让下一篇重要论文，不再错过。

JPU: Bridging Jailbreak Defense and Unlearning via On-Policy Path Rectification

Xi Wang, Songlei Jian National University of Defense Technology, Shasha Li National University of Defense Technology, Xiaopeng Li National University of Defense Technology, Zhaoye Li, Bin Ji National University of Defense Technology, Baosheng Wang, Jie Yu

Despite extensive safety alignment, Large Language Models (LLMs) often fail against jailbreak attacks. While machine unlearning has emerged as a promising defense by erasing specific harmful parameters, current methods remain vulnerable to diverse jailbreaks. We first conduct an empirical study and discover that this failure mechanism is caused by jailbreaks primarily activating non-erased parameters in the intermediate layers. Further, by probing the underlying mechanism through which these circumvented parameters reassemble into the prohibited output, we verify the persistent existence of dynamic jailbreak paths and show that the inability to rectify them constitutes the fundamental gap in existing unlearning defenses. To bridge this gap, we propose Jailbreak Path Unlearning (JPU), which is the first to rectify dynamic jailbreak paths towards safety anchors by dynamically mining on-policy adversarial samples to expose vulnerabilities and identify jailbreak paths. Extensive experiments demonstrate that JPU significantly enhances jailbreak resistance against dynamic attacks while preserving the model’s utility.

尚无 AI 解读，展示原文摘要

Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表

SPPO: Sequence-Level PPO for Long-Horizon Reasoning Tasks

Tianyi Wang, Yixia Li, Long Li Infly Tech, Yibiao Chen, Shaohan Huang, Yun Chen Shanghai University of Finance and Economics, Peng Li Tsinghua University, Yang Liu 等 9 位作者

Proximal Policy Optimization (PPO) is central to aligning Large Language Models (LLMs) in reasoning tasks with verifiable rewards. However, standard token-level PPO struggles in this setting due to the instability of temporal credit assignment over long Chain-of-Thought (CoT) horizons and the prohibitive memory cost of the value model. While critic-free alternatives like GRPO mitigate these issues, they incur significant computational overhead by requiring multiple samples for baseline estimation, severely limiting training throughput. In this paper, we introduce Sequence-Level PPO (SPPO), a scalable algorithm that harmonizes the sample efficiency of PPO with the stability of outcome-based updates. SPPO reformulates the reasoning process as a Sequence-Level Contextual Bandit problem, employing a decoupled scalar value function to derive low-variance advantage signals without multi-sampling. Extensive experiments on mathematical benchmarks demonstrate that SPPO significantly surpasses standard PPO and matches the performance of computation-heavy group-based methods, offering a resource-efficient framework for aligning reasoning LLMs.

尚无 AI 解读，展示原文摘要

Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表

GIFT: Guided Fine-Tuning and Transfer for Enhancing Instruction-Tuned Language Models

Zhiwen Ruan, Yichao Du, Jianjie Zheng Southern University of Science and Technology, Longyue Wang Alibaba Group, Yun Chen Shanghai University of Finance and Economics, Peng Li Tsinghua University, Jinsong Su Xiamen University, Yang Liu 等 9 位作者

Instruction-tuned large language models (LLMs) exhibit strong instruction-following and generalization abilities, enabled by expensive post-training pipelines. However, adapting them to specific downstream tasks remains challenging: direct fine-tuning often disrupts this delicate balance, while existing adapter-based transfer methods typically treat the instruction-tuned model as a passive target that only participates at the final merging stage. We propose GIFT (Guided Fine-Tuning and Transfer), a simple and efficient framework that incorporates instruction-level guidance into task adaptation. GIFT fine-tunes a low-rank adapter on the pretrained base model using token-level confidence signals derived from the instruction-tuned model. The learned adapter is then merged into the instruction-tuned model, yielding task-specialized models that preserve general instruction-following behavior. We evaluate GIFT on mathematical reasoning and knowledge-intensive benchmarks across multiple model families and scales. Results show that GIFT consistently outperforms direct fine-tuning and representative transfer-based baselines, while maintaining robust generalization and favorable test-time scaling behavior.

尚无 AI 解读，展示原文摘要

Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表

MirageBackdoor: A Stealthy Attack that Induces Think-Well-Answer-Wrong Reasoning

Yizhe Zeng, Wei Zhang, Yunpeng Li Institute of Information Engineering, Chinese Academy of Sciences, Juxin Xiao, Xiao Wang, Yuling Liu Institute of Information Engineering，CAS

While Chain-of-Thought (CoT) prompting has become a standard paradigm for eliciting complex reasoning capabilities in Large Language Models, it inadvertently exposes a new attack surface for backdoor attacks. Existing CoT backdoor attacks typically manipulate the intermediate reasoning steps to steer the model toward incorrect answers. However, these corrupted reasoning traces are readily detected by prevalent process-monitoring defenses. To address this limitation, we introduce MirageBackdoor(MirageBD), the first backdoor attack to achieve Think Well but Answer Wrong. By unlocking the model’s post-output space alongside a tailored training procedure, MirageBD enables the triggered model to preserve clean CoTs while selectively steering the final answer toward a specific target, significantly enhancing the stealthiness of the attack. Experiments show that MirageBD generally achieves over 90% attack success rate across four datasets and five models with a poison ratio of only 5%. Moreover, even under rigorous evaluations such as trigger perturbations and CoT-based detection, MirageBD maintains robust performance and stealthiness, posing a critical challenge to existing safety guardrails.

尚无 AI 解读，展示原文摘要

Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表

SPEAK: Spiking Neurons as an Entropy-Aware Tokenizer for Large Language Models

Ming Chen Zhejiang University, Wenyao Li, Chao Liang Hubei Academy of Aerospace Technology General Design Institute, Shi Gu Zhejiang University, Peng Lin Zhejiang University, De Ma Zhejiang University, Huajin Tang Zhejiang University, Qian Zheng 等 9 位作者

Tokenizers play a critical role in large language model studies. Despite recent advances, existing tokenizers fail to explicitly leverage historical tokenization results when making subsequent token decisions, nor do they selectively utilize such history based on contextual relevance. We propose SPEAK, a tokenizer that integrates spiking neurons to explicitly leverage historical tokenization results. Furthermore, we introduce an entropy-aware reset mechanism that selectively leverages history based on contextual relevance, which is determined by token-level entropy. High-entropy tokens are treated as contextual boundaries, whereas low-entropy tokens between consecutive such boundaries exhibit strong contextual relevance. Accordingly, we induce hard reset at high-entropy tokens to discard irrelevant historical tokenization results, and soft reset at low-entropy tokens to preserve and leverage relevant history. Experiments on 2 language models and 5 datasets spanning 16 languages demonstrate superior cross-lingual adaptability, with competitive performance and efficiency. Our code is publicly available at https://github.com/zju-bmi-lab/SPEAK.

尚无 AI 解读，展示原文摘要

Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表

Scaling External Knowledge Input Beyond Context Windows of LLMs via Multi-Agent Collaboration

Zijun Liu, Zhennan Wan, Peng Li Tsinghua University, Ming Yan Alibaba Group, Fei Huang Alibaba Group US, Yang Liu

With the rapid advancement of post-training techniques for reasoning and information seeking, large language models (LLMs) can incorporate a large quantity of retrieved knowledge to solve complex tasks. However, the limited context window of LLMs obstructs scaling the amount of external knowledge input, prohibiting further improvement. Existing context window extension methods inevitably cause information loss. LLM-based multi-agent methods emerge as a new paradigm to handle massive input in a distributional manner, where we identify two core bottlenecks in existing agent orchestration designs. In this work, we develop a multi-agent framework, ExtAgents, to overcome the bottlenecks and enable better scalability in inference-time knowledge integration without longer-context training. Benchmarked with our enhanced multi-hop question answering test, \boldsymbol\inftyBench+, and other public test sets including long survey generation, ExtAgents significantly enhances the performance over existing non-training methods with the same amount of external knowledge input, regardless of whether it falls within or exceeds the context window. Moreover, the method maintains efficiency due to high parallelism. We believe further study in the coordination of LLM agents on increasing external knowledge input could benefit real-world applications.

尚无 AI 解读，展示原文摘要

Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表

InstructDiff: Domain-Adaptive Data Selection via Contrastive Entropy for Efficient LLM Fine-Tuning

Junyou Su, He Zhu, Xiao Luo University of Wisconsin - Madison, Liyu Zhang, Hong-Yu Zhou, Yun Chen Shanghai University of Finance and Economics, Peng Li Tsinghua University, Yang Liu 等 9 位作者

Supervised fine-tuning (SFT) is fundamental to adapting large language models, yet training on complete datasets incurs prohibitive costs with diminishing returns. Existing data selection methods suffer from severe domain specificity: techniques optimized for general instruction-following fail on reasoning tasks, and vice versa. We observe that measuring contrastive entropy between base models and minimally instruction-tuned calibrated models reveals a pattern—samples with the lowest contrastive entropy consistently yield optimal performance across domains, yet this principle manifests domain-adaptively: reasoning tasks favor entropy increase (cognitive expansion), while general tasks favor entropy decrease (cognitive compression). We introduce InstructDiff, a unified framework that operationalizes contrastive entropy as a domain-adaptive selection criterion through warmup calibration, bi-directional NLL filtering, and entropy-based ranking. Extensive experiments show that InstructDiff achieves 17% relative improvement over full data training on mathematical reasoning and 52% for general instruction-following, outperforming prior baselines while using only 10% of the data.

尚无 AI 解读，展示原文摘要

Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表

Benchmarking and Enabling Efficient Chinese Medical Retrieval via Asymmetric Encoders

Angqing Jiang University of Science and Technology of China, Jianlyu Chen, Zhefang, Yongcan Wang iFLYTEK AI Research, Xinpeng Li, Keyu Ding, Defu Lian University of Science and Technology of China

Effective medical text retrieval requires both high accuracy and low latency. While LLM-based embedding models possess powerful retrieval capabilities, their prohibitive latency and high computational cost limit their application in real-time scenarios. Furthermore, the lack of comprehensive and high-fidelity benchmarks hinders progress in Chinese medical text retrieval. In this work, we introduce the Chinese Medical Text Embedding Benchmark (CMedTEB), a benchmark spanning three kinds of practical embedding tasks: retrieval, reranking, and semantic textual similarity (STS). Distinct from purely automated datasets, CMedTEB is curated via a rigorous multi-LLM voting pipeline validated by clinical experts, ensuring gold-standard label quality while effectively mitigating annotation noise. On this foundation, we propose the Chinese Medical Asymmetric REtriever (CARE), an asymmetric architecture that pairs a lightweight BERT-style encoder for online query encoding with a powerful LLM-based encoder for offline document encoding. However, optimizing such an asymmetric retriever with two structurally different encoders presents distinctive challenges. To address this, we introduce a novel two-stage training strategy that progressively bridges the query and document representations. Extensive experiments demonstrate that CARE surpasses state-of-the-art symmetric models on CMedTEB, achieving superior retrieval performance without increasing inference latency.

尚无 AI 解读，展示原文摘要

Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表

SpectroStream: A Versatile Neural Codec for General Audio

Yunpeng Li, Kehang Han, Brian McWilliams, Zalan Borsos, Marco Tagliasacchi

We propose SpectroStream, a full-band multi-channel neural audio codec. Successor to the well-established SoundStream, SpectroStream extends its capability beyond 24 kHz monophonic audio and enables high-quality reconstruction of 48 kHz stereo music at bit rates of 4--16 kbps. This is accomplished with a new neural architecture that leverages audio representation in the time-frequency domain, which leads to better audio quality especially at higher sample rate. The model also uses a delayed-fusion strategy to handle multi-channel audio, which is crucial in balancing per-channel acoustic quality and cross-channel phase consistency.

尚无 AI 解读，展示原文摘要