LEAF: Growing Trees Without Branching for Speech-Aware Large Language Model Post-Training

Name: LEAF: Growing Trees Without Branching for Speech-Aware Large Language Model Post-Training
Uploaded: 2026-10-06T09:12:14.649259+00:00
Duration: 532 s
Description: LEAF addresses coarse credit assignment in speech-aware LLM post-training by using a retrospective tree-based reinforcement learning method. It recovers structural information from training rollout batches to assign localized, piecewise constant advantages over token spans instead of treating every token equally. This improves model performance and tail risk behavior.

LEAF addresses coarse credit assignment in speech-aware LLM post-training by using a retrospective tree-based reinforcement learning method. It recovers structural information from training rollout batches to assign localized, piecewise constant advantages over token spans instead of treating every token equally. This improves model performance and tail risk behavior.

Open the full episode