Research map / Multi-agent systems and games / LLM agents
Eigen-1: Scientific Reasoning through Adaptive Multi-Agent Refinement and Monitor-based RAG
Xiangru Tang, Wanghan Xu, Yujie Wang, Zijie Guo, Daniel Shao, Cixuan Zhang, Ziyi Wang, Lixin Zhang, Guancheng Wan, Zhenfei Yin, Wenlong Zhang, LEI BAI and 3 more
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Eigen-1, a framework for scientific reasoning in LLMs that combines implicit retrieval with structured multi-agent refinement to address retrieval-induced token overhead and solution dilution in multi-agent pipelines.
Problem
Scientific reasoning with LLMs suffers from explicit retrieval fragmenting reasoning (a hidden tool tax of extra tokens and steps) and multi-agent pipelines diluting strong solutions by averaging across all candidates.
Approach
Unifies implicit retrieval (monitor-based RAG) with adaptive multi-agent refinement so that strong candidate solutions are not averaged away with weaker ones; the abstract only partially describes the mechanism.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace explicit paper-retrieval RAG with Monitor-based implicit RAG (ablation) | Eigen-Agent baseline · HLE Bio/Chem Gold | tokens (K): 470.6 -> 218.4 | Improved |
| replace explicit paper-retrieval RAG with Monitor-based implicit RAG (ablation) | Eigen-Agent baseline · HLE Bio/Chem Gold | accuracy (%): 41.4 -> 34.5 | Hurt |
| add explicit paper retrieval to the no-RAG baseline (ablation) | Eigen-Agent baseline · HLE Bio/Chem Gold | accuracy (%): 25.3 -> 41.4 | Improved |
| add explicit paper retrieval to the no-RAG baseline (ablation) | Eigen-Agent baseline · HLE Bio/Chem Gold | agent steps: 43.4 -> 94.8 | Hurt |
| add Querier to Monitor-based RAG (ablation) | Monitor-based RAG · HLE Bio/Chem Gold | accuracy (%): 34.5 -> 36.8 | Improved |
| add Injector to Monitor + Querier (ablation) | Monitor-based RAG · HLE Bio/Chem Gold | accuracy (%): 36.8 -> 40.3 | Improved |
| add Hierarchical Solution Refinement (HSR) (ablation) | Monitor + Querier + Injector · HLE Bio/Chem Gold | accuracy (%): 40.3 -> 43.7 | Improved |
| add Quality-Aware Iterative Reasoning (QAIR) (ablation) | Monitor + Querier + Injector + HSR · HLE Bio/Chem Gold | accuracy (%): 43.7 -> 48.3 | Improved |
| add Monitor, Querier and Injector to the full pipeline (vs removing them) (ablation) | Eigen-Agent · HLE Bio/Chem Gold | accuracy (%): 48.5 -> 48.3 | No clear effect |
| add Monitor, Querier and Injector to the full pipeline (vs removing them) (ablation) | Eigen-Agent · HLE Bio/Chem Gold | tokens (K): 461.3 -> 218.9 | Improved |
| Eigen-Agent Pass@1 versus the strongest agent baseline on HLE | SciMaster (DeepSeek V3.1) · HLE Bio/Chem Gold | accuracy (%): 34.92 -> 48.30 | Improved |
| use Eigen-Agent instead of SciMaster on SuperGPQA hard biology | DeepSeek V3.1 · SuperGPQA Biology (hard) | accuracy (%): 66.30 -> 69.57 | Improved |
Closest papers by meaning
- MC-Search: Evaluating and Enhancing Multimodal Agentic Search with Structured Long Reasoning Chains (ICLR 2026 · similarity 0.88)
- From What to Why: A Multi-Agent System for Evidence-based Chemical Reaction Condition Reasoning (ICLR 2026 · similarity 0.88)
- Fleet of Agents: Coordinated Problem Solving with Large Language Models (ICML 2025 · similarity 0.87)
- Training a Scientific Reasoning Model for Chemistry (NeurIPS 2025 · similarity 0.87)
- CoT-Evo: Evolutionary Distillation of Chain-of-Thought for Scientific Reasoning (ICLR 2026 · similarity 0.87)
- CURIE: Evaluating LLMs on Multitask Scientific Long-Context Understanding and Reasoning (ICLR 2025 · similarity 0.87)
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.
