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.

Large language models (LLMs) are strong sequence predictors, but standard context accumulation offers no explicit mechanism for revising the memory carried between predictions. We propose LLM-as-RNN, an inference-time framework for online learning with frozen LLMs through a bounded natural-language memory state. At each timestep, the model predicts from the current observation and its memory, receives feedback, and rewrites the state to correct errors and retain task-relevant patterns without updating model parameters. Implemented as a structured system prompt, this state makes the adaptation process inspectable through human-readable memory traces. We evaluate LLM-as-RNN on sequential prediction benchmarks in healthcare, meteorology, and finance across Llama, Gemma, and GPT-oss model families. Comparisons with static-context and feedback-aligned baselines examine predictive performance, feedback sources, memory budgets, and inference cost. Our results demonstrate that feedback-driven memory rewriting supports online error correction and sequential prediction under a fixed memory budget, with performance varying across tasks and backbones.