Mengru Wang

Aakshita Chandiramani, Aaron Blakeman, Abdullahi Olaoye et al. · amazon-science, cmu

We describe the pre-training, post-training, and quantization of Nemotron 3 Super, a 120 billion (active 12 billion) parameter hybrid Mamba-Attention Mixture-of-Experts model. Nemotron 3 Super is the first model in the Nemotron 3 family to 1) be pre-trained in NVFP4, 2) leverage LatentMoE, a new Mixture-of-Experts architecture that optimizes for both accuracy per FLOP and accuracy per parameter, and 3) include MTP layers for inference acceleration through native speculative decoding. We pre-trained Nemotron 3 Super on 25 trillion tokens followed by post-training using supervised fine tuning (SFT) and reinforcement learning (RL). The final model supports up to 1M context length and achieves comparable accuracy on common benchmarks, while also achieving up to 2.2x and 7.5x higher inference throughput compared to GPT-OSS-120B and Qwen3.5-122B, respectively. Nemotron 3 Super datasets, along with the base, post-trained, and quantized checkpoints, are open-sourced on HuggingFace.

Zhoubo Li, Ningyu Zhang, Yunzhi Yao et al.

As the cost associated with fine-tuning Large Language Models (LLMs) continues to rise, recent research efforts have pivoted towards developing methodologies to edit implicit knowledge embedded within LLMs. Yet, there's still a dark cloud lingering overhead -- will knowledge editing trigger butterfly effect? since it is still unclear whether knowledge editing might introduce side effects that pose potential risks or not. This paper pioneers the investigation into the potential pitfalls associated with knowledge editing for LLMs. To achieve this, we introduce new benchmark datasets and propose innovative evaluation metrics. Our results underline two pivotal concerns: (1) Knowledge Conflict: Editing groups of facts that logically clash can magnify the inherent inconsistencies in LLMs-a facet neglected by previous methods. (2) Knowledge Distortion: Altering parameters with the aim of editing factual knowledge can irrevocably warp the innate knowledge structure of LLMs. Experimental results vividly demonstrate that knowledge editing might inadvertently cast a shadow of unintended consequences on LLMs, which warrant attention and efforts for future works. Code and data are available at https://github.com/zjunlp/PitfallsKnowledgeEditing.

Bozhong Tian, Xiaozhuan Liang, Siyuan Cheng et al.

Large Language Models (LLMs) trained on extensive corpora inevitably retain sensitive data, such as personal privacy information and copyrighted material. Recent advancements in knowledge unlearning involve updating LLM parameters to erase specific knowledge. However, current unlearning paradigms are mired in vague forgetting boundaries, often erasing knowledge indiscriminately. In this work, we introduce KnowUnDo, a benchmark containing copyrighted content and user privacy domains to evaluate if the unlearning process inadvertently erases essential knowledge. Our findings indicate that existing unlearning methods often suffer from excessive unlearning. To address this, we propose a simple yet effective method, MemFlex, which utilizes gradient information to precisely target and unlearn sensitive parameters. Experimental results show that MemFlex is superior to existing methods in both precise knowledge unlearning and general knowledge retaining of LLMs. Code and dataset are released at https://github.com/zjunlp/KnowUnDo.

Shengyu Mao, Xiaohan Wang, Mengru Wang et al.

This paper introduces an innovative task focused on editing the personality traits of Large Language Models (LLMs). This task seeks to adjust the models' responses to opinion-related questions on specified topics since an individual's personality often manifests in the form of their expressed opinions, thereby showcasing different personality traits. Specifically, we construct PersonalityEdit, a new benchmark dataset to address this task. Drawing on the theory in Social Psychology, we isolate three representative traits, namely Neuroticism, Extraversion, and Agreeableness, as the foundation for our benchmark. We then gather data using GPT-4, generating responses that align with a specified topic and embody the targeted personality trait. We conduct comprehensive experiments involving various baselines and discuss the representation of personality behavior in LLMs. Our findings uncover potential challenges of the proposed task, illustrating several remaining issues. We anticipate that our work can stimulate further annotation in model editing and personality-related research. Code is available at https://github.com/zjunlp/EasyEdit.

Haoming Xu, Weihong Xu, Zongrui Li et al.

Long-horizon interactions require language models to manage accumulating information: when to update their state, when to preserve their state, and what to ignore. We study this challenge as \textbf{Contextual Belief Management (CBM)}: maintaining a predicted belief state aligned with formal evidence while isolating task-irrelevant noise. To make CBM measurable, we introduce BeliefTrack, a closed-world benchmark spanning Rule Discovery and Circuit Diagnosis, where a finite belief space and symbolic verifiers enable exact turn-level evaluation. BeliefTrack diagnoses three failures: Failed Stay, Failed Update, and Failed Isolation. Across multiple LLMs, vanilla models exhibit severe CBM failures, while explicit belief-tracking prompts provide limited gains. In contrast, reinforcement learning with belief-state rewards reduces failure rates by 70.9\% on average. Further probing reveals latent belief-state dynamics behind these failures, and representation-level steering reduces failure rates by 46.1\% across two tasks\footnote{Code is coming soon at https://github.com/zjunlp/CBM.

Ziwen Xu, Chenyan Wu, Hengyu Sun et al.

Methods for controlling large language models (LLMs), including local weight fine-tuning, LoRA-based adaptation, and activation-based interventions, are often studied in isolation, obscuring their connections and making comparison difficult. In this work, we present a unified view that frames these interventions as dynamic weight updates induced by a control signal, placing them within a single conceptual framework. Building on this view, we propose a unified preference-utility analysis that separates control effects into preference, defined as the tendency toward a target concept, and utility, defined as coherent and task-valid generation, and measures both on a shared log-odds scale using polarity-paired contrastive examples. Across methods, we observe a consistent trade-off between preference and utility: stronger control increases preference while predictably reducing utility. We further explain this behavior through an activation manifold perspective, in which control shifts representations along target-concept directions to enhance preference, while utility declines primarily when interventions push representations off the model's valid-generation manifold. Finally, we introduce a new steering approach SPLIT guided by this analysis that improves preference while better preserving utility. Code is available at https://github.com/zjunlp/EasyEdit/blob/main/examples/SPLIT.md.

Aaron Blakeman, Aaron Grattafiori, Aarti Basant et al. · nvidia

We present Nemotron 3 Nano 30B-A3B, a Mixture-of-Experts hybrid Mamba-Transformer language model. Nemotron 3 Nano was pretrained on 25 trillion text tokens, including more than 3 trillion new unique tokens over Nemotron 2, followed by supervised fine tuning and large-scale RL on diverse environments. Nemotron 3 Nano achieves better accuracy than our previous generation Nemotron 2 Nano while activating less than half of the parameters per forward pass. It achieves up to 3.3x higher inference throughput than similarly-sized open models like GPT-OSS-20B and Qwen3-30B-A3B-Thinking-2507, while also being more accurate on popular benchmarks. Nemotron 3 Nano demonstrates enhanced agentic, reasoning, and chat abilities and supports context lengths up to 1M tokens. We release both our pretrained Nemotron 3 Nano 30B-A3B Base and post-trained Nemotron 3 Nano 30B-A3B checkpoints on Hugging Face.

Aaron Blakeman, Aaron Grattafiori, Aarti Basant et al. · nvidia

We introduce the Nemotron 3 family of models - Nano, Super, and Ultra. These models deliver strong agentic, reasoning, and conversational capabilities. The Nemotron 3 family uses a Mixture-of-Experts hybrid Mamba-Transformer architecture to provide best-in-class throughput and context lengths of up to 1M tokens. Super and Ultra models are trained with NVFP4 and incorporate LatentMoE, a novel approach that improves model quality. The two larger models also include MTP layers for faster text generation. All Nemotron 3 models are post-trained using multi-environment reinforcement learning enabling reasoning, multi-step tool use, and support granular reasoning budget control. Nano, the smallest model, outperforms comparable models in accuracy while remaining extremely cost-efficient for inference. Super is optimized for collaborative agents and high-volume workloads such as IT ticket automation. Ultra, the largest model, provides state-of-the-art accuracy and reasoning performance. Nano is released together with its technical report and this white paper, while Super and Ultra will follow in the coming months. We will openly release the model weights, pre- and post-training software, recipes, and all data for which we hold redistribution rights.

Mengru Wang, Yunzhi Yao, Ziwen Xu et al.

Understanding knowledge mechanisms in Large Language Models (LLMs) is crucial for advancing towards trustworthy AGI. This paper reviews knowledge mechanism analysis from a novel taxonomy including knowledge utilization and evolution. Knowledge utilization delves into the mechanism of memorization, comprehension and application, and creation. Knowledge evolution focuses on the dynamic progression of knowledge within individual and group LLMs. Moreover, we discuss what knowledge LLMs have learned, the reasons for the fragility of parametric knowledge, and the potential dark knowledge (hypothesis) that will be challenging to address. We hope this work can help understand knowledge in LLMs and provide insights for future research.

Peng Wang, Ningyu Zhang, Bozhong Tian et al.

Large Language Models (LLMs) usually suffer from knowledge cutoff or fallacy issues, which means they are unaware of unseen events or generate text with incorrect facts owing to outdated/noisy data. To this end, many knowledge editing approaches for LLMs have emerged -- aiming to subtly inject/edit updated knowledge or adjust undesired behavior while minimizing the impact on unrelated inputs. Nevertheless, due to significant differences among various knowledge editing methods and the variations in task setups, there is no standard implementation framework available for the community, which hinders practitioners from applying knowledge editing to applications. To address these issues, we propose EasyEdit, an easy-to-use knowledge editing framework for LLMs. It supports various cutting-edge knowledge editing approaches and can be readily applied to many well-known LLMs such as T5, GPT-J, LlaMA, etc. Empirically, we report the knowledge editing results on LlaMA-2 with EasyEdit, demonstrating that knowledge editing surpasses traditional fine-tuning in terms of reliability and generalization. We have released the source code on GitHub, along with Google Colab tutorials and comprehensive documentation for beginners to get started. Besides, we present an online system for real-time knowledge editing, and a demo video.

Mengru Wang, Zhenqian Xu, Junfeng Fang et al.

Large Language Models (LLMs) can acquire unintended biases from seemingly benign training data even without explicit cues or malicious content. Existing methods struggle to detect such risks before fine-tuning, making post hoc evaluation costly and inefficient. To address this challenge, we introduce Data2Behavior, a new task for predicting unintended model behaviors prior to training. We also propose Manipulating Data Features (MDF), a lightweight approach that summarizes candidate data through their mean representations and injects them into the forward pass of a base model, allowing latent statistical signals in the data to shape model activations and reveal potential biases and safety risks without updating any parameters. MDF achieves reliable prediction while consuming only about 20% of the GPU resources required for fine-tuning. Experiments on Qwen3-14B, Qwen2.5-32B-Instruct, and Gemma-3-12b-it confirm that MDF can anticipate unintended behaviors and provide insight into pre-training vulnerabilities.

Mengru Wang, Ningyu Zhang, Ziwen Xu et al.

This paper investigates using knowledge editing techniques to detoxify Large Language Models (LLMs). We construct a benchmark, SafeEdit, which covers nine unsafe categories with various powerful attack prompts and equips comprehensive metrics for systematic evaluation. We conduct experiments with several knowledge editing approaches, indicating that knowledge editing has the potential to detoxify LLMs with a limited impact on general performance efficiently. Then, we propose a simple yet effective baseline, dubbed Detoxifying with Intraoperative Neural Monitoring (DINM), to diminish the toxicity of LLMs within a few tuning steps via only one instance. We further provide an in-depth analysis of the internal mechanism for various detoxifying approaches, demonstrating that previous methods like SFT and DPO may merely suppress the activations of toxic parameters, while DINM mitigates the toxicity of the toxic parameters to a certain extent, making permanent adjustments. We hope that these insights could shed light on future work of developing detoxifying approaches and the underlying knowledge mechanisms of LLMs. Code and benchmark are available at https://github.com/zjunlp/EasyEdit.

Piotr Żelasko, Zhehuai Chen, Mengru Wang et al.

A rising interest in the modality extension of foundation language models warrants discussion on the most effective, and efficient, multimodal training approach. This work focuses on neural machine translation (NMT) and proposes a joint multimodal training regime of Speech-LLM to include automatic speech translation (AST). We investigate two different foundation model architectures, decoder-only GPT and encoder-decoder T5, extended with Canary-1B's speech encoder. To handle joint multimodal training, we propose a novel training framework called EMMeTT. EMMeTT improves training efficiency with the following: balanced sampling across languages, datasets, and modalities; efficient sequential data iteration; and a novel 2D bucketing scheme for multimodal data, complemented by a batch size optimizer (OOMptimizer). We show that a multimodal training consistently helps with both architectures. Moreover, SALM-T5 trained with EMMeTT retains the original NMT capability while outperforming AST baselines on four-language subsets of FLORES and FLEURS. The resultant Multimodal Translation Model produces strong text and speech translation results at the same time.

Hongze Liu, Chang Guo, Yingzeng Li et al.

The transition to open, distributed Multi-Agent Systems (MAS) promises scalable intelligence but introduces a non-trivial tension: maximizing global efficiency requires cooperative, resource-aware scheduling, yet autonomous agents may be self-interested and cannot be managed by a centralized controller. Prior approaches fall short in two key areas: they typically focus on single-query routing, neglecting long-term resource reuse (e.g., KV-caching) and the complexities of system-level many-to-many matching; furthermore, they rely on generic incentive mechanisms that ignore the distinct characteristics of LLM inference. To bridge this gap, we propose IEMAS (Incentive-Efficiency Mechanism for Multi-Agent Systems), a distributed framework that aligns economic incentives with system performance. IEMAS integrates a probabilistic predictive model to estimate Quality of Service (QoS) under uncertainty, which feeds into a VCG-based bipartite matching mechanism. This design guarantees truthful capability reporting and social optimality while explicitly leveraging KV cache-affinity to minimize computational redundancy. We implement IEMAS on top of vLLM and evaluate it via extensive simulations. Results demonstrate that our incentive-efficiency co-design reducing average service cost by 35% and end-to-end latency by up to 2.9 compared to baselines.

Shuofei Qiao, Yunxiang Wei, Jiazheng Fan et al.

The exponential growth of global academic output has confronted researchers and AI agents with an unprecedented ``information explosion,'' where fragmented and unstructured knowledge organization impedes deep interdisciplinary integration. Current academic retrieval tools predominantly rely on superficial keyword matching or vector-space semantic retrieval, which lack the topological reasoning capabilities required to navigate complex logical connections. Agentic deep-research-based frameworks are often prone to logical hallucinations and consuming high inference costs. To bridge this gap, in this report, we introduce SciAtlas, a large-scale, multi-disciplinary, heterogeneous academic resource knowledge graph designed as a panoramic scientific evolution network. By integrating over 43M papers from 26 disciplines, and a total of 157M entities and 3B triplets, SciAtlas provides a structured topological cognitive substrate that dismantles disciplinary barriers and furnishes AI agents with a global perspective. Furthermore, we develop a neuro-symbolic retrieval algorithm featuring tri-path collaborative recall and graph reranking, achieving a seamless transition from simple semantic matching to deterministic association discovery. We also present key application directions of SciAtlas, including literature review, automated research trend synthesis, idea positioning, and academic trajectory exploration, to demonstrate that SciAtlas can serve as an effective ``cognitive map'' to empower the full loop of automated scientific research while significantly reducing reasoning costs. We have released the interfaces for KG retrieval and various downstream tasks in our GitHub repo.

Haoming Xu, Ningyuan Zhao, Liming Yang et al.

Current unlearning methods for large language models usually rely on reverse optimization to reduce target token probabilities. However, this paradigm disrupts the subsequent tokens prediction, degrading model performance and linguistic coherence. Moreover, existing evaluation metrics overemphasize contextual forgetting while inadequately assessing response fluency and relevance. To address these challenges, we propose ReLearn, a data augmentation and fine-tuning pipeline for effective unlearning, along with a comprehensive evaluation framework. This framework introduces Knowledge Forgetting Rate (KFR) and Knowledge Retention Rate (KRR) to measure knowledge-level preservation, and Linguistic Score (LS) to evaluate generation quality. Our experiments show that ReLearn successfully achieves targeted forgetting while preserving high-quality output. Through mechanistic analysis, we further demonstrate how reverse optimization disrupts coherent text generation, while ReLearn preserves this essential capability. Code is available at https://github.com/zjunlp/unlearn.

Ziwen Xu, Shuxun Wang, Kewei Xu et al.

In this paper, we introduce EasyEdit2, a framework designed to enable plug-and-play adjustability for controlling Large Language Model (LLM) behaviors. EasyEdit2 supports a wide range of test-time interventions, including safety, sentiment, personality, reasoning patterns, factuality, and language features. Unlike its predecessor, EasyEdit2 features a new architecture specifically designed for seamless model steering. It comprises key modules such as the steering vector generator and the steering vector applier, which enable automatic generation and application of steering vectors to influence the model's behavior without modifying its parameters. One of the main advantages of EasyEdit2 is its ease of use-users do not need extensive technical knowledge. With just a single example, they can effectively guide and adjust the model's responses, making precise control both accessible and efficient. Empirically, we report model steering performance across different LLMs, demonstrating the effectiveness of these techniques. We have released the source code on GitHub at https://github.com/zjunlp/EasyEdit along with a demonstration notebook. In addition, we provide a demo video at https://www.youtube.com/watch?v=AkfoiPfp5rQ for a quick introduction.

Zihao Yi, Qingxuan Jiang, Ruotian Ma et al.

Large Language Models (LLMs) are increasingly tasked with creative generation, including the simulation of fictional characters. However, their ability to portray non-prosocial, antagonistic personas remains largely unexamined. We hypothesize that the safety alignment of modern LLMs creates a fundamental conflict with the task of authentically role-playing morally ambiguous or villainous characters. To investigate this, we introduce the Moral RolePlay benchmark, a new dataset featuring a four-level moral alignment scale and a balanced test set for rigorous evaluation. We task state-of-the-art LLMs with role-playing characters from moral paragons to pure villains. Our large-scale evaluation reveals a consistent, monotonic decline in role-playing fidelity as character morality decreases. We find that models struggle most with traits directly antithetical to safety principles, such as ``Deceitful'' and ``Manipulative'', often substituting nuanced malevolence with superficial aggression. Furthermore, we demonstrate that general chatbot proficiency is a poor predictor of villain role-playing ability, with highly safety-aligned models performing particularly poorly. Our work provides the first systematic evidence of this critical limitation, highlighting a key tension between model safety and creative fidelity. Our benchmark and findings pave the way for developing more nuanced, context-aware alignment methods.

Ningyu Zhang, Yunzhi Yao, Bozhong Tian et al.

Large Language Models (LLMs) have shown extraordinary capabilities in understanding and generating text that closely mirrors human communication. However, a primary limitation lies in the significant computational demands during training, arising from their extensive parameterization. This challenge is further intensified by the dynamic nature of the world, necessitating frequent updates to LLMs to correct outdated information or integrate new knowledge, thereby ensuring their continued relevance. Note that many applications demand continual model adjustments post-training to address deficiencies or undesirable behaviors. There is an increasing interest in efficient, lightweight methods for on-the-fly model modifications. To this end, recent years have seen a burgeoning in the techniques of knowledge editing for LLMs, which aim to efficiently modify LLMs' behaviors within specific domains while preserving overall performance across various inputs. In this paper, we first define the knowledge editing problem and then provide a comprehensive review of cutting-edge approaches. Drawing inspiration from educational and cognitive research theories, we propose a unified categorization criterion that classifies knowledge editing methods into three groups: resorting to external knowledge, merging knowledge into the model, and editing intrinsic knowledge. Furthermore, we introduce a new benchmark, KnowEdit, for a comprehensive empirical evaluation of representative knowledge editing approaches. Additionally, we provide an in-depth analysis of knowledge location, which can give a deeper understanding of the knowledge structures inherent within LLMs. Finally, we discuss several potential applications of knowledge editing, outlining its broad and impactful implications.

Jizhan Fang, Xinle Deng, Haoming Xu et al. · amazon-science

Despite their remarkable capabilities, Large Language Models (LLMs) struggle to effectively leverage historical interaction information in dynamic and complex environments. Memory systems enable LLMs to move beyond stateless interactions by introducing persistent information storage, retrieval, and utilization mechanisms. However, existing memory systems often introduce substantial time and computational overhead. To this end, we introduce a new memory system called LightMem, which strikes a balance between the performance and efficiency of memory systems. Inspired by the Atkinson-Shiffrin model of human memory, LightMem organizes memory into three complementary stages. First, cognition-inspired sensory memory rapidly filters irrelevant information through lightweight compression and groups information according to their topics. Next, topic-aware short-term memory consolidates these topic-based groups, organizing and summarizing content for more structured access. Finally, long-term memory with sleep-time update employs an offline procedure that decouples consolidation from online inference. Experiments on LongMemEval with GPT and Qwen backbones show that LightMem outperforms strong baselines in accuracy (up to 10.9% gains) while reducing token usage by up to 117x, API calls by up to 159x, and runtime by over 12x. The code is available at https://github.com/zjunlp/LightMem.

Yida Xue, Mingjun Mao, Xiangyuan Ru et al.

We introduce OceanGym, the first comprehensive benchmark for ocean underwater embodied agents, designed to advance AI in one of the most demanding real-world environments. Unlike terrestrial or aerial domains, underwater settings present extreme perceptual and decision-making challenges, including low visibility, dynamic ocean currents, making effective agent deployment exceptionally difficult. OceanGym encompasses eight realistic task domains and a unified agent framework driven by Multi-modal Large Language Models (MLLMs), which integrates perception, memory, and sequential decision-making. Agents are required to comprehend optical and sonar data, autonomously explore complex environments, and accomplish long-horizon objectives under these harsh conditions. Extensive experiments reveal substantial gaps between state-of-the-art MLLM-driven agents and human experts, highlighting the persistent difficulty of perception, planning, and adaptability in ocean underwater environments. By providing a high-fidelity, rigorously designed platform, OceanGym establishes a testbed for developing robust embodied AI and transferring these capabilities to real-world autonomous ocean underwater vehicles, marking a decisive step toward intelligent agents capable of operating in one of Earth's last unexplored frontiers. The code and data are available at https://github.com/OceanGPT/OceanGym.

Yuan Liang, Ruobin Zhong, Haoming Xu et al.

Current AI agents can flexibly invoke tools and execute complex tasks, yet their long-term advancement is hindered by the lack of systematic accumulation and transfer of skills. Without a unified mechanism for skill consolidation, agents frequently ``reinvent the wheel'', rediscovering solutions in isolated contexts without leveraging prior strategies. To overcome this limitation, we introduce SkillNet, an open infrastructure designed to create, evaluate, and organize AI skills at scale. SkillNet structures skills within a unified ontology that supports creating skills from heterogeneous sources, establishing rich relational connections, and performing multi-dimensional evaluation across Safety, Completeness, Executability, Maintainability, and Cost-awareness. Our infrastructure integrates a repository of over 200,000 skills, an interactive platform, and a versatile Python toolkit. Experimental evaluations on ALFWorld, WebShop, and ScienceWorld demonstrate that SkillNet significantly enhances agent performance, improving average rewards by 40% and reducing execution steps by 30% across multiple backbone models. By formalizing skills as evolving, composable assets, SkillNet provides a robust foundation for agents to move from transient experience to durable mastery.

Mengru Wang, Xingyu Chen, Yue Wang et al.

Mixture-of-Experts (MoE) architectures within Large Reasoning Models (LRMs) have achieved impressive reasoning capabilities by selectively activating experts to facilitate structured cognitive processes. Despite notable advances, existing reasoning models often suffer from cognitive inefficiencies like overthinking and underthinking. To address these limitations, we introduce a novel inference-time steering methodology called Reinforcing Cognitive Experts (RICE), designed to improve reasoning performance without additional training or complex heuristics. Leveraging normalized Pointwise Mutual Information (nPMI), we systematically identify specialized experts, termed ''cognitive experts'' that orchestrate meta-level reasoning operations characterized by tokens like ''<think>''. Empirical evaluations with leading MoE-based LRMs (DeepSeek-R1 and Qwen3-235B) on rigorous quantitative and scientific reasoning benchmarks demonstrate noticeable and consistent improvements in reasoning accuracy, cognitive efficiency, and cross-domain generalization. Crucially, our lightweight approach substantially outperforms prevalent reasoning-steering techniques, such as prompt design and decoding constraints, while preserving the model's general instruction-following skills. These results highlight reinforcing cognitive experts as a promising, practical, and interpretable direction to enhance cognitive efficiency within advanced reasoning models.

Mengru Wang, Ziwen Xu, Shengyu Mao et al.

Precise control over language model generation is vital for ensuring both safety and reliability. Although prompt engineering and steering are commonly used to intervene in model behaviors, the vast number of parameters in models often results in highly intertwined internal representations. This interdependency can limit control precision and sometimes lead to unintended side effects. Recent research has explored the use of sparse autoencoders (SAE) to disentangle knowledge in high-dimensional spaces for steering. However, these applications have been limited to toy tasks owing to the nontrivial issue of locating atomic knowledge components. In this paper, we propose Steering Target Atoms (STA), a novel method that isolates and manipulates disentangled knowledge components to enhance safety. Comprehensive experiments demonstrate the effectiveness of our approach. Further analysis reveals that steering exhibits superior robustness and flexibility, particularly in adversarial scenarios. We also apply the steering strategy to the large reasoning model, confirming its effectiveness in precise reasoning control.

Kangwei Liu, Mengru Wang, Yujie Luo et al.

Fine-tuning enables large language models (LLMs) to adapt to specific domains, but often compromises their previously established safety alignment. To mitigate the degradation of model safety during fine-tuning, we introduce LookAhead Tuning, a lightweight and effective data-driven approach that preserves safety during fine-tuning. The method introduces two simple strategies that modify training data by previewing partial answer prefixes, thereby minimizing perturbations to the model's initial token distributions and maintaining its built-in safety mechanisms. Comprehensive experiments demonstrate that LookAhead Tuning effectively maintains model safety without sacrificing robust performance on downstream tasks. Our findings position LookAhead Tuning as a reliable and efficient solution for the safe and effective adaptation of LLMs.

Lyucheng Wu, Mengru Wang, Ziwen Xu et al.

Recent progress in Multimodal Large Language Models (MLLMs) has unlocked powerful cross-modal reasoning abilities, but also raised new safety concerns, particularly when faced with adversarial multimodal inputs. To improve the safety of MLLMs during inference, we introduce a modular and adaptive inference-time intervention technology, AutoSteer, without requiring any fine-tuning of the underlying model. AutoSteer incorporates three core components: (1) a novel Safety Awareness Score (SAS) that automatically identifies the most safety-relevant distinctions among the model's internal layers; (2) an adaptive safety prober trained to estimate the likelihood of toxic outputs from intermediate representations; and (3) a lightweight Refusal Head that selectively intervenes to modulate generation when safety risks are detected. Experiments on LLaVA-OV and Chameleon across diverse safety-critical benchmarks demonstrate that AutoSteer significantly reduces the Attack Success Rate (ASR) for textual, visual, and cross-modal threats, while maintaining general abilities. These findings position AutoSteer as a practical, interpretable, and effective framework for safer deployment of multimodal AI systems.

Zhengliang Shi, Ruotian Ma, Jen-tse Huang et al. · pku, tencent-ai

Large language models (LLMs) are increasingly entrusted with high-stakes decisions that affect human welfare. However, the principles and values that guide these models when distributing scarce societal resources remain largely unexamined. To address this, we introduce the Social Welfare Function (SWF) Benchmark, a dynamic simulation environment where an LLM acts as a sovereign allocator, distributing tasks to a heterogeneous community of recipients. The benchmark is designed to create a persistent trade-off between maximizing collective efficiency (measured by Return on Investment) and ensuring distributive fairness (measured by the Gini coefficient). We evaluate 20 state-of-the-art LLMs and present the first leaderboard for social welfare allocation. Our findings reveal three key insights: (i) A model's general conversational ability, as measured by popular leaderboards, is a poor predictor of its allocation skill. (ii) Most LLMs exhibit a strong default utilitarian orientation, prioritizing group productivity at the expense of severe inequality. (iii) Allocation strategies are highly vulnerable, easily perturbed by output-length constraints and social-influence framing. These results highlight the risks of deploying current LLMs as societal decision-makers and underscore the need for specialized benchmarks and targeted alignment for AI governance.

Xinbei Ma, Ruotian Ma, Xingyu Chen et al. · pku, tencent-ai

LLM-based multi-agent systems demonstrate great potential for tackling complex problems, but how competition shapes their behavior remains underexplored. This paper investigates the over-competition in multi-agent debate, where agents under extreme pressure exhibit unreliable, harmful behaviors that undermine both collaboration and task performance. To study this phenomenon, we propose HATE, the Hunger Game Debate, a novel experimental framework that simulates debates under a zero-sum competition arena. Our experiments, conducted across a range of LLMs and tasks, reveal that competitive pressure significantly stimulates over-competition behaviors and degrades task performance, causing discussions to derail. We further explore the impact of environmental feedback by adding variants of judges, indicating that objective, task-focused feedback effectively mitigates the over-competition behaviors. We also probe the post-hoc kindness of LLMs and form a leaderboard to characterize top LLMs, providing insights for understanding and governing the emergent social dynamics of AI community.

Xiaohan Chen, Zhongying Pan, Quan Feng et al.

Despite Retrieval-Augmented Generation improving code completion, traditional retrieval methods struggle with information redundancy and a lack of diversity within limited context windows. To solve this, we propose a resource-optimized retrieval augmentation method, SaraCoder. It maximizes information diversity and representativeness in a limited context window, significantly boosting the accuracy and reliability of repository-level code completion. Its core Hierarchical Feature Optimization module systematically refines candidates by distilling deep semantic relationships, pruning exact duplicates, assessing structural similarity with a novel graph-based metric that weighs edits by their topological importance, and reranking results to maximize both relevance and diversity. Furthermore, an External-Aware Identifier Disambiguator module accurately resolves cross-file symbol ambiguity via dependency analysis. Extensive experiments on the challenging CrossCodeEval and RepoEval-Updated benchmarks demonstrate that SaraCoder outperforms existing baselines across multiple programming languages and models. Our work proves that systematically refining retrieval results across multiple dimensions provides a new paradigm for building more accurate and resource-optimized repository-level code completion systems.