As Large Language Models (LLMs) are deployed globally, ensuring their safety and alignment across multiple languages becomes paramount. However, safety behaviors often vary unpredictably between languages, posing significant challenges for consistent and ethical AI. In this work, we systematically investigate the dynamics of multilingual alignment, exploring whether single-language alignment transfers cross-lingually, how language consistency is preserved during training, and the resulting trade-offs with general knowledge capabilities. We introduce RefusEU a novel refusal alignment dataset covering 12 European languages, including a dedicated test set for evaluating current state-of-the-art models. Our controlled Direct Preference Optimization (DPO) experiments provide two key insights: aligning models exclusively in English is insufficient to ensure cross-lingual safety, even for the same harm categories, whereas training on multilingual datasets can improve safety without degrading general performance, as measured by the Global MMLU benchmark.
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
E-commerce search relevance is a critical component of retrieval systems. While Large Language Models (LLMs)-driven Chain-of-Thought (CoT) modeling has become the dominant paradigm and yielded significant gains, a critical gap remains: the absence of a systematic definition for comprehensive relevance reasoning, which leads to significant blind spots in current approaches. In this paper, we deconstruct the task into three core competencies: reasoning knowledge, multi-modal understanding, and rule awareness. Accordingly, we propose LoRE(Large Generative Model for Search Relevance), a novel two-stage training framework. We first employ an SFT phase to instill these capabilities via a progressive CoT synthesis pipeline, followed by a Reinforcement Learning(RL) phase, which serves as a regularizer, pruning redundant logic to achieve precise and robust adjudication. Extensive experiments validate LoRE, outperforming GPT-5 by 29.1% in Macro-F1 and achieving a 27% online gain, offering a vital reference for industrial domain-specific post-training.
Learning How to Remember: A Meta-Cognitive Management Method for Structured and Transferable Agent Memory
PDF ↗Large language model (LLM) agents increasingly rely on accumulated memory to solve long-horizon decision-making tasks. However, most existing approaches store memory in fixed representations and reuse it at a single or implicit level of abstraction, which limits generalization and often leads to negative transfer when distribution shift. This paper proposes the Meta-Cognitive Memory Abstraction method (MCMA), which treats memory abstraction as a learnable cognitive skill rather than a fixed design choice. MCMA decouples task execution from memory management by combining a frozen task model with a learned memory copilot. The memory copilot is trained using direct preference optimization; it determines how experience should be structured, abstracted, and reused. Memories are further organized into a hierarchy of abstraction levels, enabling selective reuse based on task similarity. When no memory is transferable, MCMA transfers the ability to abstract and manage memory by transferring the memory copilot. Experiments on ALFWorld, ScienceWorld, and BabyAI demonstrate substantial improvements in performance, out-of-distribution generalization, and cross-task transfer over several baselines.
Learning to Evolve: A Self-Improving Framework for Multi-Agent Systems via Textual Parameter Graph Optimization
PDF ↗Designing and optimizing multi-agent systems (MAS) is a complex, labor-intensive process of "Agent Engineering." Existing automatic optimization methods, primarily focused on flat prompt tuning, lack the structural awareness to debug the intricate web of interactions in MAS. More critically, these optimizers are static; they do not learn from experience to improve their own optimization strategies. To address these gaps, we introduce Textual Parameter Graph Optimization (TPGO), a framework that enables a multi-agent system to learn to evolve. TPGO first models the MAS as a Textual Parameter Graph (TPG), where agents, tools, and workflows are modular, optimizable nodes. To guide evolution, we derive "textual gradients," structured natural language feedback from execution traces, to pinpoint failures and suggest granular modifications. The core of our framework is Group Relative Agent Optimization (GRAO), a novel meta-learning strategy that learns from historical optimization experiences. By analyzing past successes and failures, GRAO becomes progressively better at proposing effective updates, allowing the system to learn how to optimize itself. Extensive experiments on complex benchmarks like GAIA and MCP-Universe show that TPGO significantly enhances the performance of state-of-the-art agent frameworks, achieving higher success rates through automated, self-improving optimization.
Neural Processing Units (NPUs) are critical for AI infrastructure, yet developing kernels remains a bottleneck due to the complexity of vendor-specific Domain-Specific Languages (DSLs). While LLMs excel in general coding, they fail to meet the stringent constraints of NPU development, showing a near-zero success rate on complex kernels in our preliminary study. To address these challenges, we present AscendKernelGen, the first comprehensive framework for NPU kernel development, marking a pioneering effort in this field. This framework consists of three interconnected components: (1) Ascend-CoT, the first dataset in the NPU kernel domain that incorporates chain-of-thought reasoning from real-world kernel implementations; (2) KernelGen-LM, a domain-adaptive model trained on this novel dataset using supervised fine-tuning and reinforcement learning; and (3) NPUKernelBench, the first benchmark platform designed to evaluate the compilation, correctness, and performance of generated NPU kernels. Experimental results demonstrate that our approach dramatically bridges the gap in hardware-specific coding: compilation success on complex Level-2 kernels improves from 0% to 95.5% (Pass@10), with 64% functional correctness. AscendKernGen is available at AscendKernGen and NPUKernelBench.
Multi-turn interaction remains challenging for online reinforcement learning. Current GRPO-based methods—either at the trajectory level or the step level—still suffer from fundamental challenges in multi-turn settings: they allocate sampling uniformly across tasks regardless of difficulty, propagate misleading learning signals that penalize correct intermediate actions in failed trajectories, and incur high sample-collection costs under long-horizon environments. Step-level variants (e.g., GIGPO) mitigate some interaction-cost constraints by decomposing trajectories, yet they retain GRPO’s sampling imbalance and still struggle with heterogeneous multi-turn tasks. To address these issues, we propose STEP (Success-rate-aware Trajectory-Efficient Policy Optimization), a framework that dynamically allocates sampling based on per-task success rates and performs fine-grained step-level optimization. STEP maintains a smoothed success-rate record to guide adaptive trajectory resampling, allocating more effort to harder tasks. It then computes success-rate-weighted advantages and decomposes trajectories into step-level samples, followed by a step-level GRPO augmentation that strengthens updates on low-success tasks. Experiments on OSWorld and AndroidWorld show that STEP substantially improves sample efficiency and training stability over both trajectory-level and existing step-level GRPO variants, converging faster and generalizing better under the same sampling budget.
Clustered Self-Assessment: A Simple yet Effective Method for Uncertainty Quantification in Large Language Models
PDF ↗Large language models (LLMs) demonstrate remarkable performance across diverse tasks, but they often generate responses that appear plausible while being factually incorrect. This problem is compounded by the lack of explicit uncertainty estimates, which makes it difficult for users to judge the reliability of model outputs. Existing uncertainty quantification methods typically rely on indirect signals, such as entropy across sampled generations. These signals can be difficult to interpret and do not fully leverage the model’s ability to assess its own uncertainty. We propose a simple and effective self-assessment method for uncertainty quantification in LLMs. Our approach groups sampled generations into semantically distinct clusters, converts them into answer options in a structured multiple-choice question, and uses the probability assigned by the LLM to each option as a confidence estimate. Experiments across multiple models and datasets show that our method consistently outperforms baseline approaches. Notably, it achieves competitive performance with as few as two additional samples, demonstrating both its effectiveness and efficiency.
Reinforcement Learning with Verifiable Rewards (RLVR) has demonstrated promising gains in enhancing the reasoning capabilities of large language models. However, its dependence on domain-specific verifiers significantly restricts its applicability to open and general domains. Recent efforts such as RLPR have extended RLVR to general domains, enabling training on broader datasets and achieving improvements over RLVR. However, a notable limitation of these methods is their tendency to overfit to reference answers, which constrains the model’s ability to generate diverse outputs. This limitation is particularly pronounced in open-ended tasks such as writing, where multiple plausible answers exist. To address this, we propose DARL, a simple yet effective reinforcement learning framework that encourages the generation of diverse answers within a controlled deviation range from the reference while preserving alignment with it. Our framework is fully compatible with existing general reinforcement learning methods and can be seamlessly integrated without additional verifiers. Extensive experiments on thirteen benchmarks demonstrate overall improvements in reasoning performance. Notably, DARL surpasses RLPR, achieving average gains of 1.3 points on six reasoning benchmarks and 9.5 points on seven general benchmarks, highlighting its effectiveness in improving both reasoning accuracy and output diversity.
Reasoning in a Combinatorial and Constrained World: Benchmarking LLMs on Natural-Language Combinatorial Optimization
PDF ↗While large language models (LLMs) have shown strong performance in math and logic reasoning, their ability to handle combinatorial optimization (CO)—searching high-dimensional solution spaces under hard constraints—remains underexplored. To bridge the gap, we introduce NLCO, a **N**atural **L**anguage **C**ombinatorial **O**ptimization benchmark that evaluates LLMs on end-to-end CO reasoning: given a language-described decision-making scenario, the model must output a discrete solution without writing code or calling external solvers. NLCO covers 43 CO problems and is organized using a four-layer taxonomy of variable types, constraint families, global patterns, and objective classes, enabling fine-grained evaluation. We provide solver-annotated solutions and comprehensively evaluate LLMs by feasibility, solution optimality, and reasoning efficiency. Experiments across a wide range of modern LLMs show that high-performing models achieve strong feasibility and solution quality on small instances, but both degrade as instance size grows, even if more tokens are used for reasoning. We also observe systematic effects across the taxonomy: set-based tasks are relatively easy, whereas graph-structured problems and bottleneck objectives lead to more frequent failures. The benchmark dataset and code for data generation and evaluation are publicly available.
With the widespread adoption of large language models (LLMs), understanding their personality representation mechanisms has become critical. As a novel paradigm in Personality Editing, most existing methods employ neuron-editing to locate and modify LLM neurons, requiring changes to numerous neurons and leading to significant performance degradation. This raises a fundamental question: Are all modified neurons directly related to personality representation? In this work, we investigate and quantify this specificity through assessments of general capability impact and representation-level patterns. We find that: 1) Current methods can change personalities but reduce overall performance. 2) Neurons are multifunctional, connecting personality traits and general knowledge. 3) Opposing personality traits demonstrate distinctly mutually exclusive representation patterns. Motivated by these findings, we propose DPN-LE (Dual Personality Neuron Localization and Editing), which identifies personality-specific neurons by contrasting MLP activations between high-trait and low-trait samples. DPN-LE constructs layer-wise steering vectors and applies dual-criterion filtering based on Cohen’s d effect size and activation magnitude to isolate mutually exclusive neuron subsets. Sparse linear intervention on these neurons enables precise personality control at inference time. Using only 1,000 contrastive sample pairs per trait, DPN-LE intervenes on \sim0.5% of neurons while achieving competitive personality control and substantially better capability preservation across reasoning tasks. Experiments on LLaMA-3-8B-Instruct and Qwen2.5-7B-Instruct demonstrate the effectiveness and generalizability of our approach.
Implicit In-Context Learning compresses demonstration examples into a single context vector and injects it into the model’s activation space, achieving few-shot-level performance at near zero-shot inference cost. However, existing approaches typically aggregate demonstrations from all classes into a shared, task-level context vector, capturing global task information but without explicitly preserving fine-grained, class-conditional semantic distinctions. In this work, we propose Class-Conditional Context Vectors (C3V), a simple yet effective extension to implicit in-context learning that explicitly models class-specific contextual information by constructing separate context vectors for each class, without relying on explicit prompts. These class-conditional context vectors are additively injected into the model’s residual streams in a single forward pass, enabling contextual contributions to be modulated in a class-aware manner while keeping the backbone frozen. We evaluate C3V on multiple text classification benchmarks across several families of large language models. Experimental results demonstrate that C3V consistently outperforms task-level context vector baselines, and achieves higher average accuracy than conventional few-shot learning on most models.
Learning with Noisy Labels (LNL) is a challenge where the collected training set can contain incorrect or corrupted labels. Most existing solutions distinguish clean samples from noisy samples and query human experts on noisy samples for denoising. However, these solutions often operate under the unrealistic assumption that the distribution of classes is uniform, overlooking the skewed and imbalanced distributions frequently encountered in real-world scenarios. In this case, we empirically reveal that previous solutions suffer from both selection bias and training bias, leading to distinguish clean samples from noisy samples hardly. In this paper, our work introduces the imbalanced learning with noisy labels (i-LNL) task, which seeks to let the model learn from noisy labels within imbalanced distributions. A new benchmark (ImbaLNL-Bench) comprised of some synthetic and real-world datasets is created to provide a thorough representation of practical use cases. Besides, we propose an innovative collaborative learning framework DeCo for i-LNL tasks. Specifically, we first conduct debiased sample selection, consisting of a robust expert model and a debiased-enhanced threshold strategy, to better separate clean samples from noisy samples, especially for the tail classes. Then we feed selected clean samples to active annotator large language models (LLMs) for re-annotating noisy samples using in-context learning, which can better reduce human effort. Ultimately, we employ distinct loss functions adept at managing subsets with varying degrees of label noise. Extensive experimental results on synthetic and real-world datasets show the effectiveness and superiority of our method.
Benchmarking the Fine-Grained Discriminability in Image-Text Retrieval via Controlled Contrastive Differences
PDF ↗Existing cross-modal image-text retrieval models often retrieve samples with inconsistent details. To evaluate fine-grained discriminability, we introduce MSCOCO-CCD and Flickr30k-CCD, with three key features: (1) a two-level image content taxonomy for contrastive sample generation and fine-grained evaluation; (2) annotation of numerous contrastive samples, where each sample differs from the anchor by a controlled contrastive difference (CCD), with the specific type of difference labeled; (3) a fine-grained contrastive discrimination metric to assess the ability to distinguish fine-grained nuances. Extensive experiments demonstrate that contrastive samples can significantly degrade retrieval performance. Furthermore, fine-grained evaluation reveals that current models still struggle to effectively produce discriminative representations on certain feature types, such as entity emotion and scene attribute. Our datasets and related codes will be publicly released.
Producing presentation slides automatically entails coordinating narrative structure with page-level graphic design under strict spatial constraints. For such structured multimodal tasks, a well-organized design process is essential to ensure the final quality of slides. Existing approaches rely on fixed templates or directly emit executable code, thereby both limiting the creative layout-design capabilities of LLMs and bypassing the essential slide-page design step. To address these limitations, this paper: (1) proposes a hierarchical slides generation workflow DeepSlides that systematically organizes slide design tasks without any predefined template or style, decoupling slide-page design from implementation; (2) introduces SlideDesign, a dataset tailored specifically for slides generation tasks; (3) presents a multi-agent reinforcement learning training paradigm and trains a couple of models SlideQwens for slide design and implementation. Experimental results demonstrate that our proposed framework outperforms baseline methods on evaluated metrics and achieves superior performance in human preference evaluations. The dataset and code are available at: https://anonymous.4open.science/r/DeepSlides-D14D
Good Reasoning Makes Good Demonstrations: Implicit Reasoning Quality Supervision via In-Context Reinforcement Learning
PDF ↗Reinforcement Learning with Verifiable Rewards (RLVR) improves reasoning in large language models but treats all correct solutions equally, potentially reinforcing flawed traces that arrive at correct answers by chance. We observe that better reasoning makes better demonstrations: high-quality solutions serve as more effective in-context examples than low-quality ones. We term this teaching ability Demonstration Utility, and show that the policy model’s own in-context learning ability provides an efficient way to measure it, yielding a quality signal termed Evidence Gain. To leverage this signal during training, we introduce In-Context RLVR, which prepends demonstrations before each rollout. Theoretically, we prove that this simple input modification implicitly reweights rewards by a factor approximately proportional to Evidence Gain, assigning higher weights to high-quality traces without requiring costly computation. Experiments on mathematical reasoning benchmarks demonstrate consistent improvements in both accuracy and reasoning quality over standard RLVR baselines. Our codes and datasets are available at https://github.com/Mithas-114/IC-DAPO.
Large Language Models often struggle with complex, multi-step operational tasks because they remain static during inference and cannot learn from past experience. To address this, we propose MUSE, a framework that enables iterative self-improvement through a hierarchical Memory Module. MUSE organizes cross-domain insights to facilitate the orchestration of long-horizon workflows. The core of our approach is an autonomous post-execution critique mechanism: after completing each sub-task, the system analyzes its operational logs and distills raw execution data into structured, reusable knowledge. This allows the agent to evolve dynamically rather than relying on fixed parameters. Evaluated on the rigorous TAC productivity benchmark, MUSE achieves new state-of-the-art results, significantly outperforming previous methods using only the streamlined Gemini-2.5 Flash model. Our analysis demonstrates that MUSE’s performance scales with the accumulation of insights and exhibits strong cross-task transferability, marking a key step toward autonomous systems capable of lifelong learning in professional environments. Demo videos can be found in our supplementary materials.
Mixture-of-Experts (MoE) architectures have emerged as an effective approach for scaling Large Language Models (LLMs) by activating only a subset of experts during inference. Despite their computational efficiency, MoE models incur a substantial memory bottleneck from maintaining all expert parameters during inference. To address this challenge, numerous MoE pruning methods have been proposed. However, most existing methods adopt uniform pruning across layers, which fails to capture layer-wise variations in expert importance and redundancy. In this paper, we propose COmpensated MoE Pruning with Expert-Layer distribution (COMPEL). COMPEL performs layer-adaptive expert pruning by estimating expert importance using Fisher information and deriving layer importance from layer-wise outlier distributions, enabling pruning decisions that capture layer-wise heterogeneity. Furthermore, to mitigate performance degradation resulting from expert pruning, we propose a Fisher information guided expert weight compensation method. Experimental results on the Qwen1.5-MoE-A2.7B achieve near lossless performance at 25% expert pruning and maintains performance within a 4% margin even at 50% pruning. Moreover, COMPEL consistently outperforms existing pruning methods while substantially reducing inference latency and peak GPU memory usage.
While excelling at solving complex problems, Large Reasoning Models (LRMs) are still constrained by the overthinking issue. Most current studies rely on reward shaping in Reinforcement Learning (RL) to shorten the Chain-of-Thought (CoT) of LRMs, remaining sample-inefficient and non-robust due to the absence of guided exploration and prioritized exploitation. To address these issues, we propose a novel policy optimization framework with **S**elf-**I**mitation and self-**G**uidance **M**ech**A**nisms (SIGMA), which reshapes the exploration and exploitation through two core components: (i) **self-imitation exploitation**, which enables the prioritized exploitation of high-value prompts and rollouts by introducing a self-imitated loss and a dynamic sampling strategy based on compression rate; (ii) **self-guidance exploration**, which provides a preference-aware exploration guidance through diverse and pluggable self-rewriting strategies. Experiments across various datasets indicate that our method achieves superior reasoning efficiency without compromising, and even facilitating, the overall accuracy. Furthermore, ablation studies show that the proposed mechanisms can provide flexible control interfaces for the tradeoff between the reasoning accuracy and efficiency of LRMs.
Most existing Multimodal Sentiment Analysis (MSA) methods rely on holistic fusion, treating all modalities and temporal segments equally. Such strategies often introduce redundant information and obscure the decision process, limiting both robustness and interpretability. Inspired by dual-process theory, we propose FRSR (Fast Retrieval and Slow Reasoning), an interpretable framework that decomposes multimodal sentiment modeling into two cooperative pathways. The Fast Pathway acts as a lightweight evidence selector, using context-aware convolution and auxiliary supervision to retrieve a sparse set of Top-K sentiment-relevant cues from noisy multimodal inputs. Based on these cues, the Slow Pathway performs deeper cross-modal reasoning through learnable reasoning tokens, enabling hierarchical sentiment inference. By separating salient evidence retrieval from multimodal reasoning, FRSR improves interpretability while reducing computational cost. Experiments on three benchmark datasets show that FRSR achieves competitive performance, higher efficiency, stronger robustness to noise, and clearer decision transparency than existing holistic fusion methods.
With the advancement of multimodal large language models (MLLMs), building GUI agent systems has become an increasingly promising direction—especially for mobile platforms, given their rich app ecosystems and intuitive touch interactions. Yet mobile GUI agents face a critical dilemma: truly on-device models (4B or smaller) lack sufficient performance, while capable models (starting from 7B) are either too large for mobile deployment or prohibitively costly (e.g., cloud-only closed-source MLLMs). To resolve this, we propose OpenPhone, a mobile GUI agent system that leverages device-cloud collaboration to tap the cost-efficiency of on-device models and the high capability of cloud models, while avoiding their drawbacks. Specifically, OpenPhone enhances Qwen2.5-VL-3B via two-stage SFT→GRPO training on synthetic GUI data for strong decision-making, integrates an efficient long-reasoning mechanism to utilize historical interactions under tight resources, and defaults to on-device execution—only escalating challenging subtasks to the cloud via real-time complexity assessment. Experiments on the online AndroidLab benchmark and diverse apps show OpenPhone matches or nears larger models, with a significant reduction in cloud costs.