arXiv cs.CL· Suwan Wu, Yumeng Lin, Pengcheng Yuan, Xiaolong Jiang·· 11 小时前AI 评分22
统一逐 Token 门控家族：FKL/RKL 混合与多通道、偏置系数的在线策略蒸馏
A Unified Per-Token Gating Family for On-Policy Distillation: FKL/RKL Mixing with Multi-Channel and Bias Coefficients
AI 导读
研究提出四系数参数化 λ_t = σ(a·h_t + b·u(x) + c + d·gap_t)，将 EOPD 与 ToDi 的方向对齐代理统一为其中一维特例，并新增多通道组合与显式偏置。
正文 · 原文
Abstract:Per-token gating of forward/reverse KL losses has become a standard technique for on-policy knowledge distillation (OPD), but existing methods such as EOPD (Jin et al., 2026) and ToDi (Jung et al., 2025) each fix a single gating signal and a single gating direction, and the two have never been compared directly. We introduce a four-coefficient parameterization lambda_t = sigma(a * h_t + b * u(x) + c + d * gap_t) in which direction-aligned proxies of EOPD and ToDi appear as one-dimensional (1D) restrictions, and which adds multi-channel composition and an explicit bias as further degrees of freedom. On TweetEval (Barbieri et al., 2020) emotion and hate, with a Qwen3-32B teacher and a Qwen3-4B student, configurations in the full family reach higher accuracy than the matched-magnitude single-channel (entropy-only / gap-only) 1D restrictions in 33 of 36 comparable cells, and a 26-cell mean-match isolation experiment places dynamic gating ahead of effective-KL-matched static baselines in 19 of 26 cells. Because cells share training data, models, and parameter substructure, we report both counts as exploratory aggregate directional evidence rather than as independent hypothesis tests. Targeted three-seed paired replications of the nine headline comparisons singled out by that sweep -- including a third task, offensive -- are directionally consistent, but individually smaller than the single-seed estimates and not significant at n=3. We therefore present the parameterization primarily as a shared coordinate system for comparing per-token gating designs in short-output classification OPD.
| Comments: | Accepted at the Findings of the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP 2026 Findings) |
| Subjects: | Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Machine Learning (cs.LG) |
| Cite as: | arXiv:2609.11768 [cs.AI] |
| (or arXiv:2609.11768v2 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2609.11768 arXiv-issued DOI via DataCite |
Submission history
From: Suwan Wu [view email]
[v1]
Thu, 10 Sep 2026 16:22:35 UTC (141 KB)
[v2]
Fri, 30 Oct 2026 08:23:06 UTC (150 KB)
来源：arXiv cs.CL · arxiv.org