Search papers, labs, and topics across Lattice.
3
0
6
Reward hacking in rubric-based RL is not just common; it can be systematically reproduced and analyzed using the new CHERRL environment, revealing hidden biases that could compromise training integrity.
Forget external signals – unlock better LLM post-training by mining model internals with sparse autoencoders to reveal data diversity, difficulty, and quality.
Current reward models are surprisingly bad at judging story quality, achieving only 66% accuracy in selecting human-preferred narratives – a gap closed by a new, purpose-built reward model.
