论文检索

Tianlu Zheng, Yifan Zhang, Xiang An, Ziyong Feng, Kaicheng Yang, Qichuan Ding

Although Contrastive Language-Image Pre-training (CLIP) exhibits strong performance across diverse vision tasks, its application to person representation learning faces two critical challenges: (i) the scarcity of large-scale annotated vision-language data focused on person-centric images, and (ii) the inherent limitations of global contrastive learning, which struggles to maintain discriminative local features crucial for fine-grained matching while remaining vulnerable to noisy text tokens. This work advances CLIP for person representation learning through synergistic improvements in data curation and model architecture. First, we develop a noise-resistant data construction pipeline that leverages the in-context learning capabilities of MLLMs to automatically filter and caption web-sourced images. This yields WebPerson, a large-scale dataset of 5M high-quality person-centric image-text pairs. Second, we introduce the GA-DMS (Gradient-Attention Guided Dual-Masking Synergetic) framework, which improves cross-modal alignment by adaptively masking noisy textual tokens based on the gradient-attention similarity score. Additionally, we incorporate masked token prediction objectives that compel the model to predict informative text tokens, enhancing fine-grained semantic representation learning. Extensive experiments show that GA-DMS achieves state-of-the-art performance across multiple benchmarks. The data and pre-trained models are released at https://github.com/Multimodal-Representation-Learning-MRL/GA-DMS.

Simon Münker

LLMs increasingly engage with psychological instruments, yet how they represent constructs internally remains poorly understood. We introduce a novel approach to “fingerprinting” LLMs through their factor correlation patterns on standardized psychological assessments to deepen the understanding of LLMs constructs representation. Using the Humor Style Questionnaire as a case study, we analyze how six LLMs represent and correlate humor-related constructs to survey participants. Our results show that they exhibit little similarity to human response patterns. In contrast, participants’ subsamples demonstrate remarkably high internal consistency. Exploratory graph analysis further confirms that no LLM successfully recovers the four constructs of the Humor Style Questionnaire. These findings suggest that despite advances in natural language capabilities, current LLMs represent psychological constructs in fundamentally different ways than humans, questioning the validity of application as human simulacra.

Haishuo Fang, Xiaodan Zhu, Iryna Gurevych

Deploying LLM-based agents in real-life applications often faces a critical challenge: the misalignment between agents’ behavior and user intent. Such misalignment may lead agents to unintentionally execute some critical actions that carry negative outcomes (e.g., accidentally triggering a \textit{buy-now} in web shopping), resulting in undesirable or even irreversible consequences. Although addressing these issues is crucial, the preemptive detection and correction of misaligned actions remains relatively underexplored. To fill this gap, we introduce \texttt{InferAct}, a novel approach that leverages the belief reasoning ability of LLMs, grounded in Theory-of-Mind, to detect misaligned actions. Once the misalignment is detected, \texttt{InferAct} alerts users for timely correction, preventing adverse outcomes and enhancing the reliability of LLM agents’ decision-making processes. Experiments on three widely used tasks demonstrate \texttt{InferAct} achieves up to 20% improvements on Marco-F1 against baselines in misaligned action detection. An in-depth evaluation of misalignment correction further highlights \texttt{InferAct}‘s effectiveness in improving agent alignment.

Jakub Macina, Nico Daheim, Ido Hakimi, Manu Kapur, Iryna Gurevych, Mrinmaya Sachan

Evaluating the pedagogical capabilities of AI-based tutoring models is critical for making guided progress in the field. Yet, we lack a reliable, easy-to-use, and simple-to-run evaluation that reflects the pedagogical abilities of models. To fill this gap, we present MathTutorBench, an open-source benchmark for holistic tutoring model evaluation. MathTutorBench contains a collection of datasets and metrics that broadly cover tutor abilities as defined by learning sciences research in dialog-based teaching. To score the pedagogical quality of open-ended teacher responses, we train a reward model and show it can discriminate expert from novice teacher responses with high accuracy. We evaluate a wide set of closed- and open-weight models on MathTutorBench and find that subject expertise, indicated by solving ability, does not immediately translate to good teaching. Rather, pedagogy and subject expertise appear to form a trade-off that is navigated by the degree of tutoring specialization of the model. Furthermore, tutoring appears to become more challenging in longer dialogs, where simpler questioning strategies begin to fail. We release the benchmark, code, and leaderboard openly to enable rapid benchmarking of future models.

Melanie Subbiah, Akankshya Mishra, Grace Kim, Liyan Tang, Greg Durrett, Kathleen McKeown

Determining faithfulness of a claim to a source document is an important problem across many domains. This task is generally treated as a binary judgment of whether the claim is supported or unsupported in relation to the source. In many cases, though, whether a claim is supported can be ambiguous. For instance, it may depend on making inferences from given evidence, and different people can reasonably interpret the claim as either supported or unsupported based on their agreement with those inferences. Forcing binary labels upon such claims lowers the reliability of evaluation. In this work, we reframe the task to manage the subjectivity involved with factuality judgments of ambiguous claims. We introduce LLM-generated edits of summaries as a method of providing a nuanced evaluation of claims: how much does a summary need to be edited to be unambiguous? Whether a claim gets rewritten and how much it changes can be used as an automatic evaluation metric, the Ambiguity Rewrite Metric (ARM), with a much richer feedback signal than a binary judgment of faithfulness. We focus on the area of narrative summarization as it is particularly rife with ambiguity and subjective interpretation. We show that ARM produces a 21% absolute improvement in annotator agreement on claim faithfulness, indicating that subjectivity is reduced.

Branislav Pecher, Ivan Srba, Maria Bielikova

When solving NLP tasks with limited labelled data, researchers typically either use a general large language model without further update, or use a small number of labelled samples to tune a specialised smaller model. In this work, we answer an important question – how many labelled samples are required for the specialised small models to outperform general large models, while taking the performance variance into consideration. By observing the behaviour of fine-tuning, instruction-tuning, prompting and in-context learning on 8 language models, we identify such performance break-even points across 8 representative text classification tasks of varying characteristics. We show that the specialised models often need only few samples (on average 100) to be on par or better than the general ones. At the same time, the number of required labels strongly depends on the dataset or task characteristics, with fine-tuning on binary datasets requiring significantly more samples. When performance variance is taken into consideration, the number of required labels increases on average by 100 - 200%. Finally, larger models do not consistently lead to better performance and lower variance, with 4-bit quantisation having negligible impact.

Kian Ahrabian, Pegah Jandaghi, Negar Mokhberian, Sai Praneeth Karimireddy, Jay Pujara

Reinforcement learning from human feedback (RLHF) and, at its core, reward modeling have become a crucial part of training powerful large language models (LLMs). One commonly overlooked factor in training high-quality reward models (RMs) is the effect of the base model, which is becoming more challenging to choose given the rapidly growing pool of LLMs. In this work, we present a systematic analysis of the effect of base model selection on reward modeling performance. Our results show that the performance can be improved by up to 14% compared to the most common (i.e., default) choice. Moreover, we showcase the strong statistical relation between some existing benchmarks and downstream performances. We also demonstrate that the results from a small set of benchmarks could be combined to boost the model selection (+18% on average in the top 5-10). Lastly, we illustrate the impact of different post-training steps on the final performance and explore using estimated data distributions to reduce performance prediction error.

Siqi Shen, Mehar Singh, Lajanugen Logeswaran, Moontae Lee, Honglak Lee, Rada Mihalcea

The value orientation of Large Language Models (LLMs) has been extensively studied, as it can shape user experiences across demographic groups.However, two key challenges remain: (1) the lack of systematic comparison across value probing strategies, despite the Multiple Choice Question (MCQ) setting being vulnerable to perturbations, and (2) the uncertainty over whether probed values capture in-context information or predict models’ real-world actions.In this paper, we systematically compare three widely used value probing methods: token likelihood, sequence perplexity, and text generation.Our results show that all three methods exhibit large variances under non-semantic perturbations in prompts and option formats, with sequence perplexity being the most robust overall.We further introduce two tasks to assess expressiveness: demographic prompting, testing whether probed values adapt to cultural context; and value–action agreement, testing the alignment of probed values with value-based actions.We find that demographic context has little effect on the text generation method, and probed values only weakly correlate with action preferences across all methods.Our work highlights the instability and the limited expressive power of current value probing methods, calling for more reliable LLM value representations.

David Beauchemin, Richard Khoury

Large and Transformer-based language models perform outstandingly in various downstream tasks. However, there is limited understanding regarding how these models internalize linguistic knowledge, so various linguistic benchmarks have recently been proposed to facilitate syntactic evaluation of language models across languages. This paper introduces QFrCoLA (Quebec-French Corpus of Linguistic Acceptability Judgments), a normative binary acceptability judgments dataset comprising 25,153 in-domain and 2,675 out-of-domain sentences. Our study leverages the QFrCoLA dataset and seven other linguistic binary acceptability judgment corpora to benchmark seven language models. The results demonstrate that, on average, fine-tuned Transformer-based LM are strong baselines for most languages and that zero-shot binary classification large language models perform poorly on the task. However, for the QFrCoLA benchmark, on average, a fine-tuned Transformer-based LM outperformed other methods tested. It also shows that pre-trained cross-lingual LLMs selected for our experimentation do not seem to have acquired linguistic judgment capabilities during their pre-training for Quebec French. Finally, our experiment results on QFrCoLA show that our dataset, built from examples that illustrate linguistic norms rather than speakers’ feelings, is similar to linguistic acceptability judgment; it is a challenging dataset that can benchmark LM on their linguistic judgment capabilities.

David Beauchemin, Michelle Albert-Rochette, Richard Khoury, Pierre-Luc Déziel

Simplifying text while preserving its meaning is a complex yet essential task, especially in sensitive domain applications like legal texts. When applied to a specialized field, like the legal domain, preservation differs significantly from its role in regular texts. This paper introduces FrJUDGE, a new dataset to assess legal meaning preservation between two legal texts. It also introduces JUDGEBERT, a novel evaluation metric designed to assess legal meaning preservation in French legal text simplification. JUDGEBERT demonstrates a superior correlation with human judgment compared to existing metrics. It also passes two crucial sanity checks, while other metrics did not: For two identical sentences, it always returns a score of 100%; on the other hand, it returns 0% for two unrelated sentences. Our findings highlight its potential to transform legal NLP applications, ensuring accuracy and accessibility for text simplification for legal practitioners and lay users.

Donghyun Kim, Sriram Ravula, Taemin Ha, Alex Dimakis, Daehyeok Kim, Aditya Akella

Workload traces are essential to understand complex computer systems’ behavior and manage processing and memory resources. Since real-world traces are hard to obtain, synthetic trace generation is a promising alternative. This paper proposes a first-of-a-kind approach that relies on training a large language model (LLM) to generate synthetic workload traces, specifically microservice call graphs. To capture complex and arbitrary hierarchical structures and implicit constraints in such traces, we propose to train LLMs to generate recursively, making call graph generation a sequence of more manageable steps. To further enforce learning constraints on the traces and generate uncommon situations, we apply additional instruction tuning steps to align our model with the desired trace features. With this method, we train TraceLLM, an LLM for microservice trace generation, and demonstrate that it produces diverse, realistic traces under varied conditions, outperforming existing approaches in both accuracy and validity. The synthetically generated traces can effectively replace real data to optimize important microservice management tasks. Additionally, TraceLLM adapts to downstream trace-related tasks, such as predicting key trace features and infilling missing data.

Donya Rooein, Vilém Zouhar, Debora Nozza, Dirk Hovy

Stories play a pivotal role in human communication, shaping beliefs and morals, particularly in children. As parents increasingly rely on large language models (LLMs) to craft bedtime stories, the presence of cultural and gender stereotypes in these narratives raises significant concerns. To address this issue, we present Biased Tales, a comprehensive dataset designed to analyze how biases influence protagonists’ attributes and story elements in LLM-generated stories. Our analysis uncovers striking disparities. When the protagonist is described as a girl (as compared to a boy), appearance-related attributes increase by 55.26%. Stories featuring non-Western children disproportionately emphasize cultural heritage, tradition, and family themes far more than those for Western children. Our findings highlight the role of sociocultural bias in making creative AI use more equitable and diverse.

Mohsinul Kabir, Ajwad Abrar, Sophia Ananiadou

A large number of studies rely on closed-style multiple-choice surveys to evaluate cultural alignment in Large Language Models (LLMs). In this work, we challenge this constrained evaluation paradigm and explore more realistic, unconstrained approaches. Using the World Values Survey (WVS) and Hofstede Cultural Dimensions as case studies, we demonstrate that LLMs exhibit stronger cultural alignment in less constrained settings, where responses are not forced. Additionally, we show that even minor changes, such as reordering survey choices, lead to inconsistent outputs, exposing the limitations of closed-style evaluations. Our findings advocate for more robust and flexible evaluation frameworks that focus on specific cultural proxies, encouraging more nuanced and accurate assessments of cultural alignment in LLMs.

Dominic Petrak, Thy Thy Tran, Iryna Gurevych

Although LLM-based conversational agents demonstrate strong fluency and coherence, they still produce undesirable behaviors (errors) that are challenging to prevent from reaching users during deployment. Recent research leverages large language models (LLMs) to detect errors and guide response-generation models toward improvement. However, current LLMs struggle to identify errors not explicitly specified in their instructions, such as those arising from updates to the response-generation model or shifts in user behavior. In this work, we introduce Automated Error Discovery, a framework for detecting and defining errors in conversational AI, and propose SEEED (Soft Clustering Extended Encoder-Based Error Detection), as an encoder-based approach to its implementation. We enhance the Soft Nearest Neighbor Loss by amplifying distance weighting for negative samples and introduce Label-Based Sample Ranking to select highly contrastive examples for better representation learning. SEEED outperforms adapted baselines—including GPT-4o and Phi-4—across multiple error-annotated dialogue datasets, improving the accuracy for detecting unknown errors by up to 8 points and demonstrating strong generalization to unknown intent detection.

Anastasiia Sedova, Robert Litschko, Diego Frassinelli, Benjamin Roth, Barbara Plank

One of the major aspects contributing to the striking performance of large language models (LLMs) is the vast amount of factual knowledge accumulated during pre-training. Yet, many LLMs suffer from self-inconsistency, which raises doubts about their trustworthiness and reliability. This paper focuses on entity type ambiguity, analyzing the proficiency and consistency of state-of-the-art LLMs in applying factual knowledge when prompted with ambiguous entities. To do so, we propose an evaluation protocol that disentangles knowing from applying knowledge, and test state-of-the-art LLMs on 49 ambiguous entities. Our experiments reveal that LLMs struggle with choosing the correct entity reading, achieving an average accuracy of only 85%, and as low as 75% with underspecified prompts. The results also reveal systematic discrepancies in LLM behavior, showing that while the models may possess knowledge, they struggle to apply it consistently, exhibit biases toward preferred readings, and display self-inconsistencies. This highlights the need to address entity ambiguity in the future for more trustworthy LLMs.

Mohammad Akbar-Tajari, Mohammad Taher Pilehvar

Large Language Models (LLMs) have emerged as the dominant paradigm in Natural Language Processing owing to their remarkable performance across various target tasks. However, naively fine-tuning them for specific downstream tasks often requires updating a vast number of parameters, resulting in high computational costs and overfitting when training data is limited. In this paper, we propose a novel approach, called *Stochastic Tuning*, that addresses these challenges by selectively updating a small subset of parameters in each step of the tuning process. Our approach is characterized by its customization of updates based on task-specific partial gradients with respect to stochastic sub-networks. The advantage of Stochastic Tuning over existing solutions lies in its ability to consider both parameter weights as well as forward values which guarantees a context-sensitive fine-tuning. Our experiments demonstrate that Stochastic Tuning outperforms existing lightweight fine-tuning methods, improving average performance by over two points on RoBERTa across several tasks in the GLUE benchmark while updating merely **0.08**% of the model’s parameters. The code for our implementation can be found at https://github.com/m-Tajari/StocTuning_LLMs.

Yundong Gai, Ping Li

There are many emerging quantization methods to resolve the problem that the huge demand on computational and storage costs hinders the deployment of Large language models (LLMs). However, their accuracy performance still can not satisfy the entire academic and industry community. In this work, we propose ATQ, an INT8 weight-activation quantization of LLMs, that can achieve almost lossless accuracy. We employ a mathematically equivalent transformation and a triangle inequality to constrain weight-activation quantization error to the sum of a weight quantization error and an activation quantization error. For the weight part, transformed weights are quantized along the |in-feature| dimension and the quantization error is compensated by optimizing following in-features. For the activation part, transformed activations are in the normal range and can be quantized easily. We provide comparison experiments to demonstrate that our ATQ method can achieve almost lossless in accuracy on OPT and LLaMA families in W8A8 quantization settings. The increase of perplexity is within 1 and the accuracy degradation is within 0.5 percent even in the worst case.

Anton Alexandrov, Veselin Raychev, Mark Niklas Müller, Ce Zhang, Martin Vechev, Kristina Toutanova

As open-weight large language models (LLMs) achieve ever more impressive performance across a wide range of tasks in English, practitioners aim to adapt these models to different languages. However, such language adaptation is often accompanied by catastrophic forgetting of the base model’s capabilities, severely limiting the usefulness of the resulting model. We address this issue by proposing Branch-and-Merge (BaM), a new adaptation method based on iteratively merging multiple models, fine-tuned on a subset of the available training data. BaM is based on the insight that this yields lower magnitude but higher quality weight changes, reducing forgetting of the source domain while maintaining learning on the target domain. We demonstrate in an extensive empirical study on Bulgarian and German that BaM can significantly reduce forgetting while matching or even improving target domain performance compared to both standard continued pretraining and instruction finetuning across different model architectures.

Qinyuan Cheng, Xiaonan Li, Shimin Li, Qin Zhu, Zhangyue Yin, Yunfan Shao, Linyang Li, Tianxiang Sun, Hang Yan, Xipeng Qiu

In Retrieval-Augmented Generation (RAG), retrieval is not always helpful and applying it to every instruction is sub-optimal. Therefore, determining whether to retrieve is crucial for RAG, which is usually referred to as Active Retrieval. However, existing active retrieval methods face two challenges: 1. They usually rely on a single criterion, which struggles with handling various types of instructions. 2. They depend on specialized and highly differentiated procedures, and thus combining them makes the RAG system more complicated and leads to higher response latency. To address these challenges, we propose Unified Active Retrieval (UAR). UAR contains four orthogonal criteria and casts them into plug-and-play classification tasks, which achieves multifaceted retrieval timing judgements with negligible extra inference cost. We further introduce the Unified Active Retrieval Criteria (UAR-Criteria), designed to process diverse active retrieval scenarios through a standardized procedure. Experiments on four representative types of user instructions show that UAR significantly outperforms existing work on the retrieval timing judgement and the performance of downstream tasks, which shows the effectiveness of UAR and its helpfulness to downstream tasks.

Evgeniia Razumovskaia, Ivan Vulić, Pavle Marković, Tomasz Cichy, Qian Zheng, Tsung-Hsien Wen, Paweł Budzianowski

Factual faithfulness is a crucial requirement in information-seeking dialogue: the system should respond to the user queries so that the responses are meaningful and aligned with the knowledge provided to the system. However, most modern large language models (LLMs) suffer from hallucinations, that is, they generate responses not supported by or even contradicting the knowledge source. To mitigate the issue and increase faithfulness of information-seeking dialogue systems supported by the LLMs, we introduce BeInfo, a simple yet effective method that applies ‘behavioural tuning’ on the LLMs to aid information-seeking dialogue. Relying on three standard information seeking dialogue datasets, we show that models tuned with BeInfo become considerably more faithful to the knowledge source both for datasets and domains seen during BeInfo-tuning, as well as on unseen domains, when applied in a zero-shot manner. In addition, we present a ‘real-life’ case study on conversations with real users, showcasing that the models with 3B parameters (e.g., Flan-T5) tuned with BeInfo demonstrate strong performance on data from real ‘production’ conversations: when tuned on a limited amount of such realistic in-domain dialogues, they surpass much larger LLMs used ‘off-the-shelf’, both on automatic and human evaluation metrics.