ClinSeekAgent: Autonomous Clinical Reasoning

• ClinSeekAgent is an autonomous, modular framework for clinical reasoning that dynamically integrates multimodal evidence from EHRs, imaging, and knowledge bases.
• It formalizes clinical reasoning as a Markov Decision Process, enabling iterative tool calls and evidence synthesis for optimized diagnostic decisions.
• ClinSeekAgent demonstrates significant performance gains over baseline models, validated through robust, real-world benchmarks and diverse multimodal tasks.

ClinSeekAgent is an autonomous, modular framework for clinical reasoning that enables LLMs and related agentic systems to actively seek, plan, and synthesize multimodal evidence for clinical decision support. Unlike prior approaches assuming curated, preselected context, ClinSeekAgent orchestrates tool calls over heterogeneous sources—including raw electronic health records (EHRs), medical imaging, and knowledge bases—to dynamically acquire and integrate evidence, optimizing for grounded decisions in complex clinical workflows. It further supports both inference-time operation and training-time policy distillation, and has demonstrated demonstrable improvements in real-world bench evaluations over conventional, context-fed LLM baselines (Wu et al., 19 May 2026).

1. Agentic Problem Formulation for Clinical Reasoning

ClinSeekAgent formalizes the clinical reasoning task as a sequential decision process, specifically a Markov Decision Process (MDP), where the agent iteratively selects tool actions aimed at increasing the fidelity of its final output. Formally, each task instance is $x = (p, t, q, \mathcal{M}, \mathcal{Y})$, where $p$ is a patient identifier, $t$ specifies a temporal index (e.g., cutoff date), $q$ is the clinical query or instruction, $\mathcal{M}$ encodes optional modality metadata (e.g., image identifiers), and $\mathcal{Y}$ specifies the desired output schema or label set. At each round $k$, the state $s_k$ is comprised of the task plus an action-observation history $h_{k-1} = \{(a_1, o_1), ..., (a_{k-1}, o_{k-1})\}$. The action space $A$ consists of tool calls (e.g., ehr.run_sql_query, browser.search, image.chest_xray_classifier) and a Finish action that ends the episode with prediction $p$0. Rewards are sparse, typically $p$1. The objective is to optimize the agent policy $p$2 to maximize expected reward $p$3, with practical surrogate objectives such as maximizing log-likelihood of expert agent trajectories during supervised policy distillation (Wu et al., 19 May 2026).

2. System Architecture and Tool-Calling Pipeline

The ClinSeekAgent architecture comprises three principal modules: (A) Hypothesis Generator, (B) Evidence Retriever (Tool-Calling Engine), and (C) Multimodal Integrator & Answerer. The pipeline operates as follows:

• Initialization triggers EHR loading and model state setup.
• At each decision round, the agent consults its policy to determine the next tool call, leveraging available history.
• Tool calls are mapped from natural-language intents or sub-questions to API-level arguments (20 such tools are integrated, including EHR SQL wrappers, browser search, web crawling, and medical image processing pipelines).
• Each executed tool returns a structured observation (e.g., a SQL-result table, CXR classifier vector, or web snippet), appended to history.
• The answerer module fuses all accumulated observations into a final decision, with optional early termination when sufficient evidence has been gathered. These modules are tightly orchestrated; there is no forced depth-first or breadth-first traversal of modalities—the agent actively decides the evidence acquisition trajectory (Wu et al., 19 May 2026).

3. Multimodal Evidence Acquisition and Integration

ClinSeekAgent uniquely operates in settings where no curated “given” evidence is provided; instead, the agent is responsible for discovering and integrating all raw data relevant to the clinical query. The agent may:

• Enumerate and query available EHR tables via API (schema discovery, time-windowed queries, ad-hoc SQL).
• Invoke pipelines for medical image feature extraction and classification if images are referenced in $p$4.
• Use external browser search tools for guideline or knowledge-base lookup when local EHR data is incomplete.
• Iteratively revise its working hypotheses and future tool calls based on each new evidence fragment received. Multimodal integrator modules construct unified prompts ingesting all structured, textual, or visual outputs for holistic inference (Wu et al., 19 May 2026). This architecture generalizes prior condition-specific agents, such as cohort-aware model selectors (Qu et al., 20 Aug 2025), rule-based execution frameworks (2505.23055), and GUI/interaction agents (Yu et al., 2 Jun 2026).

4. Training-Time Distillation and Model Specialization

ClinSeekAgent functions both as an inference-time orchestrator for strong LLMs (e.g., Claude Opus 4.6) and as a pipeline for policy distillation into open, efficient models. Agent-generated trajectories are harvested during simulated, full-agent runs; each trajectory records the full action-observation stream interleaved with natural language and API responses. These are then used to fine-tune smaller LLMs (e.g., Qwen3.5-35B-A3B), using series-wise cross-entropy loss over up to 52k-token input lengths. The distilled model (ClinSeek-35B-A3B) thereby learns both tool-use reasoning and answer generation (Wu et al., 19 May 2026). Policy imitation at scale enables deployment in resource-constrained environments and reproducible comparison on challenging agentic benchmarks such as AgentEHR-Bench.

5. Evaluation Protocols and Empirical Results

ClinSeekAgent is validated on the ClinSeek-Bench testbed, which systematically compares active evidence-seeking agent performance to “curated input” baselines where the correct input context is provided directly. Datasets include text-only EHR benchmarks (1,800 cases; 45 subtasks), and multimodal tasks requiring CXR presence detection, temporal change analysis, phenotype extraction, and more (989 cases). Evaluation metrics are averaged sample-wise F1, computed per task group and then aggregated. Key findings:

• On text-only tasks, pairing ClinSeekAgent with Claude Opus 4.6 yields overall F1 improvement from 60.0 (curated) to 63.2 (agentic); MiniMax M2.5 sees F1 rise from 43.1 to 47.3.
• On multimodal tasks, ClinSeekAgent raises Claude Opus 4.6 F1 from 47.5 to 62.6 (+15.1); all evaluated models improve on CXR-specific subtasks.
• Distillation with ClinSeek-35B-A3B achieves average F1=34.0 on AgentEHR-Bench, outperforming Qwen3.5-35B-A3B baseline by +11.9 and approaching closed frontier model performance (Wu et al., 19 May 2026). Bootstrap-derived 95% confidence intervals for F1 are typically within ±2–3 points, supporting statistical robustness.

6. Design Extensions, Limitations, and Future Directions

ClinSeekAgent’s modular and tool-agnostic framework supports broad extensibility, including:

• Integration with advanced retrieval-augmented model selectors for patient-level individualized risk (e.g., two-stage FAISS/LLM agents for lung cancer) (Qu et al., 20 Aug 2025).
• Incorporation of rule-based execution modules, as in CDR-Agent, for precise application of clinical decision rules in emergency settings (2505.23055).
• Extension to GUI automation and screenshot-only agents, leveraging deterministic reward shaping and paired goal protocols as in MedCUA-Bench (Yu et al., 2 Jun 2026).

Noted failure modes include occasional degradation in sequential decision tasks due to over-retrieval of irrelevant or distracting signals, limited ability of student models to filter redundant tool use, and potential inefficiency in long action-trajectory imitation. A plausible implication is that trajectory compression, reward-shaping, and post-supervised-fine-tuning reinforcement learning may further improve agentic efficiency, especially for long-horizon, cross-modal clinical reasoning. Future benchmarks emphasizing deeper interleaved chain-of-thought, latent state tracking, and generalization across unseen EHR and imaging schemas are proposed (Wu et al., 19 May 2026).

ClinSeekAgent represents a paradigm shift from passive LLM consumption of static evidence to proactive, agent-mediated evidence acquisition and integration, with demonstrated performance gains and a modular blueprint for continually extensible, audit-ready clinical AI systems.