论文检索

Social Aspects · Safety

Tongxi Wu, Jian Zhang, Yang Gao

Safety alignment in large language models (LLMs) and multimodal large language models (MLLMs) is commonly assumed to operate as a near-binary threshold mechanism. We challenge this assumption by revealing that safety behavior is governed by an \emph{instability region} where small perturbations induce stochastic refusal decisions rather than deterministic outcomes. We develop a multi-metric diagnostic framework combining external and internal signals to characterize this instability. Through systematic experiments, we identify a characteristic \emph{diagnostic signature}: inputs in unstable regimes exhibit elevated output uncertainty yet \emph{decreased} internal safety activation, a decoupling phenomenon that explains why detection-based defenses fail against sophisticated attacks. Building on this framework, we introduce \textbf{Furina}, a jailbreak attack that deliberately induces this signature through fragmented, scene-anchored prompts without model-specific optimization. Furina outperforms strong single-turn and multi-turn baselines on HarmBench and achieves competitive results on MM-SafetyBench, demonstrating that uncertainty amplification provides a principled and transferable mechanism for understanding safety vulnerabilities. Code and supplementary materials: \url{https://anonymous.4open.science/r/Furina_Jailbreak-EF7C}.

Applications · Everything Else

Jiaxing Zhao, Hongbin Xie, Yuzhen Lei, Xuan Song, Zhuoran Shi, Lianxin Li, Shuangxue Liu, Haoran Zhang

Multi-Agent Systems (MAS) powered by Large Language Models (LLMs) have recently become a strong paradigm for solving complex workflow-structured tasks through expert collaboration. However, the data that make such collaboration effective are typically distributed across organizations and cannot be centrally pooled due to privacy, intellectual property, and compliance constraints. Federated Learning preserves data locality, yet most federated paradigms treat clients as independent and fail to capture workflow dependencies that are essential for coherent multi-stage collaboration. Data locality and workflow dependency are orthogonal, and the key challenge arises where both must be satisfied, namely federated, workflow-aware collaboration. We introduce FedWave, a framework that enables LLM-based experts to solve sequential workflows under strict privacy constraints. FedWave integrates a Value Chain Layer that encodes inter-stage dependencies with communication-efficient federated LoRA adaptation, a server-side Mixture-of-Experts (MoE) router that performs input-conditioned expert fusion at inference time while retaining standard federated aggregation during training, and a Direct Preference Optimization (DPO) stage that aligns collaborative outputs using router-induced preferences. Experiments show that FedWave consistently outperforms strong federated baselines and remains competitive with centralized multi-agent systems without compromising data privacy. Code is available at https://anonymous.4open.science/r/FedWave-111A.

Deep Learning · Everything Else

Xin Yang, Yemin Wang, Mingda Liu, Letian Li, Shuaishuai Cao, ZhengXiao He, Ryan Dong

Scaling large language models (LLMs) has driven their success, yet dense Transformers couple capacity and computation: every parameter is activated for every token, making training and inference costs grow linearly with model size—a critical bottleneck as models approach trillion-parameter regimes. We aim to scale capacity through MoE-style mixture throughout the LLM pipeline rather than only the FFN. Prior pipeline-level approaches include ParaScale, which introduces virtual tokens and parallel streams but incurs substantial overhead and suffers from homogenized routing and gradient collapse, and AltUp, which uses an auxiliary prediction branch but offers limited adaptivity and slow convergence. We establish that MoE-style mixture layers can be reformulated as variable-kernel dynamic convolutions, where each expert corresponds to a $1{\times}1$ convolutional kernel and routing implements input-conditioned kernel aggregation. Building on this equivalence, we introduce cMoLLM: a convolutionally gated mixture-of-LLMs that routes over end-to-end streams through fully differentiable dynamic convolution. In GPT-2-style models trained on FineWeb, cMoLLM improves language modeling perplexity and downstream GLUE and SQuAD accuracy under matched compute, with better stream utilization, more stable optimization, and favorable scaling compared to ParaScale- and AltUp-style baselines.

Deep Learning · Large Language Models

Ido Amit, Ido Galil, Ran El-Yaniv

As LLMs generate increasingly long outputs, effective uncertainty estimation must identify errors at fine-grained levels rather than discard entire responses. While such methods exist, evaluating uncertainty at any resolution (token to an entire generation) is challenging and highly sensitive to label imperfections, making zero-noise benchmarks essential; yet, long-form generation benchmarks tend to rely on fallible labels rather than deterministic ground truth. We introduce Single-answer Atomic Long-form Target (SALT), a benchmark of six procedurally generated tasks with single deterministic long textual ground truths, enabling unit-level evaluation of correctness, calibration, and ranking without external judges. Equipped with SALT, our analysis of 50+ LLMs reveals key insights: We identify which confidence functions dominate each uncertainty aspect and show that effective ranking benefits more from coarser evaluation resolutions; SALT further facilitates precise calibration tracking throughout generation, revealing a divergence in the accuracy–calibration relationship, with high- and low-performing models exhibiting degradation ($\rho=0.87$) and improvement ($\rho=-0.92$). Finally, we demonstrate that reasoning, via Chain-of-Thought prompting or internalized through training, introduces a trade-off, improving accuracy while degrading confidence ranking. These findings directly impact risk-critical applications requiring reliable error identification and mitigation.

Applications · Everything Else

Yanfang Liu, Mingjun Wang, Peng XU, Rongliang Fu, Bei Yu, Tsung-Yi Ho

Analog circuits constitute the indispensable interface between physical reality and digital computation, underpinning safety-critical systems from autonomous driving to medical implants. Consequently, verification correctness is paramount; yet, it remains the critical bottleneck in hardware design, consuming over 50\% of engineering cycles due to a heavy reliance on the manual interpretation of unstructured, heterogeneous specifications. While Large Language Models (LLMs) offer automation potential, their probabilistic, autoregressive nature is structurally misaligned with the strict determinism required for analog verification tasks. Specifically, generic LLMs struggle to resolve semantic dispersion, latent causal dependencies, and numerical precision. To bridge this gap, we introduce AnalogVerifier, a neuro-symbolic framework that automates end-to-end testbench generation by decoupling semantic translation from logical enforcement. We propose a four-stage architecture: (1) Context-Aware Task Serialization transforms complex specifications into atomic tasks via an agentic workflow; (2) Graph-Symbolic Scheduling satisfies analog design constraints through Port Dependency Graphs (PDG) for correct-by-construction sequencing; (3) Numerical-Symbolic Grounding mitigates numerical hallucination by delegating threshold derivation to a deterministic symbolic oracle; (4) Closed-Loop Repair enables correctness and completeness of the generated testbenches by simulation feedback. Evaluation on five industrial analog circuits demonstrates that AnalogVerifier achieves 82.3\%--100\% functional pass rate, establishing a new paradigm for reliable, automated analog verification. The code and data are publicly available at \url{https://anonymous.4open.science/r/ICML26--AnalogVerifier-72EE/}.

Deep Learning · Large Language Models

Xinpeng Wang, William Cao, Andrew Wilson, Zhe Zeng

Recent studies on hallucination detection have shown that hallucination-related signals are more strongly encoded in intermediate layers than in the final layer of large language models (LLMs). While a growing body of work has sought to exploit this property for hallucination detection, the problem of how to automate the selection of high-performing layers is underexplored, and the development of principled methods for this purpose remains an open challenge. To address this gap, we first propose several hypotheses for why such signals emerge in intermediate layers and test corresponding criteria for automatic layer selection. We evaluate these criteria across two LLM architectures and five datasets, and find that none of them deliver satisfying performance. Instead, we propose a new selection criterion, First Effective Peak of Intrinsic Dimension (FEPoID), that is able to consistently identify optimal or near-optimal layers and outperforms the aforementioned criteria and existing hallucination detection baselines. This criterion is training-free and requires negligible computational overhead. Additionally, we study the generation behaviors of LLMs and introduce a simple yet effective truncation strategy, which further amplifies the hallucination-related signals and leads to substantial improvements in overall detection performance.

Deep Learning · Generative Models and Autoencoders

Julian Wustl, Philipp Haid, Yarema Okhrin, Claudius Schnörr

Codebook-based generators built on masked language model (MLM) transformers have become highly effective in text and vision, yet remain underused for tabular data. This is because codebooks typically act as information bottlenecks, whereas tabular generation requires them to generalize. We address this gap with Q-Tab, a codebook-based tabular generator based on lookup-free quantization (LFQ) with residual corruption. The resulting corruption kernel induces a moving Nadaraya–Watson–style kernel regression over a large discrete code space, which turns codebook learning into a moving-target problem. We derive necessary conditions for the learnability of such moving codebooks and show how the residual LFQ construction aligns with these conditions. Q-Tab achieves state-of-the-art downstream predictive utility and missing-value imputation, while matching the distributional fidelity of diffusion-based generators, notably without any post-hoc temperature tuning.

Applications · Time Series

Zihao Yao, Qi Zheng, Jiankai Zuo, YAYING ZHANG

Synthesizing realistic time series with generative models has wide-ranging applications in real-world scenarios. Despite recent progress, most existing methods are trained under the assumption of abundant training data, which substantially limits their effectiveness in data-scarce settings. In this paper, we propose TimeMoDE, a novel framework that integrates Diffusion Transformers with Mixture-of-Experts to exploit both domain adaptability and diffusion-stage awareness for time series generation under data scarcity. It is pre-trained on a large-scale collection of multi-domain datasets to extract domain-agnostic temporal representations and domain-specific information benefiting generalization during fine-tuning. We propose Domain Prompts to condition expert assignment for indistinguishable noised tokens, mitigating the limitations of capturing inter-dataset relationships. Moreover, we incorporate diffusion timestep signals to equip the experts with awareness of time series degradation variations, facilitating adaptive calibrate to stage-dependent denoising requirements. Extensive experiments demonstrate that TimeMoDE outperforms existing methods under diverse low-data settings. It establishes an innovative paradigm for advanced time series few-shot generation.

Social Aspects · Security

Mitchell Hermon, Rahul Gupta, Weitong Ruan, Ekraam Sabir, Haohan Wang

We identify a fundamental tension in securing LLMs: the \textbf{security--fidelity tradeoff}. While defenses against indirect prompt injection are becoming more robust, we show that they inevitably impair the model's ability to process benign, instruction-like text. Current evaluations miss this cost because they conflate utility with fidelity. We address this gap with \textsc{SecFid}, a benchmark that uses behaviorally separable probes to unambiguously distinguish between resisting an attack, succumbing to it, and faithfully processing it as data. Our evaluation reveals this tradeoff across a diverse set of models and highlights how the strongest defenses achieve security often by aggressively suppressing valid content, causing fidelity failure rates up to 50\% on translation. We ground these results in a decision-theoretic framework, proving that when benign and adversarial inputs overlap, no universal defense exists. Therefore, optimal robustness is strictly task-dependent, determined by an application’s tolerance for fidelity errors versus security failures.

Deep Learning · Large Language Models

Leyla Mirvakhabova, Babak Ehteshami Bejnordi, Gaurav Kumar, Hanxue Liang, Wanru Zhao, Paul Whatmough

Upcycling pre-trained dense models into sparse Mixture-of-Experts (MoEs) efficiently increases model capacity but often suffers from poor expert specialization due to naive weight replication. We introduce Dirichlet-Prior Shaping Loss (DPSL), a novel router regularization technique that directly shapes routing probability distributions by matching expert assignments to a target Dirichlet prior resulting in enhanced expert specialization. DPSL enables encoding of inductive biases such as encouraging experts to focus on specific modalities or tasks, without requiring manual intervention. DPSL is a general tool applicable to any module that outputs categorical probability distributions, extending its utility beyond MoE training. Experiments on upcycled MoE vision-language models show that DPSL consistently outperforms upcycling strategies and regularization techniques across standard vision-language benchmarks, addressing the critical issue of poor specialization and fostering higher-performing models.

Reinforcement Learning · Multi-agent

David Mguni, Yaqi Sun, Haojun Chen, Wanrong Yang, Amir Darabi, Larry Orimoloye, Yaodong Yang

We study robustness to agent malfunctions in cooperative multi-agent reinforcement learning (MARL), a failure mode that is critical in practice yet underexplored in existing theory. We introduce MARTA, a plug-and-play robustness layer that augments standard MARL algorithms with a {\fontfamily{cmss}\selectfont Switcher}–{\fontfamily{cmss}\selectfont Adversary} mechanism which selectively induces malfunctions in performance-critical states. This formulation defines a fault-switching $(N+2)$-player Markov game in which the {\fontfamily{cmss}\selectfont Switcher} chooses when and which agent fails, and the {\fontfamily{cmss}\selectfont Adversary} controls the resulting faulty behaviour via random or worst-case policies. We develop a Q-learning-type scheme and show that the associated Bellman operator is a contraction, yielding existence and uniqueness of the minimax value, convergence to a Markov perfect equilibrium. MARTA integrates seamlessly with MARL algorithms without architectural modification and consistently improves robustness across Traffic Junction (TJ), Level-Based Foraging (LBF), MPE SimpleTag, and SMAC (v2). In these domains, MARTA achieves large gains in final performance of up to \textbf{116.7\%} in SMAC, \textbf{21.4\%} in MPE SimpleTag, and \textbf{44.6\%} in LBF, while significantly reducing failure rates under train–test mismatched fault regimes. These results establish MARTA as a theoretically grounded and practically deployable mechanism for fault-tolerant MARL.

Deep Learning · Large Language Models

Jiaming Liang, Zhaoxin Wang, Handing Wang

Large Language Models (LLMs) exhibit significant safety disparities across languages, with low-resource languages (LRLs) often bypassing safety guardrails established for high-resource languages (HRLs) like English. Existing solutions, such as multilingual supervised fine-tuning (SFT) or Reinforcement Learning from Human Feedback (RLHF), are computationally expensive and de- pendent on scarce multilingual safety data. In this paper, we propose a novel, training-free alignment framework based on Sparse Weight Editing. Identifying that safety capabilities are localized within a sparse set of ”safety neurons,” we formulate the cross-lingual alignment problem as a constrained linear transformation. We derive a closed-form solution to optimally map the harmful representations of LRLs to the robust safety subspaces of HRLs, while preserving general utility via a null-space projection constraint. Extensive experiments across 8 languages and multiple model families (Llama-3, Qwen-2.5) demonstrate that our method significantly reduces Attack Success Rate (ASR) in LRLs with negligible impact on general reasoning capabilities, all achieved with a single, data-efficient calculation.

Qingyu Yin, Yulun Wu, Zhennan Shen, Sunbowen Lee, Zhilin Wang, Yanshu Li, Chak Tou Leong, Jiale Kang, Jinjin Gu

We systematically evaluate Parameter-Efficient Fine-Tuning (PEFT) methods under the paradigm of Reinforcement Learning with Verifiable Rewards (RLVR). RLVR incentivizes language models to enhance their reasoning capabilities through verifiable feedback; however, while methods like LoRA are commonly used, the optimal PEFT architecture for RLVR remains unidentified. In this work, we conduct the first comprehensive evaluation of over 12 PEFT methodologies across the DeepSeek-R1-Distill families on mathematical reasoning benchmarks. Our empirical results challenge the default adoption of standard LoRA with three main findings. First, we demonstrate that structural variants, such as DoRA, AdaLoRA, and MiSS, consistently outperform LoRA. Second, we uncover a spectral collapse phenomenon in SVD-informed initialization strategies (e.g., PiSSA, MiLoRA), attributing their failure to a fundamental misalignment between principal-component updates and RL optimization. Furthermore, our ablations reveal that extreme parameter reduction (e.g., VeRA, Rank-1) severely bottlenecks reasoning capacity. We further conduct ablation studies and scaling experiments to validate our findings. This work provides a definitive guide for advocating for more exploration for parameter-efficient RL methods.

Social Aspects · Alignment

Yucong Huang, Xiucheng Li, Kaiqi Zhao, Jing Li

Standard RLHF relies on transitive scalar rewards, failing to capture the cyclic nature of human preferences. While some approaches like the General Preference Model (GPM) address this, we identify a theoretical limitation: their implicit formulation entangles hierarchy with cyclicity, failing to guarantee dominant solutions. To address this, we propose the Hybrid Reward-Cyclic (HRC) model, which utilizes game-theoretic decomposition to explicitly disentangle preferences into orthogonal transitive (scalar) and cyclic (vector) components. Complementing this, we introduce Dynamic Self-Play Preference Optimization (DSPPO), which treats alignment as a time-varying game to progressively guide the policy toward the Nash equilibrium. Experiments on RewardBench 2 demonstrate that HRC consistently improves over both BT and GPM baselines (e.g., +1.23\% on Gemma-2B). In particular, its superior performance in the \textit{Ties} domain empirically validates the model's robustness in handling complex, non-strict preferences. Furthermore, extensive evaluations on downstream generation tasks (AlpacaEval 2.0 and MT-Bench) confirm the efficacy of our framework. Notably, on Gemma-2B-it, HRC+DSPPO achieves a peak length-controlled win-rate of 44.75\% against GPT-4-Turbo on AlpacaEval 2.0 (evaluated by GPT-4o-mini), significantly outperforming standard SPPO baselines trained with BT or GPM.

Applications · Social Sciences

Ying Jin, Weiqing Min, Mingyu Huang, Shuqiang Jiang

A sustainable diet represents a multi-dimensional synergy among four essential pillars: nutrition adequacy, economic affordability, cultural acceptability, and environmental respect. Despite the prevalence of population-level sustainability modeling, practical implementation relies on effective individual-level adoption. This transition is often hindered by inter-individual heterogeneity, posing a formidable challenge in aligning sustainable diet requirements with individual preferences. To address this issue, we propose a personalized sustainable diet recommendation model based on a constraint-aware decision-making mechanism, where sustainability is incorporated through learnable constraints rather than modeled as user preferences. To systematically evaluate the proposed approach, we construct a sustainable diet dataset named SusDiet with about 150k recipes, characterized by broad coverage of sustainability indicators across four dimensions. Experimental results on this dataset show that our method promotes more sustainable choices without compromising individual preference. This work establishes a framework for aligning individual dietary choices with planetary health, offering quantitative evidence to guide future sustainable diet interventions and policy-making for sustainable development.

Social Aspects · Alignment

Yuting Ning, Jaylen Jones, Zhehao Zhang, Chentao Ye, Weitong Ruan, Junyi Li, Rahul Gupta, Huan Sun

Computer-use agents (CUAs) have made tremendous progress in the past year, yet they still frequently produce misaligned actions that deviate from the user's original intent. Such misaligned actions may arise from external attacks (e.g., indirect prompt injection) or from internal limitations (e.g., erroneous reasoning). They not only expose CUAs to safety risks, but also degrade task efficiency and reliability. This work makes the first effort to define and study misaligned action detection in CUAs, with comprehensive coverage of both externally induced and internally arising misaligned actions. We further identify three common categories in real-world CUA deployment and construct MisActBench, a benchmark of realistic trajectories with human-annotated, action-level misalignment labels. Moreover, we propose DeAction, a practical and universal guardrail that detects misaligned actions before execution and iteratively corrects them through structured feedback. DeAction outperforms all existing baselines across offline and online evaluations with moderate latency overhead: (1) On MisActBench, it outperforms baselines by over 12\% absolute in F1 score; (2) In online evaluation, it reduces attack success rate by over 75\% under adversarial settings while preserving or even improving task success rate in benign environments.

Deep Learning · Sequential Models, Time series

Zinuo You, Jin Zheng, John Cartlidge

Irregular multivariate time series pose a fundamental trade-off for long-horizon forecasting: discrete methods can distort temporal structure via re-gridding, while continuous-time models often rely on sequential numerical solvers that are prone to drift. To bridge this gap, we present the Latent Laplace Diffusion (LLapDiff), a generative framework that models the target as a low-dimensional latent trajectory, enabling horizon-wide generation without step-by-step integration over physical time. We guide the reverse process using a stable modal parameterization motivated by stochastic port-Hamiltonian dynamics, and parameterize its mean evolution in the Laplace domain via learnable complex-conjugate poles, allowing for direct evaluation over irregular timestamps. Moreover, we link continuous dynamics to irregular observations through renewal-averaging analysis, which maps sampling gaps to effective event-domain poles and theoretically motivates a gap-aware history summarizer for conditioning. Extensive experiments demonstrate that LLapDiff consistently outperforms baselines in long-horizon forecasting, and its continuous-time generative nature also supports missing-value imputation by querying the same model at historical timestamps.

Applications · Computer Vision

Xingyilang Yin, Chengzhengxu Li, Jiahao Chang, Chi-Man Pun, Xiaodong Cun

Humans are born with vision-based 4D spatial-temporal intelligence, which enables us to perceive and reason about the evolution of 3D space over time from purely visual inputs. Despite its importance, this capability remains a significant bottleneck for current multimodal large language models (MLLMs). To tackle this challenge, we introduce MLLM-4D, a comprehensive framework designed to bridge the gaps in training data curation and model post-training for spatiotemporal understanding and reasoning. On the data front, we develop a cost-efficient data curation pipeline that repurposes existing stereo video datasets into high-quality 4D spatiotemporal instructional data. This results in the MLLM4D-2M and MLLM4D-R1-30k datasets for Supervised Fine-Tuning (SFT) and Reinforcement Fine-Tuning (RFT), alongside MLLM4D-Bench for comprehensive evaluation. Regarding model training, our post-training strategy establishes a foundational 4D understanding via SFT and further catalyzes 4D reasoning capabilities by employing Group Relative Policy Optimization (GRPO) with specialized Spatiotemporal Chain of Thought (ST-CoT) prompting and Spatiotemporal reward functions without involving the modification of architecture. Extensive experiments demonstrate that MLLM-4D achieves state-of-the-art spatial-temporal understanding and reasoning capabilities from purely 2D RGB inputs. Our project will be made publicly available.

Applications · Robotics

Zhixuan Lin, Gengqi Liu, Chao Zheng, Gao Lin, Jindong Yu, Song Gao, Fei Wang

In robotic imitation learning, accurately modeling the multimodality and temporal correlations of long-horizon action sequences remains challenging. Long-horizon tasks require preserving global task intent while executing precise low-level control; otherwise, local errors can accumulate and lead to failure. While recent coarse-to-fine autoregressive models have improved action generation, they struggle to maintain consistency across hierarchies, leading to suboptimal performance in long-horizon tasks. To address these shortcomings, we propose Probabilistically Consistent Multi-Scale Action Generation (MSP), a novel coarse-to-fine approach that promotes cross-scale consistency. MSP adopts a streamlined multi-scale design by directly downsampling in a continuous latent space. A scale-wise autoregressive Transformer is used to generate semantic conditions at each scale, which guide a lightweight MeanFlow model to capture multi-scale latent distributions, enabling probabilistically consistent refinement across scales. Through extensive simulation and real-world experiments, including long-horizon, multi-task, and few-shot generalization settings, we show that MSP outperforms existing coarse-to-fine methods, achieving state-of-the-art performance with high efficiency. Our code will be publicly available upon publication.

Deep Learning · Large Language Models

Alexey Gorbatovski, Boris Shaposhnikov, Viacheslav Sinii, Alexey Malakhov, Daniil Gavrilov

Direct Alignment Algorithms (DAAs) simplify LLM alignment by directly optimizing policies, bypassing reward modeling and RL. While DAAs differ in their use of SFT (one-stage vs. two-stage) and the scalar score they optimize (likelihood vs. odds ratios), the key performance drivers remain underexplored. We present a systematic comparison and analyze a previously overlooked axis - the ranking objective (pairwise vs. pointwise). To isolate this factor, we propose a unified training framework across DAAs by (i) converting one-stage methods (ORPO, ASFT) into a two-stage pipeline with an explicit SFT phase and (ii) introducing a $\beta$ parameter that places all methods in the same hyperparameter space and improves the quality of odds-ratio DAAs (ORPO, ASFT). Under this setup, the ranking objective emerges as the primary determinant of alignment quality, whereas the particular scalar score (policy–reference ratio vs. odds ratio) is secondary. We corroborate this on instruction-following tasks and further confirm it on math-reasoning benchmarks across model scales. Evidence suggests that this stems from how these objectives interact with prompt-specific biases, supported both by strictly controlled experiments and by observations on real data. Our findings underscore the need for nuanced evaluations in DAA research to avoid oversimplified claims of superiority.