acceptodds
Under review as a conference paper at ICLR 2027

Reinforcing Agents with Collective Skills

Abstract

We introduce SKILL2ENV, a scalable and aligned data recipe for modern agentic RL. Public Agent Skills provide workflow guidance across diverse domains, linking to real-world artifacts and packaging domain expertise and quality criteria. We develop a pipeline to turn 3.4k web-crawled and filtered Agent Skills into 8k executable terminal environments with both programmatic tests and behavioral rubrics. Through extensive RL experiments, we observe promising improvements in general agentic abilities. With 150 steps of RL training using a FlashREINFORCE variant, Qwen3.8-27B gains 9.0 percentage points on Terminal-Bench 2.1, and 7.1 percentage points on S2EBENCH, our hand-verified private benchmark featuring real-world agent use cases. Further analysis indicates increased behavioral alignment with original Agent Skills on similar problems.

Primary area:
reinforcement learning
Keywords:
LLM, Reinforcement Learning, Terminal Agent, RL Dataset

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_49kEOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces Skill2Env, a framework for automatically converting public Agent Skills into verifiable reinforcement learning (RL) environments. The authors argue that existing agentic RL environments primarily focus on technical workflows, such as software engineering, and lack broad coverage of real-world user needs and practitioner-authored workflows. To address this limitation, the paper leverages public Agent Skills as a source of domain expertise and task generation guidance. Through workflow decomposition, task diversification, environment construction, and automated verification, Skill2Env converts approximately 3,400 Skills into 7,971 executable training tasks spanning 13 domains. Experimental results show that RL training on Skill2Env improves the model's agentic terminal-use performance. In particular, the FlashREINFORCE variant achieves improvements of 9.0 and 7.1 percentage points on Terminal-Bench 2.1 and S2EBench, respectively.

Critical strengths
  1. The paper addresses an important and practically relevant problem. Scaling diverse and verifiable training environments is a significant challenge in agentic RL, and the proposed direction is relevant to improving general agentic capabilities.
  2. The environment generation pipeline is relatively comprehensive and has practical engineering value. The framework covers multiple stages of task generation and verification, and the paper provides detailed descriptions of its implementation and engineering design.
  3. The paper constructs a relatively large-scale and diverse training dataset. The resulting dataset contains nearly 8,000 executable tasks spanning 13 domains, providing a potentially useful resource for agentic RL research.
Critical weaknesses
  1. Lack of critical controlled comparisons to validate the effectiveness of the core approach. The central claim of the paper is that leveraging practitioner-authored Agent Skills enables the generation of RL environments that better reflect real-world user needs and incorporate domain expertise. However, the authors do not provide a direct comparison between Skill-guided environment synthesis and conventional task generation approaches that do not rely on Skills. Consequently, the current experiments do not sufficiently establish the unique contribution of using Agent Skills as the source of training environments.

  2. Insufficient independent validation of environment quality despite the large dataset size. Although the authors report generating nearly 8,000 training tasks and filtering them through automated checks, these checks primarily establish basic executability and consistency between the reference solution and programmatic tests. Passing such checks does not necessarily guarantee that the generated tasks are representative of real-world use cases, have appropriate difficulty, or are equipped with comprehensive and reliable evaluation criteria. Furthermore, the main RL experiments use only approximately 2,400 randomly sampled training instances, rather than the full dataset. The paper does not provide systematic scaling experiments to investigate how training performance changes with increasing dataset size. This limits the evidence supporting the value of the full generated corpus.

  3. Limited experimental setup and insufficient evidence of generalization. The RL experiments are primarily conducted on a single base model, Qwen3.8-27B, without validation across different model families or parameter scales. In addition, the authors' private benchmark, S2EBench, contains only 79 tasks and is constructed from the same SkillHub ecosystem as the training data. Although these tasks are manually verified, the limited benchmark size and shared source ecosystem raise concerns about the representativeness of the evaluation and the generalizability of the reported improvements.

  4. Insufficient and non-independent validation of the effectiveness of Skill-derived rubrics. Skill-derived behavioral rubrics are an important component of the proposed approach, but the corresponding experimental evidence is not fully convincing. First, the Outcome + Rubric RL checkpoint performs worse than the outcome-only RL checkpoint on the evaluated benchmarks, suggesting that incorporating rubric rewards does not provide a direct benefit in terms of objective task completion performance. Second, the reported improvements in behavioral alignment rely primarily on evaluations by an LLM judge rather than independent human assessments. The same judge is used for both rubric-based reward calibration during training and the subsequent pairwise Skill-alignment evaluation, potentially introducing evaluation bias. Furthermore, the paper does not provide a direct pairwise comparison of behavioral quality between Outcome-only RL and Outcome + Rubric RL. It also lacks systematic ablations on the rubric reward weight and alternative reward combination strategies. As a result, the current evidence is insufficient to fully establish the practical benefits of rubric-based supervision or characterize its trade-offs with outcome-based performance.

  5. The presentation emphasizes system implementation and empirical results without sufficiently highlighting the central research question. In particular, the abstract primarily summarizes the scale of data generation, the RL training setup, and benchmark improvements, without clearly articulating the fundamental distinctions between the proposed approach and existing methods or the core research hypothesis being investigated. Similarly, much of the main text is devoted to describing the data generation pipeline, infrastructure, and implementation details, while relatively limited attention is given to in-depth experimental analyses of the key design choices.

Questions for authors

See above

2 comments

We thank the reviewer for the detailed review. We ran new experiments for W1, W2 and W4 (Tables A and B in general comments).

W1. No comparison against Skill-free generation. Table A (in general comments) adds this comparison. Endless-Terminals and TMax-15K are two existing datasets synthesized without Skills. With the same checkpoint, recipe and budget, they improve TB 2.1 by 2.2 and 4.4 points and Skill2Env by 9.0. The gap on S2EBench is similar (−2.2 and +3.4 against +7.1). These two datasets also use different generators from ours, so Table A compares the datasets as a whole. To isolate the effect of the Skill itself, we would need to re-generate at least 2.4k tasks with the same pipeline but without Skill input and run RL again. We could not finish this in the rebuttal period and will state it as a limitation.

W2. Environment quality and scale. We agree that Oracle/NOP only shows a task is executable and consistent with its tests. For difficulty, GLM-5.3 averages 0.74 reward over 15,968 trajectories with a median of 19 model calls (Sec. 3.4), so the tasks are solvable but not trivial for a strong open model. For usefulness, the most direct check is downstream: in Table A, 2.4k Skill2Env tasks train a better agent than 2.4k tasks from either alternative. For scale, we added a run on the full corpus (Table A). S2EBench rises from 40.5 to 43.4 and TB 2.1 stays at the 2.4k level (57.8 against 58.4). This matches the training reward curve, which slows down after about 2,000 tasks. So the full corpus keeps helping on the broader benchmark while TB 2.1 saturates early.

W3. One model; same-source benchmark. The gain does hold under two different RL recipes (DPPO and FlashREINFORCE, Table 1). For S2EBench, it shares the pipeline with the training data but its Skills do not. We held out 100 Skills before training data synthesis, and the 79 tasks were generated from them and reviewed by 4 researchers. As for the size, 79 tasks is rather common for modern terminal-agent benchmarks. We also report widely-used TerminalBench 2.1 (89 tasks, of similar size as ours), and the gain there reflect a transferred general agentic ability. We choose Qwen3.8-27B as base model because it's commonly considered one of the strongest local LLMs, and it's a size widely used in existing data research. We did not add other model variants due to limitation of resource during rebuttal. We will state the single-model scope as a limitation.

W4. Rubric evidence. We agree that the rubric reward does not improve task completion, and the paper reports it this way (Sec. 5.2). What it changes is how closely the agent follows the Skill. Table B (in general comments) fills the two gaps the reviewer names. For independence, we re-judged with Fable 5.2, a model from another family that is used nowhere in generation or training, and it agrees with GPT-6 Astra on every row. For the direct comparison, both judges prefer the Outcome+Rubric checkpoint over the Outcome-only one (50.5 to 33.0 and 47.0 to 30.0).

W5. Presentation. Thanks for the suggestion. We include rich description for data pipelines and infrastrcutre details since they are the fundamental "methods" to reproduce experiments and artifacts in our approach. We will adopt the suggestion to rewrite the abstract and introduction around the research question: whether Skills written by practitioners are a excellent source of terminal agent RL environments.

Overall, we hope our explanation and the new results address the soundness concerns. High-quality RL data is scarce in the open community. Under the same recipe, Skill2Env gives the largest gain among the open terminal RL datasets we compared (Table A), which is crucial as base models become stronger and saturated to older datasets. We release SkillHub, full 7971 terminal agent RL dataset, and the synthesis pipeline with ablation results so researchers can continue post-training research on stronger baselines.

Thank you for the rebuttal and the additional experiments. I appreciate the new evidence on data scale, the independent evaluation of rubric-based results, and the direct comparison between Outcome RL and Outcome+Rubric RL.

However, one central concern remains unresolved. I raised the comparison against Skill-free generation because I wanted to verify whether the method’s core contribution indeed comes from practitioner-written Skills, as emphasized in the paper. However, the two additional datasets introduced in the rebuttal do not isolate the contribution of Skills themselves, as the datasets use different generation pipelines. Practitioner-written Skills serve as the upstream input to Skill2Env, but most of the subsequent environment-generation and verification pipeline does not inherently depend on Skills. I would like to understand whether the performance gains come from the Skills themselves or from Skill2Env's environment-generation pipeline.

Reviewer_dQ3JOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents SKILL2ENV, a pipeline that compiles public Agent Skills into executable Harbor terminal tasks. More than 5k Skills crawled from GitHub are filtered, leaving about 3.4k (SKILLHUB). For each Skill, a Codex planner (GPT-5.6 Sol) extracts workflows and diversification axes. A Codex creator then builds the environment, instruction, partial-credit tests (up to six metrics), a Must-do/Must-avoid/Best-practice rubric, and a reference solution. A host gate keeps a task only if static checks pass, the reference solution earns full reward on every metric, and a no-op agent earns zero. The result is 7,971 tasks across 13 domains (22.5% software engineering), generated for over $90k in API cost. One epoch of RL on a random 2.4k-task subset raises Qwen3.8-27B on Terminal-Bench 2.1 (TB2.1) from 49.4 to 54.1 (DPPO) or 58.4 (FlashREINFORCE) pass@1. On the authors' private, hand-reviewed 79-task S2EBench it rises from 33.4 to 37.7 or 40.5. Adding an LLM-judge rubric term (Eq. 3) shrinks the benchmark gains (TB2.1 50.1). It also raises a judge's pairwise preference over the base model from 54.5% (outcome-only) to 73.0%. SFT on GLM-5.3 trajectories lowers TB2.1 to 45.8.

Critical strengths
  1. A carefully engineered, concretely specified environment pipeline. Tests and rubric are frozen before the reference solution exists. Acceptance requires that the reference solution earn full reward and a no-op agent zero on every metric. Base images are pinned by digest, tasks need no network at solve time, and tests give multi-metric partial credit (Sec. 3.2; App. A). The corpus is large and diverse: 7,971 tasks in 13 domains, 51.5% of them non-technical knowledge work by the paper's own labels (Fig. 3). Released with the pipeline code, it would be a valuable resource for agentic RL.

  2. Large external-benchmark gains from modest RL, reported candidly. One epoch on 2.4k tasks improves TB2.1 by 4.7 (DPPO) and 9.0 (FlashREINFORCE) points, and S2EBench pass@1 by 4.3 and 7.1 points (Table 1). Both use end-of-epoch checkpoints. The paper also reports unfavorable results: rubric calibration reduces benchmark gains, and SFT lowers TB2.1 by 3.6 points (Tables 1 and 3).

Critical weaknesses
  1. The paper claims Skill2Env is "a scalable and aligned data recipe," but none of the following is reported:
  • RL on an existing terminal-task corpus with the same recipe and compute, although Fig. 3 already charts two such corpora (Endless-Terminals, TMax-15K).
  • Ablations of the properties the paper credits: Skill grounding, domain diversity, Oracle/NOP gating, and axis diversification.
  • A scaling curve. Only 2.4k of the 7,971 tasks are used.

The TB2.1 transfer is therefore equally consistent with "RL on any reasonable set of verified terminal tasks helps." Approximately estimating from Fig. 3, 84 of the 89 (original TB2.1 documentation) TB2.1 tasks implied by the pie fall in technical domains. The paper does not show that the non-technical half of the corpus, its main differentiator, contributes to the headline gain.

  1. Weak validity of the in-distribution and alignment evidence (fixable).
  • S2EBench is private and has 79 tasks. The same pipeline generated it from SKILLHUB, with no stated Skill-level separation from training. Its review criteria are given only at a high level, and the drop rate and reviewer details are not reported.
  • Table 2 judges Skill alignment with the same GPT-6 Astra model that produced the rubric run's training reward. Part of the 73.0% preference may therefore reflect optimization toward that judge. There is no second judge, no human check, and no head-to-head comparison of the outcome-only and rubric checkpoints.
  • The training-time judge term in Fig. 5(b) is flat. This needs reconciling with the large pairwise gain.
Questions for authors
  • Run FlashREINFORCE from the same checkpoint with identical settings (G=16, 16 prompts per update, 150 updates, δ=10⁻², Pi 0.84.2) on 2,400 tasks from a corpus you already chart (Endless-Terminals or TMax-15K). Evaluate on TB2.1 and S2EBench with 5 trials. A clear Skill2Env margin, especially on TB2.1, would raise my rating.

  • Report at least 3 training seeds for Base → FlashREINFORCE and Base → DPPO (outcome-only). At minimum, give paired task-level bootstrap 95% CIs for every delta in Tables 1 and 3, and the number of trials that errored or were retried under -r 3.

  • Are the Skills behind S2EBench and the 200-task alignment set excluded from the Skills behind the 2,400 training tasks? If not, report results on the Skill-disjoint subset. For S2EBench, how many candidate tasks were reviewed and dropped, by how many reviewers, and how were the Sec. 3.5 criteria applied?

  • For Table 2, add:

    • (a) a human-labeled subset (e.g., 50 pairs, two annotators, Cohen's κ);
    • (b) a judge from a model family not used anywhere in training;
    • (c) order-swapped judging with a reported consistency rate;
    • (d) a direct Outcome-RL vs. Outcome+Rubric-RL comparison.
  • Will SKILLHUB, the 7,971 tasks (tests, rubrics, solutions), the generation and training code, and S2EBench be released? Under which licenses, and how will attribution for MIT/Apache sources and third-party assets be preserved?

1 comment

We thank the reviewer for the careful and specific review. Tables A and B are in the shared general comment.

W1/Q1. Same recipe on comparable datasets; full-set scaling. We ran the experiment as specified: FlashREINFORCE from the same checkpoint with identical settings on 2,400 random tasks from Endless-Terminals and from TMax-15K (Table A). Each corpus helps on TB 2.1 (which could partly attribute to aligned train-eval harness), but Skill2Env gives the highest gain among its alternatives. The full corpus gives 57.8 / 43.4 against 58.4 / 40.5 for 2.4k, so S2EBench keeps improving and TB 2.1 transfer saturates at about 2.4k tasks. We could not run the property ablations of Skill2Env pipeline (Skill grounding, domain diversity, axis diversification) in the rebuttal period, since re-generating 2.4k RL data under different setups and run RL over each variant takes over a month under our resources. We will list them as limitations. With the dataset ablation in Table A, we show that our dataset is one of the most effective terminal RL data for open research community so far.

Q2. CI of Table 1 and 3. Paired task-level bootstrap 95% CIs (10k resamples over tasks) using the RL checkpoints from the paper:

CheckpointTB 2.1 ΔS2EBench pass Δ
Outcome RL (DPPO)+4.7 [−0.8, +10.2]+4.3 [−1.2, +9.8]
Outcome+Rubric RL (DPPO)+0.7 [−4.7, +6.1]+1.3 [−4.1, +6.7]
Outcome RL (FlashREINFORCE)+9.0 [+3.6, +14.5]+7.1 [+1.6, +12.7]
GLM-5.3 SFT−3.6 [−9.0, +1.8]+0.5 [−4.9, +5.9]

We will add these intervals to Tables 1 and 3. Thanks for your suggestion.

W2/Q3. S2EBench and alignment-set separation. Yes, both are separated from training at the Skill level. S2EBench was generated in a separate run from SkillHub Skills that source no training task (Sec. 3.5), and the 200 alignment tasks likewise come from Skills held out of training (Sec. 5.3). The paper says "held-out" and "disjoint from training" without naming the level; we will state it in both sections. For the question on S2EBench reviewing: we held out 100 skills before training data synthesis. About 120 task candidates were generated using these skills, and we delegated 4 researchers to review them against the criteria of Sec. 3.5, resulting in the final 79 tasks. The 200 alignment set was later generated from additional skills we collected from the web, and no manual review was conducted since it's for testing model behavioral affinity to the skill documents only. We will add these details to Appendix.

W2/Q4. Judge independence. Table B adds a judge from another family and the direct comparison. Fable 5.2 is used nowhere in generation or training and agrees with GPT-6 Astra on every row. The margin over base shrinks by about a fifth, which might come from the bias the reviewer suspected. The Fable score is a order-swapped average between two independent runs. We did not have enough resource for human labeling during rebuttal and will add it to the final version.

Q5. Release. Yes. SkillHub, the generation pipeline (Skill2Env), and the 7971 harbor-format terminal tasks are packaged for release under Apache-2.0, which permits both research and commercial use. Each Skill keeps its original licence (MIT or Apache-2.0), with a per-Skill inventory of source repository and licence. We withhold links during review to preserve anonymity. S2EBench stays private so that it remains a held-out set for future study (Sec. 3.5); its construction and review will be documented in the appendix.

Reviewer_xNc7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces SKILL2ENV, which transforms approximately 3,400 publicly available agent skills into nearly 8,000 reinforcement learning tasks spanning 13 domains, while providing both outcome-based testing and behavioral scoring criteria. In experiments, training the Qwen3.8-27B model using purely outcome point increase in pass rates across two evaluation sets, respectively. Incorporating behavioral rewards improved adherence to skill specifications but attenuated the gains in pass rates.

Critical strengths
  1. The topic is significant, and the use of "Skills" as a data source is well-justified.

Since "Skills" encompass tasks, tools, processes, and quality standards, transforming them into a training environment provides a richer foundation for generation compared to approaches based solely on task classification or code-related problems. Furthermore, extending agent training to domains such as research, documentation, and business operations offers clear practical value.

  1. There are genuine training gains, which have also transferred to external benchmarks.

The best-performing model improved from 49.4 to 58.4 on Terminal-Bench, and from 33.4 to 40.5 on S2EBench. These improvements are noteworthy.

Critical weaknesses
  1. There is a lack of key control experiments to pinpoint the primary source of the performance gains.

While the current experiments demonstrate that applying RL to this set of tasks improves model performance, the specific origin of these improvements remains unclear. For instance, whether they stem from specialized knowledge within the source skills, broader domain coverage, or the use of a powerful task-generation model.

  1. The evaluation of behavioral alignment lacks sufficient independence, and the distinct contributions of the rubric have not been disentangled.

GPT-6 Astra is involved in both the scoring process during training and the subsequent assessment of behavioral preferences. It is possible that the model has learned to adopt an expressive or execution style that aligns more closely with the preferences of this specific "judge"; current experiments cannot yet rule out this explanation.

Questions for authors

See weakness.

1 comment

We thank the reviewer for the review. Both weaknesses are fair, and we ran a new control for each (Tables A and B in the shared comment).

W1. Source of the gain. Table A (in general comments) shows that the gain is specific to the data and is not a generic effect of RL on terminal tasks. With the same checkpoint, recipe and budget, Endless-Terminals and TMax-15K improve TB 2.1 by 2.2 and 4.4 points and Skill2Env by 9.0. Both are LLM-synthesized, programmatically verified corpora built from taxonomies or task seeds, and both are closer to TB 2.1 in domain than ours (Fig. 3). Going more granular, it is difficult to show whether the margin comes from Skill content or from a stronger generator. These control requires re-generating at least 2.4k data with various generator LLMs, with and without Skill assistance. And we could not finish the process in the rebuttal period and will state this as a limitation (thanks for pointing it out). We show that this dataset is one of the most effective terminal RL data for open research community by comparing RL runs over other two options.

W2. Independence of the alignment evaluation. We agree that Table 2 could not rule out that the rubric checkpoint learned the preferences of GPT-6 Astra, since that model produced its training reward. Table B (in general comments) repeats the evaluation with Fable 5.2, a judge from a different model family that is used nowhere in generation or training, and adds the direct comparison of the two RL checkpoints. The independent judge agrees on every row. In the direct comparison, it prefers the rubric checkpoint 47.0 to 30.0. Two points from the paper also bear on this. The Outcome RL checkpoint never saw a judge and is preferred to base by both. And the evaluation judge is given the source SKILL.md, while the training judge sees the per-task rubric.md and a meta rubric (Sec. 4.2, 5.3). We will replace Table 2 with Table B.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Comment

Thank you for the detailed reviews. Our paper turns public Agent Skills into verified RL environments and shows that RL on them improves a 27B open model on Terminal-Bench 2.1 and on a hand-reviewed held-out benchmark. We are encouraged that the reviewers agreed on the strengths of the work: that using Skills as a data source is well-justified and addresses "an important and practically relevant problem" (xNc7, 49kE); that the pipeline is "carefully engineered, concretely specified" (dQ3J) and that the gains are genuine and transfer to an external benchmark (xNc7). We also appreciate reviewer dQ3J for commenting our paper "reported candidly" together with the unfavorable results.

The reviews converge on two concerns: which property of the data produces the gain, and whether the alignment evaluation is independent of the training judge. We additionally ran the requested controls and report them in Tables A and B below. In short: under the identical recipe, RL on Endless-Terminals and TMax-15K improves Terminal-Bench 2.1 by 2.2 and 4.4 points, against 9.0 for Skill2Env; and a judge from a different model family (Fable 5.2) reproduces the Skill-alignment preferences of Table 2.

Table A. Additional RL experiments over comparable datasets. FlashREINFORCE from the same Qwen3.8-27B checkpoint with the settings of Table 4 (G=16, 16 prompts per update, δ=10⁻², Pi 0.84.2), one epoch over 2.4k tasks (150 updates) per row unless stated, 5 eval trials.

Training dataTB 2.1S2EBench all-pass
None (baseline)49.433.4
Skill2Env, random 2.4k (paper)58.440.5
Skill2Env, full corpus57.843.4
Endless-Terminals, random 2.4k51.631.2
TMax-15K (v1), random 2.4k53.836.8

All three corpora improve TB 2.1 under the same recipe and budget, so RL on verified terminal tasks over aligned harness improves evaluation in general. The size of the gain depends on the corpus: +9.0 with Skill2Env against +2.2 with Endless-Terminals and +4.4 with TMax-15K, although TB 2.1 is closer in domain to those two corpora than to ours (Fig. 3). On private S2EBench, TMax-15K gives +3.4 and Endless-Terminals lowers the pass rate by 2.2, against +7.1 for Skill2Env. S2EBench is built by our pipeline with hand-verified hold-out skills, so part of that margin reflects distribution match. Scaling from 2.4k tasks to the full corpus raises S2EBench further (40.5 → 43.4) and leaves TB 2.1 flat. This matches the training mean average reward curves. On Skill2Env the reward climbs quickly and notably slow down after the equivalent of about 2,000 tasks, while the other two corpora plateau much earlier: TMax-15K at 800 tasks and Endless-Terminals at around 500.

Table B. Additional skill-alignment test with independent judge. Same 200 held-out tasks as Table 2; entries are prefers-first / prefers-second / equal (%).

ComparisonJudged by GPT-6 Astra (paper)Judged by Fable 5.2
Base vs. Outcome RL33.5 / 54.5 / 12.035.0 / 58.5 / 6.5
Base vs. Outcome+Rubric RL24.0 / 73.0 / 3.029.5 / 69.0 / 1.5
Outcome RL vs. Outcome+Rubric RL33.0 / 50.5 / 16.530.0 / 47.0 / 23.0

Fable 5.2 is from a different model family and is used nowhere in generation or training. The protocol is that of Table 2: the judge sees the source SKILL.md and two anonymized trajectories in random A/B order. The independent judge agrees with GPT-6 Astra on all three comparisons.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.