Pith. sign in

REVIEW 3 cited by

AlignDistil: Token-Level Language Model Alignment as Adaptive Policy Distillation

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2503.02832 v3 pith:H2TBWP3S submitted 2025-03-04 cs.CL cs.AIcs.LG

AlignDistil: Token-Level Language Model Alignment as Adaptive Policy Distillation

classification cs.CL cs.AIcs.LG
keywords rewardmodeltoken-leveltokensaligndistilalignmentdistillationmethods
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved
0 comments
Share X Bluesky LinkedIn Reddit HN
read the original abstract

In modern large language models (LLMs), LLM alignment is of crucial importance and is typically achieved through methods such as reinforcement learning from human feedback (RLHF) and direct preference optimization (DPO). However, in most existing methods for LLM alignment, all tokens in the response are optimized using a sparse, response-level reward or preference annotation. The ignorance of token-level rewards may erroneously punish high-quality tokens or encourage low-quality tokens, resulting in suboptimal performance and slow convergence speed. To address this issue, we propose AlignDistil, an RLHF-equivalent distillation method for token-level reward optimization. Specifically, we introduce the reward learned by DPO into the RLHF objective and theoretically prove the equivalence between this objective and a token-level distillation process, where the teacher distribution linearly combines the logits from the DPO model and a reference model. On this basis, we further bridge the accuracy gap between the reward from the DPO model and the pure reward model, by building a contrastive DPO reward with a normal and a reverse DPO model. Moreover, to avoid under- and over-optimization on different tokens, we design a token adaptive logit extrapolation mechanism to construct an appropriate teacher distribution for each token. Experimental results demonstrate the superiority of our AlignDistil over existing methods and showcase fast convergence due to its token-level distributional reward optimization.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Weak-to-Strong Generalization via Direct On-Policy Distillation

    cs.LG 2026-07 conditional novelty 7.0

    Transferring the log-ratio of a small model's pre-RL and post-RL checkpoints provides a dense implicit reward that improves stronger student models at a fraction of the cost of direct RL.

  2. Weak-to-Strong Generalization via Direct On-Policy Distillation

    cs.LG 2026-07 conditional novelty 6.0

    Transferring a weak model’s RL-induced log-ratio policy shift on a strong student’s own rollouts raises AIME accuracy more cheaply than imitating the weak teacher or running matched-step RL on the student.

  3. Evolving Knowledge Distillation for Lightweight Neural Machine Translation

    cs.CL 2026-05 conditional novelty 6.0

    EKD trains lightweight NMT students progressively from a chain of teachers with rising capacity, achieving BLEU scores within 0.08 of the largest teacher on IWSLT-14.