Evaxion EVX-01 durability holds; audit flags LLM neoantigen shortcuts

Auditing Tool Traces in Agentic LLMs: Source Localization and Counterfactual Dependence Testing

A methods paper introduces a two-stage audit for agentic LLM pipelines: Per-Tool TF-IDF Decomposition (PT2D) localizes label-predictive patterns to individual tool results, and Token-Class Redaction counterfactually tests whether the model actually depends on them. Applied to TCR-pMHC binding, tool augmentation initially looked like it beat a no-evidence baseline, but the audit traced the signal to identifiers in one tool result, and redacting them dropped performance to chance. Since TCR-pMHC prediction underpins neoantigen prioritization, this is a concrete warning that shortcut-driven gains can masquerade as biological reasoning in LLM-based discovery workflows, and an actionable template for auditing them before such results inform vaccine design or diligence claims.