Fan Li Northeastern University, Yu Gu, Zhigang Wang Guangzhou University, Fangling Leng Northeastern University, Zhenghao Liu Northeastern University, Ge Yu
Diffusion language models (DLMs) have emerged as a powerful non-autoregressive alternative to GPT-style sequential generation, but suffer from substantial computational overhead due to their iterative parallel denoising. Existing acceleration works cannot accurately detect semantically stabilized tokens and then skip computation, leading to sub-optimal speedup in practice. This paper presents the first systematic study of convergence dynamics in DLMs. Innovative observations include the misalignment between traditionally used scalar detection criterion and the semantic convergence, and the post-peak confidence score, that wastes denoising computation and degrades inference quality. To address these limitations, we propose Ada-DLM, a semantic-aware adaptive denoising framework that encodes the trajectory of scalar confidence scores into an evolution-aware feature vector and then clusters vectors proactively and adaptively identify semantically converged tokens. Furthermore, we incorporate system-level optimizations to maximize runtime efficiency. Experiments show that Ada-DLM consistently outperforms the SOTA competitor, achieving up to 2x speedup and 19% quality improvement. That offers a practical path toward efficient high-quality DLM deployment.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Zhenghao Liu Northeastern University, Zhuoyang Wu, Xinze Li Northeastern University, Yukun Yan Tsinghua University, Shuo Wang, Zulong Chen, Yu Gu, Ge Yu 等 9 位作者
Large Language Models (LLMs) have demonstrated remarkable reasoning capabilities, particularly in solving complex mathematical problems. Recent studies show that distilling long reasoning trajectories can effectively enhance the reasoning performance of small-scale student models. However, teacher-generated reasoning trajectories are often excessively long and structurally complex, making them difficult for student models to learn. This mismatch leads to a gap between the provided supervision signal and the learning capacity of the student model. To address this challenge, we propose Prefix-ALIGNment distillation (P-ALIGN), a framework that fully exploits teacher CoTs for distillation through adaptive prefix alignment. Specifically, P-ALIGN adaptively truncates teacher-generated reasoning trajectories by determining whether the remaining suffix is concise and sufficient to guide the student model. Then, P-ALIGN leverages the teacher-generated prefix to supervise the student model, encouraging effective prefix alignment. Experiments on multiple mathematical reasoning benchmarks demonstrate that P-ALIGN outperforms all baselines by over 3%. Further analysis indicates that the prefixes constructed by P-ALIGN provide more effective supervision signals, while avoiding the negative impact of redundant and uncertain reasoning components. All codes are available at https://github.com/NEUIR/P-ALIGN.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Zijian Wen, Tao Zhang University of Science and Technology of China, Shuangwu Chen, Shenghao Ye University of Science and Technology of China, Yu Guo University of Science and Technology of China, Qirui Chen University of Science and Technology of China, Jingxian Shuai, Yunpeng Hou 等 10 位作者
Retrieval-Augmented Generation (RAG) significantly enhances LLMs but faces high prefill latency during long-context processing. While KV cache reuse can mitigate this, current methods relying on shallow features or static heuristics often fail to identify critical tokens for recomputation, resulting in generation quality degradation.We have an insight that KV deviations are more pronounced in deep layers.However, directly extracting deep-layer features from the target model is computationally prohibitive. Crucially, we find that the deep-layer features of a lightweight speculative model exhibit strong consistency with the target model in the selection of critical tokens for recomputation.In light of these insights, we propose SpecCache, which employs deep-layer hidden-state norms from a speculative model as a proxy to guide the critical token selection for target large model.Experiments demonstrate that SpecCache outperforms state-of-the-art (SOTA) baselines. Compared to full KV recomputation, it reduces time-to-first-token (TTFT) by 2.17-3.95\times and increases inference throughput by 2.7-5.2\times, with negligible degradation in generation quality relative to full recomputation.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
The soft-thinking paradigm for Large Language Model (LLM) reasoning can outperform the conventional discrete-token Chain-of-Thought (CoT) reasoning in some scenarios, underscoring its research and application value. However, while the discrete-token CoT reasoning pattern can be reinforced through policy optimization algorithms such as group relative policy optimization (GRPO), extending the soft-thinking pattern with Reinforcement Learning (RL) remains challenging. This difficulty stems from the complexities of injecting stochasticity into soft-thinking tokens and updating soft-thinking policies accordingly. As a result, previous attempts to combine soft-thinking with GRPO typically underperform their discrete-token GRPO counterparts. To fully unlock the potential of soft-thinking, this paper presents a novel policy optimization algorithm, SofT-GRPO, to reinforce LLMs under the soft-thinking reasoning pattern. SofT-GRPO injects the Gumbel noise into logits, employs the Gumbel-Softmax technique to avoid soft-thinking tokens outside the pre-trained embedding space, and leverages the reparameterization trick in policy gradient. We conduct experiments across base LLMs ranging from 1.5B to 7B parameters, and results demonstrate that SofT-GRPO enables soft-thinking LLMs to slightly outperform discrete-token GRPO on Pass@1 (+0.13% on average accuracy), while exhibiting a substantial uplift on Pass@32 (+2.19% on average accuracy). Codes and weights are available on https://github.com/zz1358m/SofT-GRPO-master