Do Clinical Models Change Treatment Decisions?
Dongkyu Cho, Miao Zhang, Rumi Chunara
Abstract
Clinical foundation models are evaluated with factual or exam-style medical QA, but treatment decisions must change when patient context changes. We introduce ClinPivot, an auditable treatment-decision benchmark built from biomedical relations and pivoted patient contexts. ClinPivot asks whether models change treatment choices when new clinical constraints shift the action space. We find that strong medical QA performance does not reliably predict decision-making performance: frontier models and task-adapted Qwen variants often fail to change decisions correctly, and model rankings shift across evaluation regimes. Decision-structured supervision improves pivot-sensitive decision-making and medical QA under matched knowledge budgets, while lightweight replay reduces losses in general assistant ability.
AI Impact Assessments
(1 models)Scientific Impact Assessment: "Do Clinical Models Change Treatment Decisions?"
1. Core Contribution
The paper introduces ClinPivot, a benchmark that evaluates whether language models can update treatment decisions when patient-specific constraints (comorbidities, drug interactions, allergies, contraindications, off-label temptations) change the clinical action space. The key insight is simple but important: knowing that Drug A treats Disease D is insufficient—a model must also recognize when Drug A becomes inappropriate due to new patient context. The benchmark is constructed programmatically from PrimeKG biomedical knowledge graph relations rather than from LLM-generated labels, making it auditable and transparent in its gold-label derivation.
The paper also contributes a finding that decision-structured supervision (training on pivot-style examples) outperforms QA-style supervision even for standard medical QA benchmarks, and that experience replay mitigates catastrophic forgetting of general capabilities during clinical fine-tuning.
2. Methodological Rigor
Strengths in design: The benchmark construction pipeline is well-documented. Gold labels derive from graph operations (contraindication edges, interaction edges, etc.), not from another LLM, which avoids circular evaluation. The disease-node-based train/test split prevents leakage. The LLM consistency screen is reject-only—it cannot create or change labels, only remove incoherent examples. Five random seeds are used for fine-tuning experiments with standard errors reported.
Concerns: The methodology has notable limitations that the authors partially acknowledge. First, the "gold" treatments are graph-derived, not clinically validated. PrimeKG relations can be incomplete, noisy, or outdated, meaning some "failures" may reflect models correctly reasoning beyond the graph's knowledge. Second, the vignettes are synthetic and relatively formulaic—they may not capture the ambiguity and complexity of real clinical scenarios. Third, the pivot sensitivity metric only checks whether the model avoids the banned drug, which is a necessary but not sufficient condition for good clinical reasoning. Fourth, frontier model evaluations are point estimates without confidence intervals, making the ranking comparisons less robust. The gap between MedQA and ClinPivot (29.97 points on average) is striking but could partly reflect task format differences rather than purely reasoning deficits.
The fine-tuning experiments are reasonably controlled but limited: only Qwen-3 models at two sizes, with a single learning rate and no hyperparameter search. The claim that "decision-structured supervision improves medical QA under matched knowledge budgets" is interesting but the mechanism is unclear—is it the reasoning structure or simply more diverse training signal?
3. Potential Impact
Direct impact: ClinPivot addresses a real gap in clinical AI evaluation. Current benchmarks (MedQA, PubMedQA) test factual recall, not conditional decision-making. If adopted, ClinPivot could become a useful complement to existing evaluation suites, particularly for teams building clinical decision support tools.
Broader implications: The finding that model rankings shift between QA and decision-making regimes is practically important for model selection in healthcare applications. The observation that decision-structured training transfers to QA performance suggests potential for curriculum design in clinical model training.
Limitations on impact: The benchmark remains narrow—it tests only treatment selection from a fixed candidate set, excluding diagnosis, dosing, monitoring, shared decision-making, and temporal reasoning. The synthetic nature of vignettes limits ecological validity. The gap between graph-derived constraints and real clinical guidelines is significant: a clinician would rarely make treatment decisions based solely on the presence/absence of contraindication edges without considering severity, dosing adjustments, or risk-benefit tradeoffs.
4. Timeliness & Relevance
The paper is timely. There is growing recognition that medical QA benchmarks are insufficient proxies for clinical competence (Griot et al., 2025; Kim et al., 2025; Jiang et al., 2025). The FDA and healthcare regulators are increasingly scrutinizing AI systems for clinical decision support. Having benchmarks that test conditional reasoning rather than factual recall is relevant to safety evaluation. However, several concurrent efforts address related gaps (clinical reasoning benchmarks, counterfactual medical evaluation), so the novelty window is competitive.
5. Strengths & Limitations
Key strengths:
Notable weaknesses:
Missing analysis: Error analysis by pivot type would be informative—are models worse at drug interactions than allergies? Do models fail more on "change when constrained" vs. "don't change for wrong reason" cases? The paper presents aggregate numbers but limited diagnostic insight.
Overall Assessment
This is a competent contribution that identifies a real evaluation gap and provides a reasonable first benchmark to address it. The core finding—that QA performance doesn't predict decision-making performance—is important for the clinical AI community, though perhaps unsurprising. The benchmark design is sound but limited in ecological validity. The training experiments provide useful practical guidance but lack depth in analysis. The paper would benefit from expert clinical validation, error analysis, and broader model coverage in fine-tuning experiments.
Generated May 28, 2026
Comparison History (17)
PEAM introduces a comprehensive novel framework addressing multiple fundamental challenges in embodied AI: continual learning without catastrophic forgetting, learning from failures via contrastive objectives, and self-triggered memory consolidation. Its contributions span architecture design (MoE-LoRA), training methodology (failure-correction contrastive learning), and autonomous learning mechanisms. Paper 2, while addressing an important evaluation gap in clinical AI, is primarily a benchmark and evaluation study with more incremental contributions. PEAM's broader methodological innovations have greater potential to influence multiple research directions in embodied agents and continual learning.
Paper 1 likely has higher impact because it targets a broadly relevant, under-measured variable—agent “harness” effects—across many models and realistic tool-using workflows, providing a sizable, instrumented benchmark with traces to diagnose failure modes. This can influence evaluation standards, reporting practices, and system design across the fast-growing LLM agent ecosystem (software engineering, productivity, robotics-like tool use). Paper 2 is timely and important for clinical NLP, but its scope is narrower and more domain-specific, with real-world deployment constrained by regulation and data access.
Paper 2 likely has higher scientific impact due to its direct clinical relevance and broad real-world applicability: it introduces ClinPivot, a decision-focused benchmark that tests context-sensitive treatment changes, exposing a key gap between medical QA and actionable decision-making. The benchmark is auditable and grounded in biomedical relations, supporting methodological rigor and reproducibility. Its findings affect evaluation practice across clinical AI, foundation model alignment, and safety, and it proposes practical training interventions (decision-structured supervision, replay) with implications for deployment and regulation. Paper 1 is innovative but narrower in application.
Paper 2 addresses a critical gap in a high-stakes domain (healthcare) by introducing a dynamic benchmark for clinical decision-making. While Paper 1 offers timely insights into LLM reasoning mechanics, Paper 2 has higher potential impact because it challenges the standard evaluation paradigm (static QA) in medical AI, directly impacting patient safety, deployment policies, and future clinical model evaluation. Benchmarks in critical safety fields often drive broader methodological shifts and widespread adoption, giving it stronger real-world applicability.
Paper 1 introduces a novel, concrete benchmark (ClinPivot) addressing a critical gap between medical QA performance and actual clinical decision-making, with empirical findings showing strong QA models fail at context-sensitive treatment decisions. This has immediate practical implications for clinical AI evaluation and model development. Paper 2 proposes a governance framework (OADA) that, while addressing important deployment concerns, is primarily conceptual and incremental over existing governance literature, with limited empirical validation. Paper 1's methodological rigor, novelty of the evaluation paradigm, and direct relevance to patient safety give it broader and more lasting impact.
While both propose valuable benchmarks, Paper 1 addresses a critical safety and efficacy gap in clinical AI—ensuring models adapt treatment decisions to shifting patient contexts rather than just reciting medical knowledge. This has immediate, high-stakes implications for real-world healthcare deployment, patient safety, and medical model evaluation, giving it a broader and more critical societal impact than the specialized spatial biology focus of Paper 2.
Paper 1 addresses a critical safety gap in the high-stakes domain of clinical AI by evaluating how models adapt to changing patient contexts. Its focus on actual treatment decisions over static medical QA has immediate, life-saving implications for healthcare deployment, giving it higher potential impact than the broader, but less critical, evaluation of emotional intelligence in Paper 2.
Paper 1 addresses a critical gap in high-stakes clinical AI by evaluating whether models dynamically adapt treatment decisions to changing patient contexts. Improving AI reliability in healthcare has profound real-world implications for patient safety, medical decision support, and AI safety evaluations. Paper 2's focus on e-commerce dispute resolution, while innovative for multi-agent systems, applies to a narrower, lower-stakes commercial domain. The medical focus and safety implications of Paper 1 give it greater potential for broad scientific and societal impact.
Paper 2 likely has higher scientific impact due to clearer, high-stakes real-world applicability (clinical treatment decisions) and strong timeliness as clinical foundation models move toward deployment. ClinPivot targets a core failure mode—context-dependent decision changes—beyond exam-style QA, offering an auditable benchmark and actionable training insights (decision-structured supervision, replay) that can directly influence model development and evaluation practices in medicine and beyond. Paper 1 is novel and important for agentic safety, but its simulation-based findings may generalize less directly to regulated deployment settings compared to clinically grounded decision benchmarks.
Paper 1 offers a concrete, technically novel solution to a well-defined agent failure mode (action-grammar destruction) and demonstrates large, robust gains across many environment/backbone/method settings, with clear ablations and an efficient inference-free design (real-world deployability at scale). Its contributions (step-level compression, structural floors, counterfactual action-change labels) are broadly applicable to LLM agents, memory, and systems. Paper 2 introduces a valuable benchmark and insight for clinical decision robustness, but impact may be narrower (evaluation-centric) and constrained by clinical deployment barriers and data/validation requirements.
Paper 2 introduces a broader paradigm shift—learning to learn from multimodal experience with adaptive memory design—that has wider applicability across AI agents, multimodal reasoning, and meta-learning. Its framework addresses a fundamental challenge (how to structure experience for learning) relevant across many domains. Paper 1, while addressing an important clinical AI evaluation gap with ClinPivot, is more narrowly scoped to medical decision-making benchmarking. Paper 2's contributions to adaptive memory mechanisms and multimodal agent learning have greater potential to influence multiple research communities.
Paper 2 likely has higher scientific impact due to its direct clinical relevance and a clearer path to real-world deployment: it introduces an auditable benchmark (ClinPivot) targeting a critical failure mode—context-sensitive treatment decision changes—that current medical QA metrics miss. The findings can influence evaluation standards, model training protocols (decision-structured supervision), and safety/regulatory discussions across healthcare AI. While Paper 1 is theoretically interesting and may advance graph+LLM methods, its applications are more niche and its impact depends on broader adoption of the proposed interpretation. Paper 2 is timely and broadly relevant.
Paper 1 has higher estimated impact due to a more novel and broadly applicable contribution: it formalizes “relevance-sensitive” evaluation (should-change vs should-not-change) and proposes LexGuard, a solver-grounded, adversarial multi-agent framework that integrates formal statute constraints and SMT verification—strong methodological rigor and a clear path to trustworthy deployment in high-stakes legal settings. Its ideas generalize beyond law to robustness, invariance testing, and neuro-symbolic verification. Paper 2 introduces a valuable clinical benchmark and training signal, but is more incremental (benchmarking + supervision) and narrower in cross-field methodological innovation.
Paper 2 introduces a broadly applicable failure mode (reward bias substitution) in preference optimization, with formal results showing why common audits can be fundamentally non-identifying even with oracle reward access. It offers a taxonomy, proofs, and actionable evaluation prescriptions, and demonstrates the phenomenon in RLHF/GRPO plus reanalysis of prior mitigation work—high novelty, rigor, and cross-field relevance (alignment, RL, evaluation, fairness). Paper 1 is useful and timely for clinical decision benchmarks, but its impact is narrower to medical LLM evaluation and less foundational than Paper 2’s general theoretical critique and methodology guidance.
Paper 2 introduces a novel benchmark and evaluation paradigm that challenges the current standard of static medical QA, revealing critical flaws in how clinical models handle shifting patient contexts. By exposing a fundamental gap between QA performance and clinical decision-making, it has the potential to broadly steer future research in clinical AI evaluation and model development, giving it a higher potential for widespread scientific impact than the specific algorithmic improvements proposed in Paper 1.
Paper 2 has higher likely impact: it introduces a concrete, auditable benchmark (ClinPivot) targeting a timely, safety-critical gap—whether clinical models appropriately pivot treatment decisions under changing constraints—showing QA metrics can mislead. The methodological contribution (context-pivot construction from biomedical relations, controlled evaluation regimes, matched knowledge budgets, and supervision/replay ablations) is broadly applicable to LLM evaluation, alignment, and medical AI governance. Real-world relevance is immediate for clinical decision support and regulation. Paper 1 is innovative for finance workflows, but is more domain-specific, design/architecture-centric, and less likely to generalize across fields.
Paper 2 addresses a critical gap in medical AI by shifting evaluation from static QA to dynamic, context-dependent treatment decisions. Because reliable evaluation is a major bottleneck for the real-world deployment of clinical foundation models, introducing a benchmark that reveals fundamental flaws in current models is highly likely to drive significant subsequent research and paradigm shifts in high-stakes medical AI. Paper 1 offers a strong, practical systems contribution for speech translation, but Paper 2's fundamental methodological shift in a critical domain provides higher broad scientific impact.