Tag
A technique called spurious probes uses adaptive API-level search to find questions that reveal the latent state of blackbox LLMs, such as distinguishing capability evaluations via specific prompts.
The paper introduces LLAMIA-Bench, a benchmark for collaborative chess tasks between language models and non-language agents, and proposes latent state internalization to outperform text-based verbalization, with a 14B model matching or exceeding frontier models like GPT-5.1.
StateSight is a new benchmark for evaluating spatial-state reconstruction in vision-language models, showing that models like GPT-5.5 and Claude Sonnet 5 struggle with spatial reasoning tasks compared to human performance.
This paper introduces a white-box instrument using hidden deterministic finite automata to separately measure reward success and latent-state learning in reinforcement learning agents, finding that high reward does not imply task understanding.
This paper introduces textual latent states and factorized GRPO (fGRPO) to enforce strict mediation in text-based world models, addressing the identifiability problem and achieving up to 57% gains in representation quality and 98% improvements in rollout performance.
A researcher presents evidence that strong target text can induce a measurable latent-state shift in Gemma 3 12B IT before final output, distinct from lexical or content overlaps, and discusses implications for AI safety beyond output-only evaluation.