Knowledge editing aims to update or correct factual knowledge in a language model. A widely used approach, locate-then-edit, first localizes a fact within the model and then edits the weights there. To date, such methods have been developed exclusively for autoregressive models (ARMs). Whether they work for masked diffusion models (MDMs), which model text bidirectionally and generate by iterative denoising rather than next-token prediction, remains an open question. We address it by transferring locate-then-edit to MDMs and comparing multiple MDMs with their matched ARMs. Our central finding has two parts. First, where an edit should be applied transfers between them: the same early-to-mid-layer MLP at the last subject token is most effective for both. Second, this shared location does not guarantee a shared outcome. Single-token edits succeed in both, but as targets grow longer, editing degrades far more sharply in the MDMs than in the ARMs. The failure stems from how the edited fact is generated: producing a multi-token target passes through intermediate states in which the target is partially unmasked, for which the edit was never optimized. Guided by this diagnosis, we introduce a simple correction that optimizes the edit including such states, substantially restoring multi-token performance. Our code is available at https://github.com/holi-lab/MDM-KE.
LLM judges now score most open-ended NLP output, and their mutual agreement is routinely read as evidence that the scores can be trusted. That reading is unsafe: judges may agree because they capture quality, or because they share the same blind spots, and agreement statistics alone cannot tell these apart. We develop a geometric test that can. Treating each judge's scores as a vector, we measure spread, effective rank, the angle to human scores, and the judge-judge, judge-human, and human-human agreement triple for 42 judges on two community-built Indic benchmarks covering four domains and eight languages. Every comparison is reference-matched: a judge and a held-out rater are scored against the same two-rater mean, since an averaged reference otherwise flatters judges by several degrees. On subjective rubrics, judges agree with one another as much as humans do yet reach only 58-66% of human agreement and often concentrate on an axis humans do not weight. On the one rubric with a verifiable answer, most of that gap closes. Ensembles converge on the judges' shared axis rather than the human one, and training widens scores without rotating them. Inter-judge agreement is evidence of human alignment only after this check passes.
Sourabrata Mukherjee, Hamna Hamna, Kalika Bali et al.· 0 citations
We built a dataset for the task of poem-to-prose translation from 13th-17th Century Telugu classical poetry to contemporary Telugu and English prose, which we call Padyam2Gadyam. The dataset consists of 600 poems and their human-verified Telugu and English prose translations. We evaluated 2 machine translation systems and 5 contemporary Large Language Models (LLMs) on their ability to do zero-shot poem-to-prose translation into Telugu and English using this dataset. Our results indicate that while the general purpose LLMs are better than the machine translation systems, there are systematic issues with the generation and evaluation of prose translation in both languages across systems.
Post-training has greatly improved reasoning in frontier vision-language models, yet its gains for perception remain comparatively limited, creating a bottleneck for end-to-end visual reasoning. To investigate this gap, we introduce a controlled diagnostic framework with two synthetic tasks that disentangle perception from reasoning. Our analysis reveals a consistent perception-reasoning asymmetry: post-training improves reasoning more substantially than perception, though the underlying mechanism differs across training paradigms. For supervised fine-tuning (SFT), this asymmetry stems from token imbalance, with perception occupying a smaller fraction of tokens in chain-of-thought supervision. Reweighting the loss boosts end-to-end performance by up to 18.2 points. For reinforcement learning (RL), the asymmetry instead arises from reward coupling, as outcome rewards correlate more strongly with reasoning than perception. Adding a perception-aware reward improves end-to-end accuracy by up to 6.0 points; when ground-truth perception rewards are unavailable, a reliable surrogate provides useful signal, yielding gains of 2.2 points. Beyond the controlled setting, these strategies also improve real-world visual reasoning, with gains of up to 3.3 points across three benchmarks. Overall, we diagnose the causes of asymmetric optimization and provide actionable guidance that benefits both synthetic and realistic settings.
Xueqing Wu, Yu-Chi Lin, Kai-Wei Chang et al.· 0 citations
Reach audiences
Advertise in front of researchers, engineers, and readers.
Vision-language models (VLMs) are increasingly evaluated for whether they identify the right visual content, but little is known about whether they express such content in a discourse-appropriate form. We address this research gap using information structure (IS), testing whether VLMs distinguish discourse-old Topics from discourse-new Foci in visually grounded question answering. We exploit Hungarian, a language in which Topic and Focus map onto dedicated syntactic positions, making IS choices observable in text. Comparing six VLMs with human participants, we find that models produce IS-relevant constructions, but over-regularise this sensitivity. Under the interacting pressures of discourse status, grammatical role (preference for subject Topics) and definiteness (preference for indefinite Foci), humans choose variable strategies for IS realisation. VLMs, by contrast, collapse onto narrow response templates, resembling mode collapse (Kirk et al., 2024). Our findings suggest that VLM evaluation should look beyond content accuracy to how content is packaged for the discourse.
Marcell Fekete, Johannes Bjerva, Tam\'as K\'aldi· 0 citations
Preference optimization is widely used to align Large Language Models (LLMs) with preference feedback. However, most existing methods train on a single positive-negative pair per prompt, discarding additional supervision available in preference datasets that typically contain multiple candidate responses. Motivated by this limitation, recent work explores group-wise preference optimization, which jointly contrasts multiple responses for the same prompt, but its empirical behavior and scalability remain underexplored due to the memory overhead of group-coupled objectives. In this work, we present a unified empirical and systems study of group-wise preference optimization and develop a memory-efficient implementation for group-coupled objectives. By instantiating first-order linearization with objective-specific per-response coefficients, our implementation preserves first-order gradients while decoupling samples during backpropagation, substantially reducing peak memory usage and enabling scalable training with larger groups. Across offline and online settings, we show that leveraging multiple responses consistently outperforms single-pair training. Furthermore, incorporating a negative log-likelihood (NLL) term on positive responses is critical for both performance gains and training stability.
Jixuan Leng, Si Si, Hsiang-Fu Yu et al.· 0 citations
Peer review plays a central role in the NLP publication process, but is susceptible to various biases. Here, we study language-of-study (LoS) bias: the tendency for reviewers to evaluate a paper differently based on the language(s) it studies, rather than its scientific merit. Despite being explicitly flagged in reviewing guidelines, such biases are poorly understood. Prior work treats such comments as part of broader categories of weak or unconstructive reviews without defining them as a distinct form of bias. We present the first systematic characterization of LoS bias, distinguishing negative and positive forms, and introduce the human-annotated dataset LOBSTER (Language-Of-study Bias in ScienTific pEer Review) and an LLM-based detection pipeline achieving 87.37 macro F1. We analyze 15,645 reviews to estimate how negative and positive biases differ with respect to the LoS, and find that non-English papers face substantially higher bias rates than English-only ones, with negative bias consistently outweighing positive bias. Finally, we identify four subcategories of negative bias, and find that demanding unjustified cross-lingual generalization is the most dominant form. We publicly release all resources to support work on fairer reviewing practices in NLP and beyond.
Ehsan Barkhordar, Abdulfattah Safa, Verena Blaschke et al.· 0 citations
While multilingual language models successfully transfer factual and syntactic knowledge across languages, it remains unclear whether they process culture-specific pragmatic registers, such as slang, as isolated language-specific memorizations or as unified, abstract concepts. We study this by probing the internal representations of Gemma-2-9B-IT using Sparse Autoencoders (SAEs) across three typologically diverse source languages: English, Hebrew, and Russian. To definitively isolate pragmatic register processing from trivial lexical sensitivity, we introduce a novel dataset in which every target term is polysemous, appearing in both literal and informal contexts. We find that while much of the informal-register signal is distributed across language-specific features, a small but highly robust cross-linguistic core consistently emerges. This shared core forms a geometrically coherent ``informal register subspace'' that sharpens in the model's deeper layers. Crucially, these shared representations are not merely correlational: activation steering with these features causally shifts output formality across all source languages and transfers zero-shot to six unseen languages spanning diverse language families and scripts. Together, these results provide the first mechanistic evidence that multilingual LLMs internalize informal register not just as surface-level heuristics, but as a portable, language-agnostic pragmatic abstraction.
Uri Z. Kialy, Avi Shtarkberg, Ayal Klein· 0 citations
We propose a method that represents language models by log-likelihood vectors over prompt-response pairs and constructs model maps for comparing their conditional distributions. In this space, squared Euclidean distances between models are approximately proportional to the KL divergence between the corresponding conditional distributions. Experiments on a large collection of publicly available language models show that the maps capture meaningful global structure, including relationships to model attributes and task performance. The representation also captures systematic shifts induced by prompt modifications and their approximate additive compositionality; we use the vectors to predict downstream task scores and leverage their additive structure to approximate the effects of composite prompt operations without directly observing the corresponding log-likelihood vectors. We further introduce PMI vectors to reduce the influence of unconditional distributions; in some cases, PMI-based model maps better reflect training-data-related differences. Overall, the framework supports the analysis and prediction of input-dependent model behavior.
The ``LLM-as-a-judge'' paradigm has become a standard method for evaluating open-ended generation. To address the quadratic scalability costs of pairwise comparisons, popular benchmarks like Arena-Hard and AlpacaEval compare all models against a single anchor. However, despite its widespread use, the impact of anchor selection on the reliability of the results remains largely unexplored. In this work, we systematically investigate the effect of anchor selection by evaluating 22 different anchors on the Arena-Hard-v2.0 dataset. We find that the choice of anchor is critical: a poor anchor can dramatically reduce correlation with human rankings. We identify that common anchor choices (best-performing and worst-performing models) make poor anchors. Because these extreme anchors are consistently better or worse than all other models, they are seldom indicative of the relative ranking of the models. We further quantify the effect size of anchor selection, showing it is comparable to the selection of a judge model. We conclude with actionable recommendations. First, we conduct a power analysis, and compute sufficient benchmark sizes for anchor-based evaluation, finding that standard benchmark sizes are insufficient for pairwise evaluation and fail to distinguish between competitive models reliably. Second, we provide guidelines for selecting informative anchors to ensure reliable and efficient evaluation practices.
Shachar Don-Yehiya, Asaf Yehudai, Leshem Choshen et al.· 0 citations
Recent work has identified a subset of attention heads in Transformer as retrieval heads, which are responsible for retrieving information from the context. In this work, we first investigate retrieval heads in multilingual contexts. In multilingual language models, we find that retrieval heads are often shared across multiple languages. Expanding the study to cross-lingual setting, we identify Retrieval-Transition heads(RTH), which govern the transition to specific target-language output. Our experiments reveal that RTHs are distinct from retrieval heads and more vital for Chain-of-Thought reasoning in multilingual LLMs. Across four multilingual benchmarks (MMLU-ProX, MGSM, MLQA, and XQuaD) and two model families (Qwen-2.5 and Llama-3.1), we demonstrate that masking RTH induces bigger performance drop than masking Retrieval Heads (RH). Our work advances understanding of multilingual LMs by isolating the attention heads responsible for mapping to target languages.
Shaswat Patel, Vishvesh Trivedi, Yue Han et al.· 0 citations
Legal text generated by large language models (LLMs) can usually achieve reasonable factual accuracy, but it frequently fails to adhere to the specialised stylistic norms and linguistic conventions of legal writing. In order to improve stylistic quality, a crucial first step is to establish a reliable evaluation method. However, having legal experts manually develop such a metric is impractical, as the implicit stylistic requirements in legal writing practice are difficult to formalise into explicit rubrics. Meanwhile, existing automatic evaluation methods also fall short: reference-based metrics conflate semantic accuracy with stylistic fidelity, and LLM-as-a-judge evaluations suffer from opacity and inconsistency. To address these challenges, we introduce CLASE (Chinese LegAlese Stylistic Evaluation), a hybrid evaluation method that focuses on the stylistic performance of legal text. The method incorporates a hybrid scoring mechanism that combines 1) linguistic feature-based scores and 2) experience-guided LLM-as-a-judge scores. Both the feature coefficients and the LLM scoring experiences are learned from contrastive pairs of authentic legal documents and their LLM-restored counterparts. This hybrid design captures both surface-level features and implicit stylistic norms in a transparent, reference-free manner. Experiments on 200 Chinese legal documents show that CLASE achieves substantially higher alignment with human judgments than traditional metrics and pure LLM-as-a-judge methods. Beyond improved alignment, CLASE provides interpretable score breakdowns and suggestions for improvements, offering a scalable and practical solution for professional stylistic evaluation in legal text generation (Code and data for CLASE is available at: https://github.com/rexera/CLASE).
A new method, called CW-Net, translates the reasoning process of an autonomous vehicle’s AI system into understandable concepts that explain its behavior.
MIT News · Artificial Intelligence· news.mit.eduAug 31, 2026
With millions of users across the world, Julia has been used to conduct cutting-edge research and to design new drugs, jet engines, heat pumps, and more.
A new machine-learning framework aims to improve the success rate of computational protein design while moving away from results that reproduce sequences found in nature.
New MIT research could lead to better materials for a fossil-fuel-free process for making the chemical that's essential to fertilizer and other products.
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
