PaperlayerView exact revision on arXiv

Source-linked AI summary

ReflecTool: Towards Reflection-Aware Tool-Augmented Clinical Agents

Yusheng Liao, Shuyang Jiang, Yanfeng Wang, Yu Wang

arXiv:2410.17657v3cs.CL

TL;DR

Clinical agents remain limited in handling diverse clinical information and often target narrow scenarios. The paper introduces CAB and ReflecTool, which combines long-term memory with tool-wise verification to improve tool use. ReflecTool outperforms existing clinical agents, while benchmark coverage and clinical reliability remain bounded by stated limitations.

  • Problem

    LLMs are restricted to text-based responses, while existing clinical agents typically address specific scenarios with narrow tool ranges, limiting holistic evaluation and diverse information interaction.

  • Method

    The paper builds CAB with 18 tasks across five dimensions and introduces ReflecTool, which stores successful trajectories and tool-wise experience before retrieving and verifying tool usage during inference.

  • Results

    ReflecTool outperforms existing clinical agents on CAB, with Iterative Refinement better for lower-capability models and Candidate Selection better for stronger models.

  • Takeaways & Limitations

    CAB provides a holistic evaluation framework, while ReflecTool demonstrates adaptability across diverse clinical tasks and varying model strengths.

  • Takeaways & Limitations

    CAB may not fully capture the complexity of continuously evolving real-world clinical scenarios, and its collected tools do not perfectly align with evaluation tasks.

Abstract

from arXiv · show

Large Language Models (LLMs) have shown promising potential in the medical domain, assisting with tasks like clinical note generation and patient communication. However, current LLMs are limited to text-based communication, hindering their ability to interact with diverse forms of information in clinical environments. Despite clinical agents succeeding in diverse signal interaction, they are oriented to a single clinical scenario and hence fail for broader applications. To evaluate clinical agents holistically, we propose ClinicalAgent Bench~(CAB), a comprehensive medical agent benchmark consisting of 18 tasks across five key realistic clinical dimensions. Building on this, we introduce ReflecTool, a novel framework that excels at utilizing domain-specific tools within two stages. The first optimization stage progressively enlarges a long-term memory by saving successful solving processes and tool-wise experience of agents in a tiny pre-defined training set. In the following inference stage, ReflecTool can search for supportive successful demonstrations from already built long-term memory to guide the tool selection strategy, and a verifier improves the tool usage according to the tool-wise experience with two verification methods--iterative refinement and candidate selection. Extensive experiments on ClinicalAgent Benchmark demonstrate that ReflecTool surpasses the pure LLMs with more than 10 points and the well-established agent-based methods with 3 points, highlighting its adaptability and effectiveness in solving complex clinical tasks.

1 Introduction

The paper introduces CAB for holistic evaluation of clinical agents and ReflecTool for improving domain-specific tool selection and usage across clinical scenarios.

  • Motivation: Clinical agents can use specialized tools for complex information, but existing systems remain limited to specific scenarios and narrow tool ranges.These limitations hinder interaction with the diverse information found in clinical environments.
  • ClinicalAgent Bench: CAB comprises 18 tasks across five dimensions, providing a holistic benchmark for clinical-agent capabilities across multiple clinical scenarios.The dimensions are Knowledge & Reasoning, MultiModal, Numerical Analysis, Data Understanding, and Trustworthiness.
  • ReflecTool: ReflecTool uses long-term memory and tool-wise verification to improve domain-specific tool selection and usage across diverse clinical scenarios.Successful trajectories and tool-wise suggestions are stored during optimization, then retrieved and applied during inference.
  • Verification methods: Iterative Refinement is more effective for lower-capability models, whereas Candidate Selection performs better for more intelligent models.Both methods are designed to optimize tool usage through tool-wise verification.
  • Results: Experiments on CAB compare ReflecTool with a diverse array of established methods and report superior performance in clinical tool utilization.The supplied passages state the overall result but do not provide the comparison values.

2 ClinicalAgent Bench

CAB is designed as a broad benchmark for clinical agents, covering diverse data formats and capability requirements across five clinical dimensions.

  • Data diversity: The benchmark includes tasks involving medical images, electronic health records, clinical text, and multiple clinical documents.These inputs require agents to integrate information from diverse clinical data sources.
  • Benchmark scope: CAB evaluates clinical agents across diverse medical data and five capability dimensions rather than a single clinical scenario.Its dimensions are Knowledge & Reasoning, MultiModal, Numerical Analysis, Data Understanding, and Trustworthiness.
  • Clinical toolbox: CAB is paired with a toolbox containing 15 tool types, including knowledge databases and calculators, without restricting tool types or counts for a task.This design supports flexible tool use across diverse tasks.

3 REFLECTOOL

REFLECTOOL combines long-term memory with tool-wise verification to improve clinical agents’ selection and use of domain-specific tools. It optimizes from successful and failed trajectories, retrieves similar demonstrations during inference, and applies iterative refinement or candidate selection.

  • REFLECTOOL uses long-term memory and tool-wise verification to help agents select and use domain-specific tools across clinical tasks.Its action space includes pre-built tool actions and inner actions for planning, thinking, and finishing.
  • Optimization Stage: During optimization, the agent reflects on an initial trajectory, regenerates a refined trajectory, and stores successful trajectories and action-wise tool experience.The framework compares successful and failed solving processes to build reusable experience for tool types and inner actions.
  • Inference Stage: During inference, REFLECTOOL retrieves similar successful cases from long-term memory and uses them to guide subsequent actions.Similarity retrieval uses BM25 in the implementation, returning the top-k most similar memory elements.
  • Inference Stage: Tool-wise reflection evaluates each action against stored experience using either iterative refinement or candidate selection.Iterative refinement repeatedly updates an action up to a maximum number of steps, whereas candidate selection samples actions and chooses the most effective one with a verifier.

4 Experiments

Experiments evaluate REFLECTOOL against LLM, MLLM, and agent-based baselines on CAB, showing gains over both base models and a strong reflection-based agent.

  • Baselines: The benchmark compares methods from three families: LLMs, MLLMs, and agent-based systems.The baselines are selected for strong instruction-following capacity because CAB spans diverse task formats.
  • Experimental Setup: The experiments use Qwen2 models with 7B and 72B parameters as REFLECTOOL backbones.Complete results are reported alongside dimension-level averages.
  • Main Results: REFLECTOOL surpasses Reflexion by at least 3 points with both Qwen2-7B and Qwen2-72B.Both methods use self-reflection, highlighting REFLECTOOL's advantage in domain-tool use.
  • Baselines: Qwen2-72B and GPT-4o-mini are the strongest within-group LLM and MLLM baselines, respectively.REFLECTOOL uses Qwen2-7B and Qwen2-72B backbones to examine model-size effects.
  • Main Results: REFLECTOOL exceeds its base models by more than 10 points on Qwen2-7B and Qwen2-72B.The gains hold across the evaluated model sizes.

5 Analysis

Analyses show that optimization, memory, and tool-wise verification shape performance, with Candidate Selection benefiting from memory and Iterative Refinement helping more when memory is absent.

  • 5.1 Effect of Optimization Step: Both verification methods generally improve as optimization steps increase, and Candidate Selection starts around 62.5 versus approximately 58 for Iterative Refinement.Each successful optimization task adds a memory item and tool-wise suggestions.
  • 5.2 Size of Verification methods: Candidate Selection improves steadily with verification size, while Iterative Refinement peaks at n=2 and then declines.Candidate Selection gains exceed 4 points at most as verification size increases.
  • 5.2 Size of Verification methods: Candidate Selection performs better with memory, whereas Iterative Refinement more effectively improves performance without memory.Memory demonstrations improve candidate quality, while refinement substantially helps suboptimal actions when memory is absent.
  • 5.3 Impact of the Long-Term Memory: Long-term memory improves all evaluated agent-based methods, while REFLECTOOL remains ahead of baselines under both few-shot settings.The comparison isolates successful trajectories for baselines rather than tool-wise experience.
  • Tool Selection: REFLECTOOL uses tools more often than ReAct in Knowledge & Reasoning tasks, where higher tool use is associated with better task completion.The tool-distribution analysis excludes Trustworthiness because its tasks overlap with the other dimensions.
  • Tool Selection: REFLECTOOL reduces tool-selection errors at both step and task levels, especially for tools requiring specific databases or uploaded files.Frequent error sources include SQLCoder, DBManual, and LongDocRAG, while similar task contexts can have different input formats.
  • Generalization and Efficiency: REFLECTOOL generalizes across backbones and can reduce runtime relative to Reflexion, with Candidate Selection faster than Iterative Refinement.Candidate Selection benefits from greater parallelism, while runtime comparisons use seconds per sample.

6 Conclusions

The paper introduces CAB and REFLECTOOL to evaluate and improve clinical agents across diverse tasks, reporting superior adaptability and efficacy over existing clinical agents.

  • Conclusions: CAB contains 18 tasks across five clinical dimensions, while REFLECTOOL combines long-term memory with tool-wise verification.The verification choices are Iterative Refinement and Candidate Selection.
  • Conclusions: Empirical results show that REFLECTOOL outperforms existing clinical agents across the proposed evaluation.The conclusion characterizes the framework as adaptable and effective for real-world healthcare scenarios.

Limitations

The benchmark may not capture the full complexity of evolving real-world clinical scenarios, and generated successful trajectories may not be reliably correct or optimal.

  • Limitations: CAB may not fully represent highly diverse and continuously evolving real-world clinical scenarios.The authors state that ongoing benchmark updates are needed to maintain relevance.
  • Limitations: REFLECTOOL's generated trajectories can produce correct results without guaranteeing that the underlying processes are correct or consistently optimal.The authors identify trajectory validation as an area requiring further work.

Ethic Considerations

The paper acknowledges ethical risks from deploying REFLECTOOL in healthcare, especially misleading outputs and hallucinations. It therefore frames the system as an assistive tool requiring supervision rather than a replacement for clinicians or a source of definitive decisions.

  • REFLECTOOL may generate misleading information or hallucinations that pose risks in clinical settings.
  • REFLECTOOL is intended to assist healthcare providers under appropriate supervision, not replace medical professionals or provide definitive clinical decisions.
  • Patient data was anonymized, informed consent was obtained, and data use received ethics-committee approval.

A.1 Medical Agentic Methods

This section reviews medical-agent methods and presents the benchmark resources used to evaluate broad clinical capabilities. It spans knowledge reasoning, multimodal understanding, and diverse clinical datasets and tools.

  • CAB evaluates medical agents across knowledge and reasoning, multimodal understanding, numerical analysis, data understanding, and trustworthiness dimensions.
  • Knowledge&Reasoning: The benchmark uses medical question-answering datasets including PubMedQA, MMLU, BioASQ, and MedQA to assess knowledge and reasoning.
  • MultiModal: Its multimodal evaluation covers open-ended and closed-ended medical-image question answering across 12 modalities using SLAKE, VQA-RAD, and OmniMedVQA.

B.1.3 Numerical Analysis

The numerical-analysis dimension covers both medical calculations and clinical database querying, while the surrounding benchmark resources extend evaluation to clinical data understanding, long-context processing, trustworthiness, and tool implementation.

  • Numerical Analysis: Numerical analysis evaluates equation- and rule-based calculations alongside EHR database operations and understanding.
  • Numerical Analysis: MedCalc contains over 1,000 manually reviewed instances spanning 55 medical calculation tasks with patient notes, questions, answers, and step-by-step explanations.
  • Numerical Analysis: EHRSQL provides 24,411 natural questions from 222 hospital staff and includes answerable and unanswerable queries for trustworthy EHR question answering.
  • Numerical Analysis: MIMIC-III and eICU supply structured critical-care data, including demographics, clinical measurements, medications, diagnoses, and treatment information.
  • The benchmark additionally covers clinical data understanding with MedMentions, emrQA, and LongHealthQA, including long fictional cases of 5,090 to 6,754 words.
  • Its trustworthiness resources test hallucination handling in medical reasoning, visual QA, EHR SQL, and long-document settings.

C.1 Parameter Error in Tool Usage

Tool-use errors include selecting an inappropriate tool and supplying unsuitable parameters to an otherwise appropriate tool. The case study shows reflection can improve query formulation and retrieval effectiveness.

  • Parameter errors occur when an agent selects the correct tool but provides inappropriate inputs, producing unexpected results.
  • The example distinguishes tool selection from tool-wise reflection for improving parameters during tool invocation.
  • An unsuitable knowledge-tool query can retrieve low-relevance information and fail to support the task.
  • In the case study, ReflecTool-Qwen2-72B corrected an initially ineffective Medrag query after reflection and retrieved effective information.

D Experiments Details

The experiments use distinct zero-shot and one-shot settings, restrict multimodal tools to multimodal tasks, and report results across four model types on Clinical Agent Bench. The appendix also documents ReflecTool’s optimization and inference prompts, including tool-wise suggestions and two verifier types.

  • Experimental Setup: Experiments place LLMs and MLLMs in zero-shot settings, agent-based methods in one-shot settings, and multimodal tools only in multimodal tasks.All experiments run on two NVIDIA A100 80GB GPUs.
  • Case Study: A case study compares the reasoning trajectories of two agent-based methods for the same task to examine parameter errors in tool usage.The comparison is presented in Figure 7.
  • Tool-wise Suggestion Generation: Tool-wise suggestion generation compares original and reflective retrial action chains to summarize improved action usage and input-parameter annotations.The prompt gives more comprehensive search queries as an example of better parameter annotation.
  • Verifier Design: Two verifier types are documented for the inference stage, with their prompts shown in Figure 11.The verifier prompt materials are presented alongside the inference-stage prompt documentation.
  • Results Reporting: Table 10 reports four model types on Clinical Agent Bench, marking unavailable task capabilities with “-” and identifying the best result for each task type in bold.The table also distinguishes the COT setting without pre-built tools and marks 4-bit GPTQ-quantized models with “*”.
Loading 2410.17657v3…