acceptodds
Under review as a conference paper at ICLR 2027

Recovery Training for Terminal Agents

Abstract

Terminal agents have shown strong abilities in complex coding tasks. However, they often struggle to recover from errors during long-running execution (e.g., command errors, faulty implementations), eventually leading to task failures. To this end, we present ReTerm, a terminal agent with strong recovery capability, trained by learning how to recover from its own failed terminal tasks. To support its training, we construct a customized data pipeline and produce ReTerm-10K, a high-quality dataset that provides effective recovery demonstrations for SFT cold start. Furthermore, we propose ReTerm-GRPO, an online RL algorithm that forks from the agent’s own failure states to generate multiple independent recovery attempts. By learning from whether these attempts successfully complete the task, the agent can continuously improve its recovery capability. Built on Qwen3.5-9B, our ReTerm-9B achieves 34.8% Pass@1 on Terminal-Bench 2.0, improving over the base model by 14.6 points. Further analysis shows that our method achieves consistently higher recovery rates across diverse error categories compared with standard GRPO training. All code, data and model will be released.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Terminal Agents, Reinforcement Learning, Error Recovery

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_vPavOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents ReTerm, a recovery-oriented training framework designed to improve terminal agents' ability to recover from execution failures in long-horizon tasks. The framework consists of two main components. First, the authors construct ReTerm-10K, a dataset of successful recovery trajectories generated by having a stronger teacher model continue from the terminal environment states of failed attempts produced by a weaker base model. These trajectories are filtered and used for SFT. Second, the authors propose ReTerm-GRPO, which extends standard GRPO by constructing recovery tasks from failed rollouts generated by the evolving policy and jointly optimizing original and recovery task groups. A recovery task queue is introduced to reuse failure states across training iterations.

Experiments using Qwen3.5-9B demonstrate improvements across several benchmarks, including 34.8% Pass@1 on Terminal-Bench 2.0 and 58.4% on SWE-bench Verified, outperforming the base model by 14.6 and 13.2 percentage points, respectively. Ablation studies show that ReTerm-GRPO improves Terminal-Bench 2.0 performance from 28.0% with standard GRPO to 34.8%. Additional evaluations across agent harnesses, out-of-domain benchmarks, and different error categories provide further evidence of the effectiveness and generalization of the proposed approach.

Critical strengths
  1. Strong empirical performance on terminal agent benchmarks.

The proposed framework achieves substantial performance improvements on challenging terminal agent benchmarks. In particular, ReTerm-9B improves Terminal-Bench 2.0 performance from 20.2% to 34.8%, with both recovery-oriented SFT and ReTerm-GRPO contributing meaningful gains. The improvements on SWE-bench Verified and OpenThoughts-TBLite further demonstrate the practical effectiveness of the proposed training framework.

  1. Comprehensive experimental evaluation.

The paper provides a relatively comprehensive evaluation, including ablation studies, cross-harness generalization, out-of-domain benchmarks, and recovery analysis across different error categories. These experiments help establish that the observed improvements are not limited to a single evaluation setting and provide useful evidence for the effectiveness of recovery-oriented training.

Critical weaknesses
  1. Limited methodological novelty.

The overall methodology appears relatively straightforward. The SFT stage follows a natural pipeline of collecting failed trajectories, using a stronger teacher model to generate successful recovery continuations, and filtering them for supervised learning. Similarly, ReTerm-GRPO mainly augments standard GRPO with recovery tasks constructed from failed rollouts, while retaining the original GRPO objective and advantage estimation. The recovery task queue also appears to be a relatively standard sampling mechanism. While these design choices are reasonable and practically effective, the paper does not clearly establish what constitutes a substantial methodological advance beyond existing recovery-oriented training and standard RL techniques.

  1. Insufficient analysis of the sources of performance improvements.

Although the reported improvements are substantial, the current experiments do not sufficiently isolate the factors responsible for these gains. In particular, ReTerm-GRPO changes the training task distribution by introducing recovery tasks, but the paper does not compare this strategy against alternative sampling approaches, such as training from ordinary intermediate states or oversampling difficult original tasks. It is therefore unclear whether the improvements arise specifically from learning to recover from failures or from more general benefits of task distribution augmentation. Similarly, the paper does not compare recovery-specific SFT data against conventional successful trajectories under matched settings, nor does it independently evaluate the contribution of the recovery task queue.

Questions for authors
  1. Fairness of the GRPO Comparison

Could the authors clarify whether the standard GRPO baseline uses the same number of rollouts, optimization steps, and overall training compute as ReTerm-GRPO? In particular, ReTerm-GRPO uses 8 original task groups and 2 recovery task groups per step. How is the standard GRPO baseline configured?

  1. Context Length and Computational Overhead

Since recovery rollouts are conditioned on the complete failed interaction history, could the authors report the typical length distribution of these histories and clarify how histories exceeding the context window are handled? How much additional computational overhead does this introduce?

  1. Attribution of Recovery Improvements

Since recovery rollouts receive a fresh generation and interaction budget after the original attempt has failed, could the authors clarify how they distinguish improvements in recovery capability from the benefits of additional execution opportunities? Is there a controlled comparison with standard GRPO under a comparable total execution budget?

4 comments

Q1: The paper does not clearly establish what constitutes a substantial methodological advance beyond existing recovery-oriented training and standard RL techniques.

A1: Thanks for your comments. This work represents an early exploration of self-recovery training for terminal agents, aiming to improve their robustness in long-horizon terminal interactions.

Regarding existing recovery-oriented training, TermiGen [1] constructs recovery trajectories by injecting hand-crafted errors for supervised fine-tuning. These errors may differ in distribution from the failures actually encountered by the student model. To address this, we carefully construct ReTerm-10K from the student's own execution failures and extend this idea to the design of our online RL algorithm.

Regarding general RL techniques, works such as TMax [2] primarily adopt suitable general RL algorithms to improve terminal-task performance. Building on GRPO, we explore a recovery-training algorithm tailored to terminal environments, allowing recovery capabilities to adapt to the evolving failure distribution and co-evolve with task-solving capabilities. Our ablation studies and recovery evaluations further support the effectiveness of the proposed method.

We hope our work provides valuable insights and a practical framework for the research community, inspiring further research on improving terminal agents' capabilities in long-horizon tasks.

[1] TermiGen: High-Fidelity Environment and Robust Trajectory Synthesis for Terminal Agents.

[2] TMax: A Simple Recipe for Terminal Agents.

Q2: ReTerm-GRPO changes the training task distribution by introducing recovery tasks, but the paper does not compare this strategy against alternative sampling approaches, such as training from ordinary intermediate states or oversampling difficult original tasks. It is therefore unclear whether the improvements arise specifically from learning to recover from failures or from more general benefits of task distribution augmentation.

A2: Thanks for your constructive suggestion.

For training from ordinary intermediate states, we have added a control experiment that replaces the recovery tasks in ReTerm-GRPO with ordinary intermediate-state tasks. Within each original task group of 16 trajectories, we designate one trajectory at random in advance and use reservoir sampling to uniformly select an intermediate state at an interaction boundary. When the original task group is accepted for training, we add the selected environment state and history to the queue for subsequent training steps. All other training settings remain the same as those of ReTerm-GRPO. The results are as follows:

MethodTerminal-Bench 2.0
GRPO28.0%
GRPO + ordinary intermediate states29.2%
ReTerm-GRPO34.8%

Training from ordinary intermediate states performs similarly to standard GRPO and does not reach the performance of ReTerm-GRPO.

For oversampling difficult tasks, we clarify that, following terminal-agent training work such as TMax [1], ReTerm-GRPO uses the dynamic sampling strategy introduced by DAPO [2]. Groups with identical rewards are filtered out, and sampling continues until the training batch is filled. Under this setting, ReTerm-GRPO constructs a recovery task from each original task group accepted for policy optimization. Rejected groups, including all-failure groups, do not contribute tasks to the recovery task queue. Therefore, we do not additionally increase the sampling proportion of difficult original tasks; instead, recovery training uses failed states from tasks within the model's current capabilities. We will explore combining difficult-task oversampling with recovery training in future work to further expand terminal agents' capabilities.

[1] TMax: A Simple Recipe for Terminal Agents.

[2] DAPO: An Open-Source LLM Reinforcement Learning System at Scale.

Q3: The paper does not compare recovery-specific SFT data against conventional successful trajectories under matched settings.

A3: Thanks for your comments. We have added an SFT control experiment using the same task set and sample count. We subsample 800 recovery trajectories from ReTerm-10K. For the same task set, we ask the teacher model to execute directly from the initial task environments and use rejection sampling to obtain 800 ordinary trajectories. Both models are initialized from the same Qwen3.5-9B base model and fine-tuned with identical settings: 2 epochs, a learning rate of 5×10−65\times10^{-6}, and a global batch size of 16. The results are as follows:

SFT DataTerminal-Bench 2.0
Ordinary trajectories22.5%
Recovery trajectories24.7%

Under these controlled settings, SFT on recovery trajectories improves Pass@1 by 2.2 points over SFT on ordinary trajectories, further supporting the effectiveness of recovery trajectories as fine-tuning data.

Q4: The paper does not independently evaluate the contribution of the recovery task queue.

A4: Thanks for your suggestion. We have added an ablation experiment removing the recovery task queue. In this setting, recovery tasks are constructed only from failure states generated by original tasks in the current training step and are not retained across steps. Each step still contains 8 original task groups and targets 2 recovery task groups. If the recovery groups are not filled, no additional sampling is performed, and training proceeds with the recovery groups actually obtained. The results are as follows:

MethodTerminal-Bench 2.0
ReTerm-GRPO34.8%
w/o recovery task queue30.3%

Removing the recovery task queue reduces performance by 4.5 points, supporting the effectiveness of the recovery task queue in maintaining a stable supply of recovery tasks.

Q5: Could the authors clarify whether the standard GRPO baseline uses the same number of rollouts, optimization steps, and overall training compute as ReTerm-GRPO? In particular, ReTerm-GRPO uses 8 original task groups and 2 recovery task groups per step. How is the standard GRPO baseline configured?

A5: Thanks for pointing this out. The original manuscript did not fully specify the configuration of the standard GRPO baseline. We will explicitly include these details in the revised manuscript.

Standard GRPO and ReTerm-GRPO are initialized from the same SFT checkpoint and use the same number of training steps, optimizer, and learning rate. Standard GRPO replaces the two recovery task groups with two additional original task groups, using 10 original task groups per step. Each group contains 16 sampled rollouts, so both methods target 160 rollouts per training step. Both methods also use the same dynamic sampling strategy.

Q6: Since recovery rollouts are conditioned on the complete failed interaction history, could the authors report the typical length distribution of these histories and clarify how histories exceeding the context window are handled? How much additional computational overhead does this introduce?

A6: Thanks for raising this question. We have collected additional statistics on the lengths of the failed interaction histories:

StatisticValue
Mean length18,054
Median14,753
P9034,705
P9542,175
Maximum length56,812
Histories shorter than 32K88.3%
Histories shorter than 48K98.0%

Both original and recovery attempts use a 65K context window, and the inherited history counts toward this limit. We retain the history without additional summarization or sliding-window truncation. Once the context window is exhausted, generation stops and the attempt proceeds directly to evaluation.

Recovery training primarily introduces additional context-processing overhead for the inherited histories. With the same number of rollouts, ReTerm-GRPO has approximately 23.3% more cumulative input tokens than standard GRPO. To reduce repeated context computation, we enable prefix caching so that recovery attempts branching from the same failure state can reuse the KV cache of their shared prefix.

Q7: Since recovery rollouts receive a fresh generation and interaction budget after the original attempt has failed, could the authors clarify how they distinguish improvements in recovery capability from the benefits of additional execution opportunities? Is there a controlled comparison with standard GRPO under a comparable total execution budget?

A7: Thanks for your constructive comments. Within each training batch, both methods use the same number of rollouts, maximum generation length, and interaction-turn limit. Nevertheless, the original and recovery attempts together allow longer cumulative interaction on the same task. To examine this factor, we have added a standard GRPO control experiment with the interaction-turn limit doubled to 64, giving the baseline twice the per-rollout interaction budget of ReTerm-GRPO. The results are as follows:

MethodTerminal-Bench 2.0
GRPO (32 turns)28.0%
GRPO (64 turns)31.5%
ReTerm-GRPO (32 turns + 32 turns)34.8%

Increasing the interaction-turn limit improves standard GRPO’s performance, but ReTerm-GRPO still outperforms this baseline by 3.3 points, supporting the benefit of recovery training beyond simply extending the interaction horizon. This also motivates us to explore recursive self-recovery in future work: extending the current single-round recovery to multiple rounds, allowing the agent to continue diagnosing and correcting errors when a recovery attempt still fails. Through this extension, we hope to further improve the agent’s performance on long-horizon terminal tasks.

Thank you again for your time in reviewing our work and for your constructive feedback.

We have responded to your concerns point by point and hope that our explanations and additional results have addressed them. As the discussion deadline of 19 Nov approaches, we would appreciate any further feedback or remaining questions you may have.

If our responses have resolved your concerns, we would also appreciate it if you could consider updating your score accordingly.

Reviewer_f9T8OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how to train terminal agents to recover from their own execution errors.

The pipeline has two components, first, ReTerm-10K (SFT). The authors use Qwen3.5-9B to finish tasks from TMax-15K, for those failed attempts, the environments are saved as VM snapshots for further RL rollouts. Then, a stronger teacher (DeepSeek-V4-Pro) tries to continue from these failure states given the previous attempt. Successful recoveries are filtered and used for SFT the policy model.

Second, the authors implement ReTerm-GRPO. Starting from the SFT model, the authors add recovery tasks to GRPO training. When a rollout fails during training, its failure state is saved as a recovery task. The policy then samples a group of recovery attempts from this failure state, and these groups are trained together with the groups on the original tasks.

The resulting ReTerm-9B reaches 34.8% on Terminal-Bench 2.0 (+14.6 over the base model) and 58.4% on SWE-bench Verified, with gains across several agent harnesses and out-of-domain benchmarks.

Critical strengths
  1. Training on the agent’s own failures makes sense. Prior work builds recovery data by injecting errors into trajectories, and those errors may not look like the mistakes the model actually makes. Here the failure states come from the model itself, both in the SFT data and during RL.

  2. Strong results and broad evaluation. The 9B model achieves strong results on both terminal and software-engineering benchmarks. The evaluation covers in-domain and out-of-domain benchmarks as well as four agent harnesses.

Critical weaknesses
  1. The comparison between ReTerm-GRPO and GRPO is underspecified. It is unclear whether the GRPO baseline is initialized from the same SFT checkpoint and uses the same number of rollouts per step as ReTerm-GRPO, which uses 160 (8 original and 2 recovery groups).

  2. All results are from a single run. The results would be more convincing with multiple training seeds and repeated evaluations. In addition, the out-of-domain and cross-harness experiments compare only against the base model, not against GRPO, so it is hard to tell whether the gains come from the proposed method or simply from additional training.

Questions for authors
  1. What exactly are the “4K task instances for RL” (line 305)?
3 comments

Q1: The comparison between ReTerm-GRPO and GRPO is underspecified. It is unclear whether the GRPO baseline is initialized from the same SFT checkpoint and uses the same number of rollouts per step as ReTerm-GRPO, which uses 160 (8 original and 2 recovery groups).

A1: Thanks for pointing this out. The original manuscript did not fully specify the configuration of the standard GRPO baseline. We will explicitly include these details in the revised manuscript.

Standard GRPO and ReTerm-GRPO are initialized from the same SFT checkpoint and use the same number of training steps, optimizer, and learning rate. Standard GRPO replaces the two recovery-task groups with two additional original-task groups, using 10 original-task groups per step. Each group contains 16 sampled rollouts, so both methods target 160 rollouts per training step.

Q2: All results are from a single run. The results would be more convincing with multiple training seeds and repeated evaluations.

A2: Thanks for your suggestion. RL training for long-horizon terminal tasks is computationally expensive: a full training run takes approximately 7 days on 16 A100 80GB GPUs. Due to the limited rebuttal time, we have not completed additional full training runs with multiple training seeds. Evaluation is also time-consuming. Within our available GPU budget, we have repeated the main evaluations with different evaluation seeds. We report the mean and sample standard deviation across 3 evaluation runs:

ModelOpenThoughts-TBLiteTerminal-Bench 2.0SWE-bench Verified
Qwen3.5-9B46.8 ± 3.220.6 ± 3.944.9 ± 1.6
ReTerm-9B63.7 ± 2.834.0 ± 2.358.7 ± 0.3

Q3: The out-of-domain and cross-harness experiments compare only against the base model, not against GRPO, so it is hard to tell whether the gains come from the proposed method or simply from additional training.

A3: Thanks for your comments. We have added the standard GRPO baseline to both the out-of-domain and cross-harness generalization tables:

ModelOursTerminus-2mini-SWE-agentOpenHands
Qwen3.5-9B20.222.519.115.7
GRPO28.023.621.419.1
ReTerm-9B34.832.625.828.1
ModelAIME 2024AIME 2025Aider PolyglotSpreadsheetBench-V
Qwen3.5-9B76.773.317.337.8
GRPO80.086.721.842.0
ReTerm-9B86.793.320.448.8

ReTerm-GRPO continues to outperform standard GRPO on most benchmarks, demonstrating the effectiveness of our method.

Q4: What exactly are the “4K task instances for RL” (line 305)?

A4: We apologize for the confusion. TMax-15K divides tasks into four complexity levels: Short, Moderate, Complex, and Intricate. We use the first three levels to construct the recovery demonstration dataset ReTerm-10K. For SFT, we use only recovery trajectories from the Short and Moderate levels. The RL tasks come from the third level, Complex, which contains approximately 4K tasks.

Reviewer_f9T8OpenReview ↗

Response to rebuttal

Thank you for the detailed response. The clarification of the GRPO baseline setup, the added GRPO comparisons, and the explanation of the RL tasks address my concerns. I will maintain my positive score.

Thank you for reviewing our response. We're glad to hear that your concerns have been satisfactorily addressed.

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

ReTerm trains terminal agents to recover from errors encountered during execution. A teacher generates successful continuations from saved student failure states to construct recovery demonstrations for supervised fine-tuning. ReTerm-GRPO then jointly trains on original tasks and recovery continuations sampled from failures of the current policy. The resulting 9B model is evaluated on terminal-agent and software-engineering benchmarks, across multiple harnesses, and through a separate analysis of error recovery.

Critical strengths
  1. Training from saved environment states addresses a specific limitation of long-horizon agents: a policy must act effectively after its own earlier mistakes. The online failure queue also keeps the recovery distribution connected to the evolving policy, rather than relying exclusively on a fixed collection of teacher demonstrations.

  2. The end-to-end results are substantial and extend beyond one harness. On Terminal-Bench 2.0, the reported scores progress from 20.2 for the base model to 25.8 after SFT, 28.0 with standard GRPO, and 34.8 with ReTerm-GRPO. Improvements on SWE-bench Verified and across four terminal-agent harnesses make the practical contribution convincing enough to support acceptance.

Critical weaknesses
  1. The resource allocation in the standard-GRPO ablation needs clarification. Section 4.1 specifies the SFT initialization, 300 RL steps, and a batch containing eight original-task groups and two recovery groups; Table 4 describes standard GRPO as a replacement for ReTerm-GRPO. This reasonably suggests shared training settings, but the baseline's allocation is not separately tabulated. In particular, it is unclear whether its two recovery groups are replaced by additional original-task groups or omitted, and how the realized token and verifier-call budgets compare. The 6.8-point improvement over standard GRPO is promising. Explicit accounting would make the recovery-specific interpretation more persuasive without presuming that the existing comparison is unfair.

  2. The recovery-specific evidence is supportive but less controlled than the end-to-end success comparison. Figure 5 reports category-level recovery annotations checked manually, while Appendix B compares five cases from identical environment states and histories, selecting a representative base-policy failure from each error category. These controlled illustrations are useful. They do not by themselves estimate average recovery gains over a separately sampled failure population, and trajectories generated by different policies may encounter different errors. A larger paired-checkpoint analysis would more directly distinguish improved recovery from generally stronger clean-start task solving. This is a limitation of the mechanism evidence, rather than a reason to discount the benchmark gains.

Questions for authors
  1. Can you confirm which Section 4.1 settings and SFT checkpoint are shared by the standard-GRPO row in Table 4? In particular, how are the two recovery-group slots allocated in standard GRPO, and what are the realized generated-token and verifier-call totals for the two RL conditions?

  2. Beyond selecting one representative base-policy failure per error category, what inclusion criteria determined the five Appendix B cases? Are paired task-verifier outcomes already available for a larger held-out checkpoint set with matched histories and remaining execution budgets? The latter would be additional evidence, rather than a clarification of the five reported illustrations.

1 comment

Q1: Can you confirm which Section 4.1 settings and SFT checkpoint are shared by the standard-GRPO row in Table 4? In particular, how are the two recovery-group slots allocated in standard GRPO?

A1: Thanks for pointing this out. The original manuscript did not fully specify the configuration of the standard GRPO baseline. We will explicitly include these details in the revised manuscript.

Specifically, standard GRPO and ReTerm-GRPO are initialized from the same SFT checkpoint and use the same number of training steps, optimizer, and learning rate. Standard GRPO replaces the two recovery task groups with two additional original task groups, using 10 original task groups per step. Each group contains 16 rollouts, so both methods use a training batch size of 160 rollouts per step.

Q2: What are the realized generated-token and verifier-call totals for the two RL conditions?

A2: Thanks for your comments. Following terminal-agent training work such as TMax [1], both ReTerm-GRPO and standard GRPO use the dynamic sampling strategy introduced by DAPO [2]: groups with identical rewards are filtered out, and sampling continues until the training batch is filled. Because the original logs did not retain complete generation records for all sampled trajectories, we cannot precisely reconstruct the total number of generated tokens. However, we retained the token counts used for policy optimization and the complete verifier-call records. The statistics over the first 100 training steps are as follows:

MetricStandard GRPOReTerm-GRPO
Tokens used for policy optimization168 million179 million
Verifier calls36.0K33.2K

ReTerm-GRPO uses approximately 6.5% more tokens for policy optimization and 7.8% fewer verifier calls than standard GRPO. The verifier-call totals include all sampled trajectories, including those subsequently filtered out by dynamic sampling. They therefore exceed the 16,000 calls corresponding to fixed sampling of 160 trajectories per step over 100 steps.

[1] TMax: A Simple Recipe for Terminal Agents.

[2] DAPO: An Open-Source LLM Reinforcement Learning System at Scale.

Q3: Beyond selecting one representative base-policy failure per error category, what inclusion criteria determined the five Appendix B cases?

A3: Thanks for your comments. Our selection criterion was that each case should exhibit a single, clearly identifiable error type. In practice, however, terminal-agent errors are often intertwined: for example, an inappropriate strategy may lead to subsequent tool-use and implementation errors. Intermediate states satisfying this criterion were relatively limited in the available trajectories. We therefore selected five representative cases, one for each error category, to illustrate differences in the recovery behavior of the two models.

Q4: Are paired task-verifier outcomes already available for a larger held-out checkpoint set with matched histories and remaining execution budgets? The latter would be additional evidence, rather than a clarification of the five reported illustrations.

A4: Thanks for your suggestion. We have completed a paired evaluation on 15 additional environment checkpoints.

These checkpoints were collected from actual execution trajectories of the base model on Terminal-Bench 2.0. At each checkpoint, both models resume from exactly the same environment state and interaction history, with the same remaining execution budget and generation settings. We independently sample 8 continuations from each model and use the original task verifier to determine task success. This evaluation does not distinguish specific error categories. The results are as follows:

TaskGRPOReTerm-GRPOWinner
log-summary-date-ranges87.5%100.0%ReTerm-GRPO
sanitize-git-repo0.0%12.5%ReTerm-GRPO
configure-git-webserver62.5%75.0%ReTerm-GRPO
build-pmars0.0%25.0%ReTerm-GRPO
headless-terminal12.5%50.0%ReTerm-GRPO
largest-eigenval50.0%50.0%Tie
prove-plus-comm87.5%75.0%GRPO
pypi-server62.5%62.5%Tie
kv-store-grpc25.0%62.5%ReTerm-GRPO
nginx-request-logging25.0%62.5%ReTerm-GRPO
cancel-async-tasks0.0%12.5%ReTerm-GRPO
bn-fit-modify12.5%0.0%GRPO
merge-diff-arc-agi-task75.0%50.0%GRPO
fix-code-vulnerability87.5%100.0%ReTerm-GRPO
git-leak-recovery87.5%100.0%ReTerm-GRPO
Total45.0%55.8%ReTerm-GRPO

GRPO and ReTerm-GRPO achieve recovery success rates of 45.0% (54/120) and 55.8% (67/120), respectively, an improvement of 10.8 points. ReTerm-GRPO performs better at 10 of the 15 checkpoints, worse at 3, and ties at 2. These results further demonstrate the effectiveness of our method.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.