CureAgent: A Training-Free Executor-Analyst Framework for Clinical Reasoning

🕒 Published (v1): 2025-12-05 09:56 UTC · Source: Arxiv · Venue: NEURIPS 2025 · link

Ask a follow-up

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

💬 Ask ChatGPT✦ Ask Claude

TL;DR

CureAgent proposes a training-free Executor-Analyst framework for clinical reasoning that decouples fine-tuned tool-calling agents (Executors) from large-context foundation models (Analysts) to overcome the Context Utilization Failure in compact medical LLMs. A Stratified Ensemble (late fusion) topology preserves retrieval diversity across parallel pipelines, outperforming both single-agent baselines and early-fusion alternatives. The system achieves 83.803% on CURE-Bench (NeurIPS 2025) without any end-to-end fine-tuning.

Problem

Small fine-tuned clinical agents (e.g., TxAgent, Llama-3.1-8B) exhibit a Context Utilization Failure: they successfully retrieve biomedical evidence via tool calls but fail to ground their final diagnosis in that evidence, leading to hallucination. Larger foundation models have stronger reasoning but lack task-specific tool-call syntax, and naively swapping backbones does not close the gap. No prior approach cleanly decouples syntactic tool execution from semantic clinical synthesis.

Method

Executor-Analyst Framework (training-free): - Executor: Multiple parallel TxAgent instances (Llama-3.1-8B fine-tuned on ToolUniverse ≥200 biomedical APIs) run at T=0.8 with a self-consistency mechanism — top-k tool outputs and majority-vote reasoning traces are aggregated across n₁ parallel runs. - Analyst: Gemini 2.5 Flash/Pro ingests the aggregated noisy evidence stream, performs supplementary web search for gaps, and synthesizes a chain-of-thought clinical rationale without handling tool syntax. - Stratified Ensemble (Config B, late fusion): Instead of pooling all Executor outputs into one context (early fusion/Config A), the Executor budget is partitioned into n₂ independent subgroups each with n₁ instances. Each subgroup feeds a dedicated Analyst; final answer is decided by majority vote across Analysts. This preserves retrieval diversity until the final voting stage. - Post-processing module: Deterministic regex-based format calibration and response deduplication.

Key Contributions

• Identification and characterization of Context Utilization Failure as the dominant failure mode (65.8%) in compact clinical agents.
• Executor-Analyst decoupling: separates syntactic tool execution from semantic reasoning without any additional training.
• Stratified Ensemble (late fusion) topology that outperforms global pooling (early fusion) by preventing premature evidence filtering.
• Empirical discovery of the Context-Performance Paradox: accuracy drops from ~94% to 87.93% when reasoning contexts exceed 12k tokens.
• Empirical evidence of the Curse of Dimensionality in tool spaces: expanding ToolUniverse from 200+ to 600+ tools degrades TxAgent accuracy from 92.0% to 87.5%.
• State-of-the-art result on CURE-Bench NeurIPS 2025 (prize-winning) without fine-tuning.

Results

• TxAgent baseline (single agent): 69.325% (phase2)
• Gemini 2.5 Flash baseline (no tools): 63.104%
• TxAgent × 30 (self-consistency only): 73.508%
• Config A (30 Executors + 3 Analysts, global pooling): 80.510%
• Config B (10×3 Stratified Ensemble, no search): 81.367%
• Config B + Gemini 2.5 Flash with search: 83.803% (leaderboard submission)
• Gemini 2.5 Pro + search (no task-specific tools): 74.806%
• Gemini 3 Pro + search (post-competition reference): 81.283%

Limitations

• Compute-intensive: achieving top performance requires 30+ parallel Executor runs plus multiple Analyst calls per query.
• Context-Performance Paradox remains unsolved — evidence compression/filtering strategies are future work.
• Scaling to ToolUniverse v2.0 (600+ tools) degrades Executor accuracy by 4.5 pp; scaling to 1000+ tools is unaddressed.
• Analyst (Gemini 2.5) is a closed-source proprietary model, limiting reproducibility and deployment flexibility.
• Evaluation confined to CURE-Bench multiple-choice format; generalization to open-ended clinical tasks is untested.
• No formal ablation of the web search contribution vs. structural gains in the final 83.803% result.

Relevance to Agentic AI / LLM Agents

CureAgent directly addresses a core failure mode in tool-augmented agents — the disconnect between successful tool retrieval and reasoning grounded in retrieved evidence — which is a general problem beyond the clinical domain. The Executor-Analyst decomposition is an instantiation of the broader principle of role specialization in multi-agent systems, demonstrating that heterogeneous agent collaboration (fine-tuned specialist + large generalist) outperforms homogeneous scaling. The Stratified Ensemble finding — that late fusion over diverse retrieval paths beats early consensus — is a transferable insight for any agentic pipeline where evidence diversity matters. The Context-Performance Paradox and Curse of Dimensionality findings are practically important failure modes that any agent designer working with long contexts or large tool libraries must account for.