Research map / Large language models / Retrieval and knowledge
ExpWeaver: LLM Agents Learn from Experience via Latent RAG
Tao Feng, Tianyang Luo, Jingjun Xu, Zhigang Hua, Yan Xie, Shuang Yang, Ge Liu, Jiaxuan You
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 latent retrieval and cross-attention instead of text-based experience retrieval (Search-R1) | ExpWeaver (Qwen2.5-3B) · ExpBench-Generic | average score: 73.27 -> 78.25 | Improved |
| use latent retrieval instead of Search-R1 text retrieval | ExpWeaver · Chem-TDC | exact match: 63.78 -> 69.58 | Improved |
| use latent retrieval instead of text retrieval | ExpWeaver · ExpBench-Rec average | NDCG@10: 28.53 -> 44.40 (vs Search-R1) | Improved |
| latent experience instead of concatenated text | ExpWeaver · token cost per query | tokens | Improved |
| train model on Chem-TDC with transferred experience bank vs zero-shot transfer | ExpWeaver · Chem-TDC transfer | accuracy: outperforms Search-R1 by over 8% in ZeroShot and about 8% in FewShot | Improved |
| increase retrieval number K from 1 to 3 (ablation) | ExpWeaver · five task categories | performance | Improved |
| increase retrieval number K from 3 to 8 (ablation) | ExpWeaver · five task categories | performance | Hurt |
| replace cross-attention with mean pooling (ablation) | ExpWeaver · five task categories | performance | Hurt |
| replace cross-attention with similarity-weighted mean (ablation) | ExpWeaver · five task categories | performance | Hurt |
| replace single-query cross-attention with Qformer (ablation) | ExpWeaver · five task categories | performance | No clear effect |
| set initial mixing coefficient lower bound a_min to 0.98 rather than 0.95 or 0.99 (ablation) | ExpWeaver · five task categories | performance | Improved |
Closest papers by meaning
- ExpeL: LLM Agents Are Experiential Learners (AAAI 2024 · similarity 0.91)
- SimpleMem: Efficient Lifelong Memory for LLM Agents (ICML 2026 · similarity 0.89)
- Learn-by-interact: A Data-Centric Framework For Self-Adaptive Agents in Realistic Environments (ICLR 2025 · similarity 0.89)
- Self-Generated In-Context Examples Improve LLM Agents for Sequential Decision-Making Tasks (NeurIPS 2025 · similarity 0.88)
- C-3PO: Compact Plug-and-Play Proxy Optimization to Achieve Human-like Retrieval-Augmented Generation (ICML 2025 · similarity 0.88)
- Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions (ICLR 2026 · similarity 0.88)
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.
