TraceLite: Efficient Trace Reduction for SWE Agents via Optical Compression and Redundancy Removal
[Inżynieria oprogramowania] Autorzy: Jingyu Xiao, Qianxi Liu, Yintong Huo i 2 więcej
Software engineering agents solve repository-level tasks through iterative tool use, but their continuously growing interaction traces incur substantial token consumption and inference overhead. Existing compression approaches often require model training, auxiliary LLM calls, or assumptions about specific content such as source code. We present TraceLite, a simple yet effective, training-free framework for compressing heterogeneous SWE-agent observations. Our preliminary study shows that optical rendering preserves substantial textual and semantic information under moderate compression, while agent trajectories exhibit both cross-step and phase-level redundancy. Motivated by these findings, TraceLite preserves recent observations as high-fidelity text, renders historical observations as compact images, removes overlapping content, and discards phase-expired observation payloads after c…
- Kategoria arxiv
- cs.SE
- Kategoria
- Inżynieria oprogramowania