paper-with-me

홈 › Papers

Beyond Entropy: Learning from Token-Level Distributional Deviations for LLM Reasoning

2026-06-18 · Xuanzhi Feng, Zhengyang Li, Zeyu Liu, Haoxi Li, Yuming Jiang, Bing Guo, Jingcai Guo, Jie Zhang, Song Guo arxiv

Reinforcement Learning with Verifiable Rewards (RLVR) has significantly advanced Large Language Model (LLM) reasoning; however, it faces a fundamental optimization instability: uniform token updates precipitate entropy collapse, leading to premature convergence to suboptimal strategies, whereas excessive Shannon Entropy maximization can cause entropy explosion, driving blind exploration toward incoherent reasoning chains. To resolve this dichotomy, we introduce the Independent Combinatorial Tokens (ICT) framework, which shifts the optimization focus from scalar uncertainty to the distributional properties of token logits. By leveraging the Jensen-Shannon (JS) divergence between token logits distributions, ICT identifies tokens with distinctive distributional patterns as critical branching points for guiding effective exploration in LLM reasoning. Our theoretical analysis, grounded in both Shannon and second-order Rényi entropy, proves that selectively updating on these tokens regulates policy concentration: it reduces the overall distribution uncertainty measured by Shannon entropy, while controlling probability concentration captured by second-order Rényi entropy. This dual effect prevents over-concentrated token generation from weakening exploration and effectively stabilizes the training landscape. Empirical results demonstrate that updating only the top 10% of unique tokens on Qwen2.5 (0.5B/1.5B/7B) models yields an average pass@4 improvement of 4.58%, with a maximum gain of 14.9%, over GRPO, 20-Entropy, and STAPO baselines across seven benchmarks spanning math, commonsense, and Olympiad-level problems.

📄 PDF Abstract BibTeX arXiv:2606.19771

Code (0)

등록된 구현이 없습니다.

Tasks

Reinforcement Learning

Similar Papers 제목 키워드 기반

Beyond Token-Level Cross-Entropy: Fréchet Distributional Post-Training for Autoregressive Image Generation

2026-08-01 · Jinhua Zhang, Yisong Lin, Wei Long, Shuhang Gu arxiv

Autoregressive image generators are commonly pretrained with token-level cross-entropy under teacher forcing, yet evaluated by the distributional quality of decoded images. This creates an objective mismatch, because cat…

Image Generation

Sparse but Critical: A Token-Level Analysis of Distributional Shifts in RLVR Fine-Tuning of LLMs

2026-03-23 · Haoming Meng, Kexin Huang, Shaohang Wei, Chiyu Ma 외 arxiv

Reinforcement learning with verifiable rewards (RLVR) has significantly improved reasoning in large language models (LLMs), yet the token-level mechanisms underlying these improvements remain unclear. We present a system…

Reinforcement Learning

DC-MBR: Distributional Cooling for Minimum Bayesian Risk Decoding

2022-12-08 · Jianhao Yan, Jin Xu, Fandong Meng, Jie zhou 외

Minimum Bayesian Risk Decoding (MBR) emerges as a promising decoding algorithm in Neural Machine Translation. However, MBR performs poorly with label smoothing, which is surprising as label smoothing provides decent impr…

Machine TranslationNMT

LP-SFT: Local-Preserving Supervised Fine-Tuning via Multimodal Entropy Structure

2026-07-06 · Yueyang Wang, Baolong Bi, Shuo Lu, Jingyuan Zhang arxiv

Supervised fine-tuning (SFT) is the standard approach for adapting pretrained language models to downstream domains, yet it often improves target-domain behavior at the cost of degrading pre-existing capabilities. Standa…

Beyond Uniform Token-Level Trust Region in LLM Reinforcement Learning

2026-06-09 · Renjie Mao, Xiangxin Zhou, Lvfang Tao, Yixin Ding 외 arxiv

Reinforcement learning with verifiable rewards (RLVR) has become standard for improving LLM reasoning. However, existing PPO-style trust-region mechanisms remain position-agnostic by enforcing uniform thresholds across a…

Reinforcement Learning