acceptodds
Under review as a conference paper at ICLR 2027

EvoIn: Bridging Evolution and Internalization for Agent Fine-Tuning

Abstract

Recent work has explored improving agents by jointly evolving their harnesses and models, but often takes a “potpourri” approach that bundles together new tools, new decision-making procedures, and model adaptation to the evolved harness under a single notion of agent improvement. In this paper, we instead investigate how agents can improve their decision-making procedures. In particular, we propose EvoIn, an agent fine-tuning framework that bridges evolution and internalization. EvoIn first analyzes agent execution traces to evolve and validate new decision-making procedures by temporarily instantiating them in the harness. The validated procedures guide the agent to generate improved reasoning traces. These traces are then rewritten into self-contained reasoning traces, removing explicit references to harness instructions while expressing the induced decision logic as the model's own reasoning. Finally, EvoIn fine-tunes the model on the rewritten traces, internalizing these procedures so that the improved decision-making persists without the evolved harness at inference time. We evaluate EvoIn on diverse benchmarks and find that it consistently enables agents to learn stronger decision-making procedures, raising the pass rate by 10.9 points in-domain and by 9.2 points out-of-domain. Results further show that the internalized decision procedures generalize to unseen tasks. Case studies show that agents can learn to decide how to solve a task before solving it, for example by checking a document's length to choose between reading it in full and searching it. EvoIn is also broadly applicable, showing consistent improvements on another model family.

Primary area:
foundation or frontier models, including LLMs
Keywords:
language agent, agent fine-tuning, agent evolution, harness

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

This paper proposes EvoIn, a framework that combines harness evolution with model fine-tuning to improve agents' decision-making capabilities. Specifically, a stronger model (Claude Opus 4.7) analyzes failed trajectories and proposes improved decision-making procedures through harness evolution. The target model (Qwen3.5-35B-A3B) then generates successful trajectories under the evolved harness. These trajectories are rewritten by TAILOR to remove dependencies on the evolved harness and are subsequently used for supervised fine-tuning. Experiments demonstrate substantial improvements of 10.9 and 9.2 percentage points on in-domain and out-of-domain benchmarks, respectively.

Critical strengths
  1. The paper addresses a meaningful question: how to internalize effective decision-making procedures discovered through harness evolution into the model itself, rather than relying on external instructions at inference time.

  2. Significant empirical improvements. EvoIn achieves substantial performance gains across multiple benchmarks. The ablation studies also demonstrate the effectiveness of harness evolution and trajectory rewriting, particularly for out-of-domain generalization.

  3. Clear presentation. The paper is well-written and easy to follow. The motivation and overall framework are clearly explained, and the case studies provide intuitive examples of the proposed approach.

Critical weaknesses
  1. The idea of internalizing external strategies or reasoning procedures into model parameters has been extensively explored. Recent works such as HarnessForge[1] and Co-Harness[2] have also investigated harness evolution followed by model fine-tuning using trajectories generated under improved harnesses. Therefore, the main novelty of EvoIn appears to lie in the TAILOR trajectory rewriting stage. However, TAILOR not only removes harness-specific instructions but also repairs invalid tool calls, reasoning inconsistencies, and other trajectory quality issues. It remains unclear whether the observed improvements come from genuine procedure internalization or simply from higher-quality SFT data.

  2. EvoIn relies on Claude Opus 4.7 to discover effective decision-making procedures, which are then expressed as harness instructions to guide trajectory generation by a weaker model. However, if Opus 4.7 already possesses strong agentic decision-making capabilities, why not directly use it to generate high-quality trajectories for distillation? The proposed pipeline introduces additional stages of harness evolution, trajectory generation, and rewriting, but the paper does not demonstrate whether these stages provide benefits over direct distillation from a strong teacher. This makes it difficult to assess the necessity of the proposed framework.

[1] Chen, Ming-Ju et al. “HarnessForge: Joint Harness and Policy Evolution for Adaptive Agent Systems.” [2] Chen, Zheng-Yu et al. “Co-Harness: Co-Evolving Harnesses and Model Weights for LLM Agents.”

Questions for authors

What is the key methodological novelty of EvoIn compared with existing harness evolution and distillation methods? In particular, is TAILOR the main contribution? Could the authors provide more fine-grained ablations separating the effects of removing harness-specific dependencies from general trajectory cleaning and quality improvement?

5 comments
AuthorsOpenReview ↗

Response to reviewer EVdk (part 1)

Thank you for the careful review and constructive suggestions. Both of your concerns ask where EvoIn's gains actually come from, which gives us a good opportunity to make this clear. Weakness 1 and the Question concern the same issue, so we answer them together.

Weakness 1 & Question: The main novelty seems to lie in TAILOR, and it is unclear whether the gains come from genuine procedure internalization or simply from higher-quality SFT data.

Answer: We thank the reviewer for raising this question. We would like to clarify that EvoIn differs from existing harness–model co-evolution methods in its core motivation and corresponding methodological design. HarnessForge [1] and Co-Harness [2] jointly optimize models and harnesses to improve performance on target tasks. This optimization direction faces two generalization risks. First, models may over-adapt to the evolved harness, limiting the transfer of their gains to other harnesses [1]. Second, the system may increasingly exploit patterns specific to the target tasks without acquiring strategies that generalize to unseen tasks [3]. Unlike these works, our core insight is that models can discover and validate effective decision-making and reasoning strategies through harness evolution, then internalize the reusable logic to improve their own capabilities. Here, the harness actually provides a space for exploration, while the ultimate goal is to improve the model's own reasoning and decision-making abilities beyond a specific harness or task domain. Moreover, the OOD improvements reported in Table 2 of our manuscript further support our motivation.

To address your concerns about Tailor, we conduct a 2×2 ablation separating harness-dependency removal from general trajectory quality improvement. The four settings are no rewriting, dependency removal only, quality improvement only, and full Tailor. We use the same task set, training-set size, and training recipe across these settings, with results reported in the table below. Neither dependency removal alone nor quality improvement alone yields a statistically significant Pass improvement over no rewriting. However, full Tailor significantly outperforms all three ablated settings on OOD Pass by 4.94–6.40 percentage points. These results show that general trajectory cleaning alone is insufficient to explain the full method's performance. Removing either component from full Tailor significantly reduces OOD performance, supporting the contribution of both components when used together.

MethodID PassID ScoreHeld-out ID PassHeld-out ID ScoreOOD PassOOD Score
Base35.9644.5834.1743.0718.9931.02
No Tailor44.2253.9741.3550.9816.2842.67
Dependency removal only42.8752.3742.2251.7116.2842.34
Quality improvement only43.7853.6542.1751.8017.7342.00
Full Tailor44.8754.4143.0452.7522.6748.73
AuthorsOpenReview ↗

Response to reviewer EVdk (part 2)

Weakness 1 & Question (continued)

Our response to Reviewer nbkV reports case studies of how internalization changes the behavior of the model. For your convenience we reproduce two of them here, both on OOD benchmarks, where Base and EvoIn answer the same question under the same seed harness. If the model itself did not have this decision-making ability, it would have to rely on the harness to spell out the procedure for each task. Such decisions vary widely across tasks, and it is hard to summarize all of these behaviors into a single general harness. More case studies can be found in our response to Reviewer nbkV.

(1) Giving each stage a scope, and repairing a failed method. On Table-Longer, an out-of-domain benchmark, a 2,482-line table document is queried in two stages. The first stage collects the companies that have a project which is both delayed and of environmental grade 3, and the second keeps those of them that also have a safety project with expected ROI strictly above 20. The trap is that one qualifying safety project is recorded as completed rather than delayed, so the second condition has to be checked against all records of each candidate, not only the records that made it a candidate. Both models stated the two stages in their reasoning. Base then tested all conditions on single records and lost that company. EvoIn wrote the two stages into its notes with a scope for each. When its first parser returned no records, it re-read the raw lines, found the company name on the line after the record tag, and rewrote it. In the second stage it searched all records of the 15 candidates, which is how it recovered the company Base missed.

(2) Producing the missing labels instead of searching for them. Oolong, another out-of-domain benchmark, gives the model a 3,193-line log in which each line holds one question with its date and user id, but no class label. Asked whether "numeric value" is more or less common than "abbreviation" among one user's entries, the model must therefore classify those entries itself before it can count them. Base retrieved all lines of that user in its second turn, then spent 27 further turns re-running the same search and looking for a label field that does not exist, and its 30 turns ran out with no answer. EvoIn read the hits in segments and wrote one ledger entry per record for the 13 records, each with its line number and the class assigned to it, counted the two classes from the ledger, and answered at turn 10.

AuthorsOpenReview ↗

Response to reviewer EVdk (part 3)

Weakness 1 & Question (continued)

We also train on trajectories sampled from GLM-5.3 directly under the seed harness, which can be regarded as high-quality SFT data. With the same rollout model and the same training questions, EvoIn still outperforms this direct distillation on both ID and OOD.

MethodID PassID ScoreOOD PassOOD Score
Base + seed harness35.9644.5818.9931.02
Direct distillation from GLM-5.3 under the seed harness49.2658.3822.8747.10
EvoIn with GLM-5.3 rollouts52.9661.8324.5648.26

Weakness 2: If Opus 4.7 already possesses strong agentic decision-making capabilities, why not directly use it to generate high-quality trajectories for distillation?

Answer: We thank the reviewer for raising this question. The Claude Opus 4.7 API returns only summarized thinking, so the complete reasoning process is not accessible and cannot be used directly for SFT. To address the reviewer's concern, we instead use GLM-5.3, whose full reasoning traces are available, to generate high-quality trajectories for distillation. Collecting GLM-5.3 trajectories under the evolved harness outperforms collecting them directly under the seed harness, which indicates that EvoIn is more effective than directly learning from high-quality trajectories. Besides the ID test set used in the evolution stage, the table also reports an independent ID test set of 2,300 held-out questions that we constructed in response to Reviewer nbkV and that was used neither in harness evolution nor in training, and the three settings keep the same ranking on both.

MethodDdevD_{\text{dev}} (2,300)Held-out ID (2,300)Gain on DdevD_{\text{dev}}Gain on held-out
Base + seed harness44.5843.07——
Direct distillation from GLM-5.3 under the seed harness58.3857.87+13.80+14.80
EvoIn with Qwen rollouts54.4152.75+9.83+9.68
EvoIn with GLM-5.3 rollouts61.8358.38+17.25+15.31

Moreover, EvoIn is not limited to knowledge transfer from a stronger teacher to a weaker student. Rather, our insight is that a model can discover better decision-making procedures, then internalize them to improve its own capabilities. Actually, we also evaluate EvoIn using the same model for both evolution and internalization. Specifically, as reported in Section 5 of our manuscript, we use Qwen3.5-35B-A3B to drive the EvoIn pipeline and fine-tune the same model, achieving performance gains over the baseline. From this perspective, EvoIn offers a potential path toward autonomous model improvement and motivates further exploration of recursive self-improvement.

[1] HarnessForge: Joint Harness and Policy Evolution for Adaptive Agent Systems, 2026.

[2] Co-Harness: Co-Evolving Harnesses and Model Weights for LLM Agents, 2026.

[3] Rethinking the Evaluation of Harness Evolution for Agents, 2026.

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

This paper proposes EvoIn, an agent fine-tuning framework that bridges harness evolution and model internalization to improve agents' decision-making capabilities. EvoIn first iteratively evolves and evaluates agent harnesses to discover effective decision-making procedures. The evolved harnesses are then used to generate successful agent trajectories, which are rewritten into self-contained trajectories that remove dependencies on the evolved harness while preserving useful decision-making strategies. Finally, these trajectories are used to fine-tune the original model, enabling it to internalize the improved procedures without requiring the evolved harness at inference time. Experiments across 23 in-domain task categories and six out-of-domain benchmarks demonstrate consistent performance improvements, with pass rates increasing by 10.9 and 9.2 percentage points, respectively. Ablation studies further investigate the contributions of harness evolution and trajectory rewriting.

Critical strengths

This paper has several clear strengths:

  1. Clear presentation. The paper is well-written and clearly organized. The method is easy to understand, and the overall framework is presented in a logical and accessible manner.

  2. Straightforward and easy-to-follow methodology. The proposed method is clearly described, with each stage serving a well-defined purpose. The integration of harness evolution, trajectory rewriting, and model fine-tuning is intuitive and easy to follow.

  3. Relatively comprehensive experimental evaluation. The experiments cover a diverse set of tasks, including 23 in-domain task categories and six out-of-domain benchmarks, providing evidence of both performance improvements and generalization to unseen tasks. The paper also includes ablation studies and additional analyses to investigate the contributions of key components, making the experimental evaluation relatively thorough.

Critical weaknesses
  1. Insufficiently justified motivation for internalizing decision-making procedures into the original harness. Although the proposed method is clearly presented and offers some interesting insights, I find the central motivation insufficiently justified. The authors emphasize the importance of internalizing evolved decision-making procedures into the model so that it can operate under the original harness. However, it remains unclear why this is a desirable objective in the first place, rather than simply retaining and using the evolved harness. One possible motivation is to consolidate capabilities discovered through different task-specific harnesses into a single model operating under a shared, lightweight harness. While the authors may indeed perform joint fine-tuning on trajectories collected across different task categories (This is my interpretation of the experimental setup, and the authors should explicitly clarify this point in the paper.), this motivation is not clearly articulated or sufficiently investigated. In particular, the paper lacks a comparison against directly using evolved harnesses or developing a unified harness across tasks. Such comparisons would help clarify the practical benefits of internalization and better justify the necessity of the proposed evolution–internalization framework.

  2. Concerns about the experimental setting and task selection. Although the experimental results demonstrate encouraging improvements, I have some concerns about the choice of training tasks. The in-domain tasks appear somewhat repetitive, with several categories focusing on similar long-context retrieval, document understanding, and structured information processing capabilities. In contrast, some of the out-of-domain benchmarks are more representative and challenging. This raises the question of why the authors do not directly apply their proposed framework to these more realistic and challenging tasks. For example, BrowseComp represents a more challenging exploration-oriented setting, although the paper evaluates its long-context adaptation rather than the original interactive browsing task. It would be interesting and potentially more convincing to investigate whether EvoIn can directly evolve and internalize decision-making procedures in such environments, rather than primarily relying on the current collection of relatively similar in-domain tasks.

Additionally, I noticed a potential issue with the evaluation protocol: the evolution-test split used to select the best harness also serves as the final in-domain test set (line 265, 266). Although these examples are not directly used for SFT, repeatedly evaluating and selecting harnesses on this split may introduce selection bias. I would encourage the authors to evaluate the fine-tuned model on an additional held-out test set that is entirely independent of the harness evolution and selection process.

Questions for authors
  1. Could the authors extend the evaluation to more challenging and representative agent tasks, such as the original interactive BrowseComp setting? Although the paper evaluates BrowseComp-LongContext as an OOD benchmark, it would be valuable to investigate whether EvoIn can directly evolve and internalize decision-making procedures in more complex, exploration-oriented environments.

  2. Could the authors further clarify the central motivation for internalizing evolved procedures into the original harness? In particular, why is returning to the original harness preferable to directly using the evolved harness? A clearer discussion of the intended benefits and practical scenarios would help strengthen the motivation.

  3. Could the authors provide more diverse case studies demonstrating that the fine-tuned model has internalized meaningful changes in its decision-making workflow? It would be valuable to see whether the model can autonomously adopt more sophisticated planning or exploration strategies, rather than primarily improving existing retrieval and verification procedures.

  4. The evolution-test split used for selecting the best harness also serves as the final in-domain evaluation set. Could the authors clarify this design choice and, if possible, evaluate the fine-tuned model on an additional held-out test split that is not involved in either harness evolution or harness selection?

5 comments
AuthorsOpenReview ↗

Response to reviewer nbkV (part 1)

Thank you for the careful review and for recognizing the clarity of the presentation, the design of the method and the coverage of the experiments. Several of the weaknesses and questions concern the same issue, so we group the related ones and answer them together below.

Weakness 1 & Question 2: Insufficient motivation for internalizing into the original harness, and missing comparisons with directly using evolved or unified harnesses.

Answer: We thank the reviewer for raising this question. EvoIn's core motivation is to use harness evolution to discover and validate effective decision-making procedures, then internalize strategies discovered across tasks into a single model to improve its decision-making and reasoning capabilities.

Directly using evolved harnesses is a viable option. In practical applications, however, we want a single agent to handle diverse tasks, including previously unseen ones. Evolved harnesses are optimized separately for specific task categories, and deploying them by category requires additional task identification and routing mechanisms. For an unseen OOD task, which harness to select and how to adapt it are not immediately clear. Internalization provides a way to consolidate these strategies, allowing one model to draw on decision-making experience from different tasks under a shared harness. Table 2 in our manuscript reports improvements on OOD tasks, supporting the transfer of capabilities acquired through internalization without requiring harnesses specifically designed for those new tasks.

Furthermore, our baseline seed harness is itself a unified harness shared across tasks, and our experiments show that EvoIn improves the model's capabilities within this unified framework. Evolving harnesses separately allows us to explore strategies tailored to each task without having to accommodate different tasks' requirements within a single external workflow. Internalization then consolidates these strategies into one model for reuse and transfer across tasks. Moreover, a unified harness can certainly be further optimized, and a stronger unified harness can likewise serve as EvoIn's starting point for discovering and internalizing additional effective strategies.

Overall, EvoIn is a model optimization method that aims to consolidate strategies discovered across tasks into the model, enabling their flexible reuse and transfer to improve its reasoning capabilities. Directly using the evolved harnesses remains a more task-specific solution, while real-world tasks are often unpredictable and may not fit predefined task categories. Our model optimization approach is also complementary to further improving a unified harness.

In addition, in our response to Reviewer QdfF we have added results for applying evolved harnesses directly to the ID and OOD evaluations. We take three of the strongest ones, namely the harness with the highest accuracy on its own evolution test set and the two with the largest gains over the seed harness. For the reviewer's convenience, we report the main results in the table below. Each of the three outperforms the seed harness on both the ID and the OOD sets, but all of them still fall short of EvoIn. We hope this experiment further alleviates the reviewer's concern about the performance of EvoIn.

MethodID PassID ScoreOOD PassOOD Score
Base + seed harness35.9644.5818.9931.02
Base + 3 evolved harnesses (mean)40.2349.3224.6939.75
EvoIn + seed harness46.8356.4028.2054.49
AuthorsOpenReview ↗

Response to reviewer nbkV (part 2)

Weakness 2 & Question 1: Repetitive in-domain tasks and lack of evaluation or evolution on interactive exploration tasks such as original BrowseComp.

Answer: We thank the reviewer for this suggestion, and we agree that evaluating on the original BrowseComp helps examine how EvoIn performs on interactive browsing tasks. We would first like to clarify that the 23 training categories are not homogeneous retrieval tasks: they cover multi-document cross-section reasoning, dynamic log state tracking, programmatic table computation, group-chat statistics, few-shot imitation, and multi-constraint generation, which require different decision-making procedures. To alleviate the reviewer's concern about how the EvoIn-tuned model performs in more complex interactive browsing environments, we evaluated it on the questions of the original BrowseComp that overlap with BrowseComp-LongContext. As shown in the table below, although EvoIn has never been trained with web tools, its accuracy still improves from 7.12% for Base to 9.49%, indicating that the internalized planning and budget-control procedures transfer to interactive exploration environments. Our evaluation setting follows the original BrowseComp paper. We will add this experiment and the corresponding analysis to the experiments section, and provide the detailed execution protocol in the appendix.

SettingBaseEvoIn
Interactive (live web)7.129.49
Given documents12.5413.56

Additional Weakness & Question 4: Reusing the evolution-test split (DdevD_{\text{dev}}) for final ID evaluation may introduce selection bias; clarify this choice and evaluate on an independent test set.

Answer: We thank the reviewer for raising this question. In EvoIn, DdevD_{\text{dev}} is used only to score candidate harnesses: its accuracy tracks the optimization progress of the harness during evolution, and the highest-scoring harness of each category is kept as the evolved harness H⋆H^\star. We reuse DdevD_{\text{dev}} as the ID test set so that the evolved harness and the internalized model are compared on the same questions, which shows how much of the gain from harness evolution is internalized into the model. Under this design, harness selection can influence the ID results only indirectly. The model is never trained on DdevD_{\text{dev}}, harness selection only decides which harness generates the training trajectories on DtrainD_{\text{train}}, and the fine-tuned model is evaluated with the seed harness H0H_0 rather than the selected harness. Moreover, EvoIn also improves clearly on the six OOD benchmarks, which play no role in harness selection, so its gains are not limited to the questions used for selection.

To alleviate the reviewer's concern about the performance of EvoIn, we constructed an independent ID test set of 2,300 questions by sampling 100 questions from each topic, none of which was used during harness evolution or during training. The table below reports Base and three models fine-tuned on the same shared set of training questions, on both this held-out ID test set and the ID test set used in the evolution stage. The results show that every setting performs similarly on the two test sets. Finally, we thank the reviewer again for the careful and constructive comments. We hope these results can alleviate the concern about the performance of EvoIn, and we will add this experiment and the corresponding discussion to the experiments section, and clearly distinguish the roles of the two ID test sets in the paper to avoid potential confusion for readers.

MethodDdevD_{\text{dev}} (2,300)Held-out ID (2,300)Gain on DdevD_{\text{dev}}Gain on held-out
Base + seed harness44.5843.07——
Direct distillation from GLM-5.3 under the seed harness58.3857.87+13.80+14.80
EvoIn with Qwen rollouts54.4152.75+9.83+9.68
EvoIn with GLM-5.3 rollouts61.8358.38+17.25+15.31

Question 3: Provide diverse case studies showing internalized workflow changes, especially autonomous planning or exploration strategies.

Answer: Thank you for this question. It gives us a good opportunity to show with concrete cases how internalization changes the decision-making behavior of the model and how those changes turn into better performance. Evidence about the fine-tuned model must come from the model alone, so we ran Base and EvoIn on identical questions under the identical seed harness and compared their trajectories.

AuthorsOpenReview ↗

Response to reviewer nbkV (part 3)

Question 3 (continued)

The five cases below are the differences we found most often in those trajectories, and they come from five different task families, three in domain and two out of domain.

(1) Fixing the structure before searching. Asked whether the last key of doc1 occurs in doc2, Base read the input end to end and took a colon-labelled distractor for that key. EvoIn located both documents with grep("^doc"), enumerated every key with one key-format regex, and wrote the boundaries and key list into its notes before looking anything up. Over 200 multi-document questions, boundaries are determined first on 91.0%, against Base's 27.0%.

(2) Testing a hypothesis when retrieval cannot answer. The item to classify does not appear among the exemplars, and Base answered that the label cannot be determined, as it did on 17 of 100 such questions (none for EvoIn). EvoIn conjectured that the labels name relation types, searched for exemplars with the same entity-time structure, confirmed that the only exemplar sharing the entity expresses a different relation, and took the label of the closest parallel one.

(3) Choosing the computation, and noticing a rule inside the question. Among 162 messages from 49 speakers, Base counted inside its reasoning (23,678 characters) and ranked them wrongly. EvoIn checked the line format, tallied every speaker with one shell pipeline, saw a tie for third place, and resolved it by the question's tie rule, looking up both speakers' first messages. Over 200 group-chat questions, bash is used for full-text counting on 78.5%, against 35.5%.

(4) Giving each stage a scope, and repairing a failed method. On Table-Longer, an out-of-domain benchmark, a 2,482-line table document is queried in two stages. The first stage collects the companies that have a project which is both delayed and of environmental grade 3, and the second keeps those of them that also have a safety project with expected ROI strictly above 20. The trap is that one qualifying safety project is recorded as completed rather than delayed, so the second condition has to be checked against all records of each candidate, not only the records that made it a candidate. Both models stated the two stages in their reasoning. Base then tested all conditions on single records and lost that company. EvoIn wrote the two stages into its notes with a scope for each. When its first parser returned no records, it re-read the raw lines, found the company name on the line after the record tag, and rewrote it. In the second stage it searched all records of the 15 candidates, which is how it recovered the company Base missed.

(5) Producing the missing labels instead of searching for them. Oolong, another OOD benchmark, gives the model a 3,193-line log in which each line holds one question with its date and user id, but no class label. Asked whether "numeric value" is more or less common than "abbreviation" among one user's entries, the model must therefore classify those entries itself before it can count them. Base retrieved all lines of that user in its second turn, then spent 27 further turns re-running the same search and looking for a label field that does not exist, and its 30 turns ran out with no answer. EvoIn read the hits in segments and wrote one ledger entry per record for the 13 records, each with its line number and the class assigned to it, counted the two classes from the ledger, and answered at turn 10.

Why do we read this as internalization? The seed prompt suggests scan, grep and read_lines, notes when useful, bash for aggregation, and a closing self-check. It never asks for boundary tables, stage scopes, parallel exemplars, tie rules or item-by-item ledgers. Every behavior above except the parser repair matches a rule that harness evolution wrote for one training category and that Tailor then stripped of its wording while keeping the actions. Table-Longer and Oolong had no evolved harness at all, so (4) and (5) are transfer. Two ablations agree. Fine-tuning on the same number of successful seed-harness trajectories gives +1.87 in-domain Pass and no out-of-domain gain, and removing Tailor drops out-of-domain Pass from 28.20 to 19.67.

We will add the five trajectories, with each step marked as prescribed by the seed harness or not, and these statistics to the appendix.

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

This paper proposes EvoIn, a method that internalizes decision-making procedures discovered through harness evolution into the model. Experiments show that EvoIn improves agent performance and that the improvements generalize to unseen tasks.

Critical strengths

The paper proposes a method for internalizing decision-making procedures discovered through harness evolution into the model, addressing a research problem with practical significance.

The experiments cover multiple in-domain task categories and out-of-domain benchmarks, demonstrating substantial performance improvements. The ablation studies also highlight the importance of harness evolution and trajectory rewriting.

Critical weaknesses

W1. Insufficient Motivation for Internalization The paper does not sufficiently justify why the evolved decision-making procedures need to be internalized into the model. The harness is an important part of an agent system, and keeping decision-making procedures external is not necessarily a drawback. More importantly, the paper does not systematically compare the performance of the evolved harness with that of the fine-tuned model, making it unclear how much of the performance gain is preserved after internalization.

W2. Insufficient Ablation of TAILOR The paper uses the ablation without TAILOR to demonstrate the importance of trajectory rewriting for procedure internalization and OOD generalization. However, TAILOR also includes trajectory repair, error checking, tool-call validation, and training data quality filtering. I suggest evaluating the contributions of harness dependency removal and trajectory quality improvement separately to better validate the core mechanism.

W3. Limited Comparisons with Alternative Training Strategies The paper lacks direct comparisons with other high-quality trajectory generation and distillation strategies. In particular, the method itself relies on stronger models such as Claude Opus 4.7 and GLM-5.2 to assist in generating training data, so these comparisons are important for evaluating the contribution of harness evolution. For example, the authors could add a baseline where a stronger model directly generates training trajectories under the original harness, without harness evolution.

W4. Reuse of the Development Set for ID Evaluation The paper uses D_dev to evaluate and select the evolved harness, but then reuses the same dataset for the final ID evaluation. Although these examples are not directly used for SFT, they have already been involved in harness selection, which may introduce selection bias into the reported ID results. I suggest evaluating ID performance on an independent test set.

Questions for authors

Please see the weaknesses above.

4 comments
AuthorsOpenReview ↗

Response to reviewer QdfF (part 1)

Thank you for the careful review and constructive suggestions. This is a great opportunity to explain the motivation behind EvoIn and its contributions more clearly. We will address your main concerns below.

Weakness 1: Insufficient Motivation for Internalization. Missing the comparison between evolved harness and fine-tuned model, so it is unclear how much gain is retained after internalization.

Answer: We thank the reviewer for raising this issue. We agree that the harness is part of the agent system. In a real deployment, however, one does not maintain a separate harness for every category of data; a single harness has to handle inputs of all types. EvoIn rewrites the trajectories collected under many category-specific harnesses into trajectories that are valid under one single harness, and thereby internalizes into that single harness the decision-making ability acquired on all of these different data types. EvoIn is therefore not merely a way of internalizing a harness into the model: it is a training method that improves the decision-making ability the model exhibits under a harness across a wide range of tasks.

We thank the reviewer for pointing out the missing comparison between the evolved harness and the fine-tuned model, which we report in the tables below. Note that each harness produced during EvoIn's evolution stage is specialized to one task category, and these categories differ in form from the tasks in the OOD evaluation sets, so we cannot select an evolved harness that specifically targets the OOD test tasks. In our experiments we therefore take three of the strongest evolved harnesses, namely the one with the highest accuracy on its own evolution test set and the two with the largest gains over the seed harness, and evaluate each of them on both the ID and the OOD sets. The baseline is the seed harness with the original Qwen model. As the results show, EvoIn with the seed harness achieves a clear gain over the baseline on both ID and OOD. Moreover, each of these three harnesses falls short of EvoIn on both ID and OOD (second table). This shows that EvoIn not only improves ID performance, but that the improvement in the model's decision-making ability also generalizes reliably to OOD tasks.

MethodID PassID ScoreOOD PassOOD Score
Base + seed harness35.9644.5818.9931.02
Base + 3 evolved harnesses (mean)40.2349.3224.6939.75
EvoIn + seed harness46.8356.4028.2054.49
Evolved harnessID PassID ScoreOOD PassOOD Score
Multi-Doc Key Lookup39.3948.5725.5836.00
Structured Chat Analysis39.6148.6324.2744.32
Multi-Step Structured Reasoning41.7050.7724.2238.94

Weakness 2: Insufficient Ablation of Tailor. Evaluating harness dependency removal and trajectory quality improvement separately would better validate the core mechanism.

Answer: We thank the reviewer for this suggestion. To show clearly what each part of Tailor contributes, we divide the operations of Tailor into two categories, harness dependency removal and trajectory quality improvement, and evaluate the contribution of each category separately.

  • Harness dependency removal: this category deletes the rule references and notepad template fields that exist only in the evolved harness, replaces the tools and workspace paths specific to the evolved harness with their counterparts in the seed harness, shortens the trajectory to fit within the step limit of the seed harness, and converts the final answer into the output format required by the seed harness. It is checked only by whether the edited trajectory can be replayed unchanged under the seed harness, not by the quality of the trajectory itself.
  • Trajectory quality improvement: this category deletes tool calls that fail, repeat earlier calls, or return nothing useful, and fixes reasoning that does not match the evidence, incorrect line citations, and formatting errors in the text. Its edits go through three stages in sequence. A fixed set of rules first flags each problem, targeted rewriting then fixes the flagged problems, and finally a model that took no part in the rewriting independently reviews the whole trajectory. All three stages count as part of trajectory quality improvement.
AuthorsOpenReview ↗

Response to reviewer QdfF (part 2)

Weakness 2 (continued)

As the table below shows, Full Tailor clearly outperforms applying only one category or not rewriting at all, both of which give up most of the OOD gain and even fall below Base on OOD Pass, because each category leaves behind what the other removes. Dependency removal alone leaves reasoning that does not match the evidence and citations pointing to the wrong lines, and quality improvement alone leaves the templates, tools and output formats of the evolved harness. Neither alone yields a self-contained trajectory that could have been produced under the original harness, which is what the internalized procedure needs in order to transfer to unseen tasks, so the two categories have to be applied together.

MethodID PassID ScoreHeld-out ID PassHeld-out ID ScoreOOD PassOOD Score
Base35.9644.5834.1743.0718.9931.02
No Tailor44.2253.9741.3550.9816.2842.67
Dependency removal only42.8752.3742.2251.7116.2842.34
Quality improvement only43.7853.6542.1751.8017.7342.00
Full Tailor44.8754.4143.0452.7522.6748.73

The table below lists every problem Tailor checks for and how often it was found and fixed. Tailor intercepts a large number of problems, which is what lets the fine-tuned model carry out the decision procedure under the original harness instead of citing instructions and tools that no longer exist there. 76.7% of the trajectories pass every check and become SFT data.

CategoryProblem found and fixedShare of trajectories
Dependency removalWording or notepad fields that exist only in the evolved harness42.0%
Final answer not in the output format the seed harness requires18.9%
Tools or workspace paths that the seed harness does not have6.2%
Trajectory longer than the step limit of the seed harness0.6%
A replayed tool call returns a different observation under the seed harness26.2%
Replay differs enough that the trajectory is discarded and re-sampled5.7%
Cannot be made valid under the seed harness at all, and is discarded1.1%
Quality improvementTool calls that fail, repeat an earlier call, or return nothing32.3%
Reasoning that does not match the evidence, or a wrong line citation31.3%
Text problems such as newline encoding or decimal precision5.6%
Independent review finds the conclusion contradicts the cited text5.1%
Independent review finds a claim the document does not support4.0%
Independent review finds a citation pointing to the wrong line3.0%

Weakness 3: The paper lacks direct comparisons with alternative trajectory generation and distillation strategies.

Answer: We thank the reviewer for this suggestion. We agree that a baseline in which a stronger model generates training trajectories directly under the original harness, without any harness evolution, is important for clarifying the value of the EvoIn pipeline itself. We therefore sampled trajectories from GLM-5.3 directly under the seed harness without harness evolution or Tailor, and trained both of the settings below on the questions that GLM-5.3 was able to solve in this way, about 85% of the paired training set.

MethodID PassID ScoreOOD PassOOD Score
Base + seed harness35.9644.5818.9931.02
Direct distillation from GLM-5.3 under the seed harness49.2658.3822.8747.10
EvoIn with GLM-5.3 rollouts52.9661.8324.5648.26

With the same rollout model and the same training questions, EvoIn outperforms direct distillation on both ID and OOD. We thank the reviewer again for the suggestion, and we will add this experiment and a discussion of it to Table 3 of the paper.

AuthorsOpenReview ↗

Response to reviewer QdfF (part 3)

Weakness 4: Reusing DdevD_{\text{dev}} to both select the evolved harness and evaluate final ID performance may introduce selection bias; ID performance should be evaluated on an independent test set.

Answer: We thank the reviewer for raising this issue. In EvoIn, DdevD_{\text{dev}} is used only to score candidate harnesses during evolution, and the highest-scoring harness of each category is kept as the evolved harness. We then reuse DdevD_{\text{dev}} as the ID test set so that the evolved harness and the internalized model are measured on the same questions, which is what shows how much of the gain from evolution survives internalization. Under this design, harness selection can affect the ID results only indirectly. The model is never trained on DdevD_{\text{dev}}, selection only decides which harness generates the training trajectories on DtrainD_{\text{train}}, and the fine-tuned model is evaluated under the seed harness rather than under the selected one. EvoIn also improves clearly on the six OOD benchmarks, which shows that its gains generalize to tasks beyond the training data.

To address the reviewer's concern directly, we built an independent ID test set of 2,300 questions, 100 drawn from each topic, none of which was used during harness evolution or during training. The table below reports Base and three models fine-tuned on the same shared set of training questions, on this held-out set and on the ID test set used during evolution, and every setting performs similarly on the two. We will add this experiment to the experiments section and state clearly what each of the two ID test sets is for.

MethodDdevD_{\text{dev}} (2,300)Held-out ID (2,300)Gain on DdevD_{\text{dev}}Gain on held-out
Base + seed harness44.5843.07——
Direct distillation from GLM-5.3 under the seed harness58.3857.87+13.80+14.80
EvoIn with Qwen rollouts54.4152.75+9.83+9.68
EvoIn with GLM-5.3 rollouts61.8358.38+17.25+15.31
AuthorsOpenReview ↗

Gentle Follow-up as Discussion Ends

Thank you for your valuable feedback. As the discussion period draws to a close, we would like to know whether our responses have adequately addressed your concerns, or whether there are any remaining questions we can help resolve. We remain available to provide any additional information or clarification that would be helpful.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.