Skip to content

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

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

Monitor-based RAG with multi-agent scientific reasoning reduces tool overhead

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Eigen-Agent is a multi-agent scientific reasoning framework that replaces explicit RAG tool calls with a token-level Monitor-based implicit retrieval system, and replaces democratic multi-agent aggregation with Hierarchical Solution Refinement (HSR) and Quality-Aware Iterative Reasoning (QAIR). On Humanity's Last Exam Bio/Chem Gold, it achieves 48.3% accuracy—the highest reported—while cutting token usage by 53.5% and agent steps by 43.7% versus the strongest agent baseline.

Problem

Two bottlenecks limit LLM scientific reasoning: (1) explicit RAG creates a "tool tax"—each retrieval suspends the reasoning state, introduces query reformulation overhead, and risks failure to reintegrate the retrieved knowledge; (2) democratic multi-agent pipelines treat all candidate solutions equally, wasting computation on weak candidates and diluting strong solutions through uniform averaging. Error analysis on 149 HLE problems shows 92.8% of failures involve reasoning errors and 88.7% involve knowledge gaps, with >85% co-occurrence, indicating the two problems are intertwined.

Method

Monitor-based RAG operates at the token level during inference without explicit tool invocation: - Monitor: scans the reasoning stream in 512-character windows with 128-character overlap, detecting semantic uncertainty and triggering retrieval only when needed. - Querier: converts the uncertain reasoning fragment into minimal, targeted keyword queries. - Injector: compresses raw RAG results and rewrites them into the existing reasoning context, preserving narrative flow rather than appending a foreign block.

Hierarchical Solution Refinement (HSR): given candidates \(S = \{s_1, \ldots, s_n\}\), each solution \(s_i\) is rotated as an anchor and refined via peer references \(R = S \setminus \{s_i\}\): \(s'_i = \text{Refine}(s_i, R)\). Refinement applies four targeted repair types: logical completion, numerical correction, method replacement, and expression refinement.

Quality-Aware Iterative Reasoning (QAIR): LLM-based evaluator scores each refined candidate on logic (\(q_{\text{logic}}\)), answer (\(q_{\text{answer}}\)), and explanation (\(q_{\text{explanation}}\)) on a 0–5 scale, combined as: $\(q(s') = 0.2 \cdot q_{\text{logic}}(s') + 0.6 \cdot q_{\text{answer}}(s') + 0.2 \cdot q_{\text{explanation}}(s')\)$ Candidates below threshold \(\tau = 3\) are re-sent to a Corrector with textual suggestions; already-passing solutions skip re-evaluation. Iteration continues until all pass or \(T_{\max}\) is reached.

Key Contributions

  • Token-level implicit RAG (Monitor + Querier + Injector) that eliminates reasoning-flow interruption without sacrificing retrieval adaptivity.
  • HSR: anchor-rotation cross-solution refinement replacing democratic averaging with hierarchical expert-collaboration structure.
  • QAIR: quality-thresholded iterative refinement with early stopping, avoiding computation on already-adequate solutions.
  • Empirical finding: retrieval-heavy tasks benefit from solution diversity; reasoning-heavy tasks benefit from consensus—a dichotomy that motivates adaptive aggregation.
  • 48.3% Pass@1 on HLE Bio/Chem Gold, the highest reported result on this benchmark at publication time.

Results

  • HLE Bio/Chem Gold (149 problems, Pass@1): 48.3% vs. SciMaster (DeepSeek V3.1) 34.9% (+13.4 pp), Grok-4 30.2% (+18.1 pp), GPT-5 22.8% (+25.5 pp), Claude Opus 4.1 21.5% (+26.8 pp).
  • HLE Bio/Chem Gold (Pass@5): 61.74%.
  • SuperGPQA Hard Biology (92 problems): 69.6% (Pass@1), exceeding all baselines including SciMaster (DeepSeek V3.1) at 66.3%.
  • TRQA Literature (172 problems): 54.65% (Pass@1), 79.07% (Pass@5).
  • Efficiency vs. SciMaster: −53.5% token consumption, −43.7% agent steps.
  • Error analysis: 92.8% of failures involve reasoning errors; 88.7% involve knowledge gaps; >85% co-occur.

Limitations

  • Evaluation is restricted to three benchmarks (HLE Bio/Chem, SuperGPQA Biology, TRQA Literature); generalization to non-scientific domains is asserted but not systematically tested.
  • Base model is fixed to DeepSeek-V3.1; model-agnosticism of Monitor-based RAG is claimed but not ablated across multiple backbone LLMs.
  • The Monitor's fixed 512-character / 128-overlap scanning interval is a heuristic; sensitivity to this hyperparameter is not reported.
  • QAIR threshold \(\tau = 3\) and \(T_{\max}\) are not ablated in the excerpted text.
  • RAG database construction details are deferred to appendices; retrieval corpus scope and freshness are not described in the main text.
  • Pass@1 results depend on o3-mini as auto-judge, which may introduce systematic biases for chemistry/biology answer formats.

Relevance to Agentic AI / LLM Agents

Eigen-Agent directly addresses two canonical failure modes in LLM agent design: the cost of explicit tool calls and the inefficiency of flat multi-agent voting. The Monitor-based RAG is a concrete architecture for implicit tool use—a direction that challenges the standard ReAct/explicit-call paradigm and may generalize to any agentic loop where knowledge retrieval must not interrupt long-horizon reasoning chains. HSR and QAIR together formalize hierarchical, quality-driven refinement, offering a template for structured inter-agent collaboration that goes beyond critic-rewrite pipelines. The paper's diversity-vs-consensus finding for retrieval vs. reasoning tasks has direct implications for how agentic systems should aggregate outputs depending on task type.