论文检索

Junlong Tong, Jinlan Fu, Zixuan Lin, Yingqi Fan, Anhao Zhao, Hui Su, Xiaoyu Shen

Large Language Models (LLMs) are primarily designed for batch processing. Existing methods for adapting LLMs to streaming rely either on expensive re-encoding or specialized architectures with limited scalability. This work identifies three key mismatches in adapting batch-oriented LLMs to streaming: (1) input-attention, (2) output-attention, and (3) position-ID mismatches. While it is commonly assumed that the latter two mismatches require frequent re-encoding, our analysis reveals that only the input-attention mismatch significantly impacts performance, indicating re-encoding outputs is largely unnecessary. To better understand this discrepancy with the common assumption,we provide the first comprehensive analysis of the impact of position encoding on LLMs in streaming, showing that preserving relative positions within source and target contexts is more critical than maintaining absolute order. Motivated by the above analysis, we introduce a group position encoding paradigm built on batch architectures to enhance consistency between streaming and batch modes. Extensive experiments on cross-lingual and cross-modal tasks demonstrate that our method outperforms existing approaches. Our method requires no architectural modifications, exhibits strong generalization in both streaming and batch modes. The code is available at repository https://github.com/EIT-NLP/StreamingLLM.

Gurusha Juneja, Gautam Jajoo, Hua Li, Jian Jiao, Nagarajan Natarajan, Amit Sharma

Given a task in the form of a basic description and its training examples, prompt optimization is the problem of synthesizing the given information into a text prompt for a large language model. Humans solve this problem by also considering the different facets that define a task (e.g., counter-examples, explanations, analogies) and including them in the prompt. However, it is unclear whether existing algorithmic approaches, based on iteratively editing a given prompt or automatically selecting a few in-context examples, can cover the multiple facets required to solve a complex task. In this work, we view prompt optimization as that of learning multiple facets of a task from a set of training examples. We exploit structure in the prompt optimization problem and break down a prompt into loosely coupled semantic sections. The proposed algorithm, UniPrompt, (1) clusters the input space and uses clustered batches so that each batch likely corresponds to a different facet of the task, and (2) utilizes a feedback mechanism to propose adding, editing or deleting a section, which in turn is aggregated over a batch to capture generalizable facets. Empirical evaluation on multiple datasets and a real-world task shows that prompts generated using UniPrompt obtain higher accuracy than human-tuned prompts and those from state-of-the-art methods. In particular, our algorithm can generate long, complex prompts that existing methods are unable to generate.

Avinash Baidya, Kamalika Das, Xiang Gao

Large Language Model (LLM)-based agents have significantly impacted Task-Oriented Dialog Systems (TODS) but continue to face notable performance challenges, especially in zero-shot scenarios. While prior work has noted this performance gap, the behavioral factors driving the performance gap remain under-explored. This study proposes a comprehensive evaluation framework to quantify the behavior gap between AI agents and human experts, focusing on discrepancies in dialog acts, tool usage, and knowledge utilization. Our findings reveal that this behavior gap is a critical factor negatively impacting the performance of LLM agents. Notably, as task complexity increases, the behavior gap widens (correlation: 0.963), leading to a degradation of agent performance on complex task-oriented dialogs. For the most complex task in our study, even the GPT-4o-based agent exhibits low alignment with human behavior, with low F1 scores for dialog acts (0.464), excessive and often misaligned tool usage with a F1 score of 0.139, and ineffective usage of external knowledge. Reducing such behavior gaps leads to significant performance improvement (24.3% on average). This study highlights the importance of comprehensive behavioral evaluations and improved alignment strategies to enhance the effectiveness of LLM-based TODS in handling complex tasks.

Jan Trienes, Jörg Schlötterer, Junyi Jessy Li, Christin Seifert

Large Language Models (LLMs) excel at text summarization, a task that requires models to select content based on its importance. However, the exact notion of salience that LLMs have internalized remains unclear. To bridge this gap, we introduce an explainable framework to systematically derive and investigate information salience in LLMs through their summarization behavior. Using length-controlled summarization as a behavioral probe into the content selection process, and tracing the answerability of Questions Under Discussion throughout, we derive a proxy for how models prioritize information. Our experiments on 13 models across four datasets reveal that LLMs have a nuanced, hierarchical notion of salience, generally consistent across model families and sizes. While models show highly consistent behavior and hence salience patterns, this notion of salience cannot be accessed through introspection, and only weakly correlates with human perceptions of information salience.

Metehan Oğuz, Yavuz Faruk Bakman, Duygu Nur Yaldiz

Large Language Models (LLMs) have demonstrated impressive performances in tasks related to coreference resolution. However, previous studies mostly assessed LLM performance on coreference resolution with nouns and third person pronouns. This study evaluates LLM performance on coreference resolution with indexical like I, you, here and tomorrow which come with unique challenges due to their linguistic properties. We present the first study examining how LLMs interpret indexicals in English, releasing the English Indexical Dataset with 1600 multiple-choice questions. We evaluate pioneering LLMs, including GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Pro, and DeepSeek V3. Our results reveal that LLMs exhibit an impressive performance with some indexicals (I), while struggling with others (you, here, tomorrow), and that syntactic cues (e.g. quotation) contribute to LLM performance with some indexicals, while they reduce performance with others. Code and data are available at: https://github.com/metehanoguzz/LLMs-Indexicals-English

Xiaoxin Lu, Ranran Haoran Zhang, Yusen Zhang, Rui Zhang

People get informed of a daily task plan through diverse media involving both texts and images. However, most prior research only focuses on LLM’s capability of textual plan generation. The potential of large-scale models in providing text-image plans remains understudied. Generating high-quality text-image plans faces two main challenges: ensuring consistent alignment between two modalities and keeping coherence among visual steps. To address these challenges, we propose a novel framework that generates and refines text-image plans step-by-step. At each iteration, our framework (1) drafts the next textual step based on the prediction history; (2) edits the last visual step to obtain the next one; (3) extracts PDDL-like visual information; and (4) refines the draft with the extracted visual information. The textual and visual step produced in stage (4) and (2) will then serve as inputs for the next iteration. Our approach offers a plug-and-play improvement to various backbone models, such as Mistral-7B, Gemini-1.5, and GPT-4o. To evaluate the effectiveness of our approach, we collect a new benchmark consisting of 1,100 tasks and their text-image pair solutions covering 11 daily topics. We also design and validate a new set of metrics to evaluate the multimodal consistency and coherence in text-image plans. Extensive experiment results show the effectiveness of our approach on a range of backbone models against competitive baselines.

Yiran Zhao, Guizhen Chen, Kenji Kawaguchi, Lidong Bing, Wenxuan Zhang

Large Language Models (LLMs) have transformed natural language processing, yet their substantial model sizes often demand significant computational resources. To preserve computing resources and accelerate inference speed, it is crucial to prune redundant parameters, especially for experienced users who often need expert models tailored to specific downstream scenarios. However, current pruning methods primarily focus on maintaining models’ general capabilities, either requiring extensive post-training or performing poorly due to coarse-grained pruning. In this work, we design a \underline{Cus}tom \underline{Prun}ing method (\texttt{Cus-Prun}) to prune a large general model into a smaller lightweight expert model, which is positioned along the “language”, “domain” and “task” dimensions. By identifying and pruning irrelevant neurons of each dimension, \texttt{Cus-Prun} creates expert models without any post-training. Our experiments demonstrate that \texttt{Cus-Prun} consistently outperforms other methods, achieving minimal loss in both expert and general capabilities across various models from different model families and sizes.

Fei Yuan, Yinquan Lu, Lei Li, Jingjing Xu

It is a critical challenge to learn a single model for massive languages. Prior methods focus on increasing the model size and training data size. However, large models are difficult to optimize efficiently even with distributed parallel training and translation capacity can interfere among languages. To address the challenge, we propose LegoMT2, an efficient training approach with an asymmetric multi-way model architecture for massive multilingual neural machine translation. LegoMT2 shards 435 languages into 8 language-centric groups and attributes one local encoder for each group’s languages and a mix encoder-decoder for all languages. LegoMT2 trains the model through local data parallel and asynchronous distributed updating of parameters. LegoMT2 is 16.2\times faster than the distributed training method for M2M-100-12B (which only for 100 languages) while improving the translation performance by an average of 2.2 BLEU on Flores-101, especially performing better for low-resource languages .

Yuji Zhang, Sha Li, Cheng Qian, Jiateng Liu, Pengfei Yu, Chi Han, Yi R. Fung, Kathleen McKeown, ChengXiang Zhai, Manling Li 等

Hallucination is a persistent challenge in large language models (LLMs), where even with rigorous quality control, models often generate distorted facts. This paradox, in which error generation continues despite high-quality training data, calls for a deeper understanding of the underlying LLM mechanisms. To address it, we propose a novel concept: knowledge overshadowing, where model’s dominant knowledge can obscure less prominent knowledge during text generation, causing the model to fabricate inaccurate details. Building on this idea, we introduce a novel framework to quantify factual hallucinations by modeling knowledge overshadowing. Central to our approach is the log-linear law, which predicts that the rate of factual hallucination increases linearly with the logarithmic scale of (1) Knowledge Popularity, (2) Knowledge Length, and (3) Model Size. The law provides a means to preemptively quantify hallucinations, offering foresight into their occurrence even before model training or inference. Built on overshadowing effect, we propose a new decoding strategy CoDa, to mitigate hallucinations, which notably enhance model factuality on Overshadow (27.9%), MemoTrap (13.1%) and NQ-Swap (18.3%). Our findings not only deepen understandings of the underlying mechanisms behind hallucinations but also provide actionable insights for developing more predictable and controllable language models.

Sigang Luo, Yinan Liu, Dongying Lin, Yingying Zhai, Bin Wang, Xiaochun Yang, Junpeng Liu

Event temporal reasoning (ETR) aims to model and reason about the relationships between events and time, as well as between events in the real world. Proficiency in ETR is a significant indicator that a large language model (LLM) truly understands the physical world. Previous question-answering datasets available for evaluating the ETR ability lack a systematic taxonomy and pay limited attention to compound questions. In this paper, we propose a unified taxonomy for event temporal questions and construct a comprehensive benchmark ETRQA, to evaluate the ETR abilities of LLMs based on this taxonomy. ETRQA not only inherits and expands the evaluation content of existing datasets but also contains multiple categories of compound questions. We evaluate two leading LLM series, Llama and Qwen, on ETRQA across various settings. Our experimental results indicate that large-scale LLMs exhibit certain ETR abilities. Yet they do not perform well in solving specific types of reasoning tasks, including reasoning involving time spans, reasoning for compound questions, and reasoning with fine temporal granularity. Additionally, we hope ETRQA can benefit the temporal reasoning research community for future studies.

Fengqing Jiang, Zhangchen Xu, Yuetai Li, Luyao Niu, Zhen Xiang, Bo Li, Bill Yuchen Lin, Radha Poovendran

Emerging large reasoning models (LRMs), such as DeepSeek-R1 models, leverage long chain-of-thought (CoT) reasoning to generate structured intermediate steps, enhancing their reasoning capabilities. However, long CoT does not inherently guarantee safe outputs, potentially leading to harmful consequences such as the introduction of security vulnerabilities in code or the spread of misinformation. Current research on large language model (LLM) safety usually focuses on short-answer responses, overlooking the long CoT style outputs of LRMs. To bridge this gap, we conduct a systematic study of LRM safety. First, we investigate safety evaluators calibrated against human annotations. Using our newly developed metrics, we thoroughly assess the safety of 13 state-of-the-art LRMs on StrongReject and WildJailbreak datasets. Our results show that LRMs are not safe compared to their reasoning advance. Further, we perform a fine-grained analysis of the reasoning trace and final answer. We find that three decoding strategies-ZeroThink, LessThink, and MoreThink-can improve model safety without additional training. However, these strategies either use constrained reasoning traces or incur high inference costs. To better strengthen LRM safety, we introduce SafeChain, the first-of-its-kind safety training dataset in CoT style. We fine-tune two LRMs with SafeChain, showing that it not only enhances model safety but also preserves performance across 6 reasoning benchmarks.

Yuchen Wu, Liang Ding, Li Shen, Dacheng Tao

Knowledge editing allows for efficient adaptation of large language models (LLMs) to new information or corrections without requiring full retraining. However, prior methods typically focus on either single-language editing or basic multilingual editing, failing to achieve true cross-linguistic knowledge synchronization. To address this, we present a simple and practical state-of-the-art (SOTA) recipe Cross-Lingual Knowledge Democracy Edit (X-KDE), designed to propagate knowledge from a dominant language to other languages effectively. Our X-KDE comprises two stages: (i) Cross-lingual Edition Instruction Tuning (XE-IT), which fine-tunes the model on a curated parallel dataset to modify in-scope knowledge while preserving unrelated information, and (ii) Target-language Preference Optimization (TL-PO), which applies advanced optimization techniques to ensure consistency across languages, fostering the transfer of updates. Additionally, we contribute a high-quality, cross-lingual dataset, specifically designed to enhance knowledge transfer across languages. Extensive experiments on the Bi-ZsRE and MzsRE benchmarks show that X-KDE significantly enhances cross-lingual performance, achieving an average improvement of +8.19%, while maintaining high accuracy in monolingual settings.

Xiaoqing Zhang, Yuhan Liu, Flood Sung, Xiuying Chen, Shuo Shang, Rui Yan

Code generation is crucial in software engineering for automating the coding process efficiently. While test-time computation methods show promise, they suffer from high latency due to multiple computation rounds.To overcome this, we introduce ThinkCoder, a framework that combines thorough exploration with optimal refinement.The exploration phase diversifies the solution space by searching for potential solutions, followed by a refinement phase that enhances precision.This approach allows us to select the best solution through careful consideration before taking action, avoiding excessive trial and error.To further minimize test-time computation overhead, we introduce preference-driven optimization with Reinforced Self-Training (ReST), which uses exploration trajectories from ThinkCoder to guide LLM’s evolution.This approach enhances LLM’s exploration efficiency via preference learning, cutting costs while maintaining accuracy.ThinkCoder boosts the performance with a single LLM, excelling on benchmarks like HumanEval and MBPP. Compared to SOTA models, it improves Pass@1 by 3.0% over MapCoder with just 6.4% of the computation cost.Against AgentCoder, ThinkCoder achieves a 0.5% higher Pass@1 after 2 rounds, outperforming AgentCoder’s 5 rounds.Additionally, ReST with success trajectories enhances efficiency, allowing models like LLaMA2-7B to achieve competitive results using only 20% of the computational resources. These results highlight the framework’s effectiveness and scalability.

Zeyu Gao, Yuxin Cui, Hao Wang, Siliang Qin, Yuanda Wang, Zhang Bolun, Chao Zhang

Decompilers are fundamental tools for critical security tasks, from vulnerability discovery to malware analysis, yet their evaluation remains fragmented. Existing approaches primarily focus on syntactic correctness through synthetic micro-benchmarks or subjective human ratings, failing to address real-world requirements for semantic fidelity and analyst usability. We present **DecompileBench**, the first comprehensive framework that enables effective evaluation of decompilers in reverse engineering workflows through three key components: real-world function extraction (comprising 23,400 functions from 130 real-world programs), runtime-aware validation, and automated human-centric assessment using LLM-as-Judge to quantify the effectiveness of decompilers in reverse engineering workflows. Through a systematic comparison between six industrial-strength decompilers and six recent LLM-powered approaches, we demonstrate that LLM-based methods surpass commercial tools in code understandability despite 52.2% lower functionality correctness. These findings highlight the potential of LLM-based approaches to transform human-centric reverse engineering. We open source **DecompileBench** to provide a framework to advance research on decompilers and assist security experts in making informed tool selections based on their specific requirements.

Dylan Zhang, Justin Wang, Francois Charton

Instruction-tuned language models excel in knowledge, reasoning, and instruction-following. While knowledge and reasoning are well-explored, the factors enabling generalization to unseen instructions remain underexplored due to challenges in isolating instruction-following dynamics.In this work, we model instruction-following as a computational process and design controlled experiments inspired by the Turing-complete Markov algorithm to disentangle its dynamics. Our findings reveal that the ability to generalize to instructions with unseen semantics emerges only when training data is strategically diversified across rich semantics. This finding gives us the hammer that breaks down the wall separating training instructions from unseen ones encountered in the wild. For specialist models, a balanced mix of in-domain and diverse out-of-domain tasks enhances performance more effectively than simply increasing in-domain data. For generalist models, domain diversification consistently outweighs the costs of reduced task-specific data, regardless of data budgets. Furthermore, we show that proper diversification with a lower data budget can outperform simply scaling up data volume. These findings highlight strategic data diversification as key to optimizing instruction-following and improving model performance across applications.

Ming Wang, Peidong Wang, Lin Wu, Xiaocui Yang, Daling Wang, Shi Feng, Yuxin Chen, Bixuan Wang, Yifei Zhang

Constrained by the cost and ethical concerns of involving real seekers in AI-driven mental health, researchers develop LLM-based conversational agents (CAs) with tailored configurations, such as profiles, symptoms, and scenarios, to simulate seekers. While these efforts advance AI in mental health, achieving more realistic seeker simulation remains hindered by two key challenges: dynamic evolution and multi-session memory. Seekers’ mental states often fluctuate during counseling, which typically spans multiple sessions. To address this, we propose **AnnaAgent**, an emotional and cognitive dynamic agent system equipped with tertiary memory. AnnaAgent incorporates an emotion modulator and a complaint elicitor trained on real counseling dialogues, enabling dynamic control of the simulator’s configurations. Additionally, its tertiary memory mechanism effectively integrates short-term and long-term memory across sessions. Evaluation results, both automated and manual, demonstrate that AnnaAgent achieves more realistic seeker simulation in psychological counseling compared to existing baselines. The ethically reviewed and screened code can be found on [https://github.com/sci-m-wang/AnnaAgent](https://github.com/sci-m-wang/AnnaAgent).

Yanfang Zhou, Xiaodong Li, Yuntao Liu, Yongqiang Zhao, Xintong Wang, Zhenyu Li, Jinlong Tian, Xinhai Xu

Open-world planning poses a significant challenge for general artificial intelligence due to environmental complexity and task diversity, especially in long-term tasks and lifelong learning. Inspired by cognitive theories, we propose M2PA, an open-world multi-memory planning agent. M2PA innovates by combining Large Language Models (LLMs) with human-like multi-memory systems, aiming to fully leverage the strengths of both while mitigating their respective limitations. By integrating the expansive world knowledge and language processing capabilities of LLMs with the perception and experience accumulation abilities of the human memory system, M2PA exhibits situation awareness, and experience generalization capabilities, as well as the potential for lifelong learning. In experiments, M2PA significantly outperforms current state-of-the-art agents across 50 Minecraft tasks in zero-shot learning. In exploratory lifelong learning experiments, M2PA demonstrates its continuous learning ability, achieving a 38.33% success rate in the “ObtainDiamond” task. Our findings provide a novel paradigm for constructing more effective agents in open-world environments.

Hanin Atwany, Abdul Waheed, Rita Singh, Monojit Choudhury, Bhiksha Raj

Speech foundation models trained at a massive scale, both in terms of model and data size, result in robust systems capable of performing multiple speech tasks, including automatic speech recognition (ASR). These models transcend language and domain barriers, yet effectively measuring their performance remains a challenge. Traditional metrics like word error rate (WER) and character error rate (CER) are commonly used to evaluate ASR performance but often fail to reflect transcription quality in critical contexts, particularly when detecting fabricated outputs. This phenomenon, known as hallucination, is especially concerning in high-stakes domains such as healthcare, legal, and aviation, where errors can have severe consequences. In our work, we address this gap by investigating hallucination in ASR models. We examine how factors such as distribution shifts, model size, and model architecture influence the hallucination error rate (HER), a metric we introduce to quantify hallucinations. Our analysis of over 20 ASR models reveals key insights: (1) High WERs can mask low hallucination rates, while low WERs may conceal dangerous hallucinations. (2) Synthetic noise, both adversarial and common perturbations like white noise, pitch shift, and time stretching, increase HER. (3) Distribution shift correlates strongly with HER (\alpha = 0.91). Our findings highlight the importance of incorporating HER alongside traditional metrics like WER to better assess ASR model performance, particularly in high-stakes domains.

Xinxin Li, Huiyao Chen, Chengjun Liu, Jing Li, Meishan Zhang, Jun Yu, Min Zhang

Semantic role labeling (SRL) is a crucial task of natural language processing (NLP). Although generative decoder-based large language models (LLMs) have achieved remarkable success across various NLP tasks, they still lag behind state-of-the-art encoder-decoder (BERT-like) models in SRL. In this work, we seek to bridge this gap by equipping LLMs for SRL with two mechanisms: (a) retrieval-augmented generation and (b) self-correction. The first mechanism enables LLMs to leverage external linguistic knowledge such as predicate and argument structure descriptions, while the second allows LLMs to identify and correct inconsistent SRL outputs. We conduct extensive experiments on three widely-used benchmarks of SRL (CPB1.0, CoNLL-2009, and CoNLL-2012). Results demonstrate that our method achieves state-of-the-art performance in both Chinese and English, marking the first successful application of LLMs to surpass encoder-decoder approaches in SRL.

Yipeng Kang, Junqi Wang, Yexin Li, Mengmeng Wang, Wenming Tu, Quansen Wang, Hengli Li, Tingjun Wu, Xue Feng, Fangwei Zhong 等

As large language models (LLMs) become increasingly integrated into critical applications, aligning their behavior with human values presents significant challenges. Current methods, such as Reinforcement Learning from Human Feedback (RLHF), typically focus on a limited set of coarse-grained values and are resource-intensive. Moreover, the correlations between these values remain implicit, leading to unclear explanations for value-steering outcomes. Our work argues that a latent causal value graph underlies the value dimensions of LLMs and that, despite alignment training, this structure remains significantly different from human value systems. We leverage these causal value graphs to guide two lightweight value-steering methods: role-based prompting and sparse autoencoder (SAE) steering, effectively mitigating unexpected side effects. Furthermore, SAE provides a more fine-grained approach to value steering. Experiments on Gemma-2B-IT and Llama3-8B-IT demonstrate the effectiveness and controllability of our methods.