论文检索

Akshay Paruchuri, Maryam Aziz, Rohit Vartak, Ayman Ali, Best Uchehara, Xin Liu, Ishan Chatterjee, Monica Agrawal

People are increasingly seeking healthcare information from large language models (LLMs) via interactive chatbots, yet the nature and inherent risks of these conversations remain largely unexplored. In this paper, we filter large-scale conversational AI datasets to achieve HealthChat-11K, a curated dataset of 11K real-world conversations composed of 25K user messages. We use HealthChat-11K and a clinician-driven taxonomy for how users interact with LLMs when seeking healthcare information in order to systematically study user interactions across 21 distinct health specialties. Our analysis reveals insights into the nature of how and why users seek health information, such as common interactions, instances of incomplete context, affective behaviors, and interactions (e.g., leading questions) that can induce sycophancy, underscoring the need for improvements in the healthcare support capabilities of LLMs deployed as conversational AI. We release code and artifacts to retrieve our analyses and combine them into a curated dataset for further research.

Yating Huang, Ziyan Huang, Lintao Xiang, Qijun Yang, Hujun Yin

Accurate analysis of pathological images is essential for automated tumor diagnosis but remains challenging due to high structural similarity and subtle morphological variations in tissue images. Current vision-language (VL) models often struggle to capture the complex reasoning required for interpreting structured pathological reports. To address these limitations, we propose PathoHR-Bench, a novel benchmark designed to evaluate VL models’ abilities in hierarchical semantic understanding and compositional reasoning within the pathology domain. Results of this benchmark reveal that existing VL models fail to effectively model intricate cross-modal relationships, hence limiting their applicability in clinical setting. To overcome this, we further introduce a pathology-specific VL training scheme that generates enhanced and perturbed samples for multimodal contrastive learning. Experimental evaluations demonstrate that our approach achieves state-of-the-art performance on PathoHR-Bench and six additional pathology datasets, highlighting its effectiveness in fine-grained pathology representation.

Junda Wu, Yuxin Xiong, Xintong Li, Yu Xia, Ruoyu Wang, Yu Wang, Tong Yu, Sungchul Kim, Ryan A. Rossi, Lina Yao 等

Recent MLLMs have demonstrated strong visual understanding and reasoning after large-scale multimodal pre-training. However, instruction-tuning is typically text-driven with limited visual supervision, leading to significant visual forgetting and degradation of pre-trained visual knowledge. Existing fine-tuning and continual learning methods compress visual representations and emphasize task alignment over visual retention, failing to address this challenge. We present a novel perspective using effective rank to quantify the loss of visual representation richness, framing visual forgetting as excessive compression under the information bottleneck principle. To address this, we propose modality-decoupled gradient descent (MDGD), which regulates gradient updates to preserve the effective rank of visual features and explicitly disentangles visual learning from task-specific alignment. We further introduce a memory-efficient fine-tuning variant using gradient masking for parameter-efficient adaptation. Extensive experiments show that MDGD effectively mitigates visual forgetting across downstream tasks and models, maintaining pre-trained visual knowledge while supporting strong task adaptation.

Weiqi Wang, Tianqing Fang, Haochen Shi, Baixuan Xu, Wenxuan Ding, Liyu Zhang, Wei Fan, Jiaxin Bai, Haoran Li, Xin Liu 等

Conceptualization, a fundamental element of human cognition, plays a pivotal role in human generalizable reasoning.Generally speaking, it refers to the process of sequentially abstracting specific instances into higher-level concepts and then forming abstract knowledge that can be applied in unfamiliar or novel situations. This enhances models’ inferential capabilities and supports the effective transfer of knowledge across various domains.Despite its significance, the broad nature of this term has led to inconsistencies in understanding conceptualization across various works, as there exists different types of instances that can be abstracted in a wide variety of ways.There is also a lack of a systematic overview that comprehensively examines existing works on the definition, execution, and application of conceptualization to enhance reasoning tasks.In this paper, we address these gaps by first proposing a categorization of different types of conceptualizations into four levels based on the types of instances being conceptualized, in order to clarify the term and define the scope of our work.Then, we present the first comprehensive survey of over 150 papers, surveying various definitions, resources, methods, and downstream applications related to conceptualization into a unified taxonomy, with a focus on the entity and event levels.Furthermore, we shed light on potential future directions in this field and hope to garner more attention from the community.

Jiseung Hong, Grace Byun, Seungone Kim, Kai Shu

Large Language Models (LLMs) are expected to provide helpful and harmless responses, yet they often exhibit sycophancy—conforming to user beliefs regardless of factual accuracy or ethical soundness. Prior research on sycophancy has primarily focused on single-turn factual correctness, overlooking the dynamics of real-world interactions. In this work, we introduce SYCON Bench (SYcophantic CONformity benchmark), a novel evaluation suite that assesses sycophantic behavior in multi-turn, free-form conversational settings. Our benchmark measures how quickly a model conforms to the user (Turn of Flip) and how frequently it shifts its stance under sustained user pressure (Number of Flip). Applying SYCON Bench to 17 LLMs across three real-world scenarios, we find that sycophancy remains a prevalent failure mode. Our analysis shows that alignment tuning amplifies sycophantic behavior, whereas model scaling and reasoning optimization strengthen the model’s ability to resist undesirable user views. Reasoning models generally outperform instruction-tuned models but often fail when they over-index on logical exposition instead of directly addressing the user’s underlying beliefs. Finally, we evaluate four additional prompting strategies and demonstrate that adopting a third-person perspective reduces sycophancy by up to 63.8% in debate scenario.

Hao Mark Chen, Wayne Luk, Yiu Ka Fai Cedric, Rui Li, Konstantin Mishchenko, Stylianos Venieris, Hongxiang Fan

The auto-regressive decoding of Large Language Models (LLMs) results in significant overheads in their hardware performance. While recent research has explored various speculative decoding techniques for multi-token generation, these methods introduce high memory costs from the additional weights and KV cache of separate draft models, limiting efficiency in edge and long-context scenarios. To overcome these limitations in edge-scale LLMs, we propose a novel parallel prompt decoding that requires only runtime memory overhead by employing a unified single model for both speculation and verification. Inspired by the human natural language generation process, PPD approximates outputs generated at future timesteps in parallel by using multiple prompt tokens. Furthermore, we present a hardware-aware two-stage tree pruning algorithm that adaptively optimizes this decoding scheme to fully leverage the computational capacities on different GPUs. Through extensive experiments across LLMs ranging from MobileLlama to Vicuna-13B on a wide range of benchmarks, our approach demonstrates up to 2.49 times speedup. Moreover, our parallel prompt decoding can serve as an orthogonal optimization for synergistic integration with existing speculative decoding, showing up to 1.22 times further speed improvement. To support future development, we have included our code implementation with this submission.

Yaochen Zhu, Harald Steck, Dawen Liang, Yinhan He, Nathan Kallus, Jundong Li

Large language models (LLMs) have demonstrated impressive zero-shot capabilities in conversational recommender systems (CRS). However, effectively utilizing historical conversations remains a significant challenge. Current approaches either retrieve few-shot examples or extract global rules to enhance the prompt, which fail to capture the implicit and preference-oriented knowledge. To address this challenge, we propose LLM-based Conversational Recommendation Agents with Collaborative Verbalized Experience, abbreviated as CRAVE. CRAVE begins by sampling trajectories of LLM-based CRS agents on historical queries and establishing verbalized experience banks by reflecting the agents’ actions on user feedback. Additionally, we introduce a collaborative retriever network fine-tuned with item content-parameterized multinomial likelihood on query-item pairs to retrieve preference-oriented verbal experiences for new queries. Furthermore, we developed a debater-critic agent (DCA) system where each agent maintains an independent collaborative experience bank and works together to enhance the CRS recommendations. We demonstrate that the open-ended debate and critique nature of DCA benefits significantly from the collaborative experience augmentation with CRAVE. The code is available at https://github.com/yaochenzhu/CRAVE.

Yiming Zeng, Wanhao Yu, Zexin Li, Tao Ren, Yu Ma, Jinghan Cao, Xiyan Chen, Tingting Yu

Large Language Models (LLMs) have significantly advanced natural language processing, demonstrating strong capabilities in tasks such as text generation, summarization, and reasoning. Recently, their potential for automating precise text editing tasks across specialized domains, such as programming code, LaTeX, and structured database languages, has gained attention. However, current state-of-the-art LLMs still struggle with executing precise, instruction-driven edits, particularly when structural accuracy and strict adherence to domain conventions are required.To address these challenges, we introduce InstrEditBench, an automated benchmark dataset comprising over 30,000 structured editing tasks spanning diverse domains, including Wikipedia articles, LaTeX documents, source code, and database languages. Using this benchmark, we develop FineEdit, a specialized editing model explicitly trained for accurate, context-aware text modifications. Experimental evaluations demonstrate that FineEdit outperforms state-of-the-art models, achieving improvements of approximately 10% over Gemini models on single-turn edits, up to 30% over Llama-3.2-3B, and exceeding Mistral-7B-OpenOrca performance by over 40% on direct editing tasks. FineEdit also effectively generalizes to realistic multi-turn editing scenarios, highlighting its practical applicability. To facilitate further research and reproducibility, we release FineEdit at https://github.com/StuRinDQB/FineEdit and https://huggingface.co/datasets/YimingZeng/FineEdit_bench.

Haokun Chen, Sebastian Szyller, Weilin Xu, Nageen Himayat

Large language models (LLMs) are trained using massive datasets.However, these datasets often contain undesirable content, e.g., harmful texts, personal information, and copyrighted material.To address this, machine unlearning aims to remove information from trained models.Recent work has shown that soft token attacks () can successfully extract unlearned information from LLMs.In this work, we show that s can be an inadequate tool for auditing unlearning.Using common unlearning benchmarks, i.e., Who Is Harry Potter? and TOFU, we demonstrate that, in a strong auditor setting, such attacks can elicit any information from the LLM, regardless of (1) the deployed unlearning algorithm, and (2) whether the queried content was originally present in the training corpus.Also, we show that with just a few soft tokens (1-10) can elicit random strings over 400-characters long.Thus showing that s must be used carefully to effectively audit unlearning.Example code can be found at https://github.com/IntelLabs/LLMart/tree/main/examples/unlearning

Ronald Seoh, Dan Goldwasser

In this paper, we introduce EmoGist, a training-free, in-context learning method for performing visual emotion classification with LVLMs. The key intuition of our approach is that context-dependent definition of emotion labels could allow more accurate predictions of emotions, as the ways in which emotions manifest within images are highly context dependent and nuanced. EmoGist pre-generates multiple descriptions of emotion labels, by analyzing the clusters of example images belonging to each label. At test time, we retrieve a version of description based on the cosine similarity of test image to cluster centroids, and feed it together with the test image to a fast LVLM for classification. Through our experiments, we show that EmoGist allows up to 12 points improvement in micro F1 scores with the multi-label Memotion dataset, and up to 8 points in macro F1 in the multi-class FI dataset.

Andrianos Michail, Simon Clematide, Rico Sennrich

The evaluation of cross-lingual semantic search models is often limited to existing datasets from tasks such as information retrieval and semantic textual similarity. We introduce Cross-Lingual Semantic Discrimination (CLSD), a lightweight evaluation task that requires only parallel sentences and a Large Language Model (LLM) to generate adversarial distractors. CLSD measures an embedding model’s ability to rank the true parallel sentence above semantically misleading but lexically similar alternatives. As a case study, we construct CLSD datasets for German–French in the news domain. Our experiments show that models fine-tuned for retrieval tasks benefit from pivoting through English, whereas bitext mining models perform best in direct cross-lingual settings. A fine-grained similarity analysis further reveals that embedding models differ in their sensitivity to linguistic perturbations.

Jun Zhuang, Haibo Jin, Ye Zhang, Zhengjian Kang, Wenbin Zhang, Gaby G. Dagher, Haohan Wang

Intent detection, a core component of natural language understanding, has considerably evolved as a crucial mechanism in safeguarding large language models (LLMs). While prior work has applied intent detection to enhance LLMs’ moderation guardrails, showing a significant success against content-level jailbreaks, the robustness of these intent-aware guardrails under malicious manipulations remains under-explored. In this work, we investigate the vulnerability of intent-aware guardrails and demonstrate that LLMs exhibit implicit intent detection capabilities. We propose a two-stage intent-based prompt-refinement framework, IntentPrompt, that first transforms harmful inquiries into structured outlines and further reframes them into declarative-style narratives by iteratively optimizing prompts via feedback loops to enhance jailbreak success for red-teaming purposes. Extensive experiments across four public benchmarks and various black-box LLMs indicate that our framework consistently outperforms several cutting-edge jailbreak methods and evades even advanced Intent Analysis (IA) and Chain-of-Thought (CoT)-based defenses. Specifically, our “FSTR+SPIN” variant achieves attack success rates ranging from 88.25% to 96.54% against CoT-based defenses on the o1 model, and from 86.75% to 97.12% on the GPT-4o model under IA-based defenses. These findings highlight a critical weakness in LLMs’ safety mechanisms and suggest that intent manipulation poses a growing challenge to content moderation guardrails.

Yihong Liu, Mingyang Wang, Amir Hossein Kargaran, Felicia Körner, Ercong Nie, Barbara Plank, François Yvon, Hinrich Schuetze

Large Language Models (LLMs) are capable of recalling multilingual factual knowledge present in their pretraining data. However, most studies evaluate only the final model, leaving the development of factual recall and crosslingual consistency throughout pretraining largely unexplored. In this work, we trace how factual recall and crosslingual consistency evolve during pretraining, focusing on OLMo-7B as a case study. We find that both accuracy and consistency improve over time for most languages. We show that this improvement is primarily driven by the fact frequency in the pretraining corpus: more frequent facts are more likely to be recalled correctly, regardless of language. Yet, some low-frequency facts in non-English languages can still be correctly recalled. Our analysis reveals that these instances largely benefit from crosslingual transfer of their English counterparts – an effect that emerges predominantly in the early stages of pretraining. We pinpoint two distinct pathways through which multilingual factual knowledge acquisition occurs: (1) frequency-driven learning, which is dominant and language-agnostic, and (2) crosslingual transfer, which is limited in scale and typically constrained to relation types involving named entities. We release our code and data to facilitate further research at https://github.com/cisnlp/multilingual-fact-tracing.

Linxin Song, Taiwei Shi, Jieyu Zhao

Reinforcement finetuning (RFT) has become a standard approach for enhancing the reasoning capabilities of large language models (LLMs). However, its impact on model trustworthiness remains underexplored. In this work, we identify and systematically study a critical side effect of RFT, which we term the hallucination tax: a degradation in refusal behavior causing models to produce hallucinated answers to unanswerable questions confidently. To investigate this, we introduce SUM (Synthetic Unanswerable Math), a high-quality dataset of unanswerable math problems designed to probe models’ ability to recognize an unanswerable question by reasoning from the insufficient or ambiguous information. Our results show that standard RFT training could reduce model refusal rates by more than 80%, which significantly increases model’s tendency to hallucinate. We further demonstrate that incorporating just 10% SUM during RFT substantially restores appropriate refusal behavior, with minimal accuracy trade-offs on solvable tasks. Crucially, this approach enables LLMs to leverage inference-time compute to reason about their own uncertainty and knowledge boundaries, improving generalization not only to out-of-domain math problems but also to factual question answering tasks.

Philip Lippmann, Jie Yang

Context-aware embedding methods boost retrieval accuracy by conditioning on corpus statistics (e.g., term co-occurrence and topical patterns) extracted from neighboring documents. However, this context-aware approach requires access to the target corpus or requires domain-specific finetuning, posing practical barriers in privacy-sensitive or resource-constrained settings. We present ZEST, a zero-shot contextual adaptation framework that replaces real corpus access with a one-time offline synthesis of a compact proxy. Given only a handful of exemplar documents representative of the general target domain, we use a multi-step hierarchical procedure to generate a synthetic context corpus of several hundred documents that aims to emulate key domain-specific distributions. At inference, the frozen context-aware encoder uses this proxy corpus – without any finetuning or target corpus access – to produce domain-adapted embeddings. Across the MTEB benchmark, ZEST’s zero-shot synthetic context adaptation using only five example documents performs within 0.5% of models leveraging full target corpus access – demonstrating remarkable efficacy without any retraining. ZEST thus provides a practical method for deploying high-performance, adaptable embeddings in constrained environments.

Qihang Zhang, Muchen Li, Ziao Wang, Renjie Liao, Lele Wang

Lossless compression techniques are crucial in an era of rapidly growing data. Traditional universal compressors like gzip offer low computational overhead, high speed, and broad applicability across data distributions. However, they often lead to worse compression rates than modern neural compressors, which leverage large-scale training data to model data distributions more effectively.Despite their advantages, neural compressors struggle to generalize to unseen data. To address this limitation, we propose a novel framework that performs Test-Time Steering via a Weighted Product of Experts (wPoE).At inference, our method adaptively combines a universal compression model with a pretrained neural language model, ensuring the compression rate is at least as good as the best individual model.Extensive experiments demonstrate that our approach improves the performance of text compression without requiring fine-tuning. Furthermore, it seamlessly integrates with any autoregressive language model, providing a practical solution for enhancing text compression across diverse data distributions.

Zeqiang Wang, Jon Johnson, Suparna De

Cross-lingual alignment of nuanced sociological concepts is crucial for comparative cross-cultural research, harmonising longitudinal studies, and leveraging knowledge from social science taxonomies (e.g., ELSST). However, aligning these concepts is challenging due to cultural context-dependency, linguistic variation, and data scarcity, particularly for low-resource languages. Existing methods often fail to capture domain-specific subtleties or require extensive parallel data. Grounded in a Vector Decomposition Hypothesis—positing separable domain and language components within embeddings, supported by observed language-pair specific geometric structures—we propose DLIR (Dual-Branch LoRA for Invariant Representation). DLIR employs parallel Low-Rank Adaptation (LoRA) branches: one captures core sociological semantics (trained primarily on English data structured by the ELSST hierarchy), while the other learns language invariance by counteracting specific language perturbations. These perturbations are modeled by Gaussian Mixture Models (GMMs) fitted on minimal parallel concept data using spherical geometry. DLIR significantly outperforms strong baselines on cross-lingual sociological concept retrieval across 10 languages. Demonstrating powerful zero-shot knowledge transfer, English-trained DLIR substantially surpasses target-language (French/German) LoRA fine-tuning even in monolingual tasks. DLIR learns disentangled, language-robust representations, advancing resource-efficient multilingual understanding and enabling reliable cross-lingual comparison of sociological constructs.

Yunyi Zhang, Ruozhen Yang, Siqi Jiao, SeongKu Kang, Jiawei Han

Scientific paper retrieval is essential for supporting literature discovery and research. While dense retrieval methods demonstrate effectiveness in general-purpose tasks, they often fail to capture fine-grained scientific concepts that are essential for accurate understanding of scientific queries. Recent studies also use large language models (LLMs) for query understanding; however, these methods often lack grounding in corpus-specific knowledge and may generate unreliable or unfaithful content. To overcome these limitations, we propose SemRank, an effective and efficient paper retrieval framework that combines LLM-guided query understanding with a concept-based semantic index. Each paper is indexed using multi-granular scientific concepts, including general research topics and detailed key phrases. At query time, an LLM identifies core concepts derived from the corpus to explicitly capture the query’s information need. These identified concepts enable precise semantic matching, significantly enhancing retrieval accuracy. Experiments show that SemRank consistently improves the performance of various base retrievers, surpasses strong existing LLM-based baselines, and remains highly efficient.

Daryna Dementieva, Nikolay Babakov, Alexander Fraser

While Ukrainian NLP has seen progress in many texts processing tasks, emotion classification remains an underexplored area with no publicly available benchmark to date. In this work, we introduce **EmoBench-UA**, the first annotated dataset for emotion detection in Ukrainian texts. Our annotation schema is adapted from the previous English-centric works on emotion detection (Mohammad et al., 2018; Mohammad, 2022) guidelines. The dataset was created through crowdsourcing using the Toloka.ai platform ensuring high-quality of the annotation process. Then, we evaluate a range of approaches on the collected dataset, starting from linguistic-based baselines, synthetic data translated from English, to large language models (LLMs). Our findings highlight the challenges of emotion classification in non-mainstream languages like Ukrainian and emphasize the need for further development of Ukrainian-specific models and training resources.

Milan Bhan, Yann Choho, Jean-Noël Vittaut, Nicolas Chesneau, Pierre Moreau, Marie-Jeanne Lesot

This paper proposes Complete Textual Concept Bottleneck Model (CT-CBM), a novel TCBM generator building concept labels in a fully unsupervised manner using a small language model, eliminating both the need for predefined human labeled concepts and LLM annotations. CT-CBM iteratively targets and adds important and identifiable concepts in the bottleneck layer to create a complete concept basis. CT-CBM achieves striking results against competitors in terms of concept basis completeness and concept detection accuracy, offering a promising solution to reliably enhance interpretability of NLP classifiers.