Roman Vashurin, Mikhail Seleznyov, Assel Yermekova, Fedor Chernogorskii, Danila Zykov, Elena Tutubalina, Alexander Panchenko, Preslav Nakov 等 10 位作者
Classical decompositions of predictive uncertainty, such as the information-theoretic one, are defined for a fixed representation $x$ of the input object. They do not consider that $x$ may be incomplete, nor do they measure how much the missing information can affect the prediction. In Large Language Models, users can clarify an underspecified input without any retraining. The clarified input can describe the user's unobserved intent $z^*$ more accurately and change the predictive uncertainty about the outcome. In this paper, we formalize this setting probabilistically and derive a three-term decomposition of prediction risk into specification, residual aleatoric, and epistemic uncertainties. We show that at the original information level, when only $x$ is available, the aleatoric uncertainty, represented by a Bayes risk, has two components. The first, which we call specification uncertainty, is the expected reduction in this Bayes risk from revealing the intent. The second is the residual aleatoric uncertainty, i.e., the Bayes risk that remains once the intent is known, averaged over its possible values $z$ given $x$. The distinction is important, as specification uncertainty informs the decision to request clarification before making a prediction. We build approximations of these components and connect them to existing uncertainty measures. Our experiments across different domains and modalities show distinct roles for the three components. Specification uncertainty is the most informative for detecting underspecified inputs, while residual aleatoric uncertainty is the most informative for detecting in-distribution prediction errors.
Heads of Security: Learned Attention Gates Improve Prompt Injection Robustness
Igor Shilov, Xiaoxue Yang, Alexander Panfilov, Evgenii Kortukov, Matthieu Meeus, Peter Romov, Wojciech Samek, Jonas Geiping 等 9 位作者
AI 速览：提出 Heads of Security（HoS），用少量可学习的逐头注意力门实现指令与数据分离，从而在不明显损失效用的前提下提升对提示注入攻击的鲁棒性。
We introduce Heads of Security (HoS), a prompt injection defense that enforces instruction-data separation via a small set of learned per-head attention gates. HoS learns how each head should redistribute attention between trusted and untrusted tokens, enabling models to reason over untrusted content without following injected instructions. We evaluate HoS on Llama-3.3-70B-Instruct and GPT-OSS-20B. We compare it against SecAlign++, a state-of-the-art finetuning-based defense, and ASIDE, a defense that, like ours, builds instruction-data separation into the model architecture. We show that HoS offers greater robustness to adaptive attacks, better performance on standard utility benchmarks, and matches top performance on standard agentic prompt injection benchmarks. More specifically, the black-box RL-Hammer attack reaches 97% ASR against GPT-OSS-20B and 85% against SecAlign, but only 10% against HoS. Claudini, a state-of-the-art white-box attack, strengthened against HoS through autoresearch, requires 3 times more FLOPs to reach the same ASR on HoS as it does on GPT-OSS-20B or SecAlign. HoS preserves base-model utility across standard multiple-choice benchmarks and prompting setups, whereas SecAlign and ASIDE degrade in few-shot settings. On agentic benchmarks with static attacks, HoS and SecAlign perform similarly and substantially improve the security-utility trade-off over the base model. ASIDE struggles with tool use, achieving only 10% utility. These results show that prompt injection robustness can be achieved and improved through targeted attention interventions.
Nikita Afonin, Nikita Andriianov, Vahagn Hovhannisyan, Nikhil Bageshpura, Kyle Liu Algoverse AI Research, Kevin Zhu Algoverse AI Research, Sunishchal Dev RAND Corporation and Algoverse Coding Academy LLC, Ashwinee Panda TogetherAI and University of Maryland, College Park 等 12 位作者
Recent work has shown that narrow finetuning can produce broadly misaligned LLMs, a phenomenon termed emergent misalignment (EM). While concerning, these findings were limited to finetuning and activation steering, leaving out in-context learning (ICL). We therefore ask: does EM emerge in ICL? We find that it does: across four model families (Gemini, Kimi-K2, Grok, and Qwen), narrow in-context examples cause models to produce misaligned responses to benign, unrelated queries. With 16 in-context examples, EM rates range from 1% to 24% depending on model and domain, appearing with as few as 2 examples. Neither larger model scale nor explicit reasoning provides reliable protection, and larger models are typically even more susceptible. Next, we formulate and test a hypothesis, which explains in-context EM as conflict between safety objectives and context-following behavior. Consistent with this, instructing models to prioritize safety reduces EM while prioritizing context-following increases it. These findings establish ICL as a previously underappreciated vector for emergent misalignment that resists simple scaling-based solutions.
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
Lung-Hao Lee National Yang Ming Chiao Tung University, Liang-Chih Yu Yuan Ze University, Natalia V Loukachevitch, Ilseyar Alimova Kazan Federal University, Alexander Panchenko S-NLP group, Tzu-Mi Lin, Zhe-Yu Xu, Jian-Yu Zhou National Yang Ming Chiao Tung University 等 16 位作者
Aspect-Based Sentiment Analysis (ABSA) focuses on extracting sentiment at a fine-grained aspect level and has been widely applied across real-world domains. However, existing ABSA research relies on coarse-grained categorical labels (e.g., positive, negative), which limits its ability to capture nuanced affective states. To address this limitation, we adopt a dimensional approach that represents sentiment with continuous valence–arousal (VA) scores, enabling fine-grained analysis at both the aspect and sentiment levels. To this end, we introduce DimABSA, the first multilingual, dimensional ABSA resource annotated with both traditional ABSA elements (aspect terms, aspect categories, and opinion terms) and newly introduced VA scores. This resource contains 76,958 aspect instances across 42,590 sentences, spanning six languages and four domains. We further introduce three subtasks that combine VA scores with different ABSA elements, providing a bridge from traditional ABSA to dimensional ABSA. Given that these subtasks involve both categorical and continuous outputs, we propose a new unified metric, continuous F1 (cF1), which incorporates VA prediction error into standard F1. We provide a comprehensive benchmark using both prompted and fine-tuned large language models across all subtasks. Our results show that DimABSA is a challenging benchmark and provides a foundation for advancing multilingual dimensional ABSA. We publicly released the DimABSA dataset, which was used for Track A of SemEval-2026 Task 3, attracting over 300 participants.
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
Mikhail Seleznyov AXXX, Daniil Korbut Amazon, Viktor Moskvoretskii EPFL - EPF Lausanne, Oleg Somov DS-NLP Group, Moscow Independent Research Institute of Artificial Intelligence and AXXX, Alexander Panchenko S-NLP group, Elena Tutubalina
Uncertainty quantification (UQ) methods for large language models are predominantly designed by hand based on domain knowledge and heuristics, limiting their scalability and generality. We apply LLM-powered evolutionary search to automatically discover unsupervised UQ methods represented as Python programs. On the task of atomic claim verification, our evolved methods outperform strong manually-designed baselines, achieving up to 6.7% relative ROC-AUC improvement across 9 datasets while generalizing robustly out-of-distribution. Qualitative analysis reveals that different LLMs employ qualitatively distinct evolutionary strategies: Claude models consistently design high-feature-count linear estimators, while Gpt-oss-120B gravitates toward simpler and more interpretable positional weighting schemes. Surprisingly, only Sonnet 4.5 and Opus 4.5 reliably leverage increased method complexity to improve performance – Opus 4.6 shows an unexpected regression relative to its predecessor. Overall, our results hint that LLM-powered evolutionary search is a promising paradigm for automated, interpretable hallucination detector design.
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 4: Student Research Workshop) · 2026 · 正式发表
Anna Borisiuk DS-NLP Group, Andrey Savchenko HSE University, Alexander Panchenko S-NLP group, Elena Tutubalina
Machine Unlearning is a valuable ability of LLMs, enabling the removal of unsafe, outdated, or private information. Existing unlearning methods, however, are often evaluated under the assumption that all facts are equally challenging to forget. Controllable knowledge removal is essential for reliable NLP systems. In this paper, we investigate whether fact popularity influences the efficiency of LLM unlearning. To answer this question, we build UNLamb benchmark designed to systematically investigate this relationship. It consists of 11.6k question-answer pairs derived from real-world knowledge in Wikidata, explicitly partitioned into rare and popular facts. Using this benchmark, we perform a comprehensive evaluation of state-of-the-art unlearning algorithms on a set of models of different sizes. We conduct a comprehensive analysis of four unlearning methods across three validation sets and two LLMs. We show that larger models struggle more to forget popular entities, often damaging related knowledge in the process. In contrast, it is much easier to remove rare facts without side effects.
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 4: Student Research Workshop) · 2026 · 正式发表
Elisei Rykov CompSem group, Nikolay Ivanov, Maria Bandulevich, Kseniia Petrushina, Valentin Malykh International IT University, Vasily Konovalov, Alexander Panchenko S-NLP group, Ilseyar Alimova Kazan Federal University
Frequent revisions of complex regulatory documents in large organizations often introduce inconsistencies and contradictions that are difficult for lawyers and auditors to detect manually. Existing tools rely on character-level diffs and therefore miss paraphrases and semantic shifts. We introduce LegDiff, a novel benchmark for evaluating span-aware semantic comparison of legal texts, and use it to investigate the ability of large language models to detect semantic changes beyond token- and character-level matching. LegDiff comprises manually annotated pairs of legal paragraphs drawn from different documents. In addition, we present a pipeline to generate synthetic training data that aligns with the manual annotations and mirrors the structure and label distribution of the manually curated benchmark, and a visualization tool for clearly displaying detected differences and inconsistencies. The dataset, code, and a visualization tool are publicly available to facilitate reproducibility and further research (https://github.com/s-nlp/SLeDoC).
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 4: Student Research Workshop) · 2026 · 正式发表
Maria Marina, Daniil Moskovskiy, Sergey Pletenev Applied AI Institute, Mikhail Salnikov Independent, Alexander Panchenko S-NLP group, Viktor Moskvoretskii EPFL - EPF Lausanne
Self-consistency improves large language models by sampling multiple reasoning paths and selecting the most frequent answer, but majority vote often fails to recover correct answers that are already present among samples. In this work, we reformulate answer selection in self-consistency as a ranking problem. Instead of relying on a single uncertainty or confidence signal, we train a lightweight reranker to score candidate answers using five carefully designed features that capture answer-level frequency, semantic centrality, and reasoning-trace consistency. We instantiate this approach with a LambdaRank model and evaluate it on three datasets under a range of test-time budgets. Across datasets, the proposed method consistently achieves a better accuracy-efficiency trade-off than standard self-consistency and strong baselines, with particularly large gains on question answering benchmarks. Further analysis shows that the proposed features are individually useful and, more importantly, complementary, highlighting the value of learning to combine multiple informative signals for test-time answer selection.
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 4: Student Research Workshop) · 2026 · 正式发表
Representation Engineering: A Top-Down Approach to AI Transparency
Andy Zou, Long Phan, Sarah Chen, James Campbell, Phillip Guo, Richard Ren, Alexander Pan, Xuwang Yin 等 21 位作者
AI 速览：该论文提出并刻画了表征工程（RepE）这一新兴领域，以群体级表征为中心，为理解和控制深度神经网络中的高层认知现象提供新方法。
In this paper, we identify and characterize the emerging area of representation engineering (RepE), an approach to enhancing the transparency of AI systems that draws on insights from cognitive neuroscience. RepE places population-level representations, rather than neurons or circuits, at the center of analysis, equipping us with novel methods for monitoring and manipulating high-level cognitive phenomena in deep neural networks (DNNs). We provide baselines and an initial analysis of RepE techniques, showing that they offer simple yet effective solutions for improving our understanding and control of large language models. We showcase how these methods can provide traction on a wide range of safety-relevant problems, including honesty, harmlessness, power-seeking, and more, demonstrating the promise of top-down transparency research. We hope that this work catalyzes further exploration of RepE and fosters advancements in the transparency and safety of AI systems.