PoEM predicts RL post-training outcomes for a new reward by regressing the reward on existing adapters' log-ratios and mixing those policies in log-space, validated on language and image models.
Contrastive decoding: Open-ended text generation as optimization
3 Pith papers cite this work. Polarity classification is still indexing.
3
Pith papers citing it
citation-role summary
background 1
citation-polarity summary
years
2026 3roles
background 1polarities
background 1representative citing papers
citing papers explorer
-
PoEM: Predicting RL Outcomes from Existing Policies
PoEM predicts RL post-training outcomes for a new reward by regressing the reward on existing adapters' log-ratios and mixing those policies in log-space, validated on language and image models.
- MLLMs Know When Before Speaking: Revealing and Recovering Temporal Grounding via Attention Cues
- Mitigating Multimodal LLMs Hallucinations via Relevance Propagation at Inference Time