acceptodds
Withdrawn from ICLR 2027

TraceBridge: Interaction-Aware Trajectory Rewriting across Harnesses for Coding Agent Training

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Coding agents benefit from supervised fine-tuning (SFT) on high-quality interaction trajectories. However, publicly available trajectories span heterogeneous agent harnesses, complicating their reuse for training: source calls may be incompatible with the target interface, while removing unsupported interactions can discard findings needed for subsequent actions. We introduce TraceBridge, an interaction-aware trajectory rewriting system that considers source actions together with their preceding context and associated feedback to determine what becomes target-action supervision and what remains as context. Using tool contracts and recorded evidence, it compiles supported operations into target calls and retains unmapped interactions as historical context with zero direct loss. Using the same 10,000 heterogeneous trajectories for Qwen3-8B SFT, TraceBridge outperforms direct source training and Agent Data Protocol (ADP), the strongest baseline in our comparison, across OpenHands and SWE-agent on SWE-bench Verified, Pro, and Multilingual. Over three evaluation runs, mean Verified resolved rates reach 25.07% and 26.73%, improving over direct source training by 12.40 and 7.40 percentage points, respectively. Relative to ADP, OpenHands behavior in one evaluation run shows lower call rejection (3.15% to 1.47%) and fewer empty patches (39.08% to 18.07%). These results demonstrate the training value of interaction-aware trajectory rewriting. We have publicly released our code at https://anonymous.4open.science/r/trace-bridge-308E.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
coding agents, agent trajectories, supervised fine-tuning, trajectory rewriting, cross-harness transfer

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_HTooOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces TraceBridge, an interaction-aware system for reusing coding-agent training trajectories across different agent frameworks. It translates supported operations into compatible target tool calls using tool contracts and recorded evidence, while preserving unmappable interactions as context excluded from direct training loss, retaining information needed for subsequent actions. Using the same 10,000 trajectories to fine-tune Qwen3-8B, TraceBridge outperforms direct source-trajectory training and Agent Data Protocol across OpenHands and SWE-agent on SWE-bench Verified, Pro, and Multilingual. Behavioral analysis further shows fewer rejected calls, failed edits, and empty patches, demonstrating how adapting actions while preserving interaction context can improve the training value of existing agent experience.

Critical strengths
  1. The result is better than ADP, which is another traj translation method.
Critical weaknesses
  1. It's quite confusing how to deal with think/content; it seems that section 4.3 didn't deal with that. However, the thinking pattern in the traj is important for training. Also, how to deal with the tool name in think? For example, the model considers which parameter should be added in the tool call in think, and it is highly related to the tool format.

  2. Such a method seems to work only when the data volume is small. The effectiveness of traj (sampled from different harnesses ) on different harnesses is orthogonal. (The generalization can only be from SWE-Agent to Openhands; other pairs are weak, as shown in Qwen3-Coder-Next Technical Report and later unpublished research). So it's more like having more SWE-Agent data. If we just scale up more SWE-Agent data, it can be better than this method. (And the open-source community already has.)

  3. It will train some bad patterns (mismatches) in the model, which may harm the RL process later.


I think this is just a compromise method when the data is really lacking, but this is not real in SWE.

Questions for authors
  1. How to deal with the thinking part?
Reviewer_kuqfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

TraceBridge rewrites trajectories of coding agents for training under a different harness. It converts supported operations into targeted tool calls and keeps unsupported interactions as masked context. Using 10,000 trajectories, it improves Qwen3-8B over direct SFT and ADP on three benchmarks under OpenHands and SWE-agent.

Critical strengths
  • The paper addresses a practical problem: reusing agent trajectories collected with different tools and interfaces, i.e. under different harnesses.

  • Separating useful context from actions the model should imitate is sensible. The gains are consistent across the tested settings, and the behavioral analysis provides useful supporting evidence.

Critical weaknesses
  • The source of the gains is unclear. The method changes conversion rules, summaries, message roles, and loss masks together, without ablations. ADP also retains the unsupported workflow records examined in Table 13, so the benefit of the proposed context treatment needs to be isolated.
  • The rewritten trajectories are not validated by execution. Target calls are paired with source observations, which may not match what the target tools would actually produce. Static checks do not establish that the full trajectory remains valid.
  • Preserved context may contain information the deployed agent cannot obtain. For example, keeping a subagent’s findings does not teach the target agent how to discover them. If a context-only step changes files or process state, later actions may also depend on an unexecuted operation.
  • The evaluation is not enough. The full comparison uses one base model and two target harnesses. The second-model experiment does not repeat the main comparison, and its resolution gains are not statistically significant.
Questions for authors
  1. Are separate checkpoints trained for OpenHands and SWE-agent in Table 2? Is the Source SFT checkpoint shared across both evaluations?
  2. How much do summaries and LLM confirmation contribute beyond deterministic conversion with the same loss masks?
  3. Do the main results hold on another model and a very different target harness?
Reviewer_Ree1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces TraceBridge, an interaction-aware trajectory rewriting framework that enables reuse of coding-agent trajectories across heterogeneous agent harnesses. Specifically it analyzes the interaction context, tool semantics, execution conditions, and feedback associated with each action and converts compatible operations into target-harness tool calls while preserving unsupported interactions as loss-masked contextual information. Using the same 10,000-source-trajectory corpus, the authors show that models trained on TraceBridge-rewritten traces substantially outperform direct source training and the ADP baseline across SWE-bench Verified, Pro, and Multilingual benchmarks under both OpenHands and SWE-agent evaluation settings.

Critical strengths
  • Addresses an important and practical problem in coding-agent training: how to reuse valuable trajectory data collected under different agent harnesses and tool interfaces.
  • Proposes a principled interaction-aware rewriting framework that considers action semantics, execution context, tool contracts, and feedback rather than relying on simple tool renaming or serialization conversion.
  • Demonstrates consistent gains over both direct source training and ADP across multiple benchmarks and target harnesses using the same training corpus.
  • Evaluates on multiple benchmarks (Verified, Pro, and Multilingual) and across two deployment harnesses, providing stronger evidence of generality than many prior trajectory-processing works.
Critical weaknesses
  • Experiments are limited to a single base model (Qwen3-8B) for the main evaluation, leaving open questions about scalability and transferability across model families and sizes.
  • Although improvements are substantial, it remains unclear how much of the gain comes from improved action conversion versus the additional contextual information retained through loss-masked summaries.
  • The framework assumes access to detailed trajectory records, tool definitions, and execution metadata, which may not always be available in proprietary or partially logged agent systems.
  • While the paper reports rewriting costs, the computational overhead and cost-benefit tradeoff of large-scale deployment are not extensively analyzed.
Questions for authors
  • What is the relative contribution of action adaptation, candidate confirmation, and context preservation to the final performance gains? Can ablation studies isolate the impact of each component?
  • How robust is TraceBridge to the choice of rewriting LLM used for candidate confirmation and context summarization? Would the results remain stable with different models?
  • To what extent do the improvements arise from preserving unsupported interactions as context versus improving tool-call compatibility through action rewriting?
  • How does TraceBridge perform when transferring between harnesses that have substantially larger semantic gaps than OpenHands and SWE-agent?
  • Can the framework generalize beyond coding agents to other agent domains such as web agents, enterprise assistants, or tool-using research agents?
  • Given the additional LLM calls required for rewriting, how does the training benefit scale relative to rewriting cost for larger corpora?

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

Related papers

Discussion (0)

Sign in to comment.