Search papers, labs, and topics across Lattice.
5
0
9
2
Missing rewards can skew evaluation results, but this new approach effectively mitigates selection bias in off-policy evaluation.
Proactive message sanitization in multi-agent systems can drastically reduce attack success rates while maintaining system performance.
Spotting unfaithful reasoning in LLMs just got easier: a new method efficiently compares a model's internal computations against its stated rationale.
Naturalness-based data selection, a common technique for curating LLM reasoning datasets, systematically favors longer, lower-quality reasoning chains due to a previously unnoticed "step length confounding" effect.
ARISE lets language models solve math problems better by learning and reusing successful solution strategies, outperforming existing RL methods, especially on harder, out-of-distribution problems.
