acceptodds
Under review as a conference paper at ICLR 2027

Skill2Env: Capability-Oriented Environment Synthesis from Skills for General Agents

Abstract

Executable environments are critical for post-training agents on tasks that require tool use and multi-step interaction, but constructing executable tasks together with their environments remains difficult to scale. Skills provide reusable domain knowledge, operational procedures, and tool-use instructions, but a substantial gap remains between the information contained in a skill and a concrete, challenging task with a complete executable environment. To address this gap, we introduce Skill2Env, a capability-oriented framework that starts from a skill and uses agent capability demands to guide task and environment synthesis. Skill2Env represents these demands through reusable difficulty patterns and instantiates them into task blueprints that specify objectives, challenges, environment facts, information boundaries, and acceptance criteria. These blueprints guide the joint construction of task instructions, execution substrates, workspaces, and rubric-based evaluators around source skills. We further propose Iterative Task Hardening, which uses solver execution evidence to identify insufficiently challenging task designs, strengthen or extend their difficulty-pattern instantiations, and revise the corresponding blueprints and environments. Using 1.5K high-scoring trajectories generated from Skill2Env environments for supervised fine-tuning, we observe consistent improvements across a broad range of agent benchmarks, demonstrating the effectiveness of capability-oriented environment synthesis for agent post-training.

Primary area:
foundation or frontier models, including LLMs
Keywords:
environment synthesis, llm agents, large language model

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_rNdKOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Skill2Env, a framework that synthesizes executable agent-training tasks and environments from skills using capability-oriented difficulty patterns and task blueprints. It further uses solver feedback to progressively harden the tasks. The pipeline produces 2,963 tasks, and fine-tuning Qwen3.6-35B-A3B on 1.5K high-scoring teacher trajectories collected from these environments improves performance across seven benchmarks.

Critical strengths
  1. The framework provides a practical way to turn general skills into concrete training environments. It explicitly identifies 5 capability demands and designs 100 difficulty patterns.

  2. Experimental results show that Skill2Env achieves substantial improvements over seven benchmarks. Improvements also hold across four Terminal-Bench harnesses.

Critical weaknesses
  1. The contribution of each capability demand is unclear. You should include an ablation study showing whether training on tasks designed for a particular capability actually improves that capability and how each capability demand contributes to the overall gains. Do different benchmarks focus on different capabilities?

  2. The five capability categories need clearer support. For example, you could explain which models struggle with each capability, on which tasks or benchmarks, and how many execution traces were analyzed. A chart summarizing this information would be helpful.

Questions for authors
  1. Was SFT performed in thinking mode or non-thinking mode? Specifically, did the training data include the teacher model’s reasoning traces?
4 comments
AuthorsOpenReview ↗

Response (Part 1/2): Empirical Evidence for the Five Capability Demands (W2)

Thank you for recognizing the practical value of Skill2Env and its consistent improvements across benchmarks. We appreciate your suggestions on better supporting the five capability demands and clarifying their contributions to agent training. We address these concerns in two parts.

1. Motivation: From agent execution to shared capability demands.

Our framework adopts a capability-demand-oriented perspective. As described in Section 4.2, we identify five general capability demands based on the requirements of agent execution: environment understanding, planning, skill usage, long-horizon consistency, and error recovery. To operationalize these demands in task synthesis process, we introduce reusable difficulty patterns. Importantly, capability demands describe what agents need to be capable of, while difficulty patterns specify how these capabilities are challenged in executable tasks.

2. Empirical evidence from model execution failures. In our failure analysis process, we analyzed 500 failure trajectories from five agent models. We identified 848 failure reasons and categorized them into the five capability demands. A single trajectory may involve multiple failure reasons.

Table: Distribution of capability-related failure reasons across 500 trajectories.

ModelFailure TrajectoriesTotal Failure ReasonsEPSLR
Qwen3.6-35B-A3B10016853 (31.5%)49 (29.2%)10 (6.0%)35 (20.8%)21 (12.5%)
Qwen3.5-397B-A17B10017050 (29.4%)46 (27.1%)11 (6.5%)39 (22.9%)24 (14.1%)
Gemini-3.1 Pro10017049 (28.8%)35 (20.6%)21 (12.4%)48 (28.2%)17 (10.0%)
GPT-5.410017769 (39.0%)40 (22.6%)10 (5.6%)35 (19.8%)23 (13.0%)
Kimi-K2.610016352 (31.9%)49 (30.1%)18 (11.0%)32 (19.6%)12 (7.4%)
Total500848273 (32.2%)219 (25.8%)70 (8.3%)189 (22.3%)97 (11.4%)

E: Environment Understanding; P: Planning; S: Skill Usage; L: Long-Horizon Consistency; R: Error Recovery.

The analysis identifies failures associated with five capability demands across models. Representative execution traces further illustrate how these capability demands manifest in concrete failures. For example, in AutomationBench, Qwen3.6-35B-A3B misidentifies spreadsheet resources and uses unverified contacts (Environment Understanding). In VitaBench, Gemini-3.1 Pro incorrectly reschedules an unaffected Saturday dinner when modifying a separate appointment (Long-Horizon Consistency). In τ³-Banking, GPT-5.4 closes a credit-card account without first checking for pending disputes, skipping a required prerequisite (Planning). These cases illustrate distinct, observable failure mechanisms underlying the capability demands.

AuthorsOpenReview ↗

Response (Part 2/2): Capability-Conditioned Training Effects and SFT Setup (W1, Q1)

3. Capability challenges across benchmarks.

Agent tasks often require multiple capabilities, and different benchmarks emphasize different demands. We annotated 1,772 evaluation tasks across seven benchmarks, allowing multiple capability labels per task.

Table: Task-Level Distribution of Capability Challenges Across Seven Benchmarks.

BenchmarkTasksEPSLR
Terminal-Bench 2.189766112613
SWE-bench Multilingual3002122603264
SkillsBench877663773212
ClawEval199193674235
τ³-Banking979466765734
AutomationBench6005824253094140
VitaBench40027235502321
Total1,7721,5051,063467787329

E: Environment Understanding; P: Planning; S: Skill Usage; L: Long-Horizon Consistency; R: Error Recovery.

For example, SkillsBench emphasizes Skill Usage (77/87 tasks), VitaBench and AutomationBench heavily involve Planning (355/400 and 425/600) and Long-Horizon Consistency (232/400 and 414/600), while SWE-bench Multilingual emphasizes Error Recovery (264/300).

4. Capability-stratified downstream evaluation.

We agree that it is important to examine whether capability-oriented training improves the corresponding agent abilities.

However, the five demands often co-occur within a task, making a simple leave-one-capability-out ablation difficult to interpret. Removing Planning-related tasks, for example, would also remove training signals for Environment Understanding and Long-Horizon Consistency. It would therefore not isolate Planning's contribution.

To provide more fine-grained evidence, we additionally analyze performance before and after SFT across the five capability demands on all seven benchmarks. Specifically, we group evaluation tasks by their capability demands and compare the number of success tasks before and after training.

Table: Capability-level performance before and after SFT across seven benchmarks.

Capability DemandTasksSuccess Rate Before SFTSuccess Rate After SFTGain (pp)
Environment Understanding1,505601/1,505 (39.93%)715/1,505 (47.51%)+7.57
Planning1,063400/1,063 (37.63%)486/1,063 (45.72%)+8.09
Skill Usage46775/467 (16.06%)113/467 (24.20%)+8.14
Long-Horizon Consistency787243/787 (30.88%)309/787 (39.26%)+8.39
Error Recovery329189/329 (57.45%)222/329 (67.48%)+10.03

Note: A task is counted as successful if at least one evaluated rollout achieves full success; otherwise, it is counted as unsuccessful. Each task is counted only once, and partial credit does not count as success. Tasks may belong to multiple capability categories.

These results provide a capability-level view of training effectiveness. Aggregated across the seven benchmarks, task-level success rates improve for all five capability-demanding subsets.

5. Clarification on SFT training mode.

Finally, regarding the SFT setup, we conducted training in thinking mode, and the training data included the teacher model's reasoning traces.

We thank the reviewer again for the helpful suggestions.

Thank you for your response and additional experiments. All of my concerns have been addressed.

AuthorsOpenReview ↗

Thank You for Your Feedback

Thank you for your positive feedback and for confirming that your concerns have been addressed. We sincerely appreciate your thoughtful review and constructive suggestions.

Reviewer_D2b2OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper generates agent training tasks based on existing skills. The authors first design the tasks using difficulty patterns and task blueprints, and then build the corresponding environment and scoring rules. After the model attempts the tasks, the system continues to increase the difficulty of easier tasks based on the execution records. Finally, the model is fine-tuned using high-scoring teacher trajectories to check whether these data can improve performance on downstream tasks.

Critical strengths

S1. The author translated the capability requirements into specific difficulty patterns and task blueprints. The appendix provides a complete case, allowing one to clearly understand how these designs can be transformed into actual executable tasks.

S2. The author constructed nearly three thousand executable tasks, and the generated data showed training benefits on multiple benchmarks. Improvements were also observed under different agent harnesses, indicating that the effect does not solely depend on a certain execution method.

Critical weaknesses

W1. The main experiment mainly compared the performance before and after fine-tuning, and the existing environment synthesis baseline used different base models. I agree that this batch of data is useful, but it is not clear what this framework brings compared to directly having the same strong model generate the complete task based on skills. Currently, there is a lack of this direct comparison.

W2. The task hardening experiment fixed the number of training trajectories, but did not report the actual number of training tokens. After task hardening, the average number of interaction steps increased, so it is not clear whether the improvement comes from better task design or the use of more tokens during training.

Questions for authors

Q1. Can you have the same model directly generate complete tasks based on the same batch of skills to see how the training effect compares with this method?

Q2. Can you report the number of tokens in the training trajectories from each round, and compare the training effects before and after task hardening under the same token budget?

Q3. Can you explain the process of skill screening and the construction of the initial 100 difficulty patterns? Which steps are done manually and which are done by the LLM?

Q4. Can you report how many trajectories meet the selection threshold and explain how the final 1,500 were selected? Why were this selection method and training scale adopted?

5 comments
AuthorsOpenReview ↗

Response (Part 1/2): Direct Skill-to-Task Generation Baseline (W1, Q1)

Thank you for recognizing Skill2Env's structured approach to turning capability demands into executable tasks, as well as its training benefits across multiple benchmarks and agent harnesses. We appreciate your suggestions and have conducted additional experiments to address the concerns.

Controlled experimental setup.

We conducted a new experiment comparing Skill2Env with direct task generation using the same synthesis model and source skills.

In the vanilla setting, we use the same task synthesis agent, Kimi-K3, to directly generate complete executable tasks from the same 3,203 curated skills. Both settings use the same Pi harness, tools (shell execution, file operations, web search, and web fetching), and environment access. The key difference is that vanilla generation does not use difficulty patterns, blueprint-based construction, or iterative task hardening.

For downstream training, both settings use DeepSeek-V4-Flash-0731 as the trajectory teacher and Qwen3.6-35B-A3B as the student model, with the same SFT configuration.

Vanilla generation produced 3,028 tasks, from which we collected 2,969 teacher trajectories. Among them, 2,249 achieved rewards above 0.9, and we randomly sampled 1,500 for SFT.

Experimental results:

BenchmarkSkill2Env (1.5K)Vanilla (1.5K)
Terminal-Bench 2.158.456.2
SWE-bench Multilingual71.071.0
VitaBench44.843.65
SkillsBench46.941.64
AutomationBench17.714.83
Claw-Eval61.347.24
τ³-Banking15.112.37
Average45.041.0

Skill2Env outperforms vanilla generation on six benchmarks and ties on one, improving the average score by 4.0 points. This provides direct evidence that the structured synthesis framework offers benefits beyond generating tasks directly from skills with the same model.

AuthorsOpenReview ↗

Response (Part 2/2): Token-Controlled Hardening and Data Construction Details (W2, Q2–Q4)

1. Token-Controlled Task Hardening Study (W2, Q2)

We agree that longer trajectories after task hardening may introduce a larger training token budget. In the original 500-trajectory ablation, the supervised tokens per dataset pass at hardening iterations 0, 1, and 2 were 8,972,138, 10,845,187, and 11,623,891, respectively.

To disentangle the effects of task hardening from increased training token exposure, we extended the original ablation to the full 1,500-trajectory setting and conducted two controlled comparisons.

First, we keep the number of training trajectories (1,500) and training steps (230) the same. Before hardening, 2,230 teacher trajectories achieved rewards above 0.9, from which we randomly sampled 1,500.

Second, we additionally match the total supervised token budget. Since hardened trajectories are longer, we train the initial-task setting for 287 steps, matching the total supervised tokens used by the hardened setting.

Experimental results:

BenchmarkHardened (230 steps)Initial (230 steps)Initial (287 steps, token-matched)
Terminal-Bench 2.158.457.353.9
SWE-bench Multilingual71.071.070.0
VitaBench44.842.6244.38
SkillsBench46.941.1143.7
AutomationBench17.717.8316.83
Claw-Eval61.358.2958.79
τ³-Banking15.111.013.4
Average45.042.743.0
Supervised tokens per dataset pass34.559M27.693M27.693M
Training steps230230287
Total supervised tokens169.568M135.879M169.554M

At the same training-step budget, hardening improves the average score from 42.7 to 45.0 (+2.3 points). More importantly, under nearly identical supervised token budgets (169.568M vs. 169.554M), the hardened setting achieves 45.0 vs. 43.0 (+2.0 points), outperforming the initial setting on all seven benchmarks. These results show that the benefits of task hardening cannot be explained solely by the increased training token budget.

2. Skill Curation and Initial Difficulty Pattern Construction (Q3)

Skill curation. As described in Section 4.1, we downloaded over 47K public skill folders from ClawHub. Human researchers defined three screening criteria: Practical Utility, Runtime Compatibility, and Setup Feasibility. Kimi-K3 then applied these criteria to screen all skill folders, resulting in 3,203 curated skills.

Human-in-the-loop difficulty pattern construction. We analyzed 500 failure trajectories from multiple agent models across benchmarks and conducted pilot task synthesis. The trajectory analysis helped identify recurring agent failure mechanisms, while pilot synthesis revealed how these challenges could be instantiated in executable tasks. An LLM then proposed candidate difficulty patterns based on these observations. Human reviewers verified these patterns, merged redundant patterns, removed unsupported or overly task-specific patterns, refined their definitions, and added missing patterns based on execution evidence. This process resulted in 100 initial difficulty patterns that capture reusable sources of task difficulty.

3. Trajectory Selection and Training Scale (Q4)

We use a reward threshold of 0.9 to select high-quality teacher trajectories for SFT.

We chose this threshold for two reasons. First, rubric-based rewards are continuous, so a high-quality trajectory may receive slightly less than a full score. Second, there is a trade-off between trajectory quality and training data quantity. Through manual inspection of sampled trajectories, we found that those with rewards above 0.9 were generally of high quality, while lower-scoring trajectories more often contained execution errors that could introduce incorrect behavior during SFT.

In the main Skill2Env setting, 1,568 trajectories met the reward threshold, from which we randomly selected 1,500 for training. We adopted this scale to use most of the available high-quality trajectories.

We thank the reviewer again for the helpful suggestions.

AuthorsOpenReview ↗

Response to Reviewer D2b2

Dear Reviewer D2b2,

As the rebuttal discussion period is approaching its end, we would greatly appreciate your final feedback on our responses. If our clarifications and additional experiments have adequately addressed your concerns, we kindly invite you to reconsider and update your scores accordingly.

If any concerns or questions remain, please let us know. We will do our best to address them promptly during the remaining discussion period to resolve any outstanding issues.

Thank you again for your time, consideration, and valuable feedback.

Reviewer_R7LSOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Skill2Env, a framework that synthesizes executable agent-training environments from skills using capability-oriented difficulty patterns and task blueprints. An iterative hardening process uses solver feedback to increase task difficulty. Fine-tuning Qwen3.6-35B-A3B on 1.5K generated trajectories improves performance across seven agent benchmarks, raising the average score from 36.6 to 45.0.

Critical strengths
  1. The paper breaks down agent capabilities into fine-grained requirements and operationalizes them through 100 reusable difficulty patterns, providing a useful reference for analyzing agent capabilities.
  2. The task blueprint offers a unified specification for environment construction and evaluation, while Iterative Task Hardening uses execution feedback to progressively increase task difficulty.
Critical weaknesses
  1. The reliability of task construction is insufficiently quantified. What proportion of generated tasks is consistent with their blueprints, and what is the success rate at each construction stage? How does this consistency change after iterative hardening? It is also unclear how the solvability of generated tasks is reliably verified.
  2. The SFT trajectories are generated by DeepSeek-V4-Flash, which already substantially outperforms the student model on most benchmarks. Without controlled comparisons against alternative environment synthesis methods using the same teacher and training budget, it is difficult to disentangle the contribution of the proposed synthesis method from the benefits of distillation from a stronger teacher.
Questions for authors

Please refer to the weaknesses above.

5 comments
AuthorsOpenReview ↗

Response (Part 1/2): Task Construction Reliability and Solvability

Thank you for recognizing the value of our fine-grained capability demands, reusable difficulty patterns, and blueprint-based task construction and hardening. We appreciate your constructive feedback and address the two main concerns in separate responses.

1. Consistency validation and repair.

As described in Section 4.2, we use a Consistency Validation and repair process after task synthesis and solver rollout to verify blueprint–implementation consistency and identify task feasibility issues.

Specifically, the validation agent checks the consistency between the blueprint, task implementation, and evaluation implementation using actual solver trajectories and rubric results. We identify and repair three types of issues:

(1) Task Implementation Issue: If the task environment does not correctly implement the blueprint, we repair the task implementation according to the blueprint.

(2) Evaluation Implementation Issue: If the evaluator does not correctly implement the blueprint's rubric design, we repair the evaluation implementation accordingly.

(3) Blueprint Design Issue: If the blueprint itself contains infeasible task requirements or evaluation criteria, we revise the blueprint and update the corresponding implementations.

2. Solvability verification.

An important aspect of our task blueprint is its Facts & Derivations component, which specifies the reference facts and the derivations underlying the expected task outcomes. This provides explicit, author-side evidence for assessing task feasibility.

Specifically, the blueprint identifies the facts required to derive the expected outcomes, while its information boundaries specify which evidence must be accessible to the solver in the workspace. During consistency validation, we use execution evidence to identify and repair issues such as missing evidence, infeasible requirements, or misaligned evaluation criteria.

For example, the Docling case study in Appendix E provides complete reference facts and step-by-step derivations of the expected reconciliation results, together with the workspace evidence needed to obtain them. This provides a concrete, checkable basis for verifying that the expected outputs can be derived from the task materials available to the solver.

3. Quantitative reliability results.

We additionally report the consistency validation results at each construction stage:

StageValidated TasksInitially FlaggedDirect Pass RatePass Rate After Repair
Initial Construction3,15229890.55%98.51%
Hardening Iteration 12,40830687.29%97.80%
Hardening Iteration 21,94133682.69%95.41%

During initial construction, 2,854 of 3,152 tasks (90.55%) passed validation directly. Among the 298 flagged tasks, 103 involved evaluation implementation issues, 43 involved task implementation issues, and 152 involved blueprint design issues. Our repair process resolved 251 cases, increasing the number of consistency-valid tasks to 3,105 (98.51%).

Among these 3,105 tasks, 697 were already sufficiently challenging, while 2,408 underwent the first round of hardening. Validation identified 306 inconsistencies (97 evaluation, 61 task, and 148 blueprint issues). After repair, 53 remained unresolved, leaving 3,052 consistency-valid tasks.

In the second iteration, 1,941 tasks underwent further hardening. Validation identified 336 inconsistencies (113 evaluation, 54 task, and 169 blueprint issues). After repair, 89 remained unresolved and were excluded.

Overall, Skill2Env retains 2,963 consistency-valid tasks (94.00% of the initial 3,152) after two hardening iterations.

These results show that our validation-and-repair process effectively detects and corrects inconsistencies. Although iterative hardening introduces new issues, 97.80% and 95.41% of hardened tasks pass validation after repair in iterations 1 and 2, respectively. This supports the reliability of our task construction process.

AuthorsOpenReview ↗

Response (Part 2/2): Comparison with Existing Environment Synthesis Methods under the Same Teacher

We agree that using a stronger teacher alone may improve student performance. To isolate the benefit of our environment synthesis method, we conducted additional comparisons with two existing skill-based task synthesis methods, FACET (Shi et al., 2026) and NVIDIA Skill2Env (Xu et al., 2026), using the same teacher model and training setup. We select them as comparison methods because they provide publicly released executable tasks in Harbor format, enabling teacher- and training-controlled comparisons without reimplementing their synthesis pipelines.

1. Controlled experimental setup.

  • FACET: We used its 6,020 publicly released tasks and collected trajectories using DeepSeek-V4-Flash-0731. Following FACET's selection criterion of complete task success (reward = 1), we obtained 1,035 successful trajectories for SFT.
  • NVIDIA Skill2Env: We used its 7,496 publicly released tasks and collected trajectories with the same teacher, DeepSeek-V4-Flash-0731. Among them, 2,328 trajectories achieved rewards above 0.9. We randomly sampled 1,500 trajectories, following the selection strategy used in our main experiment.

For all three settings, we use DeepSeek-V4-Flash-0731 as the teacher, Qwen3.6-35B-A3B as the student, and the same SFT configuration with 230 training steps. Our comparison with NVIDIA Skill2Env matches the number of training trajectories (1.5K), while FACET provides 1,035 qualifying trajectories under its original full-success criterion.

2. Experimental results.

BenchmarkOurs (1.5K)FACET (1,035)NVIDIA Skill2Env (1.5K)
Terminal-Bench 2.158.450.657.3
SWE-bench Multilingual71.065.771.0
VitaBench44.842.6242.94
SkillsBench46.942.2243.03
AutomationBench17.713.016.33
Claw-Eval61.355.2858.79
τ³-Banking15.113.49.62
Average45.040.442.7

Skill2Env achieves the highest average score of 45.0, outperforming FACET by 4.6 points and NVIDIA Skill2Env by 2.3 points. It outperforms FACET on all seven benchmarks and NVIDIA Skill2Env on six, while tying on SWE-bench Multilingual.

Importantly, all three settings use the same strong teacher model, so the improvements cannot be explained solely by distillation from DeepSeek-V4-Flash-0731. These results provide evidence that our capability-oriented synthesis framework produces more effective training data.

In summary, our additional results show that rollout-grounded validation and repair ensure high task consistency throughout iterative hardening. Moreover, under the same teacher, student, and training-step settings, Skill2Env outperforms FACET and NVIDIA Skill2Env. These findings provide further evidence of both the reliability of our task construction process and the benefits of our capability-oriented environment synthesis framework. We thank the reviewer again for the helpful suggestions.

AuthorsOpenReview ↗

Response to Reviewer R7LS

Dear Reviewer R7LS,

As the rebuttal discussion period is approaching its end, we would greatly appreciate your final feedback on our responses. If our clarifications and additional experiments have adequately addressed your concerns, we kindly invite you to reconsider and update your scores accordingly.

If any concerns or questions remain, please let us know. We will do our best to address them promptly during the remaining discussion period to resolve any outstanding issues.

Thank you again for your time, consideration, and valuable feedback.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Summary of Rebuttal Responses

Dear Area Chair and Reviewers,

We sincerely thank all three reviewers for their thoughtful and constructive feedback, as well as their recognition of the value of our capability-oriented environment synthesis framework. We have carefully addressed the concerns raised in the reviews through detailed clarifications, additional analyses, and several new controlled experiments. We would like to summarize our key responses and findings below.

1. Task Construction Reliability and Solvability (Reviewer R7LS).

We clarified our consistency validation and repair process, including how task feasibility is assessed through blueprint-level reference facts, derivations, and execution evidence. We additionally reported quantitative reliability statistics across construction stages. After validation and repair, the consistency pass rates reach 98.51% for initial construction, 97.80% after the first hardening iteration, and 95.41% after the second iteration, resulting in 2,963 retained tasks.

2. Controlled Comparisons to Isolate the Benefits of Skill2Env (Reviewer R7LS, Reviewer D2b2).

To distinguish the contribution of our synthesis framework from strong-teacher distillation, we conducted additional comparisons against FACET and NVIDIA Skill2Env using the same teacher, student, and SFT configuration. Skill2Env achieves an average score of 45.0, compared with 40.4 for FACET and 42.7 for NVIDIA Skill2Env.

We further introduced a direct vanilla synthesis baseline using the same Kimi-K3 synthesis model, source skills, and execution setup, but without our difficulty patterns, task blueprints, and iterative hardening. Under the same downstream training setup, Skill2Env achieves 45.0 versus 41.0 for vanilla synthesis. These results provide additional evidence that our structured synthesis framework contributes beyond the benefits of using a stronger teacher or directly generating tasks from skills.

3. Effectiveness of Task Hardening Under Matched Token Budgets (Reviewer D2b2).

We reported training token statistics and extended the hardening ablation to 1,500 trajectories. Importantly, under nearly identical total supervised token budgets, the hardened setting achieves 45.0 versus 43.0 for the initial-task setting. This suggests that the improvements from iterative hardening cannot be explained solely by increased training token exposure. We also clarified the skill curation process, human-in-the-loop difficulty pattern construction, trajectory selection criteria, and training scale.

4. Empirical Support for Capability Demands (Reviewer rNdK).

To better substantiate the five capability demands, we analyzed 500 failure trajectories from five agent models, identifying 848 capability-related failure reasons. We further annotated 1,772 tasks across seven benchmarks and conducted a capability-stratified downstream evaluation, observing improvements across all five capability categories (+7.57 to +10.03 percentage points). We also clarified that SFT was conducted in thinking mode and included the teacher model's reasoning traces.

We are grateful to Reviewer rNdK for their follow-up response explicitly confirming that all of their concerns have been addressed.

For Reviewers R7LS and D2b2, we have provided detailed responses and conducted additional experiments directly targeting their concerns. We hope these new results and clarifications help strengthen the evidence for the reliability, effectiveness, and distinct contributions of Skill2Env.

We look forward to hearing from Reviewers R7LS and D2b2 and would be happy to engage in further discussion or clarify any remaining questions.

We sincerely thank the Area Chair and all reviewers again for their time, careful evaluations, and valuable suggestions.

Best regards,
The Authors

AuthorsOpenReview ↗

Update on Reviewer Feedback

Dear Area Chair and Reviewers,

We would like to provide a brief update following our rebuttal summary.

We sincerely thank Reviewer R7LS for acknowledging that our additional evidence has addressed the main concerns underlying their initial assessment and for their willingness to reconsider their score.

We greatly appreciate the reviewers' thoughtful consideration of our responses and additional experiments. We also look forward to hearing from Reviewer D2b2 and would be happy to clarify any remaining questions.

We sincerely thank the Area Chair and all reviewers for their time and consideration.

Best regards,
The Authors

AuthorsOpenReview ↗

Follow-up on Reviewer Feedback

Dear Area Chair and Reviewers,

We sincerely thank all three reviewers for their positive follow-up responses and thoughtful reconsideration of our work. We greatly appreciate their recognition that our additional experiments and clarifications have addressed their concerns.

We also sincerely appreciate the Area Chair's time and consideration.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.