reward-decomposition

S2T-RLHF proposes a sentence-to-token reward decomposition framework that improves training stability and robustness in preference-based RLHF by assigning sequence-level rewards at sentence granularity, avoiding the instability of overly fine-grained token-level refinement.

LC-ERD is a framework that mines latent logic from LLM-generated reasoning chains to decompose global rewards into step-level signals, enabling self-evolving reasoning without human annotation. It addresses label noise, coarse supervision, and distributional collapse via variational logic potential and multi-agent value decomposition.