When medical AI systems hallucinate clinical reasoning, the consequences extend beyond incorrect answers: fabricated justifications that superficially reference retrieved evidence can mislead clinicians into unsafe treatment decisions. Medical reasoning agents must therefore produce not only correct answers but also faithful justifications that clinicians can verify against cited evidence. We identify a systematic failure mode in RL-trained retrieval agents: outcome-only rewards improve accuracy while degrading faithfulness, a phenomenon we term confident hallucination. The agent learns to answer from parametric memory and backfill plausible but unsupported justifications; citation fabrication rates rise from 16.5% to 31.8% even as accuracy improves by 5 points over the supervised baseline. We address this with a faithfulness-gated reward design: accuracy credit is conditioned on evidence grounding via a hard gate, complemented by retrieval validity and conciseness signals that close exploitation paths unique to agentic retrieval. The resulting system, MedAgent-R1, reduces citation fabrication from 31.8% to 4.7% and raises evidence completeness from 58.7 to 82.6 while maintaining 75.1% accuracy, with 13.2-point gains on HealthBench Safety. Under the same agentic retrieval setup, MedAgent-R1 outscores GPT-4o on faithfulness-specific dimensions (Factual Support 4.55 vs. 4.25; Overclaiming 4.40 vs. 4.15) while remaining below GPT-4o in overall accuracy, suggesting that explicit faithfulness training yields evidence-grounding gains not achieved by scaling alone.
Clinicians read chain-of-thought (CoT) rationales as evidence of medical reasoning, but whether the visible chain plays that role is rarely tested. General-domain CoT-faithfulness probes ignore clinical cost, and medical LLM evaluations treat the chain as a black box. We close this gap with a medical perturbation audit: a 30-operator battery edits both the chain and the question with clinically motivated operators (severity reversal, negation flip, demographic swap, evidence ablation), paired with a chain-update times answer-flip joint analysis that classifies each model by its failure mode. Applied to 14 LLMs on four medical QA benchmarks, three independent tests converge: the Chain-Decoupling Rate (CDR; chain does not register the edit and the answer does not flip) is 72.9% panel-wide on clinically meaningful destructive edits, chain corruption leaves accuracy unchanged, and removing CoT prompting does not reduce accuracy. Two board-certified clinicians re-annotate N=197 perturbed questions; 98.5% leave the gold defensible. The pattern holds across medical and reasoning fine-tuning and scale; on the closed-source tier, where the chain text is unavailable, the answer-side signals are consistent with the same decoupling. Our framework and CDR provide a reusable yardstick for auditing whether medical CoT is faithful or merely documentation.
Detecting hallucinations in clinical text is increasingly framed as an explainability problem: systems should not merely flag an unreliable response but point to the offending span. Architectures built around knowledge-graph (KG) triple decomposition are marketed for exactly this auditability, yet their localization ability is typically assumed rather than measured. We introduce MedHal-Loc, a benchmark and metric for localization faithfulness -- whether a detector's top-ranked error unit actually overlaps the erroneous span. The controlled subset comprises 300 PubMedQA-derived statements with single, span-level errors injected across four localizable types (entity substitution, relation error, mechanism misattribution, invention), yielding gold spans by construction; a complementary natural subset documents that real hallucinations are dominated by diffuse conclusion-flips that resist span localization (a human expert accepted 1/18 candidate spans). Evaluating four fine-grained paradigms, we find that NLI-per-clause, consistency-per-sentence, and the dedicated span detector FAVA all localize well above chance, whereas an elaborate KG-triple pipeline localizes no better than chance (+3.3pp, n.s.), bottlenecked by ~59% entity-extraction coverage -- despite competitive detection F1 (0.609). Detection competence does not imply faithful localization; architectural explainability must be validated, not presumed.