Detecting offensive language in Chinese is challenging due to homophonic substitutions used to evade detection. We propose a framework to improve large language models’ robustness against such phonetic attacks. First, we construct HED-COLD, the first large-scale and systematic homophonic dataset for Chinese offensive language detection. Additionally, we design a homophone-aware pretraining strategy that learns the mappings among orthography, phonetics, and semantics between original and perturbed text. Experimental results show that our approach achieves state-of-the-art performance on both the COLD test set and the toxicity benchmark ToxiCloakCN. Notably, it achieves greater gains in domains susceptible to homophonic attacks, such as gender and regional content. These results demonstrate improved robustness and generalization against phonetic adversarial attacks.
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
Despite bilingual speakers frequently using mixed-language queries in web searches, Information Retrieval (IR) research on them remains scarce. To address this, we introduce ***MiLQ***, ***Mi***xed-***L***anguage ***Q***uery test set, the first public benchmark of mixed-language queries, qualified as realistic and relatively preferred. Experiments show that multilingual IR models perform moderately on MiLQ and inconsistently across native, English, and mixed-language queries, also suggesting code-switched training data’s potential for robust IR models handling such queries. Meanwhile, intentional English mixing in queries proves an effective strategy for bilinguals searching English documents, which our analysis attributes to enhanced token matching compared to native queries.
Natural Language Inference (NLI) is an important task in natural language processing. NLI models are aimed at automatically determining logical relationships between pairs of sentences. However, recent studies based on gold labels assigned to sentence pairs by human experts have provided some evidence that NLI models tend to make inconsistent model decisions during inference. Previous studies have used existing NLI datasets to test the transitive consistency of language models. However, they test only variations of two transitive consistency rules out of four. To further evaluate the transitive consistency of NLI models, we propose a novel evaluation approach that allows us to test all four rules automatically by generating adversarial examples via antonym replacements. Since we are testing self-consistency, human labeling of generated adversarial examples is unnecessary. Our experiments on several benchmark datasets indicate that the examples generated by the proposed antonym replacement methodology can reveal transitive inconsistencies in the state-of-the-art NLI models.
To alleviate the hallucination problem of large language model (LLM), retrieval-augmented generation (RAG) has been proposed and widely adopted. Due to the limitations in cross-chunk summarization task of naive RAG, graph-based RAG has emerged as a promising solution. However, a close study reveals several flaws in these works. First, most graph-based RAGs suffer from less efficient indexing process, which leads to information loss and expensive costs. Second, they heavily rely on LLM for retrieval thus inference slowly, which hinders their application in industry. To build a more efficient and effective RAG, we propose the multi-semantic RAG (MS-RAG). In this work, we combine knowledge graphs with dense vector to build a multi-semantic RAG. To be specific, (i) at indexing stage, we create multiple semantic-level indexes, including chunk-level, relation-level, and entity-level, to leverage the merits of dense vector and knowledge graph. (ii) at retrieval stage, unlike the previous LLM-empowered entity extraction, we propose a novel mix recall algorithm. Finally, we employ a multi-semantic rerank module to purify the results. Extensive experiments show that MS-RAG achieves superior performance. In terms of retrieval effect, MS-RAG achieves state-of-the-art performance, which is about 10%-30% improvement than the existing methods. In terms of question-answering effect, MS-RAG still achieves promising results with faster inference speed. More analysis and experiments are provided in Appendix.
Context is Gold to find the Gold Passage: Evaluating and Training Contextual Document Embeddings
PDF ↗A limitation of modern document retrieval embedding methods is that they typically encode passages (chunks) from the same documents independently, often overlooking crucial contextual information from the rest of the document that could greatly improve individual chunk representations.In this work, we introduce ConTEB (Context-aware Text Embedding Benchmark), a benchmark designed to evaluate retrieval models on their ability to leverage document-wide context. Our results show that state-of-the-art embedding models struggle in retrieval scenarios where context is required. To address this limitation, we propose InSeNT (In-sequence Negative Training), a novel contrastive post-training approach which combined with late chunking pooling enhances contextual representation learning while preserving computational efficiency. Our method significantly improves retrieval quality on ConTEB without sacrificing base model performance. We further find chunks embedded with our method are more robust to suboptimal chunking strategies and larger retrieval corpus sizes.We open-source all artifacts at https://github.com/illuin-tech/contextual-embeddings.
Debiasing techniques such as SentDebias aim to reduce bias in large language models (LLMs). Previous studies have evaluated their cross-lingual transferability by directly applying these methods to LLM representations, revealing their limited effectiveness across languages. In this work, we therefore propose to perform debiasing in a joint latent space rather than directly on LLM representations. We construct a well-aligned cross-lingual latent space using an autoencoder trained on parallel TED talk scripts. Our experiments with Aya-expanse and two debiasing techniques across four languages (English, French, German, Dutch) demonstrate that a) autoencoders effectively construct a well-aligned cross-lingual latent space, and b) applying debiasing techniques in the learned cross-lingual latent space significantly improves both the overall debiasing performance and cross-lingual transferability.
Iterative evaluation of LLMs during training is essential to ensure expected capability development, but can be time- and compute-intensive. While NLU tasks, where the model selects from fixed answer choices, are cheap to evaluate, essential capabilities like reasoning and code generation rely on the more time-consuming NLG (token-by-token generation) format. In this work, our aim is to decrease the computational burden of NLG benchmarks in order to enable monitoring crucial LLM capabilities during model training. We reformulate generative tasks into computationally cheaper NLU alternatives. We test the performance correlation between the original and reformulated tasks using 8 LMs of various sizes and 4 capabilities: mathematical reasoning, code generation, factual knowledge and reading comprehension. Our results show a strong correlation between task formats, supporting capability assessment via cheaper alternatives and achieving over 35x average reduction in evaluation time. Our project is available at: https://github.com/Fraunhofer-IIS/EvalShortcut
We show that large language models (LLMs) exhibit an \textit{internal chain-of-thought}: they sequentially decompose and execute composite tasks layer-by-layer. Two claims ground our study: (i) distinct subtasks are learned at different network depths, and (ii) these subtasks are executed sequentially across layers. On a benchmark of 15 two-step composite tasks, we employ layer-from context-masking and propose a novel cross-task patching method, confirming (i). To examine claim (ii), we apply LogitLens to decode hidden states, revealing a consistent layerwise execution pattern. We further replicate our analysis on the real-world \text{TRACE} benchmark, observing the same stepwise dynamics. Together, our results enhance LLMs transparency by showing their capacity to internally plan and execute subtasks (or instructions), opening avenues for fine-grained, instruction-level activation steering.
Large language models (LLMs) take sequences of subwords as input, requiring them to effective compose subword representations into meaningful word-level representations. In this paper, we present a comprehensive set of experiments to probe how LLMs compose subword information, focusing on three key aspects: structural similarity, semantic decomposability, and form retention. Our analysis of the experiments suggests that these five LLM families can be classified into three distinct groups, likely reflecting difference in their underlying composition strategies. Specifically, we observe (i) three distinct patterns in the evolution of structural similarity between subword compositions and whole-word representations across layers; (ii) great performance when probing layer by layer their sensitivity to semantic decompositionality; and (iii) three distinct patterns when probing sensitivity to formal features, e.g., character sequence length. These findings provide valuable insights into the compositional dynamics of LLMs and highlight different compositional pattens in how LLMs encode and integrate subword information.
As AI systems take on collaborative roles, they must reason about shared goals and beliefs—not just generate fluent language. The Rational Speech Act (RSA) framework offers a principled approach to pragmatic reasoning, but existing extensions face challenges in scaling to multi-turn, collaborative scenarios. In this paper, we introduce Collaborative Rational Speech Act (CRSA), an information-theoretic (IT) extension of RSA that models multi-turn dialog by optimizing a gain function adapted from rate-distortion theory. This gain is an extension of the gain model that is maximized in the original RSA model but takes into account the scenario in which both agents in a conversation have private information and produce utterances conditioned on the dialog. We demonstrate the effectiveness of CRSA on referential games and template-based doctor–patient dialogs in the medical domain. Empirical results show that CRSA yields more consistent, interpretable, and collaborative behavior than existing baselines—paving the way for more pragmatic and socially aware language agents.
Flowcharts are a critical tool for visualizing decision-making processes. However, their non-linear structure and complex visual-textual relationships make it challenging to interpret them using LLMs, as vision-language models frequently hallucinate nonexistent connections and decision paths when analyzing these diagrams. This leads to compromised reliability for automated flowchart processing in critical domains such as logistics, health, and engineering. We introduce the task of Fine-grained Flowchart Attribution, which traces specific components grounding a flowchart referring LLM response. Flowchart Attribution ensures the verifiability of LLM predictions and improves explainability by linking generated responses to the flowchart’s structure. We propose FlowPathAgent, a neurosymbolic agent that performs fine-grained post hoc attribution through graph-based reasoning. It first segments the flowchart, then converts it into a structured symbolic graph, and then employs an agentic approach to dynamically interact with the graph, to generate attribution paths. Additionally, we present FlowExplainBench, a novel benchmark for evaluating flowchart attributions across diverse styles, domains, and question types. Experimental results show that FlowPathAgent mitigates visual hallucinations in LLM answers over flowchart QA, outperforming strong baselines by 10–14% on our proposed FlowExplainBench dataset.
We report on investigations into the characteristics of outgoing connections in feedforward layers of large language models. Our findings show that inner neurons with diverse outgoing connection strengths are more critical to model performance than those with uniform connections. We propose a new fine-tuning loss that takes advantage of this observation by decreasing the outgoing connection entropy in feedforward layers. Using this loss yields gains over standard fine-tuning across two different model families (PaLM-2 and Gemma-2) for downstream tasks in math, coding, and language understanding. To further elucidate the role of outgoing connection heterogeneity, we develop a data-free structured pruning method, which uses entropy to identify and remove neurons. This method is considerably more effective than removing neurons either randomly or based on their magnitude.
Temporal reasoning and planning are essential capabilities for large language models (LLMs), yet most existing benchmarks evaluate them in isolation and under limited forms of complexity. To address this gap, we introduce the Temporal Constraint-based Planning (TCP) benchmark, that jointly assesses both capabilities. Each instance in TCP features a naturalistic dialogue around a collaborative project, where diverse and interdependent temporal constraints are explicitly or implicitly expressed, and models must infer an optimal schedule that satisfies all constraints. To construct TCP, we generate abstract problem prototypes that are then paired with realistic scenarios from various domains and enriched into dialogues using an LLM. A human quality check is performed on a sampled subset to confirm the reliability of our benchmark. We evaluate state-of-the-art LLMs and find that even the strongest models may struggle with TCP, highlighting its difficulty and revealing limitations in LLMs’ temporal constraint-based planning abilities. We analyze underlying failure cases, open source our benchmark, and hope our findings can inspire future research.
Extensive research has been conducted to explore the capabilities of large language models (LLMs) in table reasoning. However, the essential task of transforming tables information into reports remains a significant challenge for industrial applications. This task is plagued by two critical issues: 1) the complexity and diversity of tables lead to suboptimal reasoning outcomes; and 2) existing table benchmarks lack the capacity to adequately assess the practical application of this task. To fill this gap, we propose the table-to-report task and construct a bilingual benchmark named T2R-bench, where the key information flow from the tables to the reports for this task. The benchmark comprises 457 industrial tables, all derived from real-world scenarios and encompassing 19 industry domains as well as four types of industrial tables. Furthermore, we propose a novel evaluation criteria to fairly measure the quality of report generation. Expeimental results show that Deepseek-R1 only achieves the best performance with 62.71% overall score, indicating that LLMs still have room for improvement on T2R-bench.
The advent of Large Language Models (LLMs) has revolutionized product recommenders, yet their susceptibility to adversarial manipulation poses critical challenges, particularly in real-world commercial applications. Our approach is the first one to tap into human psychological principles, seamlessly modifying product descriptions, making such manipulations hard to detect. In this work, we investigate cognitive biases as black-box adversarial strategies, drawing parallels between their effects on LLMs and human purchasing behavior. Through extensive evaluation across models of varying scale, we find that certain biases, such as social proof, consistently boost product recommendation rate and ranking, while others, like scarcity and exclusivity, surprisingly reduce visibility. Our results demonstrate that cognitive biases are deeply embedded in state-of-the-art LLMs, leading to highly unpredictable behavior in product recommendations and posing significant challenges for effective mitigation.
Grammatical Error Correction (GEC) for low-resource Indic languages faces significant challenges due to the scarcity of annotated data. In this work, we introduce the Mask-Translate&Fill (MTF) framework, a novel approach for generating high-quality synthetic data for GEC using only monolingual corpora. MTF leverages a machine translation system and a pretrained masked language model to introduce synthetic errors and tries to mimic errors made by second-language learners. Our experimental results on English, Hindi, Bengali, Marathi, and Tamil demonstrate that MTF consistently outperforms other monolingual synthetic data generation methods and achieves performance comparable to the Translation Language Modeling (TLM)-based approach, which uses a bilingual corpus, in both independent and multilingual settings. Under multilingual training, MTF yields significant improvements across Indic languages, with particularly notable gains in Bengali and Tamil, achieving +1.6 and +3.14 GLEU over the TLM-based method, respectively. To support further research, we also introduce the IndiGEC Corpus, a high-quality, human-written, manually validated GEC dataset for these four Indic languages, comprising over 8,000 sentence pairs with separate development and test splits.
Massive Supervised Fine-tuning Experiments Reveal How Data, Layer, and Training Factors Shape LLM Alignment Quality
PDF ↗Supervised fine-tuning (SFT) is a critical step in aligning large language models (LLMs) with human instructions and values, yet many aspects of SFT remain poorly understood. We trained a wide range of base models on a variety of datasets including code generation, mathematical reasoning, and general-domain tasks, resulting in 1,000+ SFT models under controlled conditions. We then identified the dataset properties that matter most and examined the layer-wise modifications introduced by SFT.Our findings reveal that some training–task synergies persist across all models while others vary substantially, emphasizing the importance of model-specific strategies. Moreover, we demonstrate that perplexity consistently predicts SFT effectiveness, often surpassing superficial similarity between the training data and the benchmark, and that mid-layer weight changes correlate most strongly with performance gains. We release these 1,000+ SFT models and benchmark results to accelerate further research. All resources are available at https://github.com/llm-jp/massive-sft.
Data perspectivism goes beyond majority vote label aggregation by recognizing various perspectives as legitimate ground truths.However, current evaluation practices remain fragmented, making it difficult to compare perspectivist approaches and analyze their impact on different users and demographic subgroups. To address this gap, we introduce PersEval, the first unified framework for evaluating perspectivist models in NLP. A key innovation is its evaluation at the individual annotator level and its treatment of annotators and users as distinct entities, consistently with real-world scenarios. We demonstrate PersEval’s capabilities through experiments with both Encoder-based and Decoder-based approaches, as well as an analysis of the effect of sociodemographic prompting. By considering global, text-, trait- and user-level evaluation metrics, we show that PersEval is a powerful tool for examining how models are influenced by user-specific information and identifying the biases this information may introduce.
Large Language Models have advanced significantly in complex reasoning, often leveraging external reward model to improve the reliability of their multi-step processes. However, existing process verification methods struggle with reliably assessing incomplete reasoning traces and are limited by the cost of high-quality human annotations or the inherent noise in automatically generated labels. Therefore, we present Dyve, a dynamic process verifier that enhances reasoning error detection in large language models by integrating fast and slow thinking, inspired by Kahneman’s Systems Theory. Dyve adaptively applies immediate token-level confirmation (System 1) for straightforward steps and comprehensive analysis (System 2) for complex ones. Unlike traditional verifiers that only evaluate final outputs, Dyve employs a step-wise consensus-filtered supervision strategy, leveraging Monte Carlo estimation, LLM-as-a-Judge, and specialized reasoning models to extract high-quality training signals from noisy rollouts. Experimental results on ProcessBench and the MATH dataset confirm that Dyve significantly outperforms existing process-based verifiers and boosts performance in Best-of-N settings while maintaining computational efficiency by strategically allocating verification resources.
Text embeddings are a fundamental component in many NLP tasks, including classification, regression, clustering, and semantic search. However, despite their ubiquitous application, challenges persist in interpreting embeddings and explaining similarities between them.In this work, we provide a structured overview of methods specializing in inherently interpretable text embeddings and text similarity explanation, an underexplored research area. We characterize the main ideas, approaches, and trade-offs. We compare means of evaluation, discuss overarching lessons learned and finally identify opportunities and open challenges for future research.