论文检索

Constanza Fierro, Negar Foroutan, Desmond Elliott, Anders Søgaard

Large Language Models (LLMs) store and retrieve vast amounts of factual knowledge acquired during pre-training. Prior research has localized and identified mechanisms behind knowledge recall; however, it has only focused on English monolingual models. The question of how these mechanisms generalize to non-English languages and multilingual LLMs remains unexplored. In this paper, we address this gap by conducting a comprehensive analysis of three multilingual LLMs. First, we show that previously identified recall mechanisms in English largely apply to multilingual contexts, with nuances based on language and architecture. Next, through patching intermediate representations, we localize the role of language during recall, finding that subject enrichment is language-independent, while object extraction is language-dependent. Additionally, we discover that the last token representation acts as a Function Vector (FV), encoding both the language of the query and the content to be extracted from the subject. Furthermore, in decoder-only LLMs, FVs compose these two pieces of information in two separate stages. These insights reveal unique mechanisms in multilingual LLMs for recalling information, highlighting the need for new methodologies—such as knowledge evaluation, fact editing, and knowledge acquisition—that are specifically tailored for multilingual LLMs.

Hongzhi Qi, Nan Bai, Jianqiang Li, Wei Zhai, Qing Zhao, Qi Gao, Bing Xiang Yang, Guanghui Fu

Cognitive distortion is a critical issue in psychology, with most existing studies based on Burns’ cognitive distortion theory. However, differences in annotation standards lead to variations in building analysis tools, resulting in inconsistent analyses and limiting the generalizability of findings, especially in large-scale and cross-linguistic contexts. To address this issue, we collected all publicly available datasets (four in total) and conducted a series of experiments to evaluate the generalizability of various cross-linguistic models. The results indicate that models exhibit significant performance differences across datasets, highlighting the generalization problem. To mitigate this issue, we propose two solutions. First, we propose a multi-task learning model based on teacher student architecture solution, which demonstrates improved generalization performance in our experiments. Second, we introduce a new dataset (~5,000 samples) derived from reannotating existing open datasets to ensure standardized alignment. The annotation process we provided is interpretable and grounded in psychological principles. Based on this, we constructed large language models with cognitive reasoning chains, enhancing both generalizability and interpretability. This study identifies the generalization challenge in cognitive distortion research, and our experiments show that the proposed solutions significantly improve model performance. The dataset and code are publicly available at: https://github.com/HongzhiQ/CrossLinCD.

Sara Rajaee, Kumar Pratik, Gabriele Cesa, Arash Behboodi

The most promising recent methods for AI reasoning require applying variants of reinforcement learning (RL) either on rolled out trajectories from the LLMs, even for the step-wise rewards, or large quantities of human-annotated trajectory data. The reliance on the rolled-out trajectory renders the compute cost and time prohibitively high. In particular, the correctness of a reasoning trajectory can typically only be judged at its completion, leading to sparse rewards in RL or requiring expensive synthetic data generation in expert iteration-like methods. In this work, we focus on the Automatic Theorem Proving (ATP) task and propose a novel verifier-in-the-loop design, which, unlike existing approaches that leverage feedback on the entire reasoning trajectory, employs an automated verifier to give intermediate feedback at each step of the reasoning process. Using Lean as the verifier, we empirically show that the step-by-step local verification produces a global improvement in the model’s reasoning accuracy and efficiency.

Naquee Rizwan, Seid Muhie Yimam, Daryna Dementieva, Dr. Florian Skupin, Tim Fischer, Daniil Moskovskiy, Aarushi Ajay Borkar, Robert Geislinger, Punyajoy Saha, Sarthak Roy 等

Despite regulations imposed by nations and social media platforms, e.g. (Government of India, 2021; European Parliament and Council of the European Union, 2022), inter alia, hateful content persists as a significant challenge. Existing approaches primarily rely on reactive measures such as blocking or suspending offensive messages, with emerging strategies focusing on proactive measurements like detoxification and counterspeech. In our work, which we call HATEPRISM, we conduct a comprehensive examination of hate speech regulations and strategies from three perspectives: country regulations, social platform policies, and NLP research datasets. Our findings reveal significant inconsistencies in hate speech definitions and moderation practices across jurisdictions and platforms, alongside a lack of alignment with research efforts. Based on these insights, we suggest ideas and research direction for further exploration of a unified framework for automated hate speech moderation incorporating diverse strategies.

Zihao Li, Xuekong Xu, Ziyao Chen, Lixin Zou, Haijun Wu, Qiang Chen, Chenliang Li

Multi-style outline controllable generation is crucial for multiple applications, including document semantic structuring and retrieval-augmented generation.The great success of preference alignment approaches encourages their application in controllable generation tasks.However, these attempts encounter several limitations: (1) response pair requirements, (2) substantial computation costs, and (3) insufficient exploitation of fine-grained preference signals.To address these problems, we propose a token-level preference self-alignment optimization, named TKPO, for outline controllable generation. TKPO extends the Bradley-Terry model from pair-wise to list-wise comparison, which is further applied at the token level for fine-grained preference signal utilization. In comparison to the representative methods, e.g., DPO, TKPO does not require response pairs; instead, we propose a controllable attributes-driven method to construct reject samples for self-alignment. Additionally, TKPO optimizes only the base model, thereby avoiding additional memory usage and substantial computational costs.We curate two outline controllable generation datasets with regard to language style and level-of-detail.Extensive experiments demonstrate that TKPO outperforms DPO by up to 19.28% in performance while requiring only 56.25% in training time.We release the code and datasets resources at https://github.com/WHUIR/TKPO.

Yanggan Gu, Junzhuo Li, Sirui Huang, Xin Zou, Zhenghua Li, Xuming Hu

Aligning small language models (SLMs) with human values typically involves distilling preference knowledge from large language models (LLMs). However, existing distillation methods model preference knowledge in teacher LLMs by comparing pairwise responses, overlooking the extent of difference between responses. This limitation hinders student SLMs from capturing the nuanced preferences for multiple responses. In this paper, we propose a Preference-Aligned Distillation (PAD) framework, which models teacher’s preference knowledge as a probability distribution over all potential preferences, thereby providing more nuanced supervisory signals. Our insight in developing PAD is rooted in the demonstration that language models can serve as reward functions, reflecting their intrinsic preferences. Based on this, PAD comprises three key steps: (1) sampling diverse responses using high-temperature; (2) computing rewards for both teacher and student to construct their intrinsic preference; and (3) training the student’s intrinsic preference distribution to align with the teacher’s. Experiments on four mainstream alignment benchmarks demonstrate that PAD consistently and significantly outperforms existing approaches, achieving over 20% improvement on AlpacaEval 2 and Arena-Hard, indicating superior alignment with human preferences. Notably, on MT-Bench, using the Gemma model family, the student trained by PAD surpasses its teacher, further validating the effectiveness of our PAD.

Tianpeng Bu, Minying Zhang, Hongtao Duan, Shurui Li, Lulu Hu, Yu Li

Large language models (LLMs) excel in problem-solving but require training data with diverse reasoning processes. Existing methods mainly optimize instruction-response pairs but lack a systematic design for the underlying reasoning structure. This paper proposes RSS: a Reasoning Structure driven data Synthesis method. We first proactively develop a hierarchical GFlowNet to construct reasoning structures efficiently through a coarse-to-fine directed acyclic graph (DAG) growth process. Then reasoning DAGs are leveraged to actively guide the instruction generation via an iterative suggester-editor workflow and enhance response quality using a structure-aware strategy. Experiments show that LLMs trained on our synthetic datasets achieve 48.50%, 84.00%, 79.90% for AlpacaEval2, GSM8K and HumanEval, outperforming existing data synthesis methods.

Tianyi Alex Qiu, Fanzhi Zeng, Jiaming Ji, Dong Yan, Kaile Wang, Jiayi Zhou, Yang Han, Josef Dai, Xuehai Pan, Yaodong Yang

Existing alignment methods share a common topology of information flow, where reward information is collected from humans, modeled with preference learning, and used to tune language models. However, this shared topology has not been systematically characterized, nor have its alternatives been thoroughly explored, leaving the problems of low data efficiency and unreliable generalization unaddressed. As a solution, we introduce a theory of **reward generalization** in reinforcement learning from human feedback (RLHF), focusing on the **topology of information flow** at both macro and micro levels. At the macro level, we portray the RLHF information flow as an autoencoding process over behavior distributions, formalizing the RLHF objective of distributional consistency between human preference and model behavior. At the micro level, we present *induced Bayesian networks* to model the impact of dataset topologies on reward generalization. Combining analysis on both levels, we propose **reward modeling from tree-structured preference information**. It is shown to reduce reward uncertainty by up to \Theta(\log n/\log\log n) times compared to baselines, where n is the dataset size. Validation on three NLP tasks shows that it achieves an average win rate of 65% against baselines, thus improving reward generalization *for free* via topology design, while *reducing* the amount of data requiring annotation.

Xiaojing Zhang, Bochen Lyu

Conceptual abstraction and analogy-making are crucial for human learning, reasoning, and adapting to unfamiliar domains. Recently, large language models (LLMs) have made the synthesis of analogical data possible, which, however, still heavily relies on extensive human efforts to be annotated. This paper empirically examines the LLMs’ capability to annotate story-level analogical data. Specifically, we propose a novel multi-stage progressive reasoning prompt framework \texttt{A3E} (Automated Analogy Annotation Expert), which is based on the structure mapping theory from cognitive psychology and efficiently annotates candidate story pairs across six fine-grained categories. We use \texttt{A3E} to evaluate how well the state-of-the-art LLMs can serve as analogy annotators. Experimental results demonstrate that our proposed \texttt{A3E} achieves an average performance gain of + 73% across a range of prompting baselines and base LLMs. The code and data is available at https://github.com/zhangxjohn/A3E.

Kseniia Petukhova, Ekaterina Kochmar

Recent advances in Large Language Models (LLMs) have shown promise in automating discourse annotation for conversations. While manually designing tree annotation schemes significantly improves annotation quality for humans and models, their creation remains time-consuming and requires expert knowledge. We propose a fully automated pipeline that uses LLMs to construct such schemes and perform annotation. We evaluate our approach on speech functions (SFs) and the Switchboard-DAMSL (SWBD-DAMSL) taxonomies. Our experiments compare various design choices, and we show that frequency-guided decision trees, paired with an advanced LLM for annotation, can outperform previously manually designed trees and even match or surpass human annotators while significantly reducing the time required for annotation. We release all code and resultant schemes and annotations to facilitate future research on discourse annotation.

Patomporn Payoungkhamdee, Pume Tuchinda, Jinheon Baek, Samuel Cahyawijaya, Can Udomcharoenchaikit, Potsawee Manakul, Peerat Limkonchotiwat, Ekapol Chuangsuwanich, Sarana Nutanong

Multi-step reasoning is essential for large language models (LLMs), yet multilingual performance remains challenging. While Chain-of-Thought (CoT) prompting improves reasoning, it struggles with non-English languages due to the entanglement of reasoning and execution. Program-of-Thought (PoT) prompting separates reasoning from execution, offering a promising alternative but shifting the challenge to generating programs from non-English questions. We propose a framework to evaluate PoT by separating multilingual reasoning from code execution to examine (i) the impact of fine-tuning on question-reasoning alignment and (ii) how reasoning quality affects answer correctness. Our findings demonstrate that PoT fine-tuning substantially enhances multilingual reasoning, outperforming CoT fine-tuned models. We further demonstrate a strong correlation between reasoning quality (measured through code quality) and answer accuracy, highlighting its potential as a test-time performance improvement heuristic.

Jiaen Lin, Jingyu Liu, Yingbo Liu

Retrieval-augmented generation (RAG) encounters challenges when addressing complex queries, particularly multi-hop questions. While several methods tackle multi-hop queries by iteratively generating internal queries and retrieving external documents, these approaches are computationally expensive. In this paper, we identify a three-stage information processing pattern in LLMs during layer-by-layer reasoning, consisting of extraction, processing, and subsequent extraction steps. This observation suggests that the representations in intermediate layers contain richer information compared to those in other layers. Building on this insight, we propose Layer-wise RAG (L-RAG). Unlike prior methods that focus on generating new internal queries, L-RAG leverages intermediate representations from the middle layers, which capture next-hop information, to retrieve external knowledge. L-RAG achieves performance comparable to multi-step approaches while maintaining inference overhead similar to that of standard RAG. Experimental results show that L-RAG outperforms existing RAG methods on open-domain multi-hop question-answering datasets, including MuSiQue, HotpotQA, and 2WikiMultiHopQA. The code is available in https://anonymous.4open.science/r/L-RAG-ADD5/.

Zhiyu Lin, Zhengda Zhou, Zhiyuan Zhao, Tianrui Wan, Yilun Ma, Junyu Gao, Xuelong Li

With the rapid advancement of Generative AI technology, Multimodal Large Language Models(MLLMs) have the potential to act as AI software engineers capable of executing complex web application development. Considering that the model requires a confluence of multidimensional sub-capabilities to address the challenges of various development phases, constructing a multi-view evaluation framework is crucial for accurately guiding the enhancement of development efficiency. However, existing benchmarks usually fail to provide an assessment of sub-capabilities and focus solely on webpage generation outcomes. In this work, we draw inspiration from the principles of software engineering and further propose WebUIBench, a benchmark systematically designed to evaluate MLLMs in four key areas: WebUI Perception, HTML Programming, WebUI-HTML Understanding, and WebUI-to-Code. WebUIBench comprises 21K high-quality question-answer pairs derived from over 0.7K real-world websites. The extensive evaluation of 29 mainstream MLLMs uncovers the skill characteristics and various weakness that models encountered during the development process.

Michele Luca Contalbo, Sara Pederzoli, Francesco Del Buono, Venturelli Valeria, Francesco Guerra, Matteo Paganelli

Assessing corporate environmental sustainability with Table Question Answering systems is challenging due to complex tables, specialized terminology, and the variety of questions they must handle. In this paper, we introduce GRI-QA, a test benchmark designed to evaluate Table QA approaches in the environmental domain. Using GRI standards, we extract and annotate tables from non-financial corporate reports, generating question-answer pairs through a hybrid LLM-human approach. The benchmark includes eight datasets, categorized by the types of operations required, including operations on multiple tables from multiple documents. Our evaluation reveals a significant gap between human and model performance, particularly in multi-step reasoning, highlighting the relevance of the benchmark and the need for further research in domain-specific Table QA. Code and benchmark datasets are available at https://github.com/softlab-unimore/gri_qa.

Rui Li, Heming Xia, Xinfeng Yuan, Qingxiu Dong, Lei Sha, Wenjie Li, Zhifang Sui

Recently, LLMs have garnered increasing attention across academic disciplines for their potential as human digital twins, virtual proxies designed to replicate individuals and autonomously perform tasks such as decision-making, problem-solving, and reasoning on their behalf.However, current evaluations of LLMs primarily emphasize dialogue simulation while overlooking human behavior simulation, which is crucial for digital twins.To address this gap, we introduce BehaviorChain, the first benchmark for evaluating LLMs’ ability to simulate continuous human behavior.BehaviorChain comprises diverse, high-quality, persona-based behavior chains, totaling 15,846 distinct behaviors across 1,001 unique personas, each with detailed history and profile metadata.For evaluation, we integrate persona metadata into LLMs and employ them to iteratively infer contextually appropriate behaviors within dynamic scenarios provided by BehaviorChain. Comprehensive evaluation results demonstrated that even state-of-the-art models struggle with accurately simulating continuous human behavior.

I-Fan Lin, Faegheh Hasibi, Suzan Verberne

In this paper, we propose Selection and Pooling with Large Language Models (SPILL), an intuitive, domain-adaptive method for intent clustering without fine-tuning. Existing embeddings-based clustering methods rely on a few labeled examples or unsupervised fine-tuning to optimize results for each new dataset, which makes them less generalizable to multiple datasets. Our goal is to make these existing embedders more generalizable to new domain datasets without further fine-tuning. Inspired by our theoretical derivation and simulation results on the effectiveness of sampling and pooling techniques, we view the clustering task as a small-scale selection problem. A good solution to this problem is associated with better clustering performance. Accordingly, we propose a two-stage approach: First, for each utterance (referred to as the seed), we derive its embedding using an existing embedder. Then, we apply a distance metric to select a pool of candidates close to the seed. Because the embedder is not optimized for new datasets, in the second stage, we use an LLM to further select utterances from these candidates that share the same intent as the seed. Finally, we pool these selected candidates with the seed to derive a refined embedding for the seed. We found that our method generally outperforms directly using an embedder, and it achieves comparable results to other state-of-the-art studies, even those that use much larger models and require fine-tuning, showing its strength and efficiency. Our results indicate that our method enables existing embedders to be further improved without additional fine-tuning, making them more adaptable to new domain datasets. Additionally, viewing the clustering task as a small-scale selection problem gives the potential of using LLMs to customize clustering tasks according to the user’s goals.

Zijing Zhang, Zhanpeng Chen, He Zhu, Ziyang Chen, Nan Du, Xiaolong Li

Tool learning enhances Large Language Models’ (LLMs) dynamic interaction with external tools, improving their ability to solve complex problems. However, current empirical methods, which primarily focus on isolated tools learning, still struggle with accurate multi-tool selection due to issues like confusing similar tools and neglecting dependencies. To address these challenges, we propose the Tool Experience Network (ToolExpNet), which integrates tools and trial-and-error experiences into a network characterized by semantic similarity and dependency relationships. ToolExpNet iteratively conducts simulated experiments using adaptive sampling to explore subtle differences and connections between tools, and summarizes these experiences to provide insightful guidance for LLM tool selection. Our experiments demonstrate that learning the relationships between tools helps achieve more comprehensive tool learning. Evaluations on multiple real-world API datasets show that ToolExpNet effectively addresses common challenges in multi-tool selection, significantly outperforming existing baselines across different foundation LLMs.

Jonas Wallat, Abdelrahman Abdallah, Adam Jatowt, Avishek Anand

Large Language Models (LLMs) encapsulate a surprising amount of factual world knowledge. However, their performance on temporal questions and historical knowledge is limited because they often cannot understand temporal scope and orientation or neglect the temporal aspect altogether.In this study, we aim to measure precisely how robust LLMs are for question answering based on their ability to process temporal information and perform tasks requiring temporal reasoning and temporal factual knowledge. Specifically, we design eight time-sensitiverobustness tests for factual information to check the sensitivity of six popular LLMs in the zero-shot setting.Overall, we find LLMs lacking temporal robustness, especially to temporal reformulations and the use of different granularities of temporal references. We show how a selection of these eight tests can be used automatically to judge a model’s temporal robustness for user questions on the fly. Finally, we apply the findings of this study to improve the temporal QA performance by up to 55%.

Xinyi Liu, Xiaoyi Zhang, Ziyun Zhang, Yan Lu

Recent advancements in Large Vision-Language Models are accelerating the development of Graphical User Interface (GUI) agents that utilize human-like vision perception capabilities to enhance productivity on digital devices. Compared to approaches predicated on GUI metadata, which are platform-dependent and vulnerable to implementation variations, vision-based approaches offer broader applicability.In this vision-based paradigm, the GUI instruction grounding, which maps user instruction to the location of corresponding element on the given screenshot, remains a critical challenge, particularly due to limited public training dataset and resource-intensive manual instruction data annotation.In this paper, we delve into unexplored challenges in this task including element-to-screen ratio, unbalanced element type, and implicit instruction. To address these challenges, we introduce a large-scale data synthesis pipeline \textit{UI-E2I-Synth} for generating varying complex instruction datasets using GPT-4o instead of human annotators. Furthermore, we propose a new GUI instruction grounding benchmark \textit{UI-I2E-Bench}, which is designed to address the limitations of existing benchmarks by incorporating diverse annotation aspects.Our model, trained on the synthesized data, achieves superior performance in GUI instruction grounding, demonstrating the advancements of proposed data synthesis pipeline.The proposed benchmark, accompanied by extensive analyses, provides practical insights for future research in this domain. We will release our dataset and benchmark to facilitate further development of GUI instruction grounding community.

Yujie Lin, Ante Wang, Moye Chen, Jingyao Liu, Hao Liu, Jinsong Su, Xinyan Xiao

Recently, inference-time scaling of chain-of-thought (CoT) has been demonstrated as a promising approach for addressing multi-modal reasoning tasks.While existing studies have predominantly centered on text-based thinking, the integration of both visual and textual modalities within the reasoning process remains unexplored.In this study, we pioneer the exploration of inference-time scaling with multi-modal thought, aiming to bridge this gap.To provide a comprehensive analysis, we systematically investigate popular sampling-based and tree search-based inference-time scaling methods on 10 challenging tasks spanning various domains.Besides, we uniformly adopt a consistency-enhanced verifier to ensure effective guidance for both methods across different thought paradigms.Results show that multi-modal thought promotes better performance against conventional text-only thought, and blending the two types of thought fosters more diverse thinking.Despite these advantages, multi-modal thoughts necessitate higher token consumption for processing richer visual inputs, which raises concerns in practical applications.We hope that our findings on the merits and drawbacks of this research line will inspire future works in the field. The code will be released upon acceptance.