ACL 2026 Papers

Verbal multiword expressions (VMWEs) remain difficult for machine translation because their meanings are often not recoverable from their component words. In this study, we analyze the impact of three VMWE categories—verbal idioms, verb-particle constructions, and light verb constructions—on machine translation quality from English to multiple languages. Using both established multiword expression datasets and standard machine translation datasets, we evaluate how state-of-the-art translation systems handle these expressions. Our experimental results consistently show that VMWEs negatively affect translation quality, with deeper analysis indicating that this degradation is primarily attributable to the VMWE itself rather than general sentence-level difficulty. We release our code and evaluation framework to test new MT systems for the community.

Verification is critical for improving agents: it provides the reward signal for Reinforcement Learning and enables inference-time gains through Test-Time Scaling (TTS). Despite its importance, verification in software engineering (SWE) agent settings often relies on code execution, which can be difficult to scale due to environment setup overhead. Scalable alternatives such as patch classifiers and heuristic methods exist, but they are less grounded in codebase context and harder to interpret. To this end, we explore Agentic Rubrics: an expert agent interacts with the repository to create a context-grounded rubric checklist, and candidate patches are then scored against it without requiring test execution. On SWE-Bench Verified under parallel TTS evaluation, Agentic Rubrics achieve a score of 54.2% on Qwen3-Coder-30B-A3B and 40.6% on Qwen3-32B, with at least a +3.5 percentage-point gain over the strongest baseline in our comparison set. We further analyze rubric behavior, showing that rubric scores are consistent with ground-truth tests while also flagging issues that tests do not capture. Our ablations show that agentic context gathering is essential for producing codebase-specific, unambiguous criteria. Together, these results suggest that Agentic Rubrics provide an efficient, scalable, and granular verification signal for SWE agents.

We study asymmetric vulnerability to embedding corruption in encoder-based language models, where uniform perturbations dis-proportionately degrade low-magnitude embeddings compared to high-magnitude ones. Since critical words concentrate in low-norm space, this asymmetry causes catastrophic degradation of grammatical and semantic structure even under moderate corruption. Existing robustness approaches either sacrifice clean performance or fail to generalize to higher corruption levels. To address this problem, we propose Robertha, an attention mechanism built on Modern Hopfield Networks, in which semantic patterns act as stable states (attractors) that pull corrupted embeddings toward correct representations. We introduce iterative refinement for differential recovery: heavily corrupted embeddings require multiple convergence steps, while lightly corrupted embeddings converge quickly. To strengthen this mechanism, we introduce Eigenspectrum Regularization (ESR), which enforces low-rank key structures by controlling eigenvalue entropy, creating strong, well separated attractors with wide recovery basins. Across 13 GLUE and SuperGLUE tasks, Robertha significantly outperforms existing robustness methods while maintaining competitive clean performance.

Misinformation is on the rise, and the strong writing capabilities of LLMs lower the barrier for malicious actors to produce and disseminate false information. We study how LLMs behave when prompted to spread misinformation across languages and target countries, and introduce GlobalLies, a multilingual parallel dataset of 440 misinformation generation prompt templates and 6,867 entities, spanning 8 languages and 195 countries. Using both human annotations and large-scale LLM-as-a-judge evaluations across hundreds of thousands of generations from state-of-the-art models, we show that misinformation generation varies systematically based on the country being discussed. Propagation of lies by LLMs is substantially higher in many lower-resource languages and for countries with a lower Human Development Index (HDI). We find that existing mitigation strategies provide uneven protection: input safety classifiers exhibit cross-lingual gaps, and retrieval-augmented fact-checking remains inconsistent across regions due to unequal information availability. We release GlobalLies for research purposes, aiming to support the development of mitigation strategies to reduce the spread of global misinformation: https://github.com/zohaib-khan5040/globallies

Evaluating cross-lingual knowledge transfer in large language models (LLMs) is challenging, as correct answers in a target language may arise either from genuine transfer or from prior exposure during pre-training. We present LiveCLKTBench, an automated generation pipeline specifically designed to isolate and measure cross-lingual knowledge transfer. Our pipeline identifies self-contained, time-sensitive knowledge entities from real-world domains, filters them based on temporal occurrence, and verifies them against the model’s knowledge. The documents of these valid entities are then used to generate factual questions, which are translated into multiple languages to evaluate transferability across linguistic boundaries. Using LiveCLKTBench, we evaluate several LLMs across five languages and observe that cross-lingual transfer is strongly influenced by linguistic distance and often asymmetric across language directions. While larger models improve transfer, the gains diminish with scale and vary across domains. These findings provide new insights into multilingual transfer and demonstrate the value of LiveCLKTBench as a reliable benchmark for future research.

Audio Large language models (LLMs) are increasingly deployed in the real world, where they inevitably capture speech from unintended nearby bystanders, raising privacy risks that existing benchmarks and defences did not consider. We introduce SH-Bench, the first benchmark designed to evaluate selective hearing: a model’s ability to attend to an intended main speaker while refusing to process or reveal information about incidental bystander speech. SH-Bench contains 3,968 multi-speaker audio mixtures, including both real-world and synthetic scenarios, paired with 77k multiple-choice questions that probe models under general and selective operating modes. In addition, we propose Selective Efficacy (SE), a novel metric capturing both multi-speaker comprehension and bystander-privacy protection. Our evaluation of state-of-the-art open-source and proprietary LLMs reveals substantial bystander privacy leakage, with strong audio understanding failing to translate into selective protection of bystander privacy. To mitigate this gap, we also present Bystander Privacy Fine-Tuning (BPFT), a novel training pipeline that teaches models to refuse bystander-related queries without degrading main-speaker comprehension. We show that BPFT yields substantial gains, achieving an absolute 47% higher bystander accuracy under selective mode and an absolute 16% higher SE compared to Gemini 2.5 Pro, which is the best audio LLM without BPFT. Together, SH-Bench and BPFT provide the first systematic framework for measuring and improving bystander privacy in audio LLMs.

The quadratic computational cost of traditional attention mechanisms poses a major bottleneck to the scalability and practical deployment of large language models (LLMs), particularly in long-context scenarios. To improve efficiency, existing approaches often enforce rigid structural constraints such as local attention windows. However, these strategies typically lead to substantial performance degradation on tasks requiring precise long-range recall. In this work, we propose MATCH, a scalable and efficient framework that augments sparsified attention mechanisms with dynamically integrated in-context information through an efficient retrieval system. Empirical results show that MATCH significantly improves the performance of sparse-attention models on both synthetic and real-world natural-language tasks. These findings highlight the versatility of MATCH as a general approach for enhancing in-context retrieval capabilities while maintaining the efficiency benefits of sparse attention architectures.

Large language models (LLMs) exhibit uneven performance across languages. In language-specific applications, practitioners often rely on target-language corpora or cross-lingual transfer to achieve better performance. However, traditional linguistic typology, commonly used as a transfer language selection strategy in previous studies, may not align with LLM’s perception of language similarity. This work proposes LLM-based language similarity as a novel perspective for selecting effective fine-tuning languages. We construct a framework to quantify the similarity within each language pair through both the lenses of language-specific performance patterns and cross-lingual transferability, ultimately deriving three similarity score matrices. Moreover, we observe a counter-intuitive phenomenon: super-additive transfer effect, where fine-tuning on a certain language yields higher performance than fine-tuning directly on the target language. Additionally, due to the absence of an existing dataset meeting our experimental requirements, we construct and release M4CQ-Pro dataset, which features domain-diverse distribution of 135 tasks and content consistency across 31 languages (including over 20 medium- and low-resource languages), with 61518 manually reviewed high-quality questions per language. We evaluate our approach on representative multilingual LLMs and results show that all three LLM-based similarity measures effectively guide fine-tuning language selection, outperforming traditional linguistic similarity, with the integrated measure achieving the best results. Our approach provides not only a novel perspective on language similarity, but also practical baselines for selecting fine-tuning languages.

Student mistakes in mathematics are often systematic: a learner applies a coherent but wrong procedure and repeats it across contexts. We introduce MalruleLib, a learning-science-grounded framework that translates documented misconceptions into executable procedures, drawing on 67 learning-science and mathematics education sources, and generates step-by-step traces of malrule-consistent student work. We formalize a core student-modeling problem as Malrule Reasoning Accuracy (MRA): infer a misconception from one worked mistake and predict the student’s next answer under cross-template rephrasing. Across nine language models (4B - 120B), accuracy drops from 66% on direct problem solving to 40% on cross-template misconception prediction. MalruleLib encodes 101 malrules over 498 parameterized problem templates and produces paired dual-path traces for both correct reasoning and malrule-consistent student reasoning. Because malrules are executable and templates are parameterizable, MalruleLib can generate over one million instances, enabling scalable supervision and controlled evaluation. Using MalruleLib, we observe cross-template degradations of 10 - 21%, while providing student step traces improves prediction by 3 - 15%. We release MalruleLib as infrastructure for educational AI that models student procedures across contexts, enabling diagnosis and feedback that targets the underlying misconception.

Do Large Language Models (LLMs) genuinely grasp the compositional semantics of events, or do they rely on surface-level probabilistic heuristics? We investigate the Imperfective Paradox, a logical phenomenon where the past progressive aspect entails event realization for activities (e.g., running \to ran) but not for accomplishments (e.g., building \nrightarrow built). We introduce ImperfectiveNLI, a diagnostic dataset designed to probe this distinction across diverse semantic classes. Evaluating state-of-the-art open-weight models, we uncover a pervasive Teleological Bias: models systematically hallucinate completion for goal-oriented events, even overriding explicit textual cancellation. Prompting interventions partially reduce this bias but trigger a calibration crisis, causing models to incorrectly reject valid entailments for atelic verbs. Representational analyses further show that while internal embeddings often distinguish progressive from simple past forms, inference decisions are dominated by strong priors about goal attainment. Taken together, our findings indicate that these current open-weight LLMs operate as predictive narrative engines rather than faithful logical reasoners, and that resolving aspectual inference requires moving beyond prompting toward structurally grounded alignment.

The National Violent Death Reporting System (NVDRS) documents suicides in the United States. In a demanding public health data pipeline, annotators manually extract structured information from death investigation records following extensive codebooks (i.e. annotation guidelines) painstakingly developed by experts. In this work, we facilitate data-driven insights from the NVDRS data to support the development of novel suicide interventions by leveraging language models (LM) as assistants to these (a) data annotators and (b) experts. We find that LM predictions match existing data annotations about 85% of the time across 50 NVDRS variables. Where the LM disagrees with existing annotations, our expert review identifies that 38% of these instances reveal inconsistencies between narratives and structured data. Finally, we introduce a human-in-the-loop algorithm that helps experts efficiently build and refine codebooks for new variables by having them only focus on providing feedback for incorrect LM predictions. We apply our algorithm to a real-world case study, and find that about 96K narratives contain evidence of victim interactions with legal professionals, which surfaces a substantial opportunity for upstream intervention that is not captured in the original structured data. Our findings provide evidence that LMs can serve as effective assistants to public health researchers who handle sensitive data in high-stakes scenarios.

Emotion is a central dimension of spoken communication, yet, we still lack a mechanistic account of how modern large audio-language models (LALMs) encode it internally. We present the first neuron-level interpretability study of emotion-sensitive neurons (ESNs) in LALMs and provide causal evidence supporting the existence of such units in Qwen2.5-Omni, Kimi-Audio, and Audio Flamingo 3. Across these three widely used open-source models, we compare frequency-, entropy-, mean-deviation-, and contrast-based neuron selectors on multiple emotion recognition benchmarks. Using inference-time interventions, we reveal a consistent emotion-specific signature: deactivating neurons selected for a given emotion disproportionately degrades recognition of that emotion while largely preserving other classes, whereas targeted steering amplifies these units to bias predictions toward the target emotion. These effects arise with modest amounts of identification data and scale systematically with intervention strength. We further observe that ESNs exhibit non-uniform layer-wise clustering with partial cross-dataset transfer. Taken together, our results offer a causal, neuron-level account of emotion decisions in LALMs and highlight targeted neuron interventions as an actionable handle for controllable affective behaviors.

Grammaticality and likelihood are distinct notions in human language. Pretrained language models (LMs), which are probabilistic models of language fitted to maximize corpus likelihood, generate grammatically well-formed text and discriminate well between grammatical and ungrammatical sentences in tightly controlled minimal pairs. However, their string probabilities do not sharply discriminate between grammatical and ungrammatical sentences overall. But do LMs implicitly acquire a grammaticality distinction distinct from string probability? We explore this question through studying pretrained LMs’ internal representations, by training a linear probe on a dataset of grammatical and (synthetic) ungrammatical sentences obtained by applying perturbations to a naturalistic text corpus. We find that this simple grammaticality probe generalizes to human-curated grammaticality judgment benchmarks and outperforms LM probability-based grammaticality judgments. When applied to semantic plausibility benchmarks, however, in which both members of a minimal pair are grammatical and differ in only plausibility, the probe performs worse than string probability. The English-trained probe also exhibits nontrivial cross-lingual generalization, outperforming string probabilities on grammaticality benchmarks in numerous other languages. Additionally, probe scores correlate only weakly with string probabilities. These results collectively suggest that pretrained LMs acquire to some extent an implicit grammaticality distinction within their hidden layers.

Evaluating multimodal large language models (MLLMs) is becoming increasingly expensive as benchmarks grow in scale and cross-modality complexity. Inspired by structuralism in cognitive psychology, we tackle this difficulty with an adaptive evaluation framework for efficient benchmarking, namely AutoJudger. Instead of passively scoring on a fixed test set, AutoJudger treats evaluation as an interview-like process by keeping a hypothesized ability structure of the evaluated model and actively selecting the informative questions so as to refine these ability boundaries. Specifically, AutoJudger has three core components: ability decomposition to organize evaluation along meaningful capability dimensions, ability estimation to maintain an up-to-date quantitative profile of the model competence, and adaptive question selection to choose the most informative questions. To operationalize this paradigm, we introduce A^2-Judger, a novel MLLM-based Agentic instantiation of AutoJudger equipped with semantic-aware retrieval and dynamic memory. Experiments on four representative multimodal benchmarks show that A^2-Judger significantly improves sample efficiency while maintaining reliable evaluation results.

Motion sensor time-series are central to Human Activity Recognition (HAR), yet conventional approaches are constrained to fixed activity sets and typically require costly parameter retraining to adapt to new behaviors. While Large Language Models (LLMs) offer promising open-set reasoning capabilities, applying them directly to numerical time-series often leads to hallucinations and weak grounding. To address this challenge, we propose ZARA (Zero-training Activity Reasoning Agents), a knowledge- and retrieval-augmented agentic framework for motion time-series reasoning in a training-free inference setting. Rather than relying on black-box projections, ZARA distills reference data into a statistically grounded textual knowledge base that transforms implicit signal patterns into verifiable natural-language priors. Guided by retrieved evidence, ZARA iteratively selects discriminative cues and performs grounded reasoning over candidate activities. Extensive experiments on eight benchmarks show that ZARA generalizes robustly to unseen subjects and across datasets, demonstrating strong transferability across heterogeneous sensor domains. These results mark a step toward trustworthy, plug-and-play motion understanding beyond dataset-specific artifacts. Our code is available at https://github.com/zechenli03/ZARA.

While model-based verifiers are essential for scaling Reinforcement Learning with Verifiable Rewards (RLVR), current outcome-centric verification paradigms primarily focus on the consistency between the final result and the ground truth, often neglecting potential errors in the derivation process. This leads to assigning positive rewards to correct answers produced from incorrect derivations. To bridge this gap, we introduce PRIME, a benchmark for evaluating verifiers on PRocess-outcome alignment verification In Mathematics and Engineering. Curated from a comprehensive collection of college-level STEM problems, PRIME comprises 2,530 high-difficulty samples through a consistency-based filtering pipeline. Through extensive evaluation, we find that current verifiers frequently fail to detect derivation flaws. Furthermore, we propose a process-aware RLVR training paradigm utilizing verifiers selected via PRIME. This approach substantially outperforms the outcome-only verification baseline, achieving absolute performance gains of 8.29%, 9.12%, and 7.31% on AIME24, AIME25, and Beyond-AIME, respectively, for the Qwen3-14B-Base model. Finally, we demonstrate a strong linear correlation (R^2 > 0.92) between verifier accuracy on PRIME and RLVR training effectiveness, validating PRIME as a reliable predictor for verifier selection.

Despite recent progress in context compression, we identify a fundamental memorization-utilization gap where models can compress context with near-perfect fidelity yet fail to effectively utilize these compressed representations for downstream tasks. We address this with a holistic training paradigm spanning pretraining, instruction tuning, and reinforcement learning, built upon an average pooling compression. Our key innovation uses outcome-based RL to enable implicit expansion: the model learns to adaptively unfold task-relevant details during generation, interleaving reconstruction with reasoning. We achieve near-lossless 16x context compression (≈5.3x decoder sequence-length reduction in our current implementation) across 7B and 32B models, recovering over 98% of full-context QA performance and outperforming prior methods by 11 points. Our 32B model demonstrates strong out-of-distribution and length generalization, robustly scaling to 120k-token contexts despite training on no more than 4k tokens, matching full-context performance on NIAH, LongBench v2, and multi-hop reasoning. We verify the implicit expansion behavior in experiments.

The quality of text generated by large language models depends critically on the decoding sampling strategy. While mainstream methods such as Top-k, Top-p, and Min-p achieve a balance between diversity and accuracy through probability-space truncation, they share an inherent limitation: extreme sensitivity to the temperature parameter. Recent logit-space approaches like Top-n\sigma achieve temperature invariance but rely on global statistics that are susceptible to long-tail noise, failing to capture fine-grained confidence structures among top candidates. We propose Min-k Sampling, a novel dynamic truncation strategy that analyzes the local shape of the sorted logit distribution to identify “semantic cliffs”: sharp transitions from high-confidence core tokens to uncertain long-tail tokens. By computing a position-weighted relative decay rate, Min-k dynamically determines truncation boundaries at each generation step. We formally prove that Min-k achieves strict temperature invariance and empirically demonstrate its low sensitivity to hyperparameter choices. Experiments on multiple reasoning benchmarks, creative writing tasks, and human evaluation show that Min-k consistently improves text quality, maintaining robust performance even under extreme temperature settings where probability-based methods collapse. We make our code, models, and analysis tools publicly available.

Hallucination, broadly referring to unfaithful, fabricated, or inconsistent content generated by LLMs, has wide-ranging implications. Therefore, a large body of effort has been devoted to detecting LLM hallucinations, as well as designing benchmark datasets for evaluating these detectors. In this work, we first establish a desiderata of properties for hallucination detection benchmarks (HDBs) to exhibit for effective evaluation. A critical look at existing HDBs through the lens of our desiderata reveals that none of them exhibits all the properties. We identify two largest gaps: (1) RAG-based grounded benchmarks with long context are severely lacking (partly because length impedes human annotation); and (2) Existing benchmarks do not make available realistic label noise for stress-testing detectors although real-world use-cases often grapple with label noise due to human or automated/weak annotation. To close these gaps, we build and open-source a new RAG-based HDB called TRIVIA+ that underwent a rigorous human annotation process. Notably, our benchmark exhibits all desirable properties including (1) TRIVIA+ contains samples with the longest context in the literature; and (2) we design and share three sets of noisy labels with different, sample-dependent noise schemes. Finally, we perform experiments on RAG-based HDBs, including our TRIVIA+, using popular SOTA detectors that reveal new insights: (i) ample room remains for current detectors to reach the performance ceiling on RAG-based HDBs, (ii) the basic LLM-as-a-Judge baseline performs competitively, and (iii) label noise hinders detection performance. We expect that our findings, along with our proposed benchmark, will motivate and foster needed research on hallucination detection for RAG-based tasks.