The scaling of Large Language Models (LLMs) is increasingly limited by data quality. Most methods handle data mixing and sample selection separately, which can break the structure in code corpora. We introduce UniGeM, a framework that unifies mixing and selection by treating data curation as a manifold approximation problem without training proxy models or relying on external reference datasets. UniGeM operates hierarchically: Macro-Exploration learns mixing weights with stability-based clustering; Micro-Mining filters high-quality instances by their geometric distribution to ensure logical consistency. Validated by training 8B and 16B MoE models on 100B tokens, UniGeM achieves 2.0 \times data efficiency over a random baseline and further improves overall performance compared to SOTA methods in reasoning-heavy evaluations and multilingual generalization.
论文检索
输入标题、作者或关键词，从 100,903 篇学术成果中精准定位
Watermarking provides a critical safeguard for large language model (LLM) services by facilitating the detection of LLM-generated text. Correspondingly, stealing watermark algorithms (SWAs) derive watermark information from watermarked texts generated by victim LLMs to craft highly targeted adversarial attacks, which compromise the reliability of watermarks. Existing SWAs rely on fixed strategies, overlooking the non-uniform distribution of stolen watermark information and the dynamic nature of real-world LLM generation processes. To address these limitations, we propose Adaptive Stealing (AS), a novel SWA featuring enhanced design flexibility through Position-Based Seal Construction and Adaptive Selection modules. AS operates by defining multiple attack perspectives derived from distinct activation states of contextually ordered tokens.During attack execution, AS dynamically selects the optimal perspective based on watermark compatibility, generation priority, and dynamic generation relevance. Our experiments demonstrate that AS significantly increases steal efficiency against target watermarks under identical experimental conditions.These findings highlight the need for more robust LLM watermarks to withstand potential attacks. We release our code to the community for future research[<https://github.com/DrankXs/AdaptiveStealingWatermark>].
Document summarization becomes more challenging when summaries must reflect a user’s subjective interests in addition to document salience. SOTA Large Language Models (LLMs) show strong in-context summarization capabilities. Prior works report that simply prepending long and dynamically evolving user histories leads to unstable, inconsistent personalization. To address this, we introduce PerDucer, a personalization inducer for frozen language models. Given a user interaction sequence (trajectory) and a query document, PerDucer first predicts the next likely preference signal. It then maps the latent signal to a small set of personalized keyphrases for the query document. These keyphrases serve as the control cues that steer the frozen summarizers (both LLMs and SLMs) towards user-aligned summaries. Across the PENS and OpenAI-Reddit benchmarks, PerDucer-boosted LLMs consistently outperform the strongest history-prompting baselines, yielding an average +0.18 improvement across personalization metrics (PerSEval in our case). Two PerDucer-augmented SLMs approach the top-performing evaluated LLM, with SmolLM2-1.7B reaching 97% of the best-performing DeepSeek-R1-14B score. These results indicate that short keyphrase cues can induce personalization in frozen summarizers without modifying their parameters.
While Multimodal Large Language Models (MLLMs) excel at general vision-language tasks, precise coordinate prediction remains a significant challenge, particularly as high-resolution inputs cause visual positional encodings (VPEs) to degrade. We demonstrate that these encoding failures do not result in random noise but instead trigger predictable, directional biases, suggesting that models default to internal spatial priors when grounding signals are weak. To counteract this, we introduce Vision-PE Shuffle Guidance (VPSG), a training-free, inference-time correction method. VPSG isolates position-unconditioned tendencies by shuffling VPEs and utilizes this negative evidence to steer digit decoding through a lightweight finite-state machine. Evaluation on the ScreenSpot-Pro benchmark confirms that VPSG effectively rectifies coordinate drift, yielding consistent improvements in localization accuracy across various model scales without any retraining.
Despite the remarkable evolution of Large Language Models (LLMs) from simple assistants to versatile agents, effective personalization remains a significant challenge. Existing approaches often treat user preferences as static or merely time-varying traits, overlooking the dynamic nature of human behavior: preferences can shift, and even conflict, depending on context. To address this limitation, we propose a fine-grained taxonomy to differentiate between stable preferences, which are context-agnostic, and situational preferences, which are context-dependent. Building on this taxonomy, we introduce S2Pref, a new dataset of 10k meticulously curated entries. Each entry is grounded in a multi-turn dialogue that implicitly manifests either a stable or a situational preference, as defined by our hierarchical taxonomy. We further design three complementary evaluation tasks to benchmark LLMs on their ability to prioritize contextual signals, proactively resolve ambiguity, and efficiently infer user preferences. Our dataset and diagnostic tasks provide a practical testbed for advancing dynamic, context-aware personalization in conversational agents.
Agents based on large language models have recently shown strong potential on real-world software engineering (SWE) tasks that require long-horizon interaction with repository-scale codebases. However, most existing agents rely on append-only context maintenance or passively triggered compression heuristics, which often lead to context explosion, semantic drift, and degraded reasoning in long-running interactions. We propose Cat, a new context management paradigm that elevates context maintenance to a callable tool integrated into the decision-making process of agents. Cat formalizes a structured context workspace consisting of stable task semantics, condensed long-term memory, and high-fidelity short-term interactions, and enables agents to proactively compress historical trajectories into actionable summaries at appropriate milestones. To support context management for SWE-agents, we propose a trajectory-level supervision framework, CaT-Generator, based on an offline data construction pipeline that injects context-management actions into complete interaction trajectories. Using this framework, we train a context-aware model, SWE-Compressor. Experiments on SWE-Bench-Verified demonstrate that SWE-Compressor reaches a 57.6% solved rate and significantly outperforms ReAct-based agents and static compression baselines, while maintaining stable and scalable long-horizon reasoning under a bounded context budget.
Reinforcement learning with verifiable reward (RLVR) has become a promising paradigm for post-training large language models (LLMs) to improve their reasoning capability. However, when the rollout accuracy is low on hard problems, the reward becomes sparse, limiting learning efficiency and causing exploration bottlenecks. Existing approaches either rely on teacher models for distillation or filter out difficult problems, which limits scalability or restricts reasoning improvement through exploration.We propose EvoCoT, a self-evolving curriculum learning framework based on two-stage chain-of-thought (CoT) reasoning optimization. EvoCoT constrains the exploration space by self-generating and verifying CoT trajectories, then gradually shortens CoT steps to expand the space in a controlled way. The framework enables LLMs to stably learn from initially unsolved hard problems under sparse rewards. We apply EvoCoT to multiple LLM families, including Qwen, DeepSeek, and Llama. Experiments show that EvoCoT enables LLMs to solve previously unsolved problems, improves reasoning capability without external CoT supervision, and is compatible with various RL fine-tuning methods. We release the source code to support future research.
While Large Language Models (LLMs) offer promising avenues for automated cognitive restructuring in mental health settings, current approaches predominantly focus on superficial positive reframing and lack the adaptability to balance conflicting therapeutic dimensions, such as empathy and rationality. To address these deficiencies, we propose CARE-CR, a context-aware framework that implements a decoupled optimization paradigm. We first train expert policies specialized for distinct therapeutic attributes rather than relying on a monolithic alignment strategy. To mitigate expert data scarcity, we introduce Dimension-Guided Hierarchical Monte Carlo Tree Search (DG-HMCTS) for data-efficient preference augmentation. At inference, a context-aware routing module dynamically predicts optimal preference weights to fuse expert outputs based on the user’s specific distress context. Extensive experiments demonstrate that CARE-CR achieves consistent improvements over strong baselines across multiple evaluation dimensions, including diagnostic accuracy, contextual appropriateness, task effectiveness, and overall helpfulness, while enabling controllable cognitive restructuring generation. The dataset and code are publicly available at https://github.com/HongzhiQ/CARE-CR.
Historical newspapers from the colonial period offer valuable evidence of how racializing language evolved over time. However, there are challenges in studying this type of historical data: 1) Data scarcity: acquiring large, annotated historical datasets is difficult, hindering the possibility of analyzing racialization comprehensively; 2) Digitized materials frequently contain Optical Character Recognition (OCR) errors and other types of noise that complicate text extraction and computational analysis; 3) Colonial newspapers are often multilingual and written in archaic prose, hindering the effectiveness of NLP tools developed for modern, single language texts. This paper addresses these challenges by conducting a dual-view, jointly studying multilingual event extraction and temporal semantic shift tasks. Specifically, we introduce a contextual question answering (CQA) and a visual question answering (VQA) derived from eighteenth- and nineteenth-century colonial newspapers. Content-wise, we focus on how enslaved people were described by enslavers as well as how they articulated their own condition through QA pairs of newspapers written in Dutch, English-French, and Spanish. Our results show that LLMs are still limited for low-resource VQA tasks. For temporal semantic change, we train temporal word embedding with a compass. The study concludes that racialization is a fluid process of linguistic recalibration where the decline of slavery merely shifted the language of control onto new categories of labor and identity.
LinguaGame: A Linguistically Grounded Game-Theoretic Paradigm for Multi-Agent Dialogue Generation
PDF ↗Large Language Models (LLMs) have enabled Multi-Agent Systems (MASs) where agents interact through natural language to solve complex tasks or simulate multi-party dialogues. Recent work on LLM-based MASs has mainly focused on architecture design, such as role assignment and workflow orchestration. In contrast, this paper targets the interaction process itself, aiming to improve agents’ communication efficiency by helping them convey their intended meaning more effectively through language. To this end, we propose LinguaGame, a linguistically-grounded game-theoretic paradigm for multi-agent dialogue generation. Our approach models dialogue as a signalling game over communicative intents and strategies, solved with a training-free equilibrium approximation algorithm for inference-time decision adjustment. Unlike prior game-theoretic MASs, whose game designs are often tightly coupled with task-specific objectives, our framework relies on linguistically informed reasoning with minimal task-specific coupling. Specifically, it treats dialogue as intentional and strategic communication, requiring agents to infer what others aim to achieve (intents) and how they pursue those goals (strategies). We evaluate our framework in simulated courtroom proceedings and debates, with human expert assessments showing significant gains in communication efficiency. We release our code and data on GitHub.
Syntactic movement is a core concept in generative linguistics to account for word-order variation and long-distance dependencies, but its psychological and neurobiological status remains debated. Here, we test the neural reality of movement in English and Chinese by correlating brain activity during naturalistic listening with syntactic node counts, traces and word embeddings derived from X-bar style tree annotations. We find that deep structure significantly predicts neural responses in English but not in Chinese, providing partial support for movement-based accounts while revealing clear cross-linguistic differences.
Distributional alignment enables large language models (LLMs) to predict how a target population distributes its responses across answer options, rather than collapsing disagreement into a single consensus answer. However, existing LLM-based distribution prediction is often unstable and degrades under cultural and domain shift. Token score-based estimates can change with minor option wording or formatting, response sampling-based estimates are expensive and sensitive to prompts and decoding settings, and directly generated distributions are frequently miscalibrated.We propose Evi-DA, an evidence-based alignment technique that improves the fidelity and robustness of LLM-based distribution estimation under domain and cultural shift. Given a target country and a multiple-choice question, Evi-DA retrieves related World Values Survey items and their answer distributions, predicts a coarse Welzel value signature for each option, and infers the country-conditioned answer distribution in a structured format. We train the LLMs using a two-stage pipeline, where reinforcement learning optimizes survey-derived rewards that encourage accurate intermediate value predictions, faithful final distributions, well-formed structured outputs, and reduced cultural bias. Across in-domain and out-of-domain benchmarks and multiple open-source backbones, Evi-DA reduces Jensen-Shannon divergence between predicted and gold distributions relative to strong baselines, with average relative improvements of up to 44%.
Targeting the Needle, Ignoring the Haystack: Anchoring Crucial Cues for Evolving Scam Call Detection via an LLM-Assisted Classifier
PDF ↗Automatic detection of fraudulent voice calls is essential for online service platforms but faces significant challenges due to the scarcity of labeled data and the continuous evolution of conversational contexts. Standard supervised methods often fail to generalize, as they tend to overfit to variable background narratives rather than capturing the core deceptive intent. In this paper, we propose a lightweight framework that anchors detection on Semantic Primitives, a set of stable, interpretable evidentiary cues derived from expert knowledge. Our approach decomposes the fraud detection task into two distinct stages: identifying the presence of these predefined semantic signals within the transcript, and deriving a final verdict through a logical combination of the detected cues. By explicitly prioritizing stable evidence over diverse conversational noise, this framework ensures that decisions are based on verifiable fraud tactics rather than spurious correlations. Experimental results demonstrate that our method achieves superior robustness and efficiency compared to traditional baselines, particularly in scenarios with shifting service contexts.
Large language models (LLMs) have demonstrated competitive performance in zero-shot multilingual machine translation (MT). Some follow-up works further improved MT performance via preference optimization, but they leave a key aspect largely underexplored: the order in which data samples are given during training. We address this topic by integrating curriculum learning into various state-of-the-art preference optimization algorithms to boost MT performance. We introduce a novel curriculum learning strategy with restarts (CLewR), which reiterates easy-to-hard curriculum multiple times during training to effectively mitigate the catastrophic forgetting of easy examples. We demonstrate consistent gains across several model families (Gemma2, Qwen2.5, Llama3.1) and preference optimization techniques. We publicly release our code at https://github.com/alexandra-dragomir/CLewR.
While large language models (LLMs) have achieved remarkable success, their reliability in knowledge-intensive tasks is often compromised by factual hallucinations. Integrating Knowledge Graphs (KGs) addresses this issue; however, existing approaches typically rely on simple graph traversal.This paradigm decouples topological navigation from logical operations (e.g., temporal filtering, aggregation), leading to imprecise retrieval and heavy post-processing burdens.Although semantic parsing offers a solution by grounding reasoning in logical forms, it traditionally suffers from a dependency on scarce supervised annotations.To bridge this gap, we propose Interactive Semantic Parsing, a framework that formulates reasoning as the sequential generation of executable logical clauses. This design allows logical constraints to be dynamically interleaved with graph search, while optimizing via reinforcement learning with only final answer feedback eliminates the need for gold program annotations.To tackle the sparse reward challenge in the vast symbolic space, we introduce a distance-aware process reward to evaluate intermediate steps based on their topological proximity to the answer.Experimental results on WebQSP and CWQ demonstrate that our method achieves state-of-the-art performance, particularly on complex queries, validating the effectiveness of our dense reward signal in enabling robust reasoning without supervision.Our code is available at https://github.com/NUSTM/ISP-KGR.
The rapid advancement of large language models (LLMs) has accelerated progress toward universal AI assistants. However, existing benchmarks for personalized assistants remain misaligned with real-world user-assistant interactions, failing to capture the complexity of external contexts and users’ cognitive states. To bridge this gap, we propose LifeSim, a user simulator that models user cognition through the Belief-Desire-Intention (BDI) model within physical environments for coherent life trajectories generation, and simulates intention-driven user interactive behaviors. Based on LifeSim, we introduce LifeSim-Eval, a comprehensive benchmark for multi-scenario, long-horizon personalized assistance. LifeSim-Eval covers 8 life domains and 1,200 diverse scenarios, and adopts a multi-turn interactive method to assess models’ abilities to complete explicit and implicit intentions, recover user profiles, and produce high-quality responses. Under both single-scenario and long-horizon settings, our experiments reveal that current LLMs face significant limitations in handling implicit intention and long-term user preference modeling.
With the rapid progress of large language models (LLMs), aligning a general-purpose model with downstream tasks through fine-tuning has become a central research focus. Selecting only high-quality examples for training has been shown to be one of the most effective ways to improve fine-tuning performance. However, prior work concentrates almost exclusively on data preprocessing: filtering and cleaning data before training begins. While the order and composition of training data during training have received little fine-grained attention. To fill this gap, our work proposed Fine-Grained Order Fine-Tuning, a fine-grained scheduling method of data order in epochs. Drawing on curriculum-learning principles, FOT defines data difficulty based on the relevance between the data and the model, and then performs dynamic scheduling of the training order in each epoch according to the difficulty. On both large-scale continued pre-training and small-scale supervised fine-tuning experiments, FOT has achieved an average 2.4% improvement over baselines. Our study offers a new perspective on data governance in the fine-tuning phase.
MimicLM: Zero-Shot Voice Imitation through Autoregressive Modeling of Pseudo-Parallel Speech Corpora
PDF ↗Voice imitation aims to transform *source* speech to match a *reference* speaker’s timbre and speaking style while preserving linguistic content. A straightforward approach is to train on triplets of *(source, reference, target)*, where *source* and *target* share the same content but *target* matches the *reference*’s voice characteristics, yet such data is extremely scarce. Existing approaches either employ carefully designed disentanglement architectures to bypass this data scarcity or leverage external systems to synthesize pseudo-parallel training data. However, the former requires intricate model design, and the latter faces a quality ceiling when synthetic speech is used as training *targets*. To address these limitations, we propose MimicLM, which takes a novel approach by using synthetic speech as training *sources* while retaining real recordings as *targets*. This design enables the model to learn directly from real speech distributions, breaking the synthetic quality ceiling. Building on this data construction approach, we incorporate interleaved text-audio modeling to guide the generation of content-accurate speech and apply post-training with preference alignment to mitigate the inherent distributional mismatch when training on synthetic data. Experiments demonstrate that MimicLM achieves superior voice imitation quality with a simple yet effective architecture, significantly outperforming existing methods in naturalness while maintaining competitive similarity scores across speaker identity, accent, and emotion dimensions.
Claim verification with large language models (LLMs) has recently attracted growing attention, due to their strong reasoning capabilities and transparent verification processes compared to traditional answer-only judgments. However, existing approaches to online claim verification, which requires iterative evidence retrieval and reasoning, still mainly rely on prompt engineering or pre-designed reasoning workflows, without unified training to improve necessary skills. Therefore, we introduce Veri-R1, an online reinforcement learning (RL) framework that enables an LLM to interact with a search engine and to receive reward signals that explicitly shape its planning, retrieval, and reasoning behaviors. The dynamic interaction between models and retrieval systems more accurately reflects real-world verification scenarios and fosters comprehensive verification skills. Empirical results show that Veri-R1 improves joint accuracy by up to 30% and doubles evidence score, often surpassing its larger-scale model counterparts. Ablation studies further reveal the impact of reward components, and the link between output logits and label accuracy. Our results highlight the effectiveness of online RL for precise and faithful claim verification, and provide a foundation for future research.
EmoMM: Benchmarking and Steering MLLM for Multimodal Emotion Recognition under Conflict and Missingness
PDF ↗Multimodal Emotion Recognition (MER) is critical for interpreting real-world interactions. While Multimodal Large Language Models (MLLM) have shown promise in MER, their internal decision-making mechanisms under modality conflict and missingness remain largely underexplored. In this paper, to systematically investigate these behaviors, we introduce EmoMM, a comprehensive benchmark featuring modality-aligned, conflict, and missing subsets. Through extensive evaluation, we uncover a Video Contribution Collapse (VCC) phenomenon, where MLLM marginalize video evidence due to high token redundancy and modality preferences. To address this, we propose Conflict-aware Head-level Attention Steering (CHASE), a lightweight mechanism that detects modality conflicts and performs inference-time attention steering, effectively mitigating decision bias without retraining the backbone. Experimental results demonstrate that CHASE consistently improves performance across various settings, significantly enhancing the reliability of MLLM in complex affective scenarios.