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

The gist The framework EIGEN-1 addresses bottlenecks in scientific reasoning by combining implicit retrieval through Monitor-based RAG with structured collaboration via Hierarchical Solution Refinement (HSR) and Quality-Aware Iterative Reasoning (QAIR) to achieve superior accuracy and efficiency.

How it works

EIGEN-1 is built on a unified framework that integrates global retrieval, role-based reasoning, and higher-level refinement into a single workflow as shown in Figure 3 The core mechanism involves Monitor-based RAG operating globally during reasoning to detect knowledge gaps without fragmenting the workflow This module operates by having the Monitor detect insufficiency in the reasoning stream, and then the Querier generates targeted queries, and finally the Injector seamlessly integrates retrieved evidence back into context with minimal disruption The Monitor acts as a sentinel that periodically examines the reasoning trace and determines whether external knowledge is required Once insufficiency is detected, the retrieval is immediately triggered, and to balance timeliness and efficiency, the Monitor runs in a streaming setup The Querier converts the uncertain fragment into one or more retrieval queries By ensuring that queries are as fine-grained as possible, the Querier avoids unnecessary expansion of the search space while maximizing the relevance of retrieved evidence The Injector first filters and compresses raw RAG outputs into concise, utility-focused snippets to avoid redundancy and irrelevant noise Then it rewrites and integrates the selected evidence in the Proposer’s reasoning context, ensuring coherence and preserving the natural flow of the reasoning narrative

Hierarchical Solution Refinement (HSR)

HSR challenges the assumption that all solutions should contribute equally to the final output by establishing structured relationships among solutions that mirror expert collaboration patterns Instead of democratic averaging, HSR organizes solutions into anchor–reference structures for targeted repair Each solution is iteratively designated as the anchor si and the remaining set R = S S i provides references This rotation ensures that every solution benefits from peer-informed repair, preventing premature convergence to a single trajectory The refinement process is described as s'i = Refine(si, R), where Refine(·) denotes the LLM-driven mechanism that applies multidimensional repairs to the anchor These repair strategies include Logic Completion, Numerical Correction, Method Replacement, and Expression Refinement This yields a coherent and biologically valid pathway by applying targeted repairs to yield a unified solution

Quality-Aware Iterative Reasoning (QAIR)

QAIR introduces an evaluation-driven control mechanism to refine candidate solutions after the HSR stage, replacing fixed workflows with adaptive cycles that respond dynamically to quality trajectories and problem characteristics Each solution s' in S' is evaluated by an LLM-based evaluator on three quality dimensions: logic, answer, and explanation The composite score is calculated as q(s') = 0.2 · qlogic(s') + 0.6 · qanswer(s') + 0.2 · qexplanation(s'), with a higher weight on the answer dimension Candidates meeting the threshold τ = 3 are retained, while those failing are marked non-passing and passed to the corrector for targeted revision This mechanism ensures that every revision contributes meaningfully, preventing uncontrolled exploration or redundant cycles

Performance and Efficiency Gains

The framework achieves 48.3% accuracy in Humanity’s Last Exam Bio/Chem Gold, surpassing SciMaster by +13.4 percentage points while reducing token consumption by 53.5% and agent steps by 43.7% The analysis shows that retrieval tasks benefit from diversity, whereas reasoning tasks favor consensus This dichotomy suggests that different ranking strategies are optimal for different task types The computational burden of explicit tool invocation, quantified as the tool tax, is shown to be more than doubled compared to the no-IR setting EIGEN-1 reduces token consumption by more than half (from 470.6K to 218.4K) and cuts workflow iterations nearly in half (from 94.8 to 51.3)

Error Analysis and Robustness

Analysis of failed problems reveals two dominant error modes: reasoning process errors (92.78%) and knowledge application errors (88.66%) These frequently co-occur, suggesting that successful scientific reasoning requires seamless integration of domain knowledge with logical inference The framework's success is validated across three heterogeneous benchmarks: biomedical, chemical and medical The results establish the advantage of the design in three heterogeneous benchmarks, demonstrating not only raw accuracy gains but also improved adaptability across domains

Conclusion

The experiments validate three key architectural innovations in EIGEN-1: implicit knowledge augmentation through Monitor-based RAG, HSR improving over uniform multi-agent aggregation by introducing structured anchor–reference relationships, and QAIR adaptively balancing exploration and early stopping On HLE Bio/Chem, EIGEN-1 reaches 48.3% accuracy under compute-matched settings with a 53.5% token reduction By integrating external knowledge with minimal disruption to reasoning flow, our framework addresses key limitations of prior approaches in complex scientific problem solving

--- Page 3 ---

The gist The framework EIGEN-1 addresses bottlenecks in scientific reasoning by combining implicit retrieval through Monitor-based RAG with structured collaboration via Hierarchical Solution Refinement (HSR) and Quality-Aware Iterative Reasoning (QAIR) to achieve superior accuracy and efficiency.

1. Bold header: Monitor-based RAG for Implicit Augmentation

This mechanism operates at the token level, integrating external knowledge with minimal disruption to reasoning, which reduces token usage by over half while preventing context suspension that causes a tool tax.

2. Bold header: Hierarchical Solution Refinement (HSR) for Structured Collaboration

HSR designates each candidate as an anchor to be repaired by its peers, moving beyond rigid democratic workflows to apply structured cross-solution refinement rather than uniform averaging.

3. Bold header: Quality-Aware Iterative Reasoning (QAIR) for Adaptive Convergence

QAIR utilizes a quality threshold, retaining solutions where the composite score is high and invoking the Corrector only when necessary, ensuring that every revision contributes meaningfully, preventing uncontrolled exploration or redundant cycles.

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.