Sam Agnoli, Brigid Reilly, Lily Barna, Emőke-Ágnes Horvát
Google's AI Overviews (AIOs) are the first result users encounter when searching the web, shifting the work of selecting and evaluating sources from the user to Gemini. This may be especially consequential for health information, given that Google's traditional algorithm was designed to prioritize quality sources, and inaccurate information carries health risks. We evaluated 3,611 benchmark health questions, comparing each AIO against that question's top five traditional results on source selection and quality. We also evaluated the AIOs' statement grounding in their cited sources and how accurate AIOs were against physician-written rubrics. Overviews appeared for 96.4% of questions but drew on substantially different evidence than traditional search: only 44.4% of the top five results were cited by the corresponding AIO, which instead overrepresented commercial health publishers (e.g., WebMD, Healthline) at the expense of academic and government sources (e.g., Cleveland Clinic, National Health Service). AIO-cited pages scored lower than traditional results on four of six well-established credibility criteria. The grounding of AIO statements to their sources was poor: after correcting for classifier error, an estimated 56.7% of cited statements were supported by their cited page, and a further 20.8% of all statements were factual claims with no citation at all. When AIOs were graded against physicians' rubrics, they earned 38.2% of all available points (e.g., accuracy 54.2%, completeness 21.6%), and among the subset of AIOs whose physician rubric anticipated a specific inaccurate or potentially harmful claim, 22.2% of AIOs contained the harmful content. Notably, the accuracy of AIOs was unrelated to the grounding of their statements. We recommend AIO literacy interventions for users and holding AIOs at least to Google's own health-ranking standards.
Henrique Schechter Vera, Sahil Dua, Jovan Andonov, Sara Smoot, Sebastian Russo, Ivan Llanos, Max Matveev, Jae Hyeon Yoo 等 73 位作者
We present EmbeddingGemma 2, an open multimodal embedding model which natively represents text, images, audio, video, and combinations thereof in a unified space. Our approach introduces improvements such as simplified per-layer embeddings and a multi-stage curriculum with modality and objective progression which ensure strong quality across unimodal, cross-modal, and mixed-modal evaluations. Notably, in comprehensive embedding benchmarks across all modalities, EmbeddingGemma 2 establishes a new Pareto frontier, outperforming both specialist models and models two to twelve times larger. Furthermore, our design offers flexibility for resource-constrained environments, supporting selective modality encoder loading, adjustable vision token budgets, truncatable embedding dimensions, and quantized weights. We present ablations analyzing key decisions and release EmbeddingGemma 2 to promote further research.
Alejandra Zambrano, Sara Vera Marjanovic, Imene Kerboua, Xing Han Lù, Leila Kosseim
Despite recent advances, LLM-based web agents still struggle with limited exploration, omission of critical steps, and sensitivity to task constraints. Prior work suggests that many of these failures stem from weaknesses in planning, yet the impact of alternative natural language plan representations remains unexplored. To address this, we introduce PlanAhead, a static planner-executor framework that evaluates the impact of plan representation on agent performance. We first automatically categorize WebArena tasks into 3 difficulty levels, enabling consistent difficulty grading without human annotation. Then we systematically evaluate 4 different plan representations on the tasks categorized as hard: sequential subgoals, narrative, pseudocode, and checklist; across different families of multimodal LLM-powered agents(OpenAI, Alibaba, and Google). To account for stochastic variability, we introduce a novel decomposition of agent Success Rate (SR) into 2 interpretable evaluation metrics: Achievement Rate (AR) and Solved-Task Consistency (STC). Our results show that both, the plan formulation and the underlying LLM generating the plan, significantly influence web-agent robustness and task success.
Tosin Adewumi, Martin Karlsson, Marcus Liwicki, Mikael Sjödahl, Lama Alkhaled, Rihab Gargouri, Nudrat Habib
This work presents a systematic review of the application of natural language processing (NLP) along the entire battery life cycle, instead of the review of one stage or method, and introduces a novel conceptual technical language processing (TLP) framework for the European Union (EU)'s proposed digital battery passport (DBP) and other general battery predictions, as the main objective. It follows the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) method and employs three reputable databases or search engines, including Google Scholar, Institute of Electrical and Electronics Engineers Xplore (IEEE Xplore), and Scopus, searched in October, 2025. The search criteria are 'text', 'language processing', and 'battery' while the exclusion criteria are non-relevant papers and those published before 2017, which was applied to all databases, except IEEE Xplore because it turned up fewer results. Consequently, it assesses 274 scientific papers before the critical review of the final 66 relevant papers. We publicly provide artifacts of the review for validation and possible reproducibility. The findings show that new NLP tasks are emerging in the battery domain, which facilitate materials discovery and other stages of the life cycle. Notwithstanding, challenges remain, such as the lack of standard benchmarks. The proposed TLP framework, which incorporates agentic AI and optimized prompts, may be apt for tackling some of the challenges.
We present Agaram 2.5B, a Tamil-first language model trained from scratch that converses in Tamil script, romanized Tanglish and English, and calls tools. Our first model, Agaram 1.5B, guided the new base, tokenizer and fine-tuning recipe: we pretrained a new 2.53B-parameter base on 50 billion tokens, 58% of them Tamil, with a tokenizer that needs 1.55 tokens per Tamil word where Qwen2.5's needs 9.44. A four-day public preview of our first 2.5B chat model showed where it fell short with real users, and we rebuilt supervised fine-tuning around what we saw, on 222,662 records. The released model scores above Sarvam-1 in all six translation directions we measured (FLORES chrF 56.9 into Tamil and 55.4 into English), chooses the expected tool on all 25 prompts of a tool set that was used during checkpoint selection and whose schemas appear in training, with exact arguments on 16, and was preferred over the preview model by an LLM judge on 100 sealed human-written prompts (net preference +30 points, 95% interval +14.5 to +45.0). On FLORES it also scores above Google's Gemma 4 E4B-it and our fine-tuned variant of it in both directions, with 2.53B parameters against E4B's 4.5B effective parameters. Factual knowledge and reliable refusal of harmful requests remain open problems, which we set out as future work. The base and chat weights and GGUF files are released under Apache-2.0, and this report gives a full account of the data, training and evaluation as a playbook for building language models for under-represented languages.
We measure the emotion geometry of six LLMs spanning Google's Gemma 2 family. Using an extraction pipeline established by Sofroniew et al. (2026) on Claude Sonnet 4.5 and replicated by Hollowell (2026) on Gemma 2 27B base, we compare emotion geometry across base and instruction-tuned variants of Gemma 2 2B, 9B, and 27B. First, all six model cells exhibit the core emotion geometry. A linear map fit between two cells places held-out emotions nearest their own vectors in the target model. Across all fifteen model pairs, fit in one direction each, this failed exactly once, and the median rank was 1 in every pair. Linear maps refit on cleaned embedding-table vectors recover every held-out emotion at rank 1 in all thirty ordered pairs. Second, scale spreads the geometry's variance across more directions without weakening its tie to valence. The leading component's share of variance falls from 31.6 percent (2B) to 29.8 (9B) to 26.7 (27B) in the base models and from 31.9 (2B) to 29.2 (9B) to 27.3 (27B) in the tuned ones, while its correlation with human valence ratings stays between .71 and .78 in all six models. Third, we project each emotion vector onto the output vocabulary and ask at what depth the vector's own emotion word enters the top 10,000 of the 256,000 tokens. In the 2B and 9B models the median emotion word is there from layer 0, while in the 27B models it begins near chance, stays outside the top 10,000 for half the network, and enters at layer 23 in one step. Fourth, tuning leaves every registered analysis-layer measure within its floor. At every size, valence correlation rises and arousal correlation falls by amounts within their floors, the latter against prediction. PC2's variance share also falls, with a pooled decrease significant after correction. Mid-late layer stability exceeds its floor and traces to base 27B's final layer. Code, data, and audit reports are released.
Hang Jiang, Nanxi Liu, Tiantian Wang, Cassandra Overney, Rukun Zhang, Artemisia Luk, Diyi Yang, Deb Roy 等 9 位作者
Given that the Chinese are one of the oldest ethnic groups to immigrate to the US, their cuisine have become central to the US dining scene and an interesting avenue to study how Chinese immigrants have culturally adapted into American life. In that regard, previous research has focused on analyzing the English names that Chinese restaurants adopt, e.g., finding the use of familiar and typical words. To investigate how Chinese immigrants have culturally integrated in the US depending on social, geographic and demographic contexts, we investigate in this work the under-explored relationship between the Chinese and English names of American Chinese restaurants. Specifically, we conduct a study on 3,172 Chinese restaurants (10% of the total on Google Maps) sampled across the US. We present 8-dimension name frames to capture the underlying meanings behind names and create the NameFrames dataset containing 884 annotated parallel Chinese/English names. We also built an NLP pipeline to analyze the semantic discrepancies between names with both frame-based and embedding-based distance metrics. Our frame analysis shows that English names more often emphasize Location and Specialty, while Chinese names more often emphasize Positivity and Ambiance. Furthermore, we observe that for Chinese restaurants located in urban areas, those in counties with a higher concentration of Chinese population, or those that are relatively new, the discrepancies between the English and Chinese names are noticeably larger. Finally, interviews with 10 Chinese restaurant owners reveal additional interesting insights regarding naming conventions.
Firoj Alam Qatar Computing Research Institute, Fatema Ahmad Hamad Bin Khalifa University, Ali Ezzat Shahroor Hamad Bin Khalifa University, Mohamed Bayan Kmainasi University of Qatar, Elisa Sartori, Giovanni Da San Martino University of Padua, Abul Hasnat, Raian Ali Hamad Bin Khalifa University
Misinformation on social media undermines informed decision-making and public trust. Prebunking offers a proactive complement by helping users recognize manipulation tactics before they encounter them in the wild. We present CritiSense, a mobile media-literacy app that builds these skills through short, interactive challenges with instant feedback. It is the first multilingual (supporting nine languages) and modular platform, designed for rapid updates across topics and domains. We report a usability study with 93 users: 83.9% expressed overall satisfaction and 90.1% rated the app as easy to use. Qualitative feedback indicates that CritiSense helps improve digital literacy skills. Overall, it provides a multilingual prebunking platform and a testbed for measuring the impact of microlearning on misinformation resilience. Over 6 months, we have reached 500+ active users. It is freely available on the Apple App Store and Google [Play Store](https://play.google.com/store/apps/details?id=com.critisense hl=en).
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 3: System Demonstrations) · 2026 · 正式发表
Samuel Joseph Amouyal School of Computer Science, Tel Aviv University, Aya Meltzer-Asscher Tel Aviv University, Jonathan Berant Google and Tel Aviv University
Large language models (LLMs) that fluently converse with humans are a reality - but do LLMs experience human-like processing difficulties? We systematically compare human and LLM sentence comprehension across seven challenging linguistic structures. We collect sentence comprehension data from humans and five families of state-of-the-art LLMs, varying in size and training procedure in a unified experimental framework. Our results show LLMs overall struggle on the target structures, but especially on garden path (GP) sentences. Indeed, while the strongest models achieve near perfect accuracy on non-GP structures (93.7% for GPT-5), they struggle on GP structures (46.8% for GPT-5). Additionally, when ranking structures based on average performance, rank correlation between humans and models increases with parameter count. For each target structure, we also collect data for their matched baseline without the difficult structure. Comparing performance on the target vs. baseline sentences, the performance gap observed in humans holds for LLMs, with two exceptions: for models that are too weak performance is uniformly low across both sentence types, and for models that are too strong the performance is uniformly high. Together, these reveal convergence and divergence in human and LLM sentence comprehension, offering new insights into the similarity of humans and LLMs.
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
Ding Xia The University of Tokyo, Xinyue Gui, Mark Colley University College London, University of London, Fan Gao, Zhongyi Zhou Google, Dongyuan Li The University of Tokyo and Institute of Science Tokyo, Renhe Jiang The University of Tokyo, Takeo Igarashi The University of Tokyo, Tokyo Institute of Technology
Automated vehicles lack natural communication channels with other road users, making external Human-Machine Interfaces (eHMIs) essential for conveying intent and maintaining trust in shared environments. However, most eHMI studies rely on developer-crafted message-action pairs, which are difficult to adapt to diverse and dynamic traffic contexts. A promising alternative is to use Large Language Models (LLMs) as action designers that generate context-conditioned eHMI actions, yet such designers lack perceptual verification and typically depend on fixed prompts or costly human-annotated feedback for improvement.We present See2Refine, a human-free, closed-loop framework that uses vision-language models (VLMs) for perceptual evaluation as automated visual feedback to improve an LLM-based eHMI action designer. Given a driving context and a candidate eHMI action, the VLM evaluates the perceived appropriateness of the action, and this feedback is used to iteratively revise the designer’s outputs, enabling systematic refinement without human supervision.We evaluate our framework across three eHMI modalities (lightbar, eyes, and arm) and multiple LLM model sizes. Across settings, our framework consistently outperforms prompt-only LLM designers and manually specified baselines in both VLM-based metrics and human-subject evaluations. The results further indicate that the improvements are generalized across modalities and that VLM evaluations are reasonably aligned with human preferences in our controlled settings, supporting the robustness and effectiveness of See2Refine for scalable action design.
尚无 AI 解读，展示原文摘要
cs.CL · cs.CV
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Minda Zhao, Yilun Du Google, Mengyu Wang Harvard University
As large language models (LLMs) transition from chat interfaces to integral components of stochastic pipelines and systems approaching general intelligence, the ability to faithfully sample from specified probability distributions has become a functional requirement rather than a theoretical curiosity. We present the first large-scale, statistically powered audit of native probabilistic sampling in frontier LLMs, benchmarking 11 models across 15 distributions. To disentangle failure modes, we employ a dual-protocol design: Batch Generation, where a model produces N{=}1000 samples within one response, and Independent Requests, comprising N{=}1000 stateless calls. We observe a sharp protocol asymmetry: batch generation achieves only modest statistical validity, with a 7% median pass rate, while independent requests collapse almost entirely, with 10 of 11 models passing none of the distributions. Beyond this asymmetry, we reveal that sampling fidelity degrades monotonically with distributional complexity and aggravates as the sampling horizon N increases. Finally, we demonstrate how the propagation of these failures into downstream real-world application tasks introduces systematic biases: models fail to enforce uniform answer-position constraints in Multiple Choice Question generation and systematically violate demographic targets in attribute-constrained text-to-image prompt synthesis. These findings indicate that current LLMs lack a functional internal sampler, necessitating external tools for applications requiring statistical guarantees.
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
Eunsu Kim, Junyeong Park Korea Advanced Institute of Science & Technology, Juhyun Oh Korea Advanced Institute of Science & Technology, Kiwoong Park Korea Advanced Institute of Science & Technology, Seyoung Song KAIST, A. Seza Doğruöz Ghent University, Alice Oh Google and Korea Advanced Institute of Science and Technology, Najoung Kim Boston University
As LLMs are increasingly deployed in real-world interactions, their social reasoning in interpersonal communication becomes critical. To explore their capabilities, we introduce SCRIPTS, a 1.1k-dialogue dataset in English and Korean, sourced from movie scripts and propose a social reasoning task based on SCRIPTS that evaluates the capacity of LLMs to infer the social relationships (e.g., friends, lovers) between speakers in each dialogue. Evaluating nine models on our task, current LLMs achieve around 75–80% on the English dataset and 58–69% in Korean, and models predict an Unlikely relationship in 10–25% of responses in both languages.Furthermore, we find that thinking models and chain-of-thought prompting provide minimal benefits for social reasoning and occasionally amplify social biases.In sum, there are significant limitations in current LLMs’ social reasoning capabilities, especially for Korean, highlighting the need for efforts to develop socially-aware LLMs across languages.
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
Dawei Zhu, Rui Meng Google Cloud AI Research, Jiefeng Chen Google, Sujian Li Peking University, Tomas Pfister Google, Jinsung Yoon Google
Comprehending long visual documents, where information is distributed across extensive pages of text and visual elements, is a critical but challenging task for modern Vision-Language Models (VLMs). Existing approaches falter on a fundamental challenge: evidence localization. They struggle to retrieve relevant pages and overlook fine-grained details within visual elements, leading to limited performance and model hallucination. To address this, we propose DocLens, a tool-augmented multi-agent framework that effectively “zooms in” on evidence like a lens. It first navigates from the full document to specific visual elements on relevant pages, then employs a sampling-adjudication mechanism to generate a single, reliable answer. Paired with Gemini-2.5-Pro, DocLens achieves state-of-the-art performance on MMLongBench-Doc and FinRAGBench-V, surpassing even human experts. The framework’s superiority is particularly evident on vision-centric and unanswerable queries, demonstrating the power of its enhanced localization capabilities.
尚无 AI 解读，展示原文摘要
cs.CL · cs.MA · cs.CV
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Guy Mor-Lan, Omer Goldman University of Cambridge, Matan Eyal Google, Adi Mayrav Gilady Research, Google, Sivan Eiger, Idan Szpektor Google, Avinatan Hassidim Bar-Ilan University and Google, Yossi Matias Google and Tel Aviv University 等 9 位作者
Multilingual large language models (LLMs) have minimized the fluency gap between languages. This advancement, however, exposes models to the risk of biased behavior, as knowledge and norms may propagate across languages. In this work we aim to quantify models’ inter- and intra-lingual biases, via their ability to answer locale-ambiguous questions. To this end, we present LocQA, a test set containing 2,156 questions in 12 languages, referring to various locale-dependent facts such as laws, dates, and measurements. The questions do not contain indications of the locales they relate to, other than the querying language itself. LLMs’ responses to LocQA locale-ambiguous questions thus reveal models’ implicit priors. We used LocQA to evaluate 32 models, and detected two types of structural biases. Inter-lingually, we show a global bias towards answers relevant to the US-locale, even when models are asked in languages other than English. Moreover, we discovered that this global bias is exacerbated in models that underwent instruction tuning, compared to their base counterparts. Intra-lingually, we show that when multiple locales are relevant for the same language, models act as demographic probability engines, prioritizing locales with larger populations. Taken together, insights from LocQA may help in shaping LLMs’ desired local behavior, and in quantifying the impact of various training phases on different kinds of biases.
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
Kushal Tatariya, Artur Kulmizev, Wessel Poelman KU Leuven, Esther Ploeger Utrecht University, Marcel Bollmann Linköping University, Johannes Bjerva Aalborg University, Jiaming Luo Google, Heather Lent Aalborg University, Aalborg University 等 9 位作者
Wikipedia’s perceived high quality and broad language coverage have established it as a fundamental resource in NLP. However, in recent years, such assumptions of high quality have become the subject of scrutiny in low-resource and multilingual contexts. In this study, we subject the entirety of non-English Wikipedia to a data filtering procedure typically reserved for noisy web-text — a process which removes a large percentage of the collection’s data. In analysing the removed data, we reveal numerous systematic quality issues, such as script and language contamination, repeated template and placeholder articles, and a high concentration of bot-generated content. We consolidate these findings into a 4-level quality ranking of Wikipedia, which shows strong correspondence with alternative quality measures and heuristics. Lastly, we evaluate the downstream impact of quality filtering in three practical language modelling scenarios, showing that models trained on filtered data largely match or outperform those trained on raw Wikipedia, with the largest gains observed for lower-quality language editions. Ultimately, our experiments serve as a first step in establishing quality-aware best practices for Wikipedia utilization in NLP, laying groundwork that can inform future dataset creation and curation efforts.
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
Zifan Jiang, Youngjoon Jang, Liliane Momeni Google, Gül Varol, Sarah Ebling University of Zurich, Andrew Zisserman Google DeepMind and University of Oxford
The goal of this work is to develop a universal approach for aligning subtitles (i.e., spoken language text with corresponding timestamps) to continuous sign language videos. Prior approaches typically rely on end-to-end training tied to a specific language or dataset, which limits their generality. In contrast, our method Segment, Embed, and Align (SEA) provides a single framework that works across multiple languages and domains. SEA leverages two pretrained models: the first to segment a video sequence into individual signs and the second to embed each sign video clip into a shared latent space with text. Alignment is subsequently performed with a lightweight dynamic programming procedure that runs efficiently on CPU within a minute, even for hour-long episodes. SEA is flexible and can adapt to a wide range of scenarios, utilizing resources from small lexicons to large continuous corpora. Experiments on four sign language datasets demonstrate state-of-the-art alignment performance, highlighting the potential of SEA to generate high-quality parallel data for advancing sign language processing.
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
Gouki Minegishi, Hiroki Furuta Google DeepMind, Takeshi Kojima The University of Tokyo, Yusuke Iwasawa The University of Tokyo, The University of Tokyo, Yutaka Matsuo The University of Tokyo
Emergent misalignment, where fine-tuning on narrow, non-harmful tasks induces harmful behaviors, poses a key challenge for AI safety in LLMs. Despite growing empirical evidence, its underlying mechanism remains unclear. To uncover the reason behind this phenomenon, we propose a mechanistic account based on the geometry of feature superposition. Because features are encoded in overlapping, fine-tuning that amplifies a target feature also unintentionally strengthens nearby harmful features in accordance with their similarity. We give a simple gradient-level derivation of this mechanism and empirically test it across multiple LLMs (Gemma-2 2B/9B/27B, LLaMA-3.1 8B, gpt-oss 20B). Using sparse autoencoders (SAEs), we identify features tied to misalignment-inducing data and to harmful behaviors, and show that they are geometrically closer to each other than features derived from non-inducing data. This trend generalizes across domains (e.g., health, career, legal advice). Finally, we show that a geometry-aware approach—filtering training samples nearest to toxic features—reduces misalignment by 34.5%, substantially outperforming random removal and achieving stronger mitigation than LLM-as-a-judge–based filtering. Our study explains emergent misalignment through feature superposition, providing a basis for understanding and mitigating this phenomenon.
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
Miaomiao Li The Chinese University of Hong Kong, Hao Chen Google DeepMind, Yang Wang Tencent Inc., Tingyuan Zhu Institute of Science Tokyo, Weijia Zhang, Kaijie Zhu, Kam-Fai Wong The Chinese University of Hong Kong, Jindong Wang William & Mary
Generating synthetic datasets via large language models (LLMs) has emerged as a promising approach to improve LLM performance.However, LLMs inherently reflect biases in their training data, leading to a critical challenge: when models are trained on synthetic data, they may propagate and amplify the inherent biases that can significantly impact fairness and robustness on downstream tasks—a phenomenon we term bias inheritance. This work presents the first systematic investigation in understanding, analyzing, and mitigating bias inheritance. We fine-tune LLMs with a combined dataset of real and LLM-augmented data with varied bias ratio as the proportion of augmented data. Through systematic experiments across 10 classification and generation tasks, we analyze how 6 different types of biases manifest. Our results indicate that bias inheritance harms downstream task performance in bias directly-related classification and generation tasks. Then, our analysis identifies three key misalignment factors: misalignment of values, group data, and data distributions. Based on these insights, we propose three mitigation strategies: token-based, mask-based, and loss-based approaches, which can work differently on various tasks and bias, indicating the substantial challenges to mitigate bias inheritance. We hope this work can provide insights to the research of LLM data augmentation.
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
Yupeng Hou University of California, San Diego, Jiacheng Li Facebook, Xiangjun Fu University of Pennsylvania, University of Pennsylvania, Zhankui He Google DeepMind, An Yan SalesForce, Xiusi Chen University of Illinois at Urbana-Champaign, Julian McAuley University of California, San Diego, University of California, San Diego
Feature engineering has long been central to recommender systems, yet effectively leveraging textual item features remains challenging. Recent advances in large language models (LLMs) have enabled their use as semantic encoders for recommendation, but their roles and behaviors in this setting are still not well understood. Prior studies often rely on general-purpose embedding benchmarks (e.g., MTEB) when selecting LLMs, overlooking the unique characteristics of recommendation tasks. To address this gap, we introduce BLaIR, a comprehensive benchmark for evaluating LLMs as semantic encoders in recommendation scenarios. We contribute (1) a new large-scale Amazon Reviews 2023 dataset with over 570 million reviews and 48 million items, (2) a unified benchmark covering sequential recommendation, collaborative filtering, and product search, and (3) a new complex-query product search task featuring both semi-synthetic and real-world evaluation datasets. Experiments with 11 leading LLMs show that their rankings on BLaIR show little correlation with MTEB, highlighting the unique challenges of semantic encoding in recommendation.
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
Guangya Wan, Mingyang Ling, Xiaoqi Ren, Rujun Han Google, Sheng Li Oracle and University of Virginia, Charlottesville, Zizhao Zhang Google
Long-horizon tasks that require sustained reasoning and multiple tool interactions remain challenging for LLM agents: small errors compound across steps, and even state-of-the-art models often hallucinate or lose coherence. We identify context management as the central bottleneck—extended histories cause agents to overlook critical evidence or become distracted by irrelevant information, thus failing to replan or reflect from previous mistakes. To address this, we propose COMPASS (Context-Organized Multi-Agent Planning and Strategy System), a lightweight hierarchical framework that separates tactical execution, strategic oversight, and context organization into three specialized components: (1) a Main Agent that performs reasoning and tool use, (2) a Meta-Thinker that monitors progress and issues strategic interventions, and (3) a Context Manager that maintains concise, relevant progress briefs for different reasoning stages. Across three challenging benchmarks—GAIA, BrowseComp, and Humanity’s Last Exam—COMPASS improves accuracy by up to 20% relative to both single- and multi-agent baselines. We further introduce a test-time scaling extension that elevates performance to match established DeepResearch agents, and a post-training pipeline that delegates context management to smaller models for enhanced efficiency.
尚无 AI 解读，展示原文摘要
cs.CL · cs.MA
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Wenjie Jacky Mo, Qin Liu University of California, Davis, Xiaofei Wen University of California, Davis, Dongwon Jung, Hadi Askari, Wenxuan Zhou Google Deepmind, Zhe Zhao University of California, Davis, Muhao Chen University of California, Davis
Large Language Models (LLMs) for code generation (i.e., Code LLMs) have demonstrated impressive capabilities in AI-assisted software development and testing. However, recent studies have shown that these models are prone to generating vulnerable or even malicious code under adversarial settings. Existing red-teaming approaches rely on extensive human effort, limiting their scalability and practicality, and generally overlook the interactive nature of real-world AI-assisted programming, which often unfolds over multiple turns. To bridge these gaps, we present RedCoder, a red-teaming agent that engages victim models in multi-turn conversation to elicit vulnerable code. The pipeline to construct RedCoder begins with a multi-agent gaming process that simulates adversarial interactions, yielding a set of prototype conversations and an arsenal of reusable attack strategies. We then fine-tune an LLM on these prototype conversations to serve as the backbone of RedCoder. Once deployed, RedCoder autonomously engages Code LLMs in multi-turn conversations, dynamically retrieving relevant strategies from the arsenal to steer the dialogue toward vulnerability-inducing outputs. Experiments across multiple Code LLMs show that our approach outperforms prior single-turn and multi-turn red-team methods in inducing vulnerabilities in code generation, offering a scalable and effective tool for evaluating the security boundaries of modern code-generation systems.
尚无 AI 解读，展示原文摘要
cs.CL · cs.MA
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Yu Fu University of California, Riverside, Chen Luo Amazon, Josef Valvoda Google, Xin Zhang, Xuejing Lei Amazon, Xiao Pan, Hui Liu Amazon, Yue Dong University of California, Riverside
Key-Value (KV) cache compression techniques have improved the efficiency of long-context summarization in Large Language Models (LLMs), but their impact on model hallucination remains underexplored. In this paper, we present the first systematic study of how KV cache compression affects hallucination in long-context summarization, demonstrating that aggressive compression can increase hallucination scores by up to 3.36× compared to the baseline. To mitigate this issue, we propose HalluKV, a decoding-phase strategy that selectively removes generated KV pairs from retrieval heads responsible for retrieving critical information from source context, thereby anchoring their attention on the preserved source information. Our approach maintains computational efficiency while significantly reducing hallucination across multiple models and datasets, achieving up to 5.48 average point reductions on Llama-3-8B-Instruct, enabling more trustworthy long-context summarization.
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
Parker Riley Google, Daniel Deutsch Google, Mara Finkelstein Google, Colten DiIanni Google, Juraj Juraska Google, Markus Freitag Google
Human evaluation of machine translation is in an arms race with translation model quality: as our models get better, our evaluation methods need to be improved to ensure that quality gains are not lost in evaluation noise. To improve annotation quality, we experiment with a two-stage version of the current state-of-the-art translation evaluation paradigm (MQM), which we call MQM re-annotation. In this setup, an annotator reviews and edits a set of prior MQM annotations that may have come from themselves, another human annotator, or an automatic system. We demonstrate that rater behavior in re-annotation aligns with our goals, and that re-annotation results in higher-quality annotations, mostly due to finding errors that were missed during the first pass.
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
Hyeonjeong Ha Department of Computer Science, University of Illinois at Urbana-Champaign, Qiusi Zhan, Jeonghwan Kim, Dimitrios Bralios, Saikrishna Sanniboina Microsoft, Nanyun Peng Google and University of California, Los Angeles, Kai-Wei Chang University of California, Los Angeles and Amazon, Daniel Kang 等 9 位作者
Retrieval-augmented generation (RAG) has become a common practice in multimodal large language models (MLLM) to enhance factual grounding and reduce hallucination. Yet, its reliance on retrieval exposes MLLMs to knowledge poisoning attacks, in which adversaries deliberately inject malicious multimodal content into external knowledge bases to steer models toward generating incorrect or even harmful responses. We present MM-PoisonRAG, a framework to systematically study the vulnerability of multimodal RAG under knowledge poisoning. Specifically, we design two novel attack strategies: Localized Poisoning Attack (LPA), which implants targeted, query-specific multimodal misinformation to manipulate outputs toward attacker-controlled responses, and Globalized Poisoning Attack (GPA), which uses a single, untargeted adversarial injection to broadly corrupt reasoning and collapse generation quality across all queries. Extensive experiments on diverse tasks, multimodal RAG components, and attacker access levels reveal severe vulnerabilities: LPA achieves up to 56% attack success rate even under restricted access, and transfers effectively across four different retrievers without re-optimizing the adversaries. GPA completely disrupts model generation to 0% accuracy with just one poisoned content. Moreover, both LPA and GPA bypass existing defenses, underscoring the fragility of multimodal RAG and establishing MM-PoisonRAG as a foundation for future research on securing RAG frameworks against multimodal knowledge poisoning.
尚无 AI 解读，展示原文摘要
cs.CL · cs.CV
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Yi Su, Dian Yu Tencent AI Lab, Linfeng Song Google, Juntao Li, Haitao Mi Tencent AI Lab, Zhaopeng Tu Tencent, Min Zhang Harbin Institute of Technology, Shenzhen and Harbin Institute of Technology, Dong Yu CapitalOne
Reinforcement learning with verifiable rewards (RLVR) has been effective on tasks with structured solutions like math and coding, but its reliance on simple, rule-based verifiers creates a fundamental bottleneck. We find their applicability is surprisingly narrow even in structured domains, a limitation that is compounded at scale: rule-based systems can paradoxically degrade in performance as multi-domain, free-form training data increases. To overcome these challenges, we propose a new RLVR framework that uses a generative verifier to provide soft, probabilistic rewards. Our key insight is that powerful LLMs show high agreement with human evaluators when judging answer correctness given a ground-truth reference, allowing us to automate reward generation without costly human annotation. Our experiments demonstrate the effectiveness of this approach. We show that a compact 7B generative reward model can guide a 7B policy model to decisively outperform models up to 10x its size, including the 72B Qwen2.5-Instruct (by a margin of +8.6%). This effectiveness is robust, holding true across diverse training datasets with answers sourced from experts, web users, and other LLMs, and generalizes strongly to seven out-of-distribution benchmarks. Our work provides a scalable and effective framework for extending RLVR beyond the limitations of pattern-based verification to complex, noisy, real-world domains.
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
Baichuan Huang, Ananth Balashankar Google DeepMind and Google, Amir Aminifar Lund University / Lund Institute of Technology
Fine-tuning the bias terms of large language models (LLMs) has the potential to achieve unprecedented parameter efficiency while maintaining competitive performance, particularly in low-data regimes. However, the link between fine-tuning different bias terms (i.e., \boldsymbol{b}_q, \boldsymbol{b}_k, \boldsymbol{b}_v in the query, key, or value projections) and downstream performance remains largely unclear to date. In this paper, we investigate the link between fine-tuning \boldsymbol{b}_q, \boldsymbol{b}_k, \boldsymbol{b}_v with the performance of the downstream task. Our key finding is that directly fine-tuning \boldsymbol{b}_v generally leads to higher downstream performance in low-data regimes, in comparison to \boldsymbol{b}_q and \boldsymbol{b}_k. We extensively evaluate this unique property across a wide range of LLMs spanning encoder-only and decoder-only architectures up to 6.7B parameters (including bias-free LLMs). Our results provide strong evidence for the effectiveness of directly fine-tuning \boldsymbol{b}_v across various downstream tasks. The implementation code is available at https://github.com/whubaichuan/BEFT.
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
Mohit Talreja Georgia Institute of Technology, Joshua Diao, Jim James, Radu Casapu Georgia Institute of Technology, Tejas Santanam Georgia Institute of Technology, Ethan Mendes, Alan Ritter Georgia Institute of Technology, Wei Xu Georgia Institute of Technology 等 9 位作者
Vision Language Models (VLMs) are good at recognizing the global location of a photograph – their geolocation prediction accuracy rivals the best human experts. But many VLMs are startlingly bad at explaining which image evidence led to their prediction, even when their location prediction is correct. The reasoning chains produced by VLMs frequently hallucinate scene attributes to support their location prediction (e.g. phantom writing, imagined infrastructure, misidentified flora). In this paper, we introduce the first benchmark for geolocation reasoning chains. We focus on the global location prediction task in the popular GeoGuessr game which draws from Google Street View spanning more than 100 countries. We collaborate with expert GeoGuessr players, including the reigning world champion, to produce 800 “ground truth” reasoning chains for 500 query scenes. These expert reasoning chains address hundreds of different discriminative visual attributes such as license plate shape, architecture, and soil properties to name just a few. We evaluate LLM-as-a-judge and VLM-as-a-judge strategies for scoring VLM-generated reasoning chains against our expert reasoning chains and find that Qwen 3 LLM-as-a-judge correlates best with human scoring. Our benchmark reveals that while large, closed-source VLMs such as Gemini and GPT 5 rival human experts at prediction locations, they still lag behind human experts when it comes to producing auditable reasoning chains. Open weights VLMs such as Llama and Qwen catastrophically fail on our benchmark – they perform only slightly better than a baseline in which an LLM hallucinates a reasoning chain with oracle knowledge of the photo location but no visual information at all. We believe the gap between human experts and VLMs on this task points to VLM limitations at extracting fine-grained visual attributes from high resolution images. We open source our benchmark for the community to use.
尚无 AI 解读，展示原文摘要
cs.CL · cs.CV
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Saujas Vaduguru Carnegie Mellon University, Yilun Hua Department of Computer Science, Cornell University, Yoav Artzi Google DeepMind, arXiv and Cornell University, Daniel Fried Meta AI and Carnegie Mellon University
Humans leverage shared conversational context to become increasingly successful and efficient at communicating over time. One manifestation of this is the formation of ad hoc linguistic conventions, which allow people to coordinate on short, less costly utterances that are understood using shared conversational context. We present a method to train large multimodal models to form conventions, enabling efficient communication. Our approach uses simulated reference games between models, and requires no additional human-produced data. In repeated reference games involving photographs and tangram images, our method enables models to communicate efficiently with people: reducing the message length by up to 41% while increasing success by 15% over the course of the interaction. Human listeners respond faster when interacting with our model that forms conventions. We also show that training based on success or cost alone is insufficient – both are necessary to elicit convention formation.
尚无 AI 解读，展示原文摘要
cs.CL · cs.CV
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Jiyeon Kim Korea Advanced Institute of Science & Technology, Hyunji Lee Department of Computer Science, University of North Carolina at Chapel Hill, Dylan Zhou Google, Sue Hyun Park KRAFTON Inc., Seunghyun Yoon Adobe Research, Trung Bui Adobe Research, Franck Dernoncourt Adobe Research, Sungmin Cha Meta Superintelligence Labs 等 9 位作者
LLMs operating in dynamic real-world contexts often encounter knowledge that evolves continuously or emerges incrementally. To remain accurate and effective, models must adapt to newly arriving information on the fly. We introduce Online Adaptation to Continual Knowledge Streams(OAKS) to evaluate this capability, establishing a benchmark for online adaptation over streaming, continually updating knowledge. Specifically, the benchmark is structured as a sequence of fine-grained context chunks where facts change dynamically across time intervals. OAKS comprises two datasets: OAKS-BABI and OAKS-Novel, where individual facts evolve multiple times across context chunks. These datasets include dense annotations to measure whether models track changes accurately. Evaluating 14 models with varied inference approaches, we observe significant limitations in current methodologies. Both state-of-the-art models and agentic memory systems fail to adapt robustly on OAKS, demonstrating delays in state-tracking and susceptibility to distraction within streaming environments. We will open-source the code and datasets.
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 1: Long Papers) · 2026 · 正式发表
Or David Shafran, Atticus Geiger Goodfire, Mor Geva Tel Aviv University and Google Research
A central goal for mechanistic interpretability has been to identify the right units of analysis in large language models (LLMs) that causally explain their outputs. While early work focused on individual neurons, evidence that neurons often encode multiple concepts has motivated a shift toward analyzing directions in activation space. A key question is how to find directions that capture interpretable features in an unsupervised manner. Current methods rely on dictionary learning with sparse autoencoders (SAEs), commonly trained over residual stream activations to learn directions from scratch. However, SAEs often struggle in causal evaluations and lack intrinsic interpretability, as their learning is not explicitly tied to the computations of the model. Here, we tackle these limitations by directly decomposing MLP activations with semi-nonnegative matrix factorization (SNMF), such that the learned features are (a) sparse linear combinations of co-activated neurons, and (b) mapped to their activating inputs, making them directly interpretable. Experiments on Llama 3.1, Gemma 2 and GPT-2 show that SNMF derived features outperform SAEs and a strong supervised baseline (difference-in-means) on causal steering, while aligning with human-interpretable concepts. Further analysis reveals that specific neuron combinations are reused across semantically-related features, exposing a hierarchical structure in the MLP’s activation space. Together, these results position SNMF as a simple and effective tool for identifying interpretable features and dissecting concept representations in LLMs.
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