Large Multimodal Models (LMMs) achieve state-of-the-art performance in high-stakes domains like healthcare, yet their reasoning remains opaque. Attention- and saliency-based methods often fail to faithfully represent the model’s decision process, particularly when integrating heterogeneous modalities. We introduce Tree-of-Evidence (ToE), an inference-time search algorithm that frames interpretability as a discrete optimization problem. Rather than relying on soft attention weights, ToE employs lightweight Evidence Bottlenecks that score coarse units of data (e.g., vital-sign windows, report chunks) and performs a beam search to identify the compact evidence set required to reproduce the model’s prediction. We evaluate ToE across six tasks spanning three datasets and two domains, including clinical prediction on MIMIC-IV, cross-center validation on eICU, and non-clinical fault detection on LEMMA-RCA. ToE retains over 98% of full-model AUROC with as few as five evidence units, achieves higher decision agreement and lower fidelity error than LIME, SHAP, saliency, and concept-bottleneck baselines under sparse budgets, and outperforms LLMs up to 70B parameters. ToE therefore provides a practical mechanism for auditing multimodal models.

Clinical language and vision-language models increasingly generate diagnostic and treatment rationales, but those claims remain difficult to audit. Existing benchmarks usually score final answers over fully revealed records, without testing whether each claim is grounded in valid patient evidence or whether models can seek missing information efficiently. We introduce Provena-Med, a benchmark for evaluating provenance-faithful multimodal clinical reasoning over linked EHR tables, clinical notes, chest radiographs, and guideline rules. Models must produce a staged artifact comprising a problem list, differential, workup, therapy, and monitoring, in which each claim cites typed evidence units and may abstain when information is insufficient. Provena-Med supports both full-information evaluation and an interactive evidence-seeking track. We further propose MM-AIS, which measures citation validity, objective support under evidence-removal tests, salient-evidence recall, and, for open-weight models, sensitivity to cited evidence. Across ten open-weight models and five cohorts, the best attribution yield is only 0.42, citation hallucination is common, and diagnostic accuracy, safety, and provenance do not reliably co-vary. Provena-Med provides an auditable testbed for evaluating grounded clinical reasoning beyond final-answer accuracy.

Large Language Models (LLMs) are fluent but prone to hallucinations, producing answers that appear plausible yet are unsupported by available evidence. This failure is especially problematic in high-stakes domains where decisions must be justified by verifiable information. We introduce EvidenceRL, a reinforcement learning framework that enforces evidence adherence during training. EvidenceRL scores candidate responses for grounding (entailment with retrieved evidence and context) and correctness (agreement with reference answers) and optimizes the generator using Group Relative Policy Optimization (GRPO). We evaluate across two high-stakes domains, cardiac diagnosis and legal reasoning, where EvidenceRL consistently improves evidence grounding and faithfulness without sacrificing task accuracy. On cardiac diagnosis, F1@3 increases from 37.0 to 54.5 on Llama-3.2-3B while grounding (G_max@3) rises from 47.6 to 78.2; hallucinations drop nearly 5x and evidence-supported diagnoses increase from 31.8% to 61.6%. On legal reasoning, EvidenceRL raises Faithfulness from 32.8% to 67.6% on Llama-3.1-8B, demonstrating consistent behavioral change across domains.

Verifier-guided best-of-N decoding has two stages: a generator samples candidate chains of thought, and a verifier selects among them. A subgroup disparity may already be present in the generator’s candidates or be amplified during selection. We separate the two stages using adversarial demographic counterfactuals of medical exam questions. On HuatuoGPT-o1-8B, a substantial race accuracy gap is already present in a single sampled response, before verifier selection. Across seven generators, only Gemma-3-27B-it under Skywork-RM shows significant verifier-induced amplification after multiple-testing correction. Even an oracle selector leaves a residual gap on every medical generator analyzed, showing that correct answers are less often available for one subgroup. Counterfactually paired low-rank adaptation (LoRA) of the generator reduces the single-sample gap on two medical models, although on HuatuoGPT-o1-8B the reduction does not persist after self-judge selection. Single-variant controls indicate that the single-sample reduction is not explained by additional training steps alone. A short fairness instruction modestly reduces the single-sample gap, and a direct preference optimization (DPO) variant of the paired training also reduces it and, on HuatuoGPT-o1-8B, keeps race mention near its baseline rate. For the medical generators studied, candidate generation rather than inference-time selection is the primary locus for mitigating this gap.

Expert writers calibrate their claims with qualifiers such as “likely benign“ in a radiology report, “may constitute“ in a court ruling, and “suggest“ in a scientific paper. We ask whether this calibration survives when Large Language Models (LLMs) summarize and rewrite professional text. We define epistemic erosion as any change in a source’s expressed certainty, including qualified claims that are dropped or introduced without basis. We measure it with the Epistemic Faithfulness Score (EFS), computed by aligning source and output claims one-to-one with rule-based matching and assigning each distortion to one of six types, namely escalation, attenuation, omission, fabrication, scope shift, and severity warp. On clinical notes, mean EFS for 11 open-weight LLMs ranges from 0.415 to 0.493; three of these models are also evaluated on legal and scientific text. Four hosted models (GPT-4o, GPT-5.4, Phi-4, Mistral-Large-3) reach at most 0.483 in any domain. The most frequent detected distortion is omission (66%), in which a qualified claim is dropped entirely; claims that survive tend to shift toward greater certainty. Model scale and medical adaptation bring no consistent gain, seven prompting strategies bring at most modest gains, and ROUGE-L and BERTScore correlate only weakly with EFS (|r| ≈ 0.1).