Work

Measuring cases where models stop verbalizing evaluation awareness while evaluation-conditioned behavior and internal eval/deploy signals persist.

Testing whether LLM agents build target-specific models of others, rely on generic social priors, or project from their own policy.

A benchmark for long-horizon planning in tool-calling LLMs using deterministic games with external environment simulation.

Studying whether pretrained full-attention Transformers can warm-start TTT-E2E models to reduce long-context training cost while preserving quality.