论文检索

Sofia Jamil, Kotla Sai Charan, Sriparna Saha, Koustava Goswami, Joseph K J

Poetry is an expressive form of art that invites multiple interpretations, as readers often bring their own emotions, experiences, and cultural backgrounds into their understanding of a poem. Recognizing this, we aim to generate images for poems and improve these images in a zero-shot setting, enabling audiences to modify images as per their requirements. To achieve this, we introduce a novel \textit{Weighted Prompt Manipulation (WPM)} technique, which systematically modifies attention weights and text embeddings within diffusion models. By dynamically adjusting the importance of specific words, \textit{WPM} enhances or suppresses their influence in the final generated image, leading to semantically richer and more contextually accurate visualizations. Our approach exploits diffusion models and large language models (LLMs) such as GPT in conjunction with existing poetry datasets, ensuring a comprehensive and structured methodology for improved image generation in the literary domain. To the best of our knowledge, this is the first attempt at integrating weighted prompt manipulation for enhancing imagery in poetic language.

Georgios Chochlakis, Peter Wu, Tikka Arjun Singh Bedi, Marcus Ma, Kristina Lerman, Shrikanth Narayanan

Modeling complex subjective tasks in Natural Language Processing, such as recognizing emotion and morality, is considerably challenging due to significant variation in human annotations. This variation often reflects reasonable differences in semantic interpretations rather than mere noise, necessitating methods to distinguish between legitimate subjectivity and error.We address this challenge by exploring label verification in these contexts using Large Language Models (LLMs). First, we propose a simple In-Context Learning binary filtering baseline that estimates the reasonableness of a document-label pair. We then introduce the Label-in-a-Haystack setting: the query and its label(s) are included in the demonstrations shown to LLMs, which are prompted to predict the label(s) again, while receiving task-specific instructions (e.g., emotion recognition) rather than label copying.We show how the failure to copy the label(s) to the output of the LLM are task-relevant and informative. Building on this, we propose the Label-in-a-Haystack Rectification (LiaHR) framework for subjective label correction: when the model outputs diverge from the reference gold labels, we assign the generated labels to the example instead of discarding it. This approach can be integrated into annotation pipelines to enhance signal-to-noise ratios. Comprehensive analyses, human evaluations, and ecological validity studies verify the utility of LiaHR for label correction. Code is available at https://github.com/gchochla/liahr.

Ahmed Karim, Qiao Wang, Zheng Yuan

Automated Essay Scoring (AES) systems now attain near–human agreement on some public benchmarks, yet real-world adoption—especially in high-stakes examinations—remains limited. A principal obstacle is that most models output a single score without any accompanying measure of confidence or explanation. We address this gap with conformal prediction, a distribution-free wrapper that equips any classifier with set-valued outputs enjoying formal coverage guarantees. Two open-source Large Language Models—Llama-3 8B and Qwen-2.5 3B—are fine-tuned on three diverse corpora (ASAP, TOEFL11, Cambridge-FCE) and calibrated at a 90% risk level. Reliability is assessed with UAcc, an uncertainty-aware accuracy that rewards models for being both correct and concise. To our knowledge, this is the first work to combine conformal prediction and UAcc for essay scoring. The calibrated models consistently meet the coverage target while keeping prediction sets compact, indicating that open-source, mid-sized LLMs can already support teacher-in-the-loop AES; we discuss scaling and broader user studies as future work.

Jiaxuan Zhao, Naibin Gu, Yuchen Feng, Xiyu Liu, Peng Fu, Zheng Lin, Weiping Wang

The high costs of customizing large language models (LLMs) fundamentally limit their adaptability to user-specific needs. Consequently, LLMs are increasingly offered as cloud-based services, a paradigm that introduces critical limitations: providers struggle to support personalized customization at scale, while users face privacy risks when exposing sensitive data. To address this dual challenge, we propose Customized Black-box Prompt Tuning (CBP-Tuning), a novel framework that facilitates efficient local customization while preserving bidirectional privacy. Specifically, we design a two-stage framework: (1) a prompt generator trained on the server-side to capture domain-specific and task-agnostic capabilities, and (2) user-side gradient-free optimization that tailors soft prompts for individual tasks. This approach eliminates the need for users to access model weights or upload private data, requiring only a single customized vector per task while achieving effective adaptation. Furthermore, the evaluation of CBP-Tuning in the commonsense reasoning, medical and financial domain settings demonstrates superior performance compared to baselines, showcasing its advantages in task-agnostic processing and privacy preservation.

Yuyi Huang, Runzhe Zhan, Lidia S. Chao, Ailin Tao, Derek F. Wong

As large language models (LLMs) are increasingly deployed for complex reasoning tasks, Long Chain-of-Thought (Long-CoT) prompting has emerged as a key paradigm for structured inference. Despite early-stage safeguards enabled by alignment techniques such as RLHF, we identify a previously underexplored vulnerability: reasoning trajectories in Long-CoT models can drift from aligned paths, resulting in content that violates safety constraints. We term this phenomenon Path Drift. Through empirical analysis, we uncover three behavioral triggers of Path Drift: (1) first-person commitments that induce goal-driven reasoning that delays refusal signals; (2) ethical evaporation, where surface-level disclaimers bypass alignment checkpoints; (3) condition chain escalation, where layered cues progressively steer models toward unsafe completions. Building on these insights, we introduce a three-stage Path Drift Induction Framework comprising cognitive load amplification, self-role priming, and condition chain hijacking. Each stage independently reduces refusal rates, while their combination further compounds the effect. To mitigate these risks, we propose a path-level defense strategy incorporating role attribution correction and metacognitive reflection (reflective safety cues). Our findings highlight the need for trajectory-level alignment oversight in long-form reasoning beyond token-level alignment.

Yuhao Wang, Heyang Liu, Ziyang Cheng, Ronghua Wu, Qunshan Gu, Yanfeng Wang, Yu Wang

Speech large language models (LLMs) have emerged as a prominent research focus in speech processing. In this work, we introduce VocalNet, a series of high-performance speech LLMs featuring a scalable and model-agnostic training framework as well as a novel multi-token prediction (MTP) paradigm for speech generation. We first propose an efficient two-stage training framework that enables LLMs to acquire real-time speech interaction capabilities. Through extensive experiments on various training configurations, we ensure both simplicity and effectiveness in the training strategy. Furthermore, inspired by advances in language modeling, we introduce MTP into the domain of speech LLMs—an alternative to traditional next-token prediction (NTP)—which enables the model to predict multiple future tokens at each step. Through systematic analysis and improved implementation, we show that MTP not only accelerates inference speed but also significantly enhances speech quality. Experimental results demonstrate that VocalNet achieves performance comparable to state-of-the-art Omni LLMs while outperforming existing open-source speech LLMs, despite using limited training data.

Mario Sanz-Guerrero, Minh Duc Bui, Katharina von der Wense

When evaluating large language models (LLMs) with multiple-choice question answering (MCQA), it is common to end the prompt with the string “Answer:” to facilitate automated answer extraction via next-token probabilities. However, there is no consensus on how to tokenize the space following the colon, often overlooked as a trivial choice. In this paper, we uncover accuracy differences of up to 11% due to this (seemingly irrelevant) tokenization variation as well as reshuffled model rankings, raising concerns about the reliability of LLM comparisons in prior work. Surprisingly, we are able to recommend one specific strategy – tokenizing the space together with the answer letter – as we observe consistent and statistically significant performance improvements. Additionally, it improves model calibration, enhancing the reliability of the model’s confidence estimates. Our findings underscore the importance of careful evaluation design and highlight the need for standardized, transparent evaluation protocols to ensure reliable and comparable results.

Yujie Wang, Yunwei Zhao, Jing Yang, Han Han, Shiguang Shan, Jie Zhang

Digital social media platforms frequently contribute to cognitive-behavioral fixation, a phenomenon in which users exhibit sustained and repetitive engagement with narrow content domains. While cognitive-behavioral fixation has been extensively studied in psychology, methods for computationally detecting and evaluating such fixation remain underexplored. To address this gap, we propose a novel framework for assessing cognitive-behavioral fixation by analyzing users’ multimodal social media engagement patterns. Specifically, we introduce a multimodal topic extraction module and a cognitive-behavioral fixation quantification module that collaboratively enable adaptive, hierarchical, and interpretable assessment of user behavior. Experiments on existing benchmarks and a newly curated multimodal dataset demonstrate the effectiveness of our approach, laying the groundwork for scalable computational analysis of cognitive fixation. All code in this project is publicly available for research purposes at https://github.com/Liskie/cognitive-fixation-evaluation.

Nikhil Bhendawade, Irina Belousova, Qichen Fu, Henry Mason, Antonie Lin, Mohammad Rastegari, Mahyar Najibi

Speculative decoding is a prominent technique for accelerating LLM inference by leveraging an auxiliary draft model, but its effectiveness is limited by the autoregressive nature of draft generation, where acceptance rates depend on the draft model’s size. Scaling the draft model improves acceptance but also increases speculation latency, limiting overall speedup. Furthermore, fine-tuning both the draft and target models is often necessary to achieve high acceptance rates, adding complexity to inference systems as the number of downstream tasks grows. Single-model approaches like Medusa generate speculative tokens non-autoregressively but lack token dependencies, limiting effectiveness. Alternatives like Hydra and Eagle incorporate token dependencies but rely on dedicated heads, making speculation independent of the base model and limiting the extent to which stronger base models can improve speculation.We introduce a novel speculative decoding method that integrates speculative draft generation directly within the target model using multi-stream attention. This improves acceptance rates by introducing interdependencies between speculative tokens while ensuring non-autoregressive draft generation with minimal overhead. As target models scale in size and quality, speculative generation improves naturally with our method, unlike prior approaches. Furthermore, our approach is both parameter- and FLOP-efficient, requiring over 1000X fewer additional parameters than Medusa, making it highly suitable for resource-constrained devices. We design our method to operate in two modes: (1) Lossless mode, a plug-and-play method that preserves the output of any pre-trained model; and (2) Shared mode, optimizing both speedup and downstream performance. We demonstrate a 2–3.5X speedup across diverse tasks, including summarization, translation, question answering, mathematical reasoning, SQL generation, and retrieval-augmented generation (RAG).

Yihuai Hong, Lei Yu, Haiqin Yang, Shauli Ravfogel, Mor Geva

The task of “unlearning” certain concepts in large language models (LLMs) has gained attention for its role in mitigating harmful, private, or incorrect outputs. Current evaluations mostly rely on behavioral tests, without monitoring residual knowledge in model parameters, which can be adversarially exploited to recover erased information. We argue that unlearning should also be assessed internally by tracking changes in the parametric traces of unlearned concepts. To this end, we propose a general evaluation methodology that uses vocabulary projections to inspect concepts encoded in model parameters. We apply this approach to localize “concept vectors” — parameter vectors encoding concrete concepts — and construct ConceptVectors, a benchmark of hundreds of such concepts and their parametric traces in two open-source LLMs. Evaluation on ConceptVectors shows that existing methods minimally alter concept vectors, mostly suppressing them at inference time, while direct ablation of these vectors removes the associated knowledge and reduces adversarial susceptibility. Our findings reveal limitations of behavior-only evaluations and advocate for parameter-based assessments. We release our code and benchmark at https://github.com/yihuaihong/ConceptVectors.

Kirolos Ataallah, Eslam Mohamed Bakr, Mahmoud Ahmed, Chenhui Gou, Khushbu Pahwa, Jian Ding, Mohamed Elhoseiny

Understanding long-form videos, such as movies and TV episodes ranging from tens of minutes to two hours, remains a significant challenge for multi-modal models. Existing benchmarks often fail to test the full range of cognitive skills needed to process these temporally rich and narratively complex inputs. Therefore, we introduce InfiniBench, a comprehensive benchmark designed to evaluate the capabilities of models in long video understanding rigorously.InfiniBench offers:(1) Over 1,000 hours of video content, with an average video length of 53 minutes.(2) The largest set of question-answer pairs for long video comprehension, totaling around 87.7 K.(3) Eight diverse skills that span both grounding-based (e.g., scene transitions, character actions) and reasoning-based (e.g., deep context understanding, multi-event linking).(4) Rich annotation formats, including both multiple-choice and open-ended questions.We conducted an in-depth evaluation across both commercial (GPT-4o, Gemini 2.0 Flash) and most recent open-source vision-language models, such as Qwen2.5-VL, InternVL3.0). Results reveal that:(1) Models struggle across the board: Even the best model, GPT-4o, achieves only 47.1% on grounding-based skills, with most models performing near or just above random chance.(2) Strong reliance on world knowledge: Models achieve surprisingly high scores using only metadata (e.g., video titles), highlighting a tendency to rely on pre-trained knowledge rather than actual visual or temporal understanding.(3) Multi-Modal Importance: When provided with full video and subtitle context, however, models show substantial improvements, confirming the critical role of multimodal input in video understanding.Our findings underscore the inherent challenges in long-video comprehension and point to the need for substantial advancements in both grounding and reasoning capabilities in MLLMs.

Hongxin Ding, Yue Fang, Runchuan Zhu, Xinke Jiang, Jinyang Zhang, Yongxin Xu, Weibin Liao, Xu Chu, Junfeng Zhao, Yasha Wang

Large Language Models (LLMs) excel in general language tasks, motivating their adaptation to specialized domains such as healthcare. Effective domain adaptation typically involves supervised fine-tuning (SFT) on carefully selected instruction-tuning data. Current data selection methods adopt a data-centric approach, relying on external annotations and heuristics to identify externally defined high-quality or challenging data. Our exploratory experiments highlight this approach fails to improve the model’s domain performance, due to misalignment between selected data and the model’s knowledge distribution. To tackle this, we propose Decomposed Difficulty-based Data Selection (3DS), a two-stage model-centric data selection framework that aligns data selection with the model’s distribution. 3DS employs Prompt-Driven Data Selection to filter out noise based on the model’s knowledge via explicit alignment in Stage#1, then adopts Decomposed Difficulty-based Data Selection to guide selection via three novel data difficulty metrics, including Instruction Understanding, Response Confidence, and Response Correctness in Stage#2, enhanced by an attention-based importance weighting mechanism for accurate calibration.Extensive experiments in the healthcare domain show 3DS outperforms existing methods by up to 2.97% accuracy, with additional validation in law and general domains, confirming its generalization ability. Our dataset and code are open-sourced at https://github.com/PuppyKnightUniversity/3DS.

Jiulong Wu, Zhengliang Shi, Shuaiqiang Wang, Jizhou Huang, Dawei Yin, Lingyong Yan, Min Cao, Min Zhang

Large Visual Language Models (LVLMs) have demonstrated impressive capabilities across multiple tasks. However, their trustworthiness is often challenged by hallucinations, which can be attributed to the modality misalignment and the inherent hallucinations of their underlying Large Language Models (LLMs) backbone. Existing preference alignment methods focus on aligning model responses with human preferences while neglecting image-text modality alignment, resulting in over-reliance on LLMs and hallucinations. In this paper, we propose Entity-centric Multimodal Preference Optimization (EMPO), which achieves enhanced modality alignment than existing human preference alignment methods. Besides, to overcome the scarcity of high-quality multimodal preference data, we utilize open-source instruction datasets to automatically construct high-quality preference data across three aspects: image, instruction, and response. Experiments on two human preference datasets and five multimodal hallucination benchmarks demonstrate the effectiveness of EMPO, e.g., reducing hallucination rates by 80.4% on Object HalBench and 52.6% on MM HalBench, thereby enhancing the trustworthiness of LVLMs. The code and dataset will be made publicly available.

Nicholas Deas, Kathleen McKeown

We introduce and study artificial impressions–patterns in LLMs’ internal representations of prompts that resemble human impressions and stereotypes based on language. We fit linear probes on generated prompts to predict impressions according to the two-dimensional Stereotype Content Model (SCM). Using these probes, we study the relationship between impressions and downstream model behavior as well as prompt features that may inform such impressions. We find that LLMs inconsistently report impressions when prompted, but also that impressions are more consistently linearly decodable from their hidden representations. Additionally, we show that artificial impressions of prompts are predictive of the quality and use of hedging in model responses. We also investigate how particular content, stylistic, and dialectal features in prompts impact LLM impressions.

Jian Zhang, Junyi Guo, Junyi Yuan, Huanda Lu, Yanlin Zhou, Fangyu Wu, Qiufeng Wang, Dongming Lu

Cross-modal retrieval is essential for interpreting cultural heritage data, but its effectiveness is often limited by incomplete or inconsistent textual descriptions, caused by historical data loss and the high cost of expert annotation. While large language models (LLMs) offer a promising solution by enriching textual descriptions, their outputs frequently suffer from hallucinations or miss visually grounded details. To address these challenges, we propose C^3, a data augmentation framework that enhances cross-modal retrieval performance by improving the completeness and consistency of LLM-generated descriptions. C^3 introduces a bidirectional validation mechanism to assess semantic coverage using both visual cues and language-model outputs. Furthermore, to mitigate factual inconsistencies, we formulate a Markov Decision Process to supervise Chain-of-Thought reasoning, guiding consistency verification through adaptive query control. Experiments on the cultural heritage dataset CulTi and general benchmarks MSCOCO and Flickr30K demonstrate that C^3 achieves state-of-the-art performance in both fine-tuned and zero-shot settings.

Chaojun Nie, Jun Zhou, Guanxiang Wang, Shisong Wu, Zichen Wang

Large language models (LLMs) often exhibit limited performance on domain-specific tasks due to the natural disproportionate representation of specialized information in their training data and the static nature of these datasets. Knowledge scarcity and temporal lag create knowledge gaps for domain applications. While post-training on domain datasets can embed knowledge into models, existing approaches have some limitations. Continual Pre-Training (CPT) treats all tokens in domain documents with equal importance, failing to prioritize critical knowledge points, while supervised fine-tuning (SFT) with question-answer pairs struggles to develop the coherent knowledge structures necessary for complex reasoning tasks. To address these challenges, we propose Reinforcement Learning from Augmented Generation (RLAG). Our approach iteratively cycles between sampling generations and optimizing the model through calculated rewards, effectively embedding critical and contextually coherent domain knowledge. We select generated outputs with the highest log probabilities as the sampling result, then compute three tailored reward metrics to guide the optimization process. To comprehensively evaluate domain expertise, we assess answer accuracy and the rationality of explanations generated for correctly answered questions. Experimental results across medical, legal, astronomy, and current events datasets demonstrate that our proposed method significantly outperforms baseline approaches. Our code and data are open sourced at https://github.com/ChaojunNie/RLAG.

Xiaoyu Luo, Yiyi Chen, Johannes Bjerva, Qiongxiu Li

We present the first comprehensive study of Memorization in Multilingual Large Language Models (MLLMs), analyzing 95 languages using models across diverse model scales, architectures, and memorization definitions. As MLLMs are increasingly deployed, understanding their memorization behavior has become critical. Yet prior work has focused primarily on monolingual models, leaving multilingual memorization underexplored, despite the inherently long-tailed nature of training corpora. We find that the prevailing assumption, that memorization is highly correlated with training data availability, fails to fully explain memorization patterns in MLLMs. We hypothesize that treating languages in isolation — ignoring their similarities — obscures the true patterns of memorization. To address this, we propose a novel graph-based correlation metric that incorporates language similarity to analyze cross-lingual memorization. Our analysis reveals that among similar languages, those with fewer training tokens tend to exhibit higher memorization, a trend that only emerges when cross-lingual relationships are explicitly modeled. These findings underscore the importance of a language-aware perspective in evaluating and mitigating memorization vulnerabilities in MLLMs. This also constitutes empirical evidence that language similarity both explains Memorization in MLLMs and underpins Cross-lingual Transferability, with broad implications for multilingual NLP.

Yuyang Wu, Jinhui Ye, Shuhao Zhang, Lu Dai, Yonatan Bisk, Olexandr Isayev

Large Language Models (LLMs) have shown growing potential in molecular sciences, but they often produce chemically inaccurate descriptions and struggle to recognize or justify potential errors. This raises important concerns about their robustness and reliability in scientific applications. To support more rigorous evaluation of LLMs in chemical reasoning, we present the MolErr2Fix benchmark, designed to assess LLMs on error detection and correction in molecular descriptions. Unlike existing benchmarks focused on molecule-to-text generation or property prediction, MolErr2Fix emphasizes fine-grained chemical understanding. It tasks LLMs with identifying, localizing, explaining, and revising potential structural and semantic errors in molecular descriptions. Specifically, MolErr2Fix consists of 1,193 fine-grained annotated error instances. Each instance contains quadruple annotations, i.e., (error type, span location, the explanation, and the correction). These tasks are intended to reflect the types of reasoning and verification required in real-world chemical communication. Evaluations of current state-of-the-art LLMs reveal notable performance gaps, underscoring the need for more robust chemical reasoning capabilities. MolErr2Fix provides a focused benchmark for evaluating such capabilities and aims to support progress toward more reliable and chemically informed language models. All annotations and an accompanying evaluation API will be publicly released to facilitate future research.

Chenlin Liu, Minghui Fang, Patrick Zhang, Wei Zhou, Jie Gao, Jiqing Han

Language Model (LM)-based Text-to-Speech (TTS) systems often generate hallucinated speech that deviates from input text. Existing mitigation strategies either demand excessive training resources or introduce significant inference latency. In this paper, we propose GFlOwNet-guided distribution AlignmenT (GOAT) for LM-based TTS, a post-training framework that mitigates hallucinations without relying on massive resources or inference cost. Specifically, we first conduct an uncertainty analysis, revealing a strong positive correlation between hallucination and model uncertainty. Based on this, we reformulate TTS generation as a trajectory flow optimization problem and introduce an enhanced Subtrajectory Balance objective together with a sharpened internal reward as target distribution. We further integrate reward temperature decay and learning rate optimization for stability and performance balance. Extensive experiments show that GOAT reduce over 50% character error rates on challenging test cases and lowering uncertainty by up to 58%, demonstrating its strong generalization ability and effectiveness.

Runze Li, Siyu Wu, Jun Wang, Wei Zhang

Knowledge Tracing (KT) aims to model a student’s learning state over time and predict their future performance. However, traditional KT methods often face challenges in explainability, scalability, and effective modeling of complex knowledge dependencies. While Large Language Models (LLMs) present new avenues for KT, their direct application often struggles with generating structured, explainable student representations and lacks mechanisms for continuous, task-specific refinement. To address these gaps, we propose Collaborative Iterative Knowledge Tracing (CIKT), a framework that harnesses LLMs to enhance both prediction accuracy and explainability. CIKT employs a dual-component architecture: an Analyst generates dynamic, explainable user profiles from student historical responses, and a Predictor utilizes these profiles to forecast future performance. The core of CIKT is a synergistic optimization loop. In this loop, the Analyst is iteratively refined based on the predictive accuracy of the Predictor, which conditions on the generated profiles, and the Predictor is subsequently retrained using these enhanced profiles. Evaluated on multiple educational datasets, CIKT demonstrates significant improvements in prediction accuracy, offers enhanced explainability through its dynamically updated user profiles, and exhibits improved scalability. Our work presents a robust and explainable solution for advancing knowledge tracing systems, effectively bridging the gap between predictive performance and model transparency.