Eigen-Agent: Adaptive Multi-Agent Scientific Reasoning with Monitor-Based RAG

Xiangru Tang; Wanghan Xu; Yujie Wang; Zijie Guo; Daniel Shao; Jiapeng Chen; Cixuan Zhang; Ziyi Wang

cs.CLSep 25, 2025

Eigen-Agent: Adaptive Multi-Agent Scientific Reasoning with Monitor-Based RAG

Authors: Xiangru Tang, Wanghan Xu, Yujie Wang, Zijie Guo, Daniel Shao, Jiapeng Chen, Cixuan Zhang, Ziyi Wang, +8 more

Organizations: Yale University · Shanghai Jiao Tong University · Fudan University · University of California, Los Angeles · Shanghai AI Lab · University of Oxford · Eigen AI

Abstract

Large language models (LLMs) have recently shown strong progress on scientific reasoning, yet two major bottlenecks remain. First, explicit retrieval fragments reasoning, imposing a hidden "tool tax" of extra tokens and steps. Second, multi-agent pipelines often dilute strong solutions by averaging across all candidates. We address these challenges with a unified framework that combines implicit retrieval and structured collaboration. At its foundation, a Monitor-based retrieval module operates at the token level, integrating external knowledge with minimal disruption to reasoning. On top of this substrate, Hierarchical Solution Refinement (HSR) iteratively designates each candidate as an anchor to be repaired by its peers, while Quality-Aware Iterative Reasoning (QAIR) adapts refinement to solution quality. On Humanity's Last Exam (HLE) Bio/Chem Gold, our framework achieves 48.3% accuracy - the highest reported to date, surpassing the strongest agent baseline by 13.4 points and leading frontier LLMs by up to 18.1 points, while simultaneously reducing token usage by 53.5% and agent steps by 43.7%. Results on SuperGPQA and TRQA confirm robustness across domains. Error analysis shows that reasoning failures and knowledge gaps co-occur in over 85% of cases, while diversity analysis reveals a clear dichotomy: retrieval tasks benefit from solution variety, whereas reasoning tasks favor consensus. Together, these findings demonstrate how implicit augmentation and structured refinement overcome the inefficiencies of explicit tool use and uniform aggregation. The code is available at https://github.com/tangxiangru/Eigen-1.

Figures & tables

Figure 1: HLE Bio/Chem Gold overall accuracy. On the 149-problem HLE Bio/Chem Gold split (Pass@1, auto-judged by o3-mini ), our system attains 48.3% accuracy, exceeding the strongest agent baseline (SciMaster) by +13.4 points and leading frontier LLMs by up to +18.1 points.

Figure 2: Population genetics case with two failure modes. Left (Error Type 1): the model confidently recalls an incorrect formula (

\theta=2N_{e}\mu

) and derives

N_{e}=\theta/2\mu

, yielding the wrong answer. Right (Error Type 2): the model retrieves the correct relation (

\theta=4N_{e}\mu

) via explicit RAG, but the reasoning flow is disrupted and the result is not reintegrated into the original problem, illustrating the tool tax . Our Monitor-based RAG avoids this context suspension by injecting the correct formula directly into the reasoning stream.

System	Triggering	Fine-grained	Continuity	Efficiency	Adaptivity
Single-round RAG	✗	✗	✗	✓	✗
Iterative RAG	✓	✓	✗	✗	✓
Reasoning RAG	✓	✓	✓	✗	✓
Monitor-based RAG (Ours)	✓	✓	✓	✓	✓

Table 1: RAG paradigms vs. key capabilities. Single-round RAG is efficient but inadaptable; iterative RAG improves grounding but increases latency; reasoning-aware RAG offers tighter coupling yet still relies on explicit calls. Monitor-based RAG integrates evidence implicitly at the token level, improving continuity and efficiency.

Figure 3: Framework overview. (a) Monitor-based RAG operates globally during reasoning: the Monitor detects insufficiency in the reasoning stream, the Querier generates targeted queries, and the Injector integrates retrieved evidence into context with minimal disruption. (b) Building on this substrate, the Proposer generates initial candidate solutions. Each candidate is revised individually by the Corrector , which applies local targeted fixes without access to other solutions. The improved candidates are then passed to HSR , which enables cross-solution refinement via anchor–reference relationships. Finally, QAIR evaluates overall quality and may invoke the Corrector again if needed, while the Ranker selects the strongest solution as the final answer.

Figure 4: Haplotype counting with single crossovers (F1 $\rightarrow$ F3). The Proposer exhibits uncertainty about recombination constraints; Monitor detects insufficiency, Querier issues a targeted query, and Injector integrates two retrieved facts. This enables the reasoning to exclude invalid cases and converge on correct count of 30 haplotypes.

Figure 5: Illustrative example: HSR. The system rotates anchors among candidate solutions and integrates targeted corrections from references (e.g., fixing arithmetic mistakes, filling missing steps). Instead of averaging inconsistent candidates, HSR applies targeted improvements to yield a coherent final answer.

Figure 6: Comparison of retrieval backends within Monitor-based RAG.

Figure 7: Error type distribution. Analysis of incorrect solution logs shows reasoning- and knowledge-related errors as dominant. Note that a single problem may involve multiple error types, so percentages do not sum to 100.

Figure 8: Diversity vs. consensus. Task-dependent effect of solution diversity: retrieval tasks benefit from variety, while reasoning tasks benefit from agreement. The horizontal axis reports the average pairwise consistency score among Proposers, computed by an LLM-based judge that evaluates semantic overlap between answers on a 0–1 scale. The vertical axis shows the average accuracy score , also judged by an LLM, which rates the degree of correctness of each answer relative to ground truth on a continuous 0–1 scale (rather than a binary 0/1). This continuous evaluation enables us to capture fine-grained trends between diversity and correctness across tasks. The fitted trend lines further highlight the contrast: retrieval tasks show a relatively flat slope (

\approx

0.369 ), whereas reasoning tasks exhibit a much steeper positive slope (

\approx

0.851 ), indicating a stronger dependence on consensus.

Figure 9: Quantifying the tool tax. Comparison of accuracy and coherence relative to compute cost, showing the overhead of explicit retrieval vs. implicit augmentation. Note that in this analysis, the baseline refers to the explicit RAG configuration (i.e., Proposers equipped with paper retrieval and web search), which represents the standard setup in most existing agent systems.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Biology Category (n=2029)	% of Domain	Chemistry Category (n=359)	% of Domain
Molecular and Cell Biology (777)	38.39%	Organic Chemistry (172)	47.91%
Immunology and Microbiology (482)	23.76%	Physical Chemistry (71)	19.78%
Genetics, Genomics & Computation (411)	20.26%	Materials Chemistry (68)	18.94%
Neuroscience (205)	10.11%	Analytical Chemistry (37)	10.31%
Ecology and Evolution (149)	7.35%	Inorganic Chemistry (11)	3.06%
Overall Summary

Appendix

Table A1: Paper Classification Statistics by Domain (Side-by-Side)

Hyperparameter	Value	Description
model	gpt-4.1-mini	LLM model used in RAG Monitor
query_top_k	$3$	Maximum number of retrieved documents for each query.
rag_chunk	$512$	Text chunk size for RAG monitoring.
rag_overlapping	$128$	Number of overlapping characters between consecutive chunks to maintain continuity.
max_rag	$2$	Maximum number of RAG insertions allowed in one reasoning step.
temperature	$0.5$	Controls generation diversity; higher values lead to more randomness.

Appendix

Table A2: Main Hyperparameter Settings used in RAG Monitor.

Configuration	Acc (%)
Baseline	25.3%
Baseline + HSR/QAIR (No RAG)	34.5%
Eigen-Agent (Generic Mixed-Domain Corpus)	45.6%
Eigen-Agent (Bio/Chem Domain Corpus)	48.3%

Appendix

Table A5: Evaluation of Eigen-Agent under no-retrieval, generic mixed-domain, and Bio/Chem retrieval corpora.

Explore similar work

SciOrch: Learning to Orchestrate Expert LLMs for Solving Frontier Multimodal Scientific Reasoning Tasks
Jun 14, 2026Jingru Guo, Xiangyuan Xue, Lian Zhang +6LLM Agent Orchestration Multimodal Reasoning
1Imperial College London · 2The Chinese University of Hong Kong · University of Illinois Urbana-Champaign +4
AI scientists produce results without reasoning scientifically
Apr 20, 2026Martiño Ríos-García, Nawaf Alampara, Chandan Gupta +5AI Agent Reliability Scientific Reasoning
Laboratory of Organic and Macromolecular Chemistry (IOMC), Friedrich Schiller University Jena, Humboldstr. 10, 07743 Jena, Germany · Department of Civil Engineering, Indian Institute of Technology Delhi, Hauz Khas, New Delhi 110016, India · School of Interdisciplinary Research, Indian Institute of Technology Delhi, Hauz Khas, New Delhi 110016, India +3
TopoAgent: A Self-Evolving Topological Agent for Multimodal Scientific Reasoning
Jul 16, 2026Mingze Xu, Yinghui Li, Jiayi Kuang +5AI Agents for Scientific Discovery Multimodal Reasoning
Tsinghua University · Sun Yat-sen University · 3Tencent

Jun 14, 2026J/K move · Enter open · S save

SciOrch: Learning to Orchestrate Expert LLMs for Solving Frontier Multimodal Scientific Reasoning Tasks

Jingru Guo, Xiangyuan Xue, Lian Zhang, Wanghan Xu+5

1Imperial College London · 2The Chinese University of Hong Kong · University of Illinois Urbana-Champaign+4

LLM Agent Orchestration Multimodal Reasoning Scientific Reasoning in Language Models Large Language Model-Guided Scientific Discovery RL for LLM Agents

Open paper AskPDF arXiv

Frontier scientific reasoning remains a major challenge for large language models (LLMs), where even the strongest commercial systems fall short of expert-level performance. A closer look at model behavior reveals substantial complementarity that single-model evaluation hides: different frontier models excel on different question types, and no single model captures the full picture. We present SciOrch, a framework that trains a lightweight 8B model to orchestrate frontier LLMs for scientific reasoning. The orchestrator decomposes each question, delegates sub-problems to selected commercial models through API calls, and synthesizes a final answer. Training such an orchestrator is fundamentally harder than conventional agentic RL: each action triggers an API call that is expensive in both dollar cost and latency, making standard online rollouts infeasible. We address this with MCTS-based approach, producing diverse orchestration trajectories, extracting per-node single-turn samples, and optimizing the orchestrator with GRPO-style training. On a 240-question test set spanning SGI-Reasoning and Scientists' First Exam, SciOrch reaches 56.66% average accuracy, outperforming the strongest single commercial model by 3.74% and the strongest multi-agent baseline by 3.33%. It also attains the best accuracy on both SGI and SFE with less than half the API cost of typical multi-agent methods.

Abstract

Explore similar work

SciOrch: Learning to Orchestrate Expert LLMs for Solving Frontier Multimodal Scientific Reasoning Tasks

AI scientists produce results without reasoning scientifically

TopoAgent: A Self-Evolving Topological Agent for Multimodal Scientific Reasoning