Zhengmian Hu, Tong Zheng, Vignesh Viswanathan, Ziyi Chen, Ryan A. Rossi, Yihan Wu, Dinesh Manocha, Heng Huang
Large Language Models (LLMs) have become an indispensable part of natural language processing tasks. However, autoregressive sampling has become an efficiency bottleneck. Multi-Draft Speculative Decoding (MDSD) is a recent approach where, when generating each token, a small draft model generates multiple drafts, and the target LLM verifies them in parallel, ensuring that the final output conforms to the target model distribution. The two main design choices in MDSD are the draft sampling method and the verification algorithm. For a fixed draft sampling method, the optimal acceptance rate is a solution to an optimal transport problem, but the complexity of this problem makes it difficult to solve for the optimal acceptance rate and measure the gap between existing verification algorithms and the theoretical upper bound. This paper discusses the dual of the optimal transport problem, providing a way to efficiently compute the optimal acceptance rate. For the first time, we measure the theoretical upper bound of MDSD efficiency for vocabulary sizes in the thousands and quantify the gap between existing verification algorithms and this bound. We also compare different draft sampling methods based on their optimal acceptance rates. Our results show that the draft sampling method strongly influences the optimal acceptance rate, with sampling without replacement outperforming sampling with replacement. Additionally, existing verification algorithms do not reach the theoretical upper bound for both without replacement and with replacement sampling. Our findings suggest that carefully designed draft sampling methods can potentially improve the optimal acceptance rate and enable the development of verification algorithms that closely match the theoretical upper bound.
Jesson Wang, Shawn Li, Wei Yang, Franck Dernoncourt, Ryan A. Rossi, Charith Peris, Yue Zhao
Manipulating next-token probabilities during generation can bypass the safety alignment of large language models. Existing approaches, however, rely on access to model weights or numerical token probabilities and therefore do not apply to interfaces that return only sampled text. Reconstructing probabilities from sampled outputs offers a possible alternative, but finite sampling produces sparse and noisy estimates, while repeating this process at every generation step incurs substantial query costs. Our empirical observations suggest that large distributional changes along successful jailbreak trajectories are concentrated at a small subset of positions, motivating selective control. We introduce \method{}, a framework for jailbreaking through text-only continuation interfaces that permit repeated sampling and assistant-prefix continuation. Sample-Based Distribution Reconstruction combines sampled outputs with a prior over unobserved actions to obtain a usable control signal. Risk-Gated Residual Control uses the evolving response prefix to decide when to reconstruct and modify the distribution, concentrating sampling costs at selected positions. Speculative Multi-Token Execution further amortizes target calls by verifying and accepting draft prefixes that require no intervention. Across four target endpoints and three benchmarks, \method{} achieves the highest mean score most comparisons against baselines.
Chien Van Nguyen University of Oregon, Huy Huu Nguyen, Ruiyi Zhang, Hanieh Deilamsalehy Adobe Systems, Puneet Mathur Adobe Systems, Viet Dac Lai Adobe Systems, Haoliang Wang Adobe Research, Jayakumar Subramanian Adobe Systems 等 13 位作者
We propose Lizard, a linearization framework that transforms pretrained Transformer-based Large Language Models (LLMs) into subquadratic architectures. Transformers faces severe computational and memory bottlenecks with long sequences due to the quadratic complexity of softmax attention and the growing Key-Value (KV) cache that makes inference memory-bound by context length. Lizard addresses these limitations by introducing a subquadratic attention mechanism that closely approximates softmax attention while preserving model quality. Unlike prior linearization methods constrained by fixed, non-adaptive structures, Lizard augments the architecture with compact, learnable modules that enable adaptive memory control and robust length generalization. Moreover, we introduce a hardware-aware algorithm that solves numerical instability in gated attention to accelerate training. Extensive experiments show that Lizard achieves near-lossless recovery of its teacher model’s performance, significantly outperforming previous methods by up to 9.4 - 24.5 points on the 5-shot MMLU benchmark and demonstrating superior associative recall.
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
Chien Van Nguyen University of Oregon, Ryan A. Rossi, Linh Ngo Van Hanoi University of Science and Technology, Franck Dernoncourt Adobe Research, Thien Huu Nguyen University of Oregon
Transformer inference becomes increasingly memory-bound as the Key–Value (KV) cache grows linearly with sequence length. While subquadratic architectures offer constant-memory inference, they rely on aggressive state compression that degrades performance on complex reasoning tasks. We propose Octopus, a framework that confers fixed-memory inference onto pretrained Transformers without the information loss of linearization. Octopus retrofits attention layers with Gated Selective Attention, a learnable module that enforces an adaptive sparsity policy over the context history. By dynamically scoring and retaining only high-utility KV states, this mechanism transforms the unbounded cache into a compact, evolving memory budget that filters out uninformative noise. Empirically, on the GSM8K benchmark, it outperforms state-of-the-art linearized baselines by over 36 points under identical memory constraints. Remarkably, Octopus also surpasses its own full-cache teacher, demonstrating that learned sparse retention serves as an effective regularizer for long-horizon reasoning.
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
Rui Wang Adobe Systems, Junda Wu University of California, San Diego, Yu Xia University of California, San Diego, Tong Yu Adobe Research, Ruiyi Zhang, Ryan A. Rossi, Subrata Mitra Adobe Systems, Lina Yao University of New South Wales 等 9 位作者
Large Language Models (LLMs) are susceptible to indirect prompt injection attack, where the model inadvertently responds to instructions injected into the prompt context. This vulnerability stems from LLMs’ inability to distinguish between data and instructions within a prompt. We propose CachePrune that defends against this attack by identifying and pruning neurons associated with instruction-following, during KV cache encoding of the prompt context. The pruning steers the LLM toward interpreting the context purely as data rather than as instructions to follow. To identify these neurons, we introduce a neural attribution mechanism guided by a preferential attribution loss, and theoretically connect this loss to an upper bound of the Direct Preference Optimization (DPO) objective. Further, we improve on the fidelity of neural attribution by leveraging an observed triggering effect in instruction-following. Our approach does not interfere with prompt formatting or incur test-time overhead in response generation. Experiments show that CachePrune significantly reduces the attack success rate while preserving the LLM’s ability to follow user instructions.
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
Abhinav Sharma, Brian Zhang, Deepti Guntur University of Massachusetts at Amherst, Zhiyang Zuo MavenBio, Shreyas Chaudhari University of Massachusetts at Amherst, Wenlong Zhao University of Massachusetts at Amherst, Franck Dernoncourt Adobe Research, Puneet Mathur Adobe Systems 等 10 位作者
Retrieval-Augmented Generation (RAG) systems face challenges with complex, multi-hop questions, and iterative agentic frameworks such as Search-R1 (Jin et al., 2025) have been proposed to address these complexities. However, such approaches can introduce inefficiencies, including repetitive retrieval of previously processed information and challenges in contextualizing retrieved results effectively within the current generation prompt. Such issues can lead to unnecessary retrieval turns, suboptimal reasoning, inaccurate answers, and increased token consumption. In this paper, we investigate test-time modifications to Search-R1’s open-source Qwen2.5-7B pipeline to mitigate these identified shortcomings. Specifically, we explore the integration of two components and their combination: a contextualization module to better integrate relevant information from retrieved documents into reasoning, and a de-duplication module that replaces previously retrieved documents with the next most relevant ones. We evaluate our approaches using the HotpotQA (Yang et al., 2018) and the Natural Questions (Kwiatkowski et al., 2019) datasets, reporting the exact match (EM) score, an LLM-as-a-Judge assessment of answer correctness, and the average number of turns. Our best-performing variant (contextualization) achieves a 5.6% increase in EM score and reduces the average number of turns by 10.5% compared to the Search-R1 baseline. While contextualization itself introduces additional LLM calls, our results demonstrate improved answer accuracy and reduced retrieval load.
尚无 AI 解读，展示原文摘要
cs.CL · cs.AI
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop) · 2026 · 正式发表