Tag
This paper presents a factorised study of probe-based uncertainty estimation in LLMs, showing that raw hidden states and attention features perform well in-domain but structured features are more robust under distribution shift, and provides pretrained probes as off-the-shelf baselines.
This paper introduces POISE, a method for stable policy optimization in large reasoning models by estimating baselines using the model's own internal states, reducing computational overhead compared to PPO and GRPO.
This paper challenges the assumption that LLMs can reliably distinguish between hallucinated and factual outputs through internal signals, arguing that internal states primarily reflect knowledge recall rather than truthfulness. The authors propose a taxonomy of hallucinations (associated vs. unassociated) and show that associated hallucinations exhibit hidden-state geometries overlapping with factual outputs, making standard detection methods ineffective.