Instruction tuning—supervised fine-tuning using instruction-response pairs—is a key step in making pre-trained large language models (LLMs) instructable. Meanwhile, LLMs perform multitask learning during their pre-training, acquiring extensive knowledge and capabilities. We hypothesize that the pre-training stage can enable them to develop the ability to comprehend and address instructions. To verify this, we propose Response Tuning (RT), which removes the instruction and its corresponding mapping to the response from instruction tuning. Instead, it focuses solely on establishing a response distribution. Our experiments demonstrate that RT models, trained only on responses, can effectively respond to a wide range of instructions akin to their instruction-tuned counterparts. In addition, we observe that the models can recognize and reject unsafe queries after learning a safety policy only from the response data. Furthermore, we find that these observations extend to an in-context learning setting. These findings support our hypothesis, highlighting the extensive inherent capabilities of pre-trained LLMs.
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
Large language models (LLMs) have demonstrated an impressive ability to role-play humans and replicate complex social dynamics. However, large-scale LLM-driven simulations still face significant challenges in high time and computational costs. We observe that there exists redundancy in current agent communication: when expressing the same intention, agents tend to use lengthy and repetitive language, whereas humans naturally prefer concise expressions. To this end, we propose EcoLANG: Efficient and Effective Agent Communication Language Induction for Social Simulation. Inspired by how human language evolves through interactions, we induce a more compact language by identifying and preserving core communicative concepts at the vocabulary level and evolving efficient expression patterns at the sentence level through natural selection. We apply the induced language in various social simulations. Experimental results demonstrate that EcoLANG reduces token consumption by over 20%, enhancing efficiency without sacrificing simulation accuracy.
Tag&Tab: Pretraining Data Detection in Large Language Models Using Keyword-Based Membership Inference Attack
PDF ↗Large language models (LLMs) have become essential tools for digital task assistance. Their training relies heavily on the collection of vast amounts of data, which may include copyright-protected or sensitive information. Recent studies on detecting pretraining data in LLMs have primarily focused on sentence- or paragraph-level membership inference attacks (MIAs), usually involving probability analysis of the target model’s predicted tokens. However, these methods often exhibit poor accuracy, failing to account for the semantic importance of textual content and word significance. To address these shortcomings, we propose Tag&Tab, a novel approach for detecting data used in LLM pretraining. Our method leverages established natural language processing (NLP) techniques to tag keywords in the input text, a process we term Tagging. Then, the LLM is used to obtain probabilities for these keywords and calculate their average log-likelihood to determine input text membership, a process we refer to as Tabbing. Our experiments on four benchmark datasets (BookMIA, MIMIR, PatentMIA, and the Pile) and several open-source LLMs of varying sizes demonstrate an average increase in AUC scores ranging from 5.3% to 17.6% over state-of-the-art methods. Tag&Tab not only sets a new standard for data leakage detection in LLMs, but its outstanding performance is a testament to the importance of words in MIAs on LLMs.
X-Boundary: Establishing Exact Safety Boundary to Shield LLMs from Jailbreak Attacks without Compromising Usability
PDF ↗With the widespread application of large language models (LLMs) across various domains, techniques for enhancing their security have progressed rapidly. In this paper, we reveal that although existing defense methods can improve the robustness of LLMs against jailbreaks, they compromise usability, i.e., reducing general capabilities or causing the over-refusal problem. From the perspective of LLM mechanism interpretability, we discover that these methods fail to establish a boundary that exactly distinguishes safe and harmful feature representations. Therefore, boundary-safe representations close to harmful representations are inevitably disrupted, leading to a decline in usability. To address this issue, we propose X-Boundary to push harmful representations away from boundary-safe representations and obtain an exact distinction boundary. In this way, harmful representations can be precisely erased without disrupting safe ones. Experimental results show that X-Boundary achieves state-of-the-art defense performance against both single-turn and multi-turn jailbreak attacks, while reducing the over-refusal rate by about 20% and maintaining nearly complete general capability. Furthermore, we theoretically prove and empirically verify that X-Boundary can accelerate the convergence process during training.
In competitive programming task, problem statements are often embedded within elaborate narrative backgrounds, requiring deep understanding of the underlying solutions to successfully complete the tasks. Current code generation models primarily focus on token-level semantic modeling, highly susceptible to distractions from irrelevant narrative statements. Inspired by RAG, retrieving reference code with similar solutions may help enhance model performance on difficult problems. However, existing retrieval models also emphasize surface-level semantic similarity, neglecting the deeper solution-level logical similarities that are critical in competitive programming. Therefore, designing ranking models capable of accurately identifying and retrieving problems and corresponding codes remains an urgent research problem in competitive code generation. In this paper, we propose SolveRank, a solution-aware ranking model empowered by synthetic data for competitive programming tasks. Specifically, we leverage the DeepSeek-R1 model to generate logically equivalent but differently phrased new problems, verified by GPT-4o for solution consistency. Then, we train SolveRank with these as positive samples and BM25/random-retrieved problems as negatives. During inference, SolveRank retrieves relevant problems and corresponding code from the corpus to assist a downstream code generator. Experiments on the xCodeEval dataset demonstrate that SolveRank outperforms SOTA ranking methods in precision and recall metrics, and boosts code generation performance for difficult problems.
Understanding temporal relationships and accurately reconstructing the event timeline is important for case law analysis, compliance monitoring, and legal summarization. However, existing benchmarks lack specialized language evaluation, leaving a gap in understanding how LLMs handle event ordering in legal contexts. We introduce LexTime, a dataset designed to evaluate LLMs’ event ordering capabilities in legal language, consisting of 512 instances from U.S. Federal Complaints with annotated event pairs and their temporal relations. Our findings show that (1) LLMs are more accurate on legal event ordering than on narrative texts (up to +10.5%); (2) longer input contexts and implicit events boost accuracy, reaching 80.8% for implicit-explicit event pairs; (3) legal linguistic complexities and nested clauses remain a challenge. While performance is promising, specific features of legal texts remain a bottleneck for legal temporal event reasoning, and we propose concrete modeling directions to better address them.
Retrieval-Augmented Generation (RAG) enhances large language models by grounding their outputs in external knowledge. Recent advances in Graph-based RAG (GRAG) frameworks, such as GraphRAG, LightRAG, and HippoRAG2, integrate knowledge graphs into the retrieval process to improve multi-hop reasoning and semantic coherence. While effective in monolingual settings, these methods remain underexplored in cross-lingual scenarios and face limitations in semantic granularity and entity alignment. In this work, we propose MaGiX, the first GRAG framework tailored for English–Vietnamese cross-lingual question answering. MaGiX constructs a multi-granular cross-lingual knowledge graph using fine-grained attribute descriptions and cross-synonym edges, and incorporates a custom multilingual embedding model trained with contrastive learning for semantic alignment. During retrieval, MaGiX leverages graph-based reasoning and a semantic-aware reranking strategy to enhance cross-lingual relevance. Experiments across five benchmarks show that MaGiX substantially outperforms prior GRAG systems in both retrieval accuracy and generation quality, advancing structured retrieval for multilingual QA.
Large reasoning models (LRMs) excel at solving complex tasks by leveraging long chain-of-thought (CoT) reasoning. However, this often leads to overthinking on simple tasks, resulting in unnecessary computational overhead. We observe that LRMs inherently possess the capability for efficient short CoT reasoning, which can be reliably elicited through prompt design. To leverage this capability, we propose ThinkSwitcher, a framework that enables a single LRM to dynamically switch between short and long CoT modes based on task complexity. ThinkSwitcher introduces a lightweight switching module trained with supervision signals derived from the relative performance of each reasoning mode across tasks. Experiments on multiple reasoning benchmarks show that ThinkSwitcher reduces computational cost by 20-30% while maintaining high accuracy on complex tasks. This demonstrates the effectiveness of ThinkSwitcher as a scalable and efficient solution for unified LRM deployment.
Social relationship recognition, as one of the fundamental tasks in video understanding, contributes to the construction and application of multi-modal knowledge graph. Previous works have mainly focused on two aspects: generating character graphs and multi-modal fusion. However, they often overlook the impact of cultural differences on relationship recognition. Specifically, relationship recognition models are susceptible to being misled by training data from a specific cultural context. This can result in the learning of culture-specific spurious correlations, ultimately restricting the ability to recognize social relationships in different cultures. Therefore, we employ a customized causal graph to analyze the confounding effects of culture in the relationship recognition task. We propose a Cultural Causal Intervention (CCI) model that mitigates the influence of culture as a confounding factor in the visual and textual modalities. Importantly, we also construct a novel video social relation recognition (CVSR) dataset to facilitate discussion and research on cultural factors in video tasks. Extensive experiments conducted on several datasets demonstrate that the proposed model surpasses state-of-the-art methods.
WebCoT: Enhancing Web Agent Reasoning by Reconstructing Chain-of-Thought in Reflection, Branching, and Rollback
PDF ↗Web agents powered by Large Language Models (LLMs) show promise for next-generation AI, but their limited reasoning in uncertain, dynamic web environments hinders robust deployment. In this paper, we identify key reasoning skills essential for effective web agents, i.e., reflection & lookahead, branching, and rollback, and curate trajectory data that exemplifies these abilities by reconstructing the agent’s (inference-time) reasoning algorithms into chain-of-thought rationales. We conduct experiments in the agent self-improving benchmark, OpenWebVoyager, and demonstrate that distilling salient reasoning patterns into the backbone LLM via simple fine-tuning can substantially enhance its performance. Our approach yields significant improvements across multiple benchmarks, including WebVoyager, Mind2web-live, and SimpleQA (web search), highlighting the potential of targeted reasoning skill enhancement for web agents.
Towards Personalized Conversational Sales Agents: Contextual User Profiling for Strategic Action
PDF ↗Conversational Recommender Systems (CRSs) aim to engage users in dialogue to provide tailored recommendations. While traditional CRSs focus on eliciting preferences and retrieving items, real-world e-commerce interactions involve more complex decision-making, where users consider multiple factors beyond simple attributes. To capture this complexity, we introduce Conversational Sales (CSALES), a novel task that integrates preference elicitation, recommendation, and persuasion within a unified conversational framework. To support realistic and systematic evaluation, we present CSUSER, an evaluation protocol with LLM-based user simulator grounded in real-world behavioral data by modeling fine-grained user profiles for personalized interaction. We also propose CSI, a conversational sales agent that proactively infers contextual user profiles and strategically selects actions through conversation. Comprehensive experiments show that CSI significantly improves both recommendation success and persuasive effectiveness across diverse user profiles.
Retrieval-augmented generation (RAG) has become a fundamental paradigm for addressing the challenges faced by large language models in handling real-time information and domain-specific problems. Traditional RAG systems primarily rely on the in-context learning (ICL) capabilities of the large language model itself. Still, in-depth research on the specific capabilities needed by the RAG generation model is lacking, leading to challenges with inconsistent document quality and retrieval system imperfections. Even the limited studies that fine-tune RAG generative models often lack a granular focus on RAG tasks or a deeper utilization of chain-of-thought processes. To address this, we propose that RAG models should possess three progressively hierarchical abilities (1) Filtering: the ability to select relevant information; (2) Combination: the ability to combine semantic information across paragraphs; and (3) RAG-specific reasoning: the ability to further process external knowledge using internal knowledge. Thus, we introduce our new RAG instruction fine-tuning method, Hierarchical-Thought Instruction-Tuning Retrieval-Augmented Generation (HIRAG) incorporates a “think before answering” strategy. This method enhances the model’s open-book examination capability by utilizing multi-level progressive chain-of-thought. Experiments show that the HIRAG training strategy significantly improves the model’s performance on datasets such as RGB, PopQA, MuSiQue, HotpotQA, and PubmedQA.
Listwise ranking based on Large Language Models (LLMs) has achieved state-of-the-art performance in Information Retrieval (IR).However, their effectiveness often depends on LLMs with massive parameter scales and computationally expensive sliding window processing, leading to substantial efficiency bottlenecks. In this paper, we propose a Collaborative Ranking framework (CoRanking) for LLM-based listwise ranking.Specifically, we strategically combine an efficient small reranker and an effective large reranker for collaborative ranking.The small reranker performs initial passage ranking, effectively filtering the passage set to a condensed top-k list (e.g., top-20 passages), and the large reranker (with stronger ranking capability) then reranks only this condensed subset rather than the full list, significantly improving efficiency. We further address that directly passing the top-ranked passages from the small reranker to the large reranker is suboptimal because of the LLM’s strong positional bias in processing input sequences. To resolve this issue, we propose a passage order adjuster learned by RL that dynamically reorders the top passages returned by the small reranker to better align with the large LLM’s input preferences. Our extensive experiments across three IR benchmarks demonstrate that CoRanking achieves superior efficiency, reducing ranking latency by approximately 70% while simultaneously improving effectiveness, compared to the standalone large reranker.
Typically, parametric adaptation methods such as domain-adaptive pretraining (DAP) and retrieval-augmented generation (RAG) have been considered effective approaches for adapting large language models (LLMs) to new knowledge or domains. To unify positive effects of parametric adaptation and RAG, this paper proposes \textbf{GenPoE}, i.e., “generative’’ passage-level mixture of experts (MoEs) for enhancing knowledge of LLMs. The key component is its novel \textit{MoE-generating hypernetwork} which takes in-context retrieved passages and generates their “expert’’ parameters, where these generated parameters are then integrated into LLMs by forming expert networks. With its use of “generated’’ parameters, GenPoE does not require a separate parameter training or finetuning stage, which is often costly. By parameterizing passages into expert networks, GenPoE likely exhibits robustness even when the retrieved passages are irrelevant. Experiment results in two open-domain question answering (QA) tasks present that GenPoE shows improved performances over other passage-level knowledge editing, and its combination of RAG produces superior performances over RAG. Our data and code will be available at https://github.com/Liu-Xuebing/GenPoE.
This study investigates the position bias in information retrieval, where models tend to overemphasize content at the beginning of passages while neglecting semantically relevant information that appears later. To analyze the extent and impact of position bias, we introduce a new evaluation framework consisting of two position-aware retrieval benchmarks (SQuAD-PosQ, FineWeb-PosQ) and an intuitive diagnostic metric, the Position Sensitivity Index (PSI), for quantifying position bias from a worst-case perspective. We conduct a comprehensive evaluation across the full retrieval pipeline, including BM25, dense embedding models, ColBERT-style late-interaction models, and full-interaction reranker models. Our experiments show that when relevant information appears later in the passage, dense embedding models and ColBERT-style models suffer significant performance degradation (an average drop of 15.6%). In contrast, BM25 and reranker models demonstrate greater robustness to such positional variation. These findings provide practical insights into model sensitivity to the position of relevant information and offer guidance for building more position-robust retrieval systems. Code and data are publicly available at: https://github.com/NovaSearch-Team/position-bias-in-IR.
Detecting Continuously Evolving Scam Calls under Limited Annotation: A LLM-Augmented Expert Rule Framework
PDF ↗The increasing prevalence of scam calls, particularly on online platforms for recruitment, ride-hailing, and delivery services, has become a significant social and economic issue. Traditional approaches to scam call detection rely on labeled data and assume a static distribution of scam narratives. However, scammers continuously evolve their tactics, making these methods less effective. In this paper, we propose a novel approach leveraging large language models (LLMs) to detect continuously evolving scam calls. By abstracting scam and normal call rules based on expert knowledge, we develop a hierarchical few-shot prompting framework. This framework consists of a discrimination module to identify scam characteristics, a reflection module to reduce false positives by comparing with normal call features, and a summary step to synthesize the final detection results. Our method is evaluated on real-world and synthesized datasets, demonstrating superior performance in detecting evolving scam calls with minimal labeled data. Furthermore, we show that the framework is highly adaptable to new scam detection scenarios, requiring only modifications to the expert rules.
Beyond the Scientific Document: A Citation-Aware Multi-Granular Summarization Approach with Heterogeneous Graphs
PDF ↗Scientific summarization remains a challenging task due to the complex characteristics of internal structure and its external relations to other documents. To address this, our proposed model constructs a heterogeneous graph to represent a document and its relevant external citations. This heterogeneous graph enables the model to exploit information across multiple granularities, ranging from fine-grained textual components to the global document structure, and from internal content to external citation context, which facilitates context-aware representations and effectively reduces redundancy. In addition, we develop an effective encoder based on a multi-granularity graph attention mechanism and the triplet loss objective to enhance representation learning performance. Experimental results across three different scenarios consistently demonstrate that our model outperforms existing approaches. Source code is available at: https://github.com/quocanuetcs/CiteHeteroSum.
We introduce Llama-Krikri-8B, a cutting-edge Large Language Model tailored for the Greek language, built on Meta’s Llama 3.1-8B. Llama-Krikri-8B has been extensively trained on high-quality Greek data to ensure superior adaptation to linguistic nuances. With 8 billion parameters, it offers advanced capabilities while maintaining efficient computational performance. Llama-Krikri-8B supports both Modern Greek and English, and is also equipped to handle polytonic text and Ancient Greek. The chat version of Llama-Krikri-8B features a multi-stage post-training pipeline, utilizing both human and synthetic instruction and preference data, by applying techniques such as MAGPIE. In addition, for evaluation, we propose three novel public benchmarks for Greek. Our evaluation on existing as well as the proposed benchmarks shows notable improvements over comparable Greek and multilingual LLMs in both natural language understanding and generation as well as code generation.
Cross-lingual consistency should be considered to assess cross-lingual transferability, maintain the factuality of the model knowledge across languages, and preserve the parity of language model performance. We are thus interested in analyzing, evaluating, and interpreting cross-lingual consistency for factual knowledge.To facilitate our study, we examine multiple pretrained models and tuned models with code-mixed coreferential statements that convey identical knowledge across languages. Interpretability approaches are leveraged to analyze the behavior of a model in cross-lingual contexts, showing different levels of consistency in multilingual models, subject to language families, linguistic factors, scripts, and a bottleneck in cross-lingual consistency on a particular layer. Code-switching training and cross-lingual word alignment objectives show the most promising results, emphasizing the worthiness of cross-lingual alignment supervision and code-switching strategies for both multilingual performance and cross-lingual consistency enhancement. In addition, experimental results suggest promising result for calibrating consistency on test time via activation patching.
Forewarned is Forearmed: Pre-Synthesizing Jailbreak-like Instructions to Enhance LLM Safety Guardrail to Potential Attacks
PDF ↗Despite advances in improving large language model (LLM) to refuse to answer malicious instructions, widely used LLMs remain vulnerable to jailbreak attacks where attackers generate instructions with distributions differing from safety alignment corpora. New attacks expose LLMs’ inability to recognize unseen malicious instructions, highlighting a critical distributional mismatch between training data and real-world attacks that forces developers into reactive patching cycles. To tackle this challenge, we propose **IMAGINE**, a synthesis framework that leverages embedding space distribution analysis to generate jailbreak-like instructions. This approach effectively fills the distributional gap between authentic jailbreak patterns and safety alignment corpora. IMAGINE follows an iterative optimization process that dynamically evolves text generation distributions across iterations, thereby augmenting the coverage of safety alignment data distributions through synthesized data examples. Based on the safety-aligned corpus enhanced through IMAGINE, our framework demonstrates significant decreases in attack success rate on Qwen2.5, Llama3.1, and Llama3.2 without compromising their utility.