论文检索

Reza Esfandiarpoor, George Zerveas, Ruochen Zhang, Macton Mgonzo, Carsten Eickhoff, Stephen Bach

Although synthetic data has changed various aspects of information retrieval (IR) pipelines, the main training paradigm remains: contrastive learning with binary relevance labels, where one positive document is compared against several negatives using the InfoNCE loss. This objective treats all documents that are not explicitly annotated as relevant on an equally negative footing, regardless of their actual degree of relevance, thus missing subtle nuances useful for ranking. To overcome this limitation, in this work, we forgo real documents and annotations and use large language models to directly generate synthetic documents that answer the MS MARCO queries according to _several different levels of relevance_. We also propose using Wasserstein distance as a more effective loss function for training transformer-based retrievers with graduated relevance labels. Our experiments on MS MARCO and BEIR benchmark show that our proposed approach outperforms conventional training with InfoNCE by a large margin. Without using any real documents, our method significantly improves self-supervised retrievers and is more robust to distribution shift compared to contrastive learning using real data. Our method also successfully integrates existing real data into the synthetic ranking context, further boosting the performance. Overall, we show that generating multi-level ranking contexts is a better approach to synthetic data generation for IR than just generating the standard positive and negative documents.

Najrin Sultana, Md Rafi Ur Rashid, Kang Gu, Shagufta Mehnaz

LLMs can provide substantial zero-shot performance on diverse tasks using a simple task prompt, eliminating the need for training or fine-tuning. However, when applying these models to sensitive tasks, it is crucial to thoroughly assess their robustness against adversarial inputs. In this work, we introduce Static Deceptor (StaDec) and Dynamic Deceptor (DyDec), two innovative attack frameworks designed to systematically generate dynamic and adaptive adversarial examples by leveraging the understanding of the LLMs. We produce subtle and natural-looking adversarial inputs that preserve semantic similarity to the original text while effectively deceiving the target LLM. By utilizing an automated, LLM-driven pipeline, we eliminate the dependence on external heuristics. Our attacks evolve with the advancements in LLMs, while demonstrating a strong transferability across models unknown to the attacker. Overall, this work provides a systematic approach for self-assessing the robustness of the LLMs. We release our code and data at https://github.com/Shukti042/AdversarialExample.

Aashaka Desai, Daniela Massiceti, Richard Ladner, Hal Daumé III, Danielle Bragg, Alex Xijie Lu

Like most languages, sign languages evolve over time. It is important that sign language dictionaries’ vocabularies are updated over time to reflect these changes, such as by adding new signs. However, most dictionary retrieval methods based upon machine learning models only work with fixed vocabularies, and it is unclear how they might support dictionary expansion without retraining. In this work, we explore the feasibility of dictionary expansion for sign language dictionaries using a simple representation-based method. We explore a variety of dictionary expansion scenarios, e.g., varying number of signs added as well as amount of data for these newly added signs. Through our results, we show how performance varies significantly across different scenarios, many of which are reflective of real-world data challenges. Our findings offer implications for the development & maintenance of video-based sign language dictionaries, and highlight directions for future research on dictionary expansion.

Derin Ozer, Nicolas Gutowski, Benoit Da Mota, Thomas Cauchy, Sylvain Lamprier

Natural Language Processing (NLP) has catalyzed a paradigm shift in Computer-Aided Synthesis Planning (CASP), reframing chemical synthesis prediction as a sequence-to-sequence modeling problem over molecular string representations like SMILES. This framing has enabled the direct application of language models to chemistry, yielding impressive benchmark scores on the USPTO dataset, a large text corpus of reactions extracted from US patents. However, we show that USPTO’s patent-derived data are both industrially biased and incomplete. They omit many fundamental transformations essential for practical real-world synthesis. Consequently, models trained exclusively on USPTO perform poorly on simple, pharmaceutically relevant reactions despite high benchmark scores. Our findings highlight a broader concern in applying standard NLP pipelines to scientific domains without rethinking data and evaluation: models may learn dataset artifacts rather than domain reasoning. We argue for the development of chemically meaningful benchmarks, greater data diversity, and interdisciplinary dialogue between the NLP community and domain experts to ensure real-world applicability.

Avneet Kaur

Existing evaluation of political biases in Large Language Models (LLMs) outline the high sensitivity to prompt formulation. Furthermore, Large Language Models are known to exhibit sycophancy, a tendency to align their outputs with a user’s stated belief, which is often attributed to human feedback during fine-tuning. However, such bias in the presence of explicit argumentation within a prompt remains underexplored. This paper investigates how argumentative prompts induce sycophantic behaviour in LLMs in a political context. Through a series of experiments, we demonstrate that models consistently alter their responses to mirror the stance present expressed by the user. This sycophantic behaviour is observed in both single and multi-turn interactions, and its intensity correlates with argument strength. Our findings establish a link between user stance and model sycophancy, revealing a critical vulnerability that impacts model reliability. Thus has significant implications for models being deployed in real-world settings and calls for developing robust evaluations and mitigations against manipulative or biased interactions.

Sarfraz Ahmad, Hasan Iqbal, Momina Ahsan, Numaan Naeem, Muhammad Ahsan Riaz Khan, Arham Riaz, Muhammad Arslan Manzoor, Yuxia Wang, Preslav Nakov

The rapid adoption of Large Language Models (LLMs) has raised important concerns about the factual reliability of their outputs, particularly in low-resource languages such as Urdu. Existing automated fact-checking systems are predominantly developed for English, leaving a significant gap for the more than 200 million Urdu speakers worldwide. In this work, we present UrduFactBench and UrduFactQA, two novel hand-annotated benchmarks designed to enable fact-checking and factual consistency evaluation in Urdu. While UrduFactBench focuses on claim verification, UrduFactQA targets the factuality of LLMs in question answering. These resources, the first of their kind for Urdu, were developed through a multi-stage annotation process involving native Urdu speakers. To complement these benchmarks, we introduce UrduFactCheck, a modular fact-checking framework that incorporates both monolingual and translation-based evidence retrieval strategies to mitigate the scarcity of high-quality Urdu evidence. Leveraging these resources, we conduct an extensive evaluation of twelve LLMs and demonstrate that translation-augmented pipelines consistently enhance performance compared to monolingual ones. Our findings reveal persistent challenges for open-source LLMs in Urdu and underscore the importance of developing targeted resources. All code and data are publicly available at https://github.com/mbzuai-nlp/UrduFactCheck.

Dong Won Lee, Hae Won Park, Cynthia Breazeal, Louis-Philippe Morency

We propose a large language model based reward decomposition framework for aligning dialogue agents using only a single session-level feedback signal. We leverage the reasoning capabilities of a frozen, pretrained large language model (LLM) to infer fine-grained local implicit rewards by decomposing global, session-level feedback. Our first text-only variant prompts the LLM to perform reward decomposition using only the dialogue transcript. The second multimodal variant incorporates additional behavioral cues, such as pitch, gaze, and facial affect, expressed as natural language descriptions. These inferred turn-level rewards are distilled into a lightweight reward model, which we utilize for RL-based fine-tuning for dialogue generation. We evaluate both text-only and multimodal variants against state-of-the-art reward decomposition methods and demonstrate notable improvements in human evaluations of conversation quality, suggesting that LLMs are strong reward decomposers that obviate the need for manual reward shaping and granular human feedback.

Marc Felix Brinner, Tarek Al Mustafa, Sina Zarrieß

We investigate the use of LLM-generated data for continual pretraining of transformer encoder models in specialized domains with limited training data, using the scientific domain of invasion biology as a case study. To this end, we leverage domain-specific ontologies by enriching them with LLM-generated data and pretraining the encoder model as an ontology-informed embedding model for concept definitions. To evaluate the effectiveness of this method, we compile a benchmark specifically designed for assessing model performance in invasion biology. After demonstrating substantial improvements over standard MLM pretraining, we investigate the feasibility of applying the proposed approach to domains without comprehensive ontologies by substituting ontological concepts with concepts automatically extracted from a small corpus of scientific abstracts and establishing relationships between concepts through distributional statistics. Our results demonstrate that this automated approach achieves comparable performance using only a small set of scientific abstracts, resulting in a fully automated pipeline for enhancing domain-specific understanding of small encoder models that is especially suited for application in low-resource settings and achieves performance comparable to masked language modeling pretraining on much larger datasets.

Rongzhi Zhang, Chenwei Zhang, Xinyang Zhang, Liang Qiu, Haoming Jiang, Yuchen Zhuang, Qingru Zhang, Hyokun Yun, Xian Li, Bing Yin 等

Aligning large language models (LLMs) with human preferences relies heavily on high-quality reward models. However, existing approaches struggle with two critical challenges: noisy preference labels and the varying importance of preference samples. We introduce DORM, a method that enhances reward modeling by learning to dynamically weigh preference data.DORM initializes data importance using a combination of model uncertainty and prediction disagreement, then iteratively refines them via bilevel optimization to maximize validation performance. Using only 50k samples, DORM trains a 12B reward model that achieves 90.5% accuracy on RewardBench, matching the performance of models trained on significantly larger datasets. Furthermore, downstream alignment tasks show that fine-tuned LLMs with DORM achieve a 61.2% win rate against baseline methods, highlighting its data efficiency and generalizability.

Skyler Seto, Maartje Ter Hoeve, Maureen de Seyssel, David Grangier

Bilingual and multilingual language models offer a promising path toward scaling NLP systems across diverse languages and users. However, their performance often varies wildly between languages as prior works show that adding more languages can degrade performance for some languages (such as English), while improving others (typically more data constrained languages). In this work, we investigate causes of these inconsistencies by comparing bilingual and monolingual language models. Our analysis reveals that unequal data quality, not just data quantity, is a major driver of performance degradation in bilingual settings. We propose a simple yet effective data filtering strategy to select higher-quality bilingual training data with only high quality English data. Applied to French, German, and Chinese, our approach improves monolingual performance by 2–4% and reduces bilingual model performance gaps to 1%. These results highlight the overlooked importance of data quality in multilingual pretraining and offer a practical recipe for balancing performance.

Wenyan Li, Raphael Tang, Chengzu Li, Caiqi Zhang, Ivan Vulić, Anders Søgaard

Vision–language models (VLMs) often process visual inputs through a pretrained vision encoder, followed by a projection into the language model’s embedding space via a connector component. While crucial for modality fusion, the potential information loss induced by this projection step and its direct impact on model capabilities remain understudied. We introduce two complementary approaches to examine and quantify this loss by analyzing the latent representation space. First, we evaluate semantic information preservation by analyzing changes in k-nearest neighbor relationships between image representations, before and after projection. Second, we directly measure information loss by reconstructing visual embeddings from the projected representation, localizing loss at an image patch level. Experiments reveal that connectors substantially distort the local geometry of visual representations, with k-nearest neighbors diverging by 40–60% post-projection, correlating with degradation in retrieval performance. The patch-level embedding reconstruction provides interpretable insights for model behavior on visually grounded question-answering tasks, finding that areas of high information loss reliably predict instances where models struggle.

Akshat Gupta, Phudish Prateepamornkul, Maochuan Lu, Ahmed Alaa, Thomas Hartvigsen, Gopala Anumanchipalli

Knowledge editing is a promising way to improve factuality in large language models, but recent studies have shown significant model degradation during sequential editing. In this paper, we formalize the popular locate-then-edit methods as a two-step fine-tuning process, allowing us to precisely identify the root cause of this degradation. We show that model degradation occurs due to (1) over-optimization of internal activations and (2) continuous norm-growth of edited matrices. To mitigate these issues, we introduce two regularization techniques: (1) Most-Probable Early Stopping (MPES) and (2) explicit Frobenius norm-constraint. We demonstrate that applying these simple yet effective regularization techniques at key points in the editing process can substantially mitigate model degradation. Combining these regularization methods enables scaling locate-then-edit methods to 10,000 edits while reducing editing time by 42-61%. These results show that targeted regularization is essential for lifelong knowledge editing.

Shoubin Yu, Yue Zhang, Ziyang Wang, Jaehong Yoon, Mohit Bansal

Combining pre-trained expert models offers substantial potential for scalable multimodal reasoning, but building a unified framework remains challenging due to the increasing diversity of input modalities and task complexity. For instance, medical diagnosis requires precise reasoning over structured clinical tables, while financial forecasting depends on interpreting plot-based data to make informed predictions. To tackle this challenge, we introduce MEXA, a training-free framework that performs modality- and task-aware aggregation of multiple expert models to enable effective multimodal reasoning across diverse and distinct domains. MEXA dynamically selects expert models based on the input modality and the task-specific reasoning demands (i.e., skills). Each expert model, specialized in a modality task pair, generates interpretable textual reasoning outputs. MEXA then aggregates and reasons over these outputs using a Large Reasoning Model (LRM) to produce the final answer. This modular design allows flexible and transparent multimodal reasoning across diverse domains without additional training overhead. We extensively evaluate our approach on diverse multimodal benchmarks, including Video Reasoning, Audio Reasoning, 3D Understanding, and Medical QA. MEXA consistently delivers performance improvements over strong multimodal baselines, highlighting the effectiveness and broad applicability of our expert-driven selection and aggregation in diverse multimodal reasoning tasks.

Michal Štefánik, Timothee Mickus, Michal Spiegel, Marek Kadlčík, Josef Kuchař

A large body of recent work assesses models’ generalization capabilities through the lens of performance on out-of-distribution (OOD) datasets. Despite their practicality, such evaluations build upon a strong assumption: that OOD evaluations can capture and reflect upon possible failures in a real-world deployment. In this work, we challenge this assumption and confront the results obtained from OOD evaluations with a set of specific failure modes documented in existing question-answering (QA) models, referred to as a reliance on spurious features or prediction shortcuts.We find that different datasets used for OOD evaluations in QA provide an estimate of models’ robustness to shortcuts that have a vastly different quality, some largely under-performing even a simple, in-distribution evaluation. We partially attribute this to the observation that spurious shortcuts are shared across ID+OOD datasets, but also find cases where a dataset’s quality for training and evaluation is largely disconnected. Our work underlines limitations of commonly-used OOD-based evaluations of generalization, and provides methodology and recommendations for evaluating generalization within and beyond QA more robustly.

Andreas Geert Motzfeldt, Joakim Edin, Casper L. Christensen, Christian Hardmeier, Lars Maaløe, Anna Rogers

In medical coding, experts map unstructured clinical notes to alphanumeric codes for diagnoses and procedures. We introduce ‘Code Like Humans’: a new agentic framework for medical coding with large language models. It implements official coding guidelines for human experts, and it is the first solution that can support the full ICD-10 coding system (+70K labels). It achieves the best performance to date on rare diagnosis codes. Fine-tuned discriminative classifiers retain an advantage for high-frequency codes, to which they are limited. Towards future work, we also contribute an analysis of system performance and identify its ‘blind spots’ (codes that are systematically undercoded).

Emilio Villa-Cueva, S M Masrur Ahmed, Rendi Chevi, Jan Christian Blaise Cruz, Kareem Elzeky, Fermin Cristobal, Alham Fikri Aji, Skyler Wang, Rada Mihalcea, Thamar Solorio

Understanding Theory of Mind is essential for building socially intelligent multimodal agents capable of perceiving and interpreting human behavior. We introduce MoMentS (Multimodal Mental States), a comprehensive benchmark designed to assess the ToM capabilities of multimodal large language models (LLMs) through realistic, narrative-rich scenarios presented in short films. MoMentS includes over 2,300 multiple-choice questions spanning seven distinct ToM categories. The benchmark features long video context windows and realistic social interactions that provide deeper insight into characters’ mental states. We evaluate several MLLMs and find that although vision generally improves performance, models still struggle to integrate it effectively. For audio, models that process dialogues as audio do not consistently outperform transcript-based inputs. Our findings highlight the need to improve multimodal integration and point to open challenges that must be addressed to advance AI’s social understanding.

Changyu Zeng, Yifan Wang, Zimu Wang, Wei Wang, Zhengni Yang, Muyi Bao, Jimin Xiao, Anh Nguyen, Yutao Yue

Recent advancements in 2D multimodal large language models (MLLMs) have significantly improved performance in vision-language tasks. However, extending these capabilities to 3D environments remains a distinct challenge due to the complexity of spatial reasoning. Nevertheless, existing 3D benchmarks often lack fine-grained numerical reasoning task annotations, limiting MLLMs’ ability to perform precise spatial measurements and complex numerical reasoning. To address this gap, we introduce NUMINA, the first Natural Understanding benchmark for Multi-dimensional Intelligence and Numerical reasoning Abilities to enhance multimodal indoor perceptual understanding. NUMINA features multi-scale annotations and various question-answer pairs, generated using NUMINA-Flow, an automated annotation pipeline that integrates LLM rewriting and rule-based self-verification. We evaluate the performance of various state-of-the-art LLMs on NUMINA following the Chat-Scene framework, demonstrating that current LLMs struggle with multimodal numerical reasoning, particularly in performing precise computations such as distance and volume estimation, highlighting the need for further advancements in 3D models. The dataset and source codes can be obtained from https://github.com/fengshun124/NUMINA.

Mengying Wang, Andreas Spitz

Writing assistants and large language models see widespread use in the creation of text content. While their effectiveness for individual users has been evaluated in the literature, little is known about their proclivity to change language or reduce its richness when adopted by a large user base. In this paper, we take a first step towards quantifying this risk by measuring the semantic and vocabulary change enacted by the use of rephrasing tools on a multi-domain corpus of human-generated text.

Xiaojun Wu, Junxi Liu, Huan-Yi Su, Zhouchi Lin, Yiyan Qi, Chengjin Xu, Jiajun Su, Jiajie Zhong, Fuwei Wang, Saizhuo Wang 等

As large language models (LLMs) increasingly permeate the financial sector, there is a pressing need for a standardized method to comprehensively assess their performance. Existing financial benchmarks often suffer from limited language and task coverage, low-quality datasets, and inadequate adaptability for LLM evaluation. To address these limitations, we introduce Golden Touchstone, a comprehensive bilingual benchmark for financial LLMs, encompassing eight core financial NLP tasks in both Chinese and English. Developed from extensive open-source data collection and industry-specific demands, this benchmark thoroughly assesses models’ language understanding and generation capabilities. Through comparative analysis of major models such as GPT-4o, Llama3, FinGPT, and FinMA, we reveal their strengths and limitations in processing complex financial information. Additionally, we open-source Touchstone-GPT, a financial LLM trained through continual pre-training and instruction tuning, which demonstrates strong performance on the bilingual benchmark but still has limitations in specific tasks. This research provides a practical evaluation tool for financial LLMs and guides future development and optimization.The source code for Golden Touchstone and model weight of Touchstone-GPT have been made publicly available at https://github.com/IDEA-FinAI/Golden-Touchstone.

Matteo Bortoletto, Constantin Ruhdorfer, Lei Shi, Andreas Bulling

Despite growing interest in Theory of Mind (ToM) tasks for evaluating language models (LMs), little is known about how LMs internally represent mental states of self and others. Understanding these internal mechanisms is critical - not only to move beyond surface-level performance, but also for model alignment and safety, where subtle misattributions of mental states may go undetected in generated outputs. In this work, we present the first systematic investigation of belief representations in LMs by probing models across different scales, training regimens, and prompts - using control tasks to rule out confounds. Our experiments provide evidence that both model size and fine‐tuning substantially improve LMs’ internal representations of others’ beliefs, which are structured - not mere by-products of spurious correlations - yet brittle to prompt variations. Crucially, we show that these representations can be strengthened: targeted edits to model activations can correct wrong ToM inferences.