A research framework now lets engineers statistically label what an LLM agent is "thinking" at each step of a multi-step task — and steer it away from failure mid-run.
The paper introduces what the authors call a conformal interpretability framework for temporal tasks. It pairs step-wise reward modeling with conformal prediction to classify an agent's internal representations at each step as on-track or failing. Linear probes trained on those representations then identify directions in the model's activation space that correspond to success, failure, or reasoning drift. The researchers tested the approach on two simulated environments — ScienceWorld and AlfWorld — and found that these internal signals are linearly separable, meaning the model's hidden states genuinely encode task-relevant structure that can be read and redirected.
This matters because LLM agents are being deployed in settings where silent failure is expensive. Most current monitoring is output-level: you catch a mistake after the agent has already acted. A framework that detects drift in internal representations before the action lands is a qualitatively different safety primitive — closer to an instrument panel than a crash report. The authors also show preliminary evidence that steering toward successful activation directions improves agent performance, not just detection.
The catch, as always with interpretability research, is the gap between controlled simulation and production chaos — ScienceWorld and AlfWorld are tidy benchmarks, and real agent deployments are not.