While LLMs have been extensively studied on general text generation tasks, there is less research on text rewriting, a task related to general text generation, and particularly on the behavior of models on this task. In this paper we analyze what changes LLMs make in a text rewriting setting. We focus specifically on argumentative texts and their improvement, a task named Argument Improvement (ArgImp). We present CLEAR: an evaluation pipeline consisting of 57 metrics mapped to four linguistic levels: lexical, syntactic, semantic and pragmatic. This pipeline is used to examine the qualities of LLM-rewritten arguments on a broad set of argumentation corpora and compare the behavior of different LLMs on this task and analyze the behavior of different LLMs on this task in terms of linguistic levels. By taking all four linguistic levels into consideration, we find that the models perform ArgImp by shortening the texts while simultaneously increasing average word length and merging sentences. Overall we note an increase in the persuasion and coherence dimensions.
论文检索
输入标题、作者或关键词，从 7,591 篇学术成果中精准定位
Retrieval-Augmented Generation (RAG) enhances the accuracy of Large Language Model (LLM) responses by leveraging relevant external documents during generation. Although previous studies noted that retrieving many documents can degrade performance, they did not isolate how the quantity of documents affects performance while controlling for context length. We evaluate various language models on custom datasets derived from a multi-hop QA task. We keep the context length and position of relevant information constant while varying the number of documents, and find that increasing the document count in RAG settings poses significant challenges for most LLMs, reducing performance by up to 20%. However, Qwen2 maintained consistent results across increasing document counts, indicating better multi-document handling capability. Finally, our results indicate that processing multiple documents is a separate challenge from handling long contexts. We will publicly release the datasets and code upon publication to facilitate further research in multi-document retrieval.
Sparse autoencoders (SAEs) are emerging as a key analytical tool in the field of mechanistic interpretability for large language models (LLMs). While SAEs have primarily been used for interpretability, we shift focus and explore an understudied question: “Can SAEs be applied to practical tasks beyond interpretability?” Given that SAEs are trained on billions of tokens for sparse reconstruction, we believe they can serve as effective extractors, offering a wide range of useful knowledge that can benefit practical applications. Building on this motivation, we demonstrate that SAEs can be effectively applied to in-context learning (ICL). In particular, we highlight the utility of the SAE-reconstruction loss by showing that it provides a valuable signal in ICL—exhibiting a strong correlation with LLM performance and offering a powerful unsupervised approach for prompt selection. These findings underscore the versatility of SAEs and reveal their potential for real-world applications beyond interpretability. Our code is available at https://github.com/ihcho2/SAE-GPS.
CAC-CoT: Connector-Aware Compact Chain-of-Thought for Efficient Reasoning Data Synthesis Across Dual-System Cognitive Tasks
PDF ↗Long chain-of-thought (CoT) prompting helps Large Language Models (LLMs) solve difficult problems, but very long traces often slow or even degrade performance on fast, intuitive “System-1” tasks. We introduce Connector-Aware Compact CoT (CAC-CoT) — a method that deliberately restricts reasoning to a small, fixed set of connector phrases, steering the model toward concise and well — structured explanations. Despite its simplicity, our synthetic method with general-purpose LLMs yields a high-quality training quality. CAC-CoT achieves \approx 85% on GSM8K and \approx 40% on GPQA (System-2) while also achieving \approx 85% on S1-Bench (System-1), surpassing the baseline by over 20%. Its reasoning traces average \approx 300 tokens(ART), about one-third the length of baseline traces, delivering higher efficiency without loss of accuracy.
MUG-Eval: A Proxy Evaluation Framework for Multilingual Generation Capabilities in Any Language
PDF ↗Evaluating text generation capabilities of large language models (LLMs) is challenging, particularly for low-resource languages where methods for direct assessment are scarce. We propose MUG-Eval, a novel framework that evaluates LLMs’ multilingual generation capabilities by transforming existing benchmarks into conversational tasks and measuring the LLMs’ accuracies on those tasks. We specifically designed these conversational tasks to require effective communication in the target language. Then, we simply use task success rate as a proxy for successful conversation generation. Our approach offers two key advantages: it is independent of language-specific NLP tools or annotated datasets, which are limited for most languages, and it does not rely on LLMs-as-judges, whose evaluation quality degrades outside a few high-resource languages. We evaluate 8 LLMs across 30 languages spanning high, mid, and low-resource categories, and we find that MUG-Eval correlates strongly with established benchmarks (r > 0.75) while enabling standardized comparisons across languages and models. Our framework provides a robust and resource-efficient solution for evaluating multilingual generation that can be extended to thousands of languages.
The integration of Large Language Models (LLMs) with external sources is becoming increasingly common, with Retrieval-Augmented Generation (RAG) being a prominent example. However, this integration introduces vulnerabilities of Indirect Prompt Injection (IPI) attacks, where hidden instructions embedded in external data can manipulate LLMs into executing unintended or harmful actions. We recognize that IPI attacks fundamentally rely on the presence of instructions embedded within external content, which can alter the behavioral states of LLMs. Can the effective detection of such state changes help us defend against IPI attacks? In this paper, we propose InstructDetector, a novel detection-based approach that leverages the behavioral states of LLMs to identify potential IPI attacks. Specifically, we demonstrate the hidden states and gradients from intermediate layers provide highly discriminative features for instruction detection. By effectively combining these features, InstructDetector achieves a detection accuracy of 99.60% in the in-domain setting and 96.90% in the out-of-domain setting, and reduces the attack success rate to just 0.03% on the BIPIA benchmark. The code is publicly available at https://github.com/MYVAE/Instruction-detection.
Multimodal emotion recognition in conversation (MERC) aims to identify speakers’ emotional states by utilizing text, audio, and visual modalities. Although recent large language model (LLM)-based methods have demonstrated strong performance, they typically adopt static fusion strategies that integrate all available modalities uniformly. This overlooks the fact that the necessity of multimodal cues can vary significantly across utterances. In this work, we propose an adaptive modality selection framework for MERC. The core of our approach is a modality selection module based on Group Relative Policy Optimization (GRPO), which enables a LoRA-tuned LLM to reason about the necessity of multimodal input via chain-of-thought (CoT) generation. This process does not require manually labeled modality selection data and is trained in a fully unsupervised manner. The selected modality configuration is then provided as input to a downstream emotion classifier, which is also implemented using a LoRA-tuned LLM and trained to predict emotional states. Experimental results on benchmark multimodal dialogue datasets show that our method consistently outperforms strong baselines, demonstrating the effectiveness of adaptive modality selection in improving recognition accuracy. Our code is available at https://github.com/youflyaway/Modality-Selection-Enhanced-LoRA-Tuned-LLMs.
Large vision-language models (VLMs) often struggle to generate long and factual captions. However, traditional measures for hallucination and factuality are not well suited for evaluating longer, more diverse captions and in settings where ground-truth human-annotated captions are unavailable. We introduce OVFact, a novel method for measuring caption factuality of long captions that leverages open-vocabulary visual grounding and tool-based verification without depending on human annotations. Our method improves agreement with human judgements and captures both caption descriptiveness (recall) and factual precision in the same metric. Furthermore, unlike previous metrics, our reference-free method design enables new applications towards factuality-based data filtering. We observe models trained on an OVFact-filtered (2.5-5x less) subset of a large-scale, noisy (VLM-generated) pretraining set meaningfully improve factuality precision without sacrificing caption descriptiveness across a range of downstream long caption benchmarks.
Recent advancements in Large Language Models (LLMs) have demonstrated sophisticated capabilities, including the ability to process and comprehend extended contexts. These emergent capabilities necessitate rigorous evaluation methods to effectively assess their performance in long-context understanding. In this paper, we present LC-Eval, a bilingual, multi-task evaluation benchmark designed to evaluate long-context understanding in English and Arabic, targeting context lengths ranging from 4k to over 128k tokens. LC-Eval introduces four novel and challenging tasks: multi-document question answering, bilingual question answering, claim verification within a paragraph, and multiple-choice questions based on long contexts. These tasks are designed to assess LLMs’ abilities in deep reasoning, document comprehension, information tracing, and bilingual information extraction and understanding. The benchmark includes datasets in both Arabic and English for each task, allowing for a comparative analysis of their performance across different text genres. Evaluations were conducted on both open-weight and closed LLMs, with results indicating that LC-Eval presents significant challenges. Even high-performing models, such as GPT-4o, struggled with certain tasks, highlighting the complexity and rigor of the benchmark.
As they become increasingly multilingual, Large Language Models (LLMs) exhibit more language confusion, i.e., they tend to generate answers in a language different from the language of the prompt or the answer language explicitly requested by the user. In this work, we propose ReCoVeR (REducing language COnfusion in VEctor Representations), a novel lightweight approach for reducing language confusion based on language-specific steering vectors. We first isolate language vectors with the help of multi-parallel corpus and then effectively leverage those vectors for effective LLM steering via fixed (i.e., unsupervised) as well as trainable steering functions. Our extensive evaluation, encompassing three benchmarks and 18 languages, shows that ReCoVeR effectively mitigates language confusion in both monolingual and cross-lingual setups while at the same time—and in contrast to prior language steering methods—retaining task performance. Our data code is available at https://github.com/hSterz/recover.
Attributes as Textual Genes: Leveraging LLMs as Genetic Algorithm Simulators for Conditional Synthetic Data Generation
PDF ↗Large Language Models (LLMs) excel at generating synthetic data, but ensuring its quality and diversity remains challenging. We propose Genetic Prompt, a novel framework that combines genetic algorithms with LLMs to augment synthetic data generation. Our approach treats semantic text attributes as gene sequences and leverages the LLM to simulate crossover and mutation operations. This genetic process enhances data quality and diversity by creating novel attribute combinations, yielding synthetic distributions closer to real-world data. To optimize parent selection, we also integrate an active learning scheme that expands the offspring search space. Our experiments on multiple NLP tasks reveal several key findings: Genetic Prompt not only significantly outperforms state-of-the-art baselines but also shows robust performance across various generator model sizes and scales. Moreover, we demonstrate that fusing our synthetic data with the original training set significantly boosts downstream model performance, particularly for class-imbalanced scenarios. Our findings validate that Genetic Prompt is an effective method for producing high-quality synthetic data for a wide range of NLP applications.
As large language models (LLMs) continue to grow in size, their practical deployment increasingly relies on a range of compression techniques, such as quantization, pruning, and low-rank approximation. Especially, post-training compression methods–which do not require re-training–have drawn considerable interest. Many recent methods leverage calibration data to capture magnitude or second-order characteristics of input activations. However, the role and significance of calibration data remain underexplored. In this study, we demonstrate that the sequence length of calibration data plays a crucial role in the effectiveness of post-training compression methods for LLMs. We then analyze input activations and find that, within the normalized hidden states, the embedding of the first token exhibits characteristics opposite to those of subsequent tokens. Building on this insight, we introduce state-aware length calibration, a technique that applies masking along the sequence axis, specifically targeting normalized hidden states. Experimental results show that our approach improves perplexity and zero-shot downstream tasks performance.
The scarcity of non-English language data in specialized domains significantly limits the development of effective Natural Language Processing (NLP) tools. We present TransBERT, a novel framework for pre-training language models using exclusively synthetically translated text, and introduce TransCorpus, a scalable translation toolkit. Focusing on the life sciences domain in French, our approach demonstrates that state-of-the-art performance on various downstream tasks can be achieved solely by leveraging synthetically translated data. We release the TransCorpus toolkit, the TransCorpus-bio-fr corpus (36.4GB of French life sciences text), TransBERT-bio-fr, its associated pre-trained language model and reproducible code for both pre-training and fine-tuning. Our results highlight the viability of synthetic translation in a high-resource translation direction for building high-quality NLP resources in low-resource language/domain pairs.
Over-Generation and Compaction: A Prompting Strategy for Procedural Text Adaptation with Large Language Models
PDF ↗Procedural text adaptation—such as modifying recipes or revising instructional guides—has traditionally relied on specialized models extensively fine‐tuned for specific domains. To address the scalability limitations of such approaches, recent research has increasingly turned to general‐purpose large language models (LLMs). However, existing prompting strategies for LLMs often yield superficial or erroneous adaptations due to alignment‐induced biases and the inherent complexity of procedural editing. To overcome these challenges, we propose the Over‐generation‐and‐Compaction (OC) prompting strategy, which first elicits an exhaustive set of procedural details to leverage the model’s latent knowledge, and subsequently compacts them into concise, coherent adaptations. We further introduce Recipe Consistency & Feasibility (RCF), a novel metric for systematically assessing procedural validity and practicality in cooking recipe adaptations. Experiments on public datasets demonstrate that OC significantly improves adaptation consistency and feasibility compared to baseline prompting methods, without the need for additional fine-tuning or curated training resources.
This work stems from an observed limitation of text encoders: embeddings may not be able to recognize fine-grained entities or events within encoded semantics, resulting in failed retrieval even in simple cases. To examine such behaviors, we first introduce a new evaluation dataset, CapRetrieval, in which passages are image captions and queries are phrases targeting entity or event concepts in diverse forms. Zero-shot evaluation suggests that encoders often struggle with these fine-grained matching, regardless of training sources or model size. Aiming for enhancement, we proceed to finetune encoders with our proposed data generation strategies, enabling a small 0.1B encoder to outperform the state-of-the-art 7B model. Within this process, we further uncover the granularity dilemma, a challenge for embeddings to capture fine-grained salience while aligning with overall semantics. Our dataset, code and models in this work are publicly released at https://github.com/lxucs/CapRetrieval.
BAGELS: Benchmarking the Automated Generation and Extraction of Limitations from Scholarly Text
PDF ↗In scientific research, “limitations” refer to the shortcomings, constraints, or weaknesses of a study. A transparent reporting of such limitations can enhance the quality and reproducibility of research and improve public trust in science. However, authors often underreport limitations in their papers and rely on hedging strategies to meet editorial requirements at the expense of readers’ clarity and confidence. This tendency, combined with the surge in scientific publications, has created a pressing need for automated approaches to extract and generate limitations from scholarly papers. To address this need, we present a full architecture for computational analysis of research limitations. Specifically, we (1) create a dataset of limitations from ACL, NeurIPS, and PeerJ papers by extracting them from the text and supplementing them with external reviews; (2) we propose methods to automatically generate limitations using a novel Retrieval Augmented Generation (RAG) technique; (3) we design a fine-grained evaluation framework for generated limitations, along with a meta-evaluation of these techniques. Code and datasets are available at: Code: https://github.com/IbrahimAlAzhar/BAGELS_Limitation_GenDataset: https://huggingface.co/datasets/IbrahimAlAzhar/limitation-generation-dataset-bagels
ClueAnchor: Clue-Anchored Knowledge Reasoning Exploration and Optimization for Retrieval-Augmented Generation
PDF ↗Retrieval-Augmented Generation (RAG) augments Large Language Models (LLMs) with external knowledge to improve factuality. However, existing RAG systems frequently underutilize the retrieved documents, failing to extract and integrate the key clues needed to support faithful and interpretable reasoning, especially in cases where relevant evidence is implicit, scattered, or obscured by noise. To address this issue, we propose ClueAnchor, a novel framework for enhancing RAG via clue-anchored reasoning exploration and optimization. ClueAnchor extracts key clues from retrieved content and generates multiple reasoning paths based on different knowledge configurations, optimizing the model by selecting the most appropriate reasoning path for the given context through reward-based preference optimization. Experiments show that ClueAnchor significantly outperforms prior RAG baselines in the completeness and robustness of reasoning. Further analysis confirms its strong resilience to noisy or partially relevant retrieved content, as well as its capability to identify supporting evidence even in the absence of explicit clue supervision during inference. All codes are available at https://github.com/thunlp/ClueAnchor.
IoTMigrator: LLM-driven Embedded IoT Code Migration across Different OSes for Cloud-device Integration
PDF ↗The increasing prevalence of embedded systems has necessitated manufacturers to migrate product code, transferring existing products to new embedded operating systems (OSes) for getting better compatibility and performance. Since manufacturers’ product code predominantly employs the Thing Specification Language (TSL) paradigm for cloud connectivity, migrated code consequently adheres to the same TSL standard. However, embedded code migration under the TSL paradigm proves more complex than conventional code migration. Neither outline-based code generation nor common code translation techniques can adequately address this challenge, despite their prevalence in existing systems. There exists a growing demand for a algorithm tailored to TSL paradigm embedded code migration. In response to this demand, we have developed IoTMigrator that employs a multi-agent pipeline to handle the issue. The key insight of our algorithm is the TSL enhancer, specifically designed for the characteristics of the TSL paradigm, which serves as a crucial component in the agent pipeline.To demonstrate the superiority of our algorithm, we have established our own benchmark, which includes six tasks across two OSes, RIOT and Zephyr. We adopted two key metrics: compilation pass rate and task completeness score. The experiment results show that our algorithm outperforms the baseline by an average of at least 50.5% for pass rate and 13.0% for completeness across all tasks in RIOT, and at least 83.4% for pass rate and 18.4% for completeness in Zephyr. This work will be open-sourced in the future.
Large Language Models (LLMs) achieve strong performance across diverse tasks, but their effectiveness often depends on the quality of the provided context. Retrieval-Augmented Generation (RAG) enriches prompts with external information, but its reliance on static databases constrains adaptability and can result in irrelevant demonstrations. In this work, we propose a Generative Retrieval-Aligned Demonstrator (GRAD), a dynamic demonstration-based approach where an LLM model is trained to generate input-specific concise demonstrations. By tailoring demonstrations to each input, our method offers better contextual support than traditional RAG approaches. We demonstrate the superiority of GRAD under budget constraints, where we limit both the number of tokens used per demonstration and the number of tokens used for the final output. Trained solely on a math dataset, GRAD consistently outperforms strong baselines on Qwen2.5-14B across mathematical reasoning and advanced STEM questions, highlighting GRAD’s robust generalization to out-of-distribution (OOD) domains such as physics, chemistry, and computer science. Furthermore, we show that demonstrations generated by trained smaller models can effectively guide larger target models, reducing training costs while maintaining competitive accuracy. Overall, this work introduces a scalable demonstration generator model presenting the first step toward a dynamic few-shot learning paradigm in resource-constrained settings. We release the code used for the project: https://github.com/charafkamel/GRAD-demonstration-sampler
Offline preference optimization methods are efficient for large language models (LLMs) alignment. Direct Preference optimization (DPO)-like learning, one of the most popular approaches, stands out for its efficiency in reward modeling. However, these methods typically follow the convention to use Bradley-Terry (BT) reward modeling that faces several critical assumptions, including the requirement for pairwise training data, model distribution shifting, human rationality assumption, etc. To address these limitations, we propose a general framework for offline preference optimization methods, Adaptive Preference Optimization with Utility Anchor (UAPO), which introduces an anchoring function to estimate the uncertainties brought from preference data annotation. Our method enables training even in scenarios where the data is unpaired, significantly enhancing data utilization efficiency. Moreover, the anchor design makes UAPO more robust in the training process. Experimental results demonstrate that UAPO achieves competitive outcomes without the strict dependency on data pairing, paving the way for more flexible and effective preference optimization methods.