PaperlayerView exact revision on arXiv

Source-linked AI summary

Experience Makes Skillful: Enabling Generalizable Medical Agent Reasoning via Self-Evolving Skill Memory

Haoran Sun, Wenjie Li, Yujie Zhang, Zekai Lin, Fanrui Zhang, Kaitao Chen, Xingqi He, Yichen Li, Mianxin Liu, Lei Liu, Yankai Jiang

arXiv:2606.09365v3cs.AIcs.CL

TL;DR

Medical agents need to reuse experience across dynamic, multi-step clinical decisions, but existing memories often retain noisy raw traces without distinguishing useful memories. SkeMex distills interactions into utility-aware skills and continually governs their repository, consistently improving performance across clinical tasks while generalizing across model backbones and task settings.

  • Problem

    Existing medical-agent evaluations and memories inadequately support dynamic, multi-step clinical decision making and distinguish useful experience from noisy historical traces.

  • Method

    SkeMex distills interaction trajectories into structured skills, retrieves them by context-dependent utility, and evolves the repository through a Read–Write–Assess–Govern lifecycle.

  • Results

    SkeMex consistently improves performance across diverse clinical tasks, generalizes across model backbones and task settings, and supports transferable skill memory.

  • Takeaways & Limitations

    SkeMex provides a post-deployment approach for accumulating and reusing clinical experience without updating model weights.

  • Takeaways & Limitations

    The evaluated benchmarks cannot fully capture the complexity and diversity of real clinical environments.

Abstract

from arXiv · show

Medical agent systems are increasingly expected to support interactive clinical decision making rather than only static question answering. In such settings, effective agents must reuse prior experience across evolving cases, yet existing memory mechanisms often retain raw historical traces that are redundant, noisy, and difficult to govern. More importantly, they rarely distinguish which memories are truly useful for future reasoning. This limits their ability to accumulate compact and reliable experience for long-horizon clinical reasoning. To close this gap, we propose SkeMex, a post-deployment self-evolution framework that improves medical agents through a skill-based memory without updating model weights. SkeMex distills informative interaction trajectories into structured skills that encode reusable procedural knowledge, and organizes them into a multi-branch repository spanning general, task-specific, and action-level experience. To determine which memories should be reused and retained, SkeMex estimates context-dependent utility from environment feedback and uses it to guide value-aware retrieval and repository governance. A closed-loop ``Read--Write--Assess--Govern" lifecycle further supports continual evolution by writing new skills, updating utilities, promoting useful memories, and removing harmful entries. Experiments across diverse clinical tasks show that SkeMex consistently outperforms representative memory-based agents in both offline and online settings. It also generalizes across model backbones and supports transferable skill memory. All data and code will be released publicly.

1 Introduction

SkeMex addresses the limits of raw or intra-task memory by converting informative clinical trajectories into structured, reusable skills without updating model weights. Its value-aware Read–Write–Assess–Govern process supports continual memory evolution and improves performance across clinical tasks, model backbones, and heterogeneous settings.

  • 1 Introduction: Medical agents must support multi-step clinical decision making involving interaction, evidence gathering, hypothesis revision, multimodal interpretation, and action adjustment under uncertainty.
  • 1 Introduction: SkeMex converts informative interaction trajectories into structured, actionable skills organized across general reasoning, task-specific knowledge, and action-level experience.
  • 1 Introduction: SkeMex estimates context-dependent memory utility from clinical feedback as a non-parametric reinforcement process guiding both skill retrieval and repository governance.
  • 1 Introduction: The Read–Write–Assess–Govern lifecycle converts trajectories into reusable skills while maintaining a governed memory repository.
  • 1 Introduction: SkeMex consistently improves performance across diverse clinical tasks, generalizes across model backbones, and supports transferable skill memory across heterogeneous task settings.

2 Related Works

Related work spans LLM-based medical agents that reason over heterogeneous clinical evidence and self-evolving memory mechanisms that retain and reuse experience beyond context-window limitations.

  • LLM-based Medical Agents: Medical foundation models such as Lingshu [80], Hulu-Med [19], and MedGemma [49] expand medical reasoning across text, imaging, and multimodal data.
  • LLM-based Medical Agents: Medical agents combine retrieval, tool use, and multi-agent collaboration to handle heterogeneous clinical evidence.
  • LLM-based Medical Agents: i-MedRAG and MedRAG [90] improve medical retrieval through iterative search and knowledge-guided reasoning.
  • Self-Evolving Memory: Agent memory evolved from static buffers and retrieval-augmented generation for context limits into structured stores that summarize and reuse environmental experience.

3 Method

SkeMex models post-deployment medical-agent improvement as a memory-based decision process that retrieves and evolves structured skills instead of updating model parameters. Its Read–Write–Assess–Govern lifecycle organizes value-aware retrieval, trajectory-based skill writing, utility valuation, and repository governance.

  • 3.1 Formulation: SkeMex represents memory units as structured skills with retrieval keys, reusable content, and utility statistics, enabling improvement through memory retrieval and evolution rather than parameter updates [39] [46].The agent retrieves skills before decisions, receives environment feedback, and updates the skills repository after interaction.
  • 3.2 Skills Repository: The skills repository uses general, task-level, and action-level branches to separate transferable reasoning, clinical-task patterns, and operational tool-use knowledge.This organization prevents skills at different abstraction levels from competing during retrieval and valuation.
  • 3.3 Value-Aware Retrieval: SkeMex retrieves skills once at episode onset, routing by clinical category and ranking candidates with semantic similarity, category-conditioned historical utility, and temporally decayed memory strength.Episode-level retrieval provides a stable context, reduces context fragmentation, and limits noise in utility estimation through episodic return.
  • 3.4 Skill Writing: A gated trajectory buffer retains informative multi-step reasoning and failures while filtering infrastructure errors, repetitions, and trivial successes; a two-pass writer distills retained trajectories into reusable skills.Retention scores reward longer reasoning and injected-skill use while penalizing over-represented clinical categories.
  • 3.5 Skill Assessment and Governance: Window-level valuation uses relative advantage across trajectories to update skill utility, while the closed-loop lifecycle coordinates retrieval, writing, assessment, and repository governance over successive task windows.At each window’s end, writing creates or updates skills, valuation updates skills with adoption signals, and governance manages the repository.

4 Experiments and Results

Across nine diverse medical benchmarks, SkeMex is evaluated offline and online against specialist, memory-free, reflection-based, and self-improving memory baselines. It achieves the strongest overall performance, while ablations show that gated skill encoding, adaptive utility valuation, and complementary memory branches are important for repository quality.

  • Offline Mode: SkeMex achieves the best offline performance on both backbones, improving DeepSeek-V3.2 ReAct from 48.20% to 56.08% (+7.88 points) and beating the strongest non-SkeMex memory baseline by 3.84 points.On unseen benchmark families, it gains +13.78 points over ReAct versus about +8.24 points for the strongest competing memory baseline, including a +34.11-point gain on AgentClinic-Text.
  • Online Mode: In online streaming evaluation, SkeMex performs best from epoch@1 and improves from 76.39% to 78.56% by epoch@3, exceeding Evolver’s 76.97%.It maintains stable epoch-wise gains of +0.98 and +1.19 points across text and multimodal settings.
  • Buffer management and skill encoding: Removing buffer gating reduces average performance from 53.22% to 47.56%, while single-prompt encoding and removing draft review reduce it to 50.97% and 48.82%, respectively.These results indicate that trajectory selection and encoding quality are central to preventing noisy or irrelevant experience from corrupting skill extraction.
  • Utility-driven skill valuation: Removing baseline correction causes utility-valuation drops of up to 7.00% on LiveMedBench and 6.09% on MedXpertQA-MM, while fixed learning rates broadly reduce performance.The findings support category-aware reward normalization and adaptive updates for both new and mature skills.
  • Multi-branch memory structure: The full three-branch memory design averages 53.22%, outperforming the best partial General + Action variant by 4.57 points, while two- and single-branch variants lag behind.General, task-level, and action-level branches therefore provide complementary signals, with task-specific medical knowledge remaining necessary alongside general reasoning and action guidance.

5 Conclusion … A.3 Integration of Skill Memory into the Agent Loop

SkeMex enables post-deployment medical-agent improvement through skill-based memory, utility-driven valuation, and closed-loop governance without updating model weights. Its appendix details a bounded, auditable ReAct execution backbone that injects retrieved skills at runtime and feeds behavioral evidence back into skill evolution.

  • 5 Conclusion: SkeMex improves medical agents through reusable skill distillation, utility-driven valuation, and closed-loop memory governance without updating model weights.The framework supports reliable experience accumulation and reuse across diverse clinical tasks, with experiments showing consistent improvements over ReAct and representative memory-based agents.
  • Appendix Contents: The appendix connects SkeMex’s skill memory to a bounded ReAct execution backbone while preserving the core agent loop’s responsibility for task execution.A runtime skill-injection interface links the evolution module to the execution backbone.
  • A.1 Runtime Components: Each task runs through an AgentLoop whose components coordinate model interaction, dataset-specific tools, structured reasoning, memory rendering, and context control.The language-model interface receives system and step prompts, while the tool registry exposes callable tools and renders their availability.
  • A.1 Runtime Components: Each benchmark sample resets conversation and step histories to ensure task independence and prevent information leakage across samples.For vision inputs, images are supplied only on the first model call, while later steps use textual conversation history.
  • A.2 Stepwise Execution Protocol: The structured protocol makes execution deterministic: every turn contains reasoning followed by exactly one tool call or a terminating response.Optional planning is allowed for multistep problems, but the required reasoning-and-action format bounds and audits model behavior.
  • A.2 Stepwise Execution Protocol: Valid tool outputs are parsed and executed, observations are appended to conversation memory, and response outputs terminate tasks while saving complete trajectories.This processing loop supports continued stepwise execution and later trajectory analysis.
  • A.3 Integration of Skill Memory into the Agent Loop: Before each sample, the evolution runner assigns a task category, retrieves relevant skills, and injects their rendered context and identifiers into the agent loop.The injected skill context is rendered throughout the task without replacing the agent policy.
  • A.3 Integration of Skill Memory into the Agent Loop: Recorded reasoning traces, tool calls, observations, and final answers reveal whether injected skills were adopted, ignored, or harmful, enabling subsequent skill evolution.The execution backbone supplies behavioral evidence, while the evolution loop supplies compact task-specific guidance.

A.4 Intra-Task Context Control · B Tool Suite · C Additional Details of SkeMex

SkeMex combines runtime context control, a modular clinical tool suite, and a formalized skill-memory lifecycle to support stable, reusable medical-agent reasoning. These components reduce redundant exploration, preserve relevant evidence, standardize heterogeneous tool use, and govern informative experience over time.

  • A.4 Intra-Task Context Control: Loop breaking detects repeated identical tool actions and prompts the agent to change strategy, use another tool, or answer when sufficient evidence exists.It activates only after a configured number of consecutive repetitions, making the intervention conservative.
  • A.4 Intra-Task Context Control: Confirmed-finding pinning summarizes verified early observations once per task and appends them to later histories, while trimming compresses older content when token use exceeds a configured budget fraction.Recent steps remain fuller, and notices direct the agent toward pinned findings and recent context.
  • A.4 Intra-Task Context Control: SkeMex’s context guard combines loop breaking, confirmed-finding pinning, and budget-aware history trimming to stabilize long medical-agent trajectories without changing the external task interface.The mechanisms respectively reduce redundant exploration, preserve salient early evidence, and control prompt length through runtime context injections.
  • B Tool Suite: The modular tool suite spans six categories: external evidence retrieval, structured medical knowledge, quantitative computation, multimodal perception, reasoning control, and benchmark-specific interaction.All tools receive structured payloads, return compact observations, and expose parameter schemas through a shared execution protocol.
  • B Tool Suite: Clinical tool coverage includes web and medical retrieval, drug information and interaction checks, biomedical concepts, calculators, unit conversion, clinical scores, image analysis, OCR, reflection, and verification.These tools support evidence gathering, medication safety, numerical accuracy, multimodal interpretation, reasoning control, and final-answer auditing.
  • B Tool Suite: Benchmark-specific AgentClinic tools separately provide patient history, objective examinations, and image access while preventing hidden information leakage and separating data retrieval from interpretation.Patient replies exclude hidden diagnoses and objective results, examination requests report unavailable findings rather than inventing them, and image requests return assets without interpretation.
  • C Additional Details of SkeMex: SkeMex represents skills as memory units and writes only informative trajectories, including nontrivial successes, useful failures, and evidence about whether retrieved skills helped or harmed performance.The operational formulation uses an M-MDP with external skill memory, trajectory windows, and a Read–Write–Assess–Govern lifecycle.

D Dataset Details · E Baselines & Metrics

SkeMex is evaluated on nine heterogeneous medical benchmarks spanning interactive decision making, patient-centered reasoning, medical question answering, and multimodal understanding. A unified preprocessing pipeline yields 3,278 samples across 12 dataset configurations while prioritizing difficult or reasoning-intensive cases when metadata permit.

  • D Dataset Details: Nine benchmarks cover interactive clinical decision making, patient-centered reasoning, knowledge-intensive question answering, and multimodal medical understanding across heterogeneous formats and evaluation protocols.The evaluation tests both text-only and multimodal reasoning in varied clinical settings.
  • D Dataset Details: AgentClinic [48] evaluates sequential information gathering and image-supported clinical reasoning through patient questioning, examinations, and multimodal evidence.Both text-only and multimodal settings are used.
  • D Dataset Details: LiveClin [67] uses continually updated, peer-reviewed case reports to test reasoning on recent, clinically grounded, and often complex text-only and multimodal cases.Its design aims to reduce data leakage and knowledge obsolescence while representing realistic clinical pathways.
  • D Dataset Details: The benchmark suite also evaluates patient-centered workflows, rubric-based clinical completeness, physician-authored health responses, information-seeking under uncertainty, and expert-level specialty reasoning.These dimensions are represented by MedJourney, LiveMedBench, HealthBench [3], MediQ [30], and MedXpertQA [96], respectively.
  • D Dataset Details: A unified preprocessing pipeline prioritizes difficult or reasoning-intensive samples when official difficulty labels, metadata, or subset annotations are available.The pipeline is intended to make skill accumulation informative, improve evaluation reliability, and reduce unnecessary experimental cost.
  • D Dataset Details: After preprocessing, the evaluation pool contains 3,278 samples across 12 dataset configurations derived from the nine benchmarks.AgentClinic, LiveClin, and MedXpertQA are split into text-only and multimodal configurations, while MMMU and MMMU-Pro are restricted to Health & Medicine subsets.

E.1 Baselines · E.2 Metrics

The evaluation compares SkeMex against specialist models and memory-based agents spanning no-memory, reflective, experience-distillation, skill, persistent, and clinical-memory paradigms. Metrics are dataset-specific, with deterministic accuracy, semantic-equivalence, or rubric-based scoring also supplying rewards for utility updates.

  • E.1 Baselines: Vanilla ReAct matches SkeMex’s prompts and tools while removing skill memory, retrieval, utility estimation, and governance, isolating the contribution of evolving skill memory.Lingshu [80], Hulu-Med [19], and MedGemma [49] provide specialist references for medical and multimodal reasoning.
  • E.1 Baselines: Baselines span specialist models and memory agents, including no-memory, reflective, experience-distillation, skill/workflow, persistent, and graph-structured clinical-memory methods.This taxonomy includes Vanilla ReAct, Reflexion, CRITIC, Voyager, DILU, ExPeL, GenerativeMemory, Memp, SkillWeaver, AgentWorkflowMemory, AgentKB, Evolver, DynamicCheatsheet, MobileE, and CerebraFusionMemory.
  • E.2 Metrics: Metrics follow each benchmark’s answer format: standard close-ended tasks primarily use accuracy over successfully evaluated samples.Correctness is represented by a per-sample binary indicator, with failures such as missing inputs or invalid records excluded from the evaluated sample set.
  • E.2 Metrics: Multiple-choice predictions and references are normalized to option letters, while unparseable responses count as incorrect rather than being removed from the denominator.Single-answer questions use exact option matching after normalizing formats such as A, (A), [A], A., and A :.
  • E.2 Metrics: Open-ended AgentClinic Text and MedJourney examples use a semantic-equivalence judge that converts reference-prediction comparisons into binary accuracy verdicts.The judge determines whether a prediction is correct or semantically equivalent to the reference answer.
  • E.2 Metrics: HealthBench and LiveMedBench use weighted rubric-based scoring for free-form clinical responses, including penalties for unsafe advice, unsupported claims, and clinically inappropriate content.Gemini-3-Flash evaluates all rubric criteria for each sample in one call; HealthBench reports mean sample score and LiveMedBench mean case score.
  • E.2 Metrics: The same scoring functions convert outcomes into rewards for memory evolution, aligning reported benchmark metrics with Assess-stage utility updates.Accuracy datasets provide binary rewards, whereas rubric-based datasets provide continuous rewards in [0, 1].

F Implementation Details

SkeMex is evaluated with held-out in-domain and entirely unseen out-of-domain test data, using a shared primary backbone and a fixed ReAct-style execution and retrieval setup. Its skill memory is updated through windowed experience filtering, utility valuation, and periodic repository governance.

  • F Implementation Details: In-domain samples are split approximately equally for experience accumulation and held-out testing, while out-of-domain benchmarks reserve all processed samples for testing.This protocol evaluates both performance on related held-out cases and transfer to benchmark families excluded from repository construction.
  • F Implementation Details: DeepSeek-V3.2 [32] serves as the primary backbone for reasoning, skill distillation, classification, utility judgment, and repository governance, while Qwen3.6-Plus tests transfer using the unchanged repository.The cross-model setting evaluates whether skills accumulated by DeepSeek-V3.2 can be reused without re-distillation or repository rewriting.
  • F Implementation Details: Each episode follows a bounded ReAct-style trajectory in which the agent receives tools and retrieved skills, then alternates between reasoning, structured tool calls, and final answering.Tool outputs become later-step observations and are also available to trajectory-to-skill distillation.
  • F Implementation Details: Retrieval occurs once per episode with a default budget of K = 6, combining category-aware routing, semantic similarity, utility, memory strength, maturity bonuses, and branch-aware selection.The system uses a minimum similarity threshold of 0.2 and retrieves across general, task-level, and action-level branches; pre-screening activates above 5 candidates and considers up to 5 per branch.
  • F Implementation Details: Experience learning uses 30-trajectory windows with a 20-trajectory retained capacity, filters low-reuse traces, values skills using normalized rewards and adoption-aware credit, and governs the repository every 2 windows.Governance merges similar skills, deprecates low-utility entries, promotes stable high-utility skills, and enforces branch capacities; mature status requires utility at least 0.75 and usage count at least 15.

G Sensitivity Analysis

SkeMex remains stable across broad hyperparameter ranges on HealthBench and LiveMedBench, with only modest degradation from undersized retrieval, memory-heavy weighting, or overly aggressive utility updates. The default configuration provides a balanced operating point across retrieval, learning-window, baseline, and utility-update choices.

  • Retrieval budget: 42.93 average score at K = 9 exceeds 42.80 at K = 6, while increasing retrieval to K = 12 lowers performance to 42.70.Performance rises from 42.52 at K = 3, suggesting too few skills provide insufficient guidance whereas too many add redundant or weakly relevant information.
  • Retrieval weights: 42.85 average score with increased semantic-similarity weighting slightly exceeds the default 42.80, whereas memory-heavy weighting reduces it to 42.57.Increasing utility weighting gives 42.79 and the best HealthBench score among tested settings, indicating semantic relevance and estimated utility are more reliable than memory strength alone for retrieval.
  • Learning window size: Average scores across learning-window sizes L = 10, 20, 30, 40, and 60 are 42.78, 42.82, 42.80, 42.78, and 42.64, showing no monotonic trend.The optimal window depends on the amount and distribution of training data; smaller windows react faster to recent feedback.
  • Category baseline update coefficient: 42.86 average score at α = 0.10 exceeds the default 42.80, while larger baseline-update coefficients produce slightly lower scores.The degradation remains limited, although overly responsive category baselines may track short-term variation too closely.
  • Utility update step: 42.83 average score at ηmax = 0.30 is close to the default 42.80, but increasing the maximum update step to 0.40 drops performance to 42.54.Excessively large utility updates can amplify short-term feedback noise and destabilize skill valuation.
  • Overall sensitivity: SkeMex remains stable across a broad range of hyperparameter choices, with modest variation even when retrieval is too small, memory strength is overemphasized, or utility updates are too aggressive.The default configuration balances retrieval coverage, retrieval precision, and adaptation speed.

H Ablation Study · I Further Analyses · I.1 Offline OOD Generalization

SkeMex’s value-aware retrieval and closed-loop repository governance each materially improve performance, while its frozen skill repository transfers effectively to unseen benchmark families across model backbones. Offline OOD results show consistent gains over both memory-free and competing memory-based baselines.

  • H Ablation Study: The full value-aware retrieval model reaches a 53.22% average score, outperforming memory-strength removal at 50.76% and LLM-only ranking at 50.93%.The result indicates that effective skill selection requires more than surface-level semantic matching.
  • H Ablation Study: Using only similarity lowers the average to 47.64%, while using only utility reaches 48.30%, showing that relevance and historical usefulness capture complementary aspects of skill value.Similarity can retrieve contextually mismatched skills, whereas utility alone can favor overly general procedures.
  • H Ablation Study: Prescreening improves retrieval quality: removing it reduces the average score to 50.69% by allowing irrelevant or weakly matched candidates to reach final ranking.Early filtering becomes increasingly important as the skill repository grows and noisy candidates accumulate.
  • H Ablation Study: The full closed-loop lifecycle reaches a 53.22% average score, while removing maturation causes the largest drop, to 47.60%, demonstrating the importance of repository governance after skill creation.Maturation produces a 5.62-point decline when removed.
  • H Ablation Study: Removing deprecation lowers the average to 49.27%, and disabling memory merging lowers it to 49.67%, showing that forgetting harmful skills and reducing redundancy support reliable retrieval.Obsolete or low-utility entries can introduce noise, while redundant skills increase repository storage and retrieval burden.
  • H Ablation Study: Removing capacity limits reduces the average score to 51.05%, indicating that unconstrained repository growth weakens retrieval quality even when other lifecycle operations remain active.Branch-wise capacity regulation limits low-value candidates as skills accumulate.
  • I.1 Offline OOD Generalization: SkeMex achieves the best offline OOD average performance on both backbones, improving ReAct by 13.78 points to 75.79% with DeepSeek-V3.2 and by 13.91 points to 78.18% with Qwen3.6-Plus.These gains exceed the strongest competing memory baseline by 5.44 and 2.97 points, respectively, demonstrating transfer across model families.
  • I.1 Offline OOD Generalization: SkeMex attains the best score on four of five OOD benchmarks with DeepSeek-V3.2 and all five with Qwen3.6-Plus, including MediQ, AgentClinic-Text, MMMU-Pro, and AgentClinic-MM.Its largest reported DeepSeek-V3.2 gains over ReAct are +34.11 points on AgentClinic-Text and +12.50 points on MMMU-Pro.

I.2 Cross Backbone Generalization · I.3 Cross Backbone Skill Transfer · I.4 Execution Cost and Interaction Depth

SkeMex generalizes across backbone models and transfers fixed skill repositories across architectures, achieving strong benchmark gains. This broader capability comes with greater interaction depth and runtime on average, though relevant skills can sometimes streamline execution.

  • I.2 Cross Backbone Generalization: SkeMex remains strongest across Qwen3.6-Max-Preview, Kimi-2.6, and GLM-5.1, including best results on all six benchmarks for the first two backbones.On GLM-5.1, it is best on five benchmarks, ties on AgentClinic-MM, and produces nonnegative gains across all six datasets.
  • I.2 Cross Backbone Generalization: Across 18 backbone–dataset pairs, SkeMex is best in 17 and tied once, raising pooled average performance from 48.88% to 59.13% over ReAct.The pooled improvement is 10.25 points, with positive average gains across the reported datasets.
  • I.3 Cross Backbone Skill Transfer: A fixed skill repository learned with DeepSeek-V3.2 transfers to Claude Sonnet-4.6, where SkeMex leads all seven benchmarks and improves the ReAct average from 49.59% to 60.27%.The average gain is 10.68 points, and SkeMex exceeds the strongest competing memory baseline, CFM, by 3.31 points on average.
  • I.3 Cross Backbone Skill Transfer: On Qwen3.6-35B-A3B, the transferred repository raises the ReAct average from 47.71% to 58.23% and exceeds CFM by 2.93 points on average.SkeMex ranks first on MedXpertQA-Text, HealthBench, LiveMedBench, LiveClin-MM, and MMMU, while trailing specified baselines on two datasets.
  • I.3 Cross Backbone Skill Transfer: The transfer results indicate that SkeMex stores reusable clinical procedures rather than cached answers or backbone-specific response traces.The repository is produced by DeepSeek-V3.2 but remains beneficial for both a closed-source frontier model and an open-source mixture-of-experts model.
  • I.4 Execution Cost and Interaction Depth: SkeMex uses 4.77 interaction steps on average, exceeding ReAct’s 3.17 and other memory-based methods’ 3.97–4.61 steps.Retrieved skills encourage task decomposition, intermediate verification, tool use when necessary, and avoidance of premature final answers.
  • I.4 Execution Cost and Interaction Depth: SkeMex averages 116.06 seconds per task versus 54.48 for ReAct and 81.65 for Evolver, but relevant skills can reduce HealthBench time to 52.36 seconds.Average time per step is 24.33 seconds for SkeMex, compared with 17.19 for ReAct and 20.41 for Evolver; retrieval and deeper reasoning contribute to overhead.

I.5 Impact of Training Data Order

Training-data order affects SkeMex’s in-domain skill-evolution performance on HealthBench and LiveMedBench. Random ordering performs best overall, while Hard-to-Easy ordering performs worst, suggesting that early repository updates are sensitive to trajectory mix and quality.

  • I.5 Impact of Training Data Order: The experiment evolves SkeMex’s repository on each benchmark’s training split and evaluates it on the corresponding held-out in-domain test split.HealthBench and LiveMedBench use rubric-based sample-level scores as a proxy for task difficulty.
  • I.5 Impact of Training Data Order: 45.03% average score: random ordering achieves the best overall performance across HealthBench and LiveMedBench.The result is attributed to a mixed stream of categories and difficulty levels that balances trajectories across learning windows.
  • I.5 Impact of Training Data Order: 41.41% average score: Hard-to-Easy ordering performs worst among the evaluated training-order strategies.Processing difficult cases first may expose early skill writing to noisy trajectories, incomplete reasoning, or overly specific error patterns that persist through later retrieval and governance.

J Limitations

SkeMex’s evaluation does not fully capture complex real-world clinical environments, and its skill-based reasoning incurs higher API usage and longer runtime. The framework also requires human oversight because it could reinforce errors or be misused in high-stakes settings.

  • The evaluated benchmarks cannot fully represent real clinical environments with diverse patient histories, institutional workflows, and decision constraints.
  • Skill-based reasoning increases API usage and wall-clock time compared with memory-free agents, reflecting an efficiency trade-off for structured, reliable reasoning.
  • Without sufficient human oversight, the framework could reinforce incorrect patterns or be misused in high-stakes settings, so it should support rather than replace professional medical judgment.

K Case Study … ANALYSIS_PROMPT_RUBRIC

The case studies show how retrieved skills guide clinical reasoning, tool selection, diagnostic verification, and procedure choice, while exposing failure from interaction-budget waste. The prompt appendix specifies agent interaction, retrieval, evaluation, attribution, mutation, review, and skill-merging procedures.

  • K Case Study: Across five cases, retrieved skills support successful reasoning in four scenarios, while one failure exposes insufficient diagnostic specificity from premature convergence.The figures illustrate skills shaping agent trajectories through injected repository experience.
  • K Case Study: Skills adapt reasoning by avoiding uninformative searches, selecting appropriate tools, verifying diagnostic boundaries, and comparing procedures before stopping retrieval.Examples include structured differential diagnosis, administrative-tool routing, Brunt-based cirrhosis verification, and multimodal biopsy-procedure selection.
  • K Case Study: The failure case shows that two early tool-format errors consume the interaction budget, leaving insufficient steps to confirm the specific lymphoma subtype.The ground truth is Diffuse Large B-Cell Lymphoma, despite an initially correct vignette-first differential workflow.
  • L Prompts: The core prompts enforce planning, reasoning, tool, and response formats, inject conversation history and retrieved skills, and force final answers at the step limit.Retrieved skills are grouped into general, task-level, and action-level branches.
  • L Prompts: Memory-evolution prompts classify medical queries into action-oriented categories and pre-screen semantically relevant skills before top-ranked candidates enter the scoring pipeline.The classifier supports methodology-level retrieval, while pre-screening directly affects which skills are injected into the agent.
  • ANALYSIS_PROMPT_BINARY: Binary-outcome trajectory analysis evaluates skill adoption and extracts decisive success or failure patterns for subsequent memory evolution.The prompt frames trajectory attribution as a medical knowledge-engineering task.
  • ANALYSIS_PROMPT_RUBRIC: Rubric-based analysis attributes score deductions to agent actions, then mutation drafts or patches skills, review checks novelty and anti-fragmentation, and merging consolidates overlaps.These prompts form a pipeline from trajectory attribution through skill creation, governance, and concise consolidation.
  • L Prompts: Automatic-evaluation prompts provide rubric-based grading for HealthBench and interactive clinical assessment for LiveMedBench using the query, prediction, and evaluation criteria.Separate system prompts tailor grading to each benchmark.
Loading 2606.09365v3…