Research map / Large language models / Agents, code and math
RE-TRAC: REcursive TRAjectory Compression for Deep Search Agents
jialiang zhu, Gongrui Zhang, Xiaolong Ma, Lin Xu, Miaosen Zhang, Ruiqi Yang, Song Wang, Kai Qiu, Zhirong Wu, Qi Dai, Ruichun Ma, Bei Liu and 8 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Re-TRAC (8 rounds) instead of single pass | o4-mini · BrowseComp300 | accuracy (%): 25.7 -> 46.8 | Improved |
| use Re-TRAC (8 rounds) instead of single pass | o3 · BrowseComp300 | accuracy (%): 54.9 -> 69.8 | Improved |
| use Re-TRAC (8 rounds) instead of single pass | GPT-5-medium · BrowseComp300 | accuracy (%): 48.3 -> 66.6 | Improved |
| use Re-TRAC (8 rounds) instead of single pass | DeepSeek-V3.2 · BrowseComp300 | accuracy (%): 45.3 -> 60.8 | Improved |
| use Re-TRAC instead of majority voting at 8 | GLM-4.7 · BrowseComp300 | accuracy (%): 41.7 -> 60.7 | Improved |
| use Re-TRAC instead of weighted voting at 8 | GPT-5-medium · BrowseComp300 | accuracy (%): 64.7 -> 66.6 | Improved |
| add Re-TRAC SFT on GLM-4.7 trajectories (ablation) | Qwen3-4B-Instruct · BrowseComp300 | accuracy (%): 2.7 -> 30.0 | Improved |
| add Re-TRAC SFT on GLM-4.7 trajectories (ablation) | Qwen3-4B-Instruct · GAIA | accuracy (%): 24.4 -> 70.4 | Improved |
| add instruction to freely use the summary (ablation) | Re-TRAC with o3 · BrowseComp300 round 8 | accuracy (%): 68.9 -> 71.7 | Improved |
| use GLM-4.7 as summarizer instead of the model itself (ablation) | RE-TRAC-4B · BrowseComp300 | accuracy (%): 30.0 -> 38.5 | Improved |
| use GLM-4.7 as summarizer instead of the model itself (ablation) | RE-TRAC-30B-A3B · BrowseComp300 | accuracy (%): 53.0 -> 52.4 | No clear effect |
| use majority voting instead of single pass | GLM-4.7 · BrowseComp300 | accuracy (%): 37.7 -> 41.7 | Mixed |
Closest papers by meaning
- Search More, Think Less: Rethinking Long-Horizon Agentic Search for Efficiency and Generalization (NeurIPS 2026 · similarity 0.85)
- Spinning Straw into Gold: Relabeling LLM Agent Trajectories in Hindsight for Successful Demonstrations (ICLR 2026 · similarity 0.85)
- State-Covering Trajectory Stitching for Diffusion Planners (NeurIPS 2025 · similarity 0.84)
- OpenSearch-VL: An Open Recipe for Frontier Multimodal Search Agents (NeurIPS 2026 · similarity 0.84)
- LLMs Improving LLMs: Agentic Discovery for Test-Time Scaling (NeurIPS 2026 · similarity 0.84)
- MM-DeepResearch: A Simple and Effective Multimodal Agentic Search Baseline (ICML 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
