue.
  • App. D. The revised evaluation scope and SOP construction paragraphs state the limitations of the design (W4).

Reviewer 9kLj

  • App. D (SOP construction and portability). The corrected description of how the Evolution-SOPs are generated (W1, Q1).
  • App. C (Table 9, new). Equal-information comparison, in which Reflexion, AWM, and ReasoningBank receive the same Evolution-SOP text (W1, Q1).
  • App. C (Tables 7 and 8, new). Held-out evaluation, and cross-backbone transfer of the frozen state to GLM-5 and Kimi-K3 on held-out tasks (W2, Q2).
  • Sec. 4.4. We state that Table 3 evaluates cross-backbone reuse on the same MultiWOZ-2.4 task set used for evolution (W2, Q2).
  • Table 1 caption. Living-Harness and all interactive baselines report the best observed performance over three evolution cycles under the same selection rule (Q2).
  • Tables 1 and 2, Sec. 4.2. Telecom Cycle 0 is corrected to 52.63 (60/114), the GPT-5.2 τ²-Bench average to 57.55, and the Telecom Cycle 1 gain to +14.91. Sec. 4.2 also reports that Airline peaks at 88.00 in Cycle 2 and finishes Cycle 3 at 86.00 (Q2).

Other changes

  • We also made minor wording edits throughout for clarity, which do not change any claim or result.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.