论文检索

Deep Learning · Theory

Torben Berndt, Jan Stuehmer

Equivariance is a powerful inductive bias in neural networks, improving generalisation and physical consistency. Recently, however, non-equivariant models have regained attention, due to their better runtime performance and imperfect symmetries that might arise in real-world applications. This has motivated the development of approximately equivariant models that strike a middle ground between respecting symmetries and fitting the data distribution. Existing approaches in this field either rely on sampling from a group, incurring a high sample complexity, or explicitly parameterise a model as a sum of an equivariant and non-equivariant network. This work instead approaches approximate equivariance via a projection-based regulariser which leverages a layer-wise orthogonal decomposition of a network's layers into equivariant and non-equivariant components. In contrast to existing methods, this penalises non-equivariance at an operator level across the full group orbit, rather than point-wise as in sample-based approaches. We present a mathematical framework for computing the non-equivariance penalty exactly and efficiently in both the spatial and spectral domains. In our experiments, our method consistently outperforms prior approximate equivariance approaches in both model performance and efficiency, achieving substantial runtime gains over sample-based regularisers.

Deep Learning · Large Language Models

Zhiling Zhou, Tianhao Wang, Zhuoran Yang

Large language models built on autoregressive Transformers excel at next-token prediction, but it is unclear how their internal computations capture the latent hierarchical dependencies that often underlie language. We study this question in a controlled formal-language setting based on probabilistic context-free grammars (PCFGs), where sequences are generated by a latent hierarchical process. Empirically, standard autoregressive Transformers can be trained to accurately match the grammar-induced next-token distribution. Using probing analyses, we find that Transformer hidden states contain information used by classical parsing algorithms. Moreover, this information emerges through a layer-wise progression, revealing a local-to-global mechanism: early layers accumulate local patterns, while later layers aggregate them into a compact summary for next-token prediction. Complementing these empirical findings, we provide an explicit construction of Transformers that can parse binary PCFGs with depth \emph{logarithmic} in the grammar's sequence length. Surprisingly, trained Transformers in this setting exhibit prediction behavior and internal representations that closely mirror our construction. Together, our results offer a mechanistic account of how Transformers integrate hierarchical parsing with autoregressive generation.

Deep Learning · Large Language Models

Wanlong Fang, Tianle Zhang, Wen Tao, Alvin Chan

Understanding \emph{modality interaction} in multimodal large language models (MLLMs) remains a central challenge for reliable and interpretable deployment. We introduce Partial Information Decomposition (PID) as a unified, decision-level framework that separates \emph{unique}, \emph{redundant}, and \emph{synergistic} contributions of sensory and linguistic inputs, moving beyond representation alignment and outcome-based evaluation. Across vision–language benchmarks, PID reveals stable \emph{interaction regimes}: reasoning-oriented tasks consistently exhibit high cross-modal synergy, whereas knowledge-oriented tasks are dominated by language-unique information. These regimes generalize across architectures and scales and predict causal sensitivity to modality-level interventions. We extend this framework to tri-modal systems with Sensory PID, treating language as a control variable to decompose information gain from video and audio. Applied to omni-modal models, this analysis uncovers a persistent \emph{sensory synergy bottleneck}, where decisions remain dominated by visual information even on fusion-dependent tasks. Layer-wise analysis further show that sensory integration emerges late and is instruction-gated, following early visual saturation.

Deep Learning · Attention Mechanisms

Rupert Mitchell, Kristian Kersting

We present Multipole Semantic Attention (MuSe), an efficient approximation of softmax attention for long-context transformers. MuSe clusters queries and keys separately in their learned representation spaces, computing query-specific cluster summaries that capture how each query cluster attends to each key cluster. This is combined with retrieval of high-attention clusters for exact computation. Unlike prior work that clusters only keys, our separate query clustering provides a ~9× effective cluster count advantage, enabling high approximation quality at extreme sparsity. For causal attention, we introduce a block-sparse structure with causal accumulation of cluster summaries across spatial blocks, followed by two-level retrieval. At 64k context, MuSe achieves 64× sparsity in the far-field attention with <1% relative squared error and 2× speedup over CUDNN Flash Attention on isolated attention layers. We pretrain language models up to 1B parameters at 64k context, achieving 36% wallclock speedup with <1% loss degradation.

Social Aspects · Robustness

Huiyi Chen, Jiawei Peng, Dehai Min, Changchang Sun, Kaijie Chen, Yan Yan, Xu Yang, Lu Cheng

Evaluating the robustness of Large Vision-Language Models (LVLMs) is essential for their continued development and responsible deployment in real-world applications. However, existing robustness benchmarks typically focus on hallucination or misleading textual inputs, while largely overlooking the equally critical challenge posed by misleading visual inputs in assessing visual understanding. To fill this important gap, we introduce MVI-Bench, the first comprehensive benchmark specially designed for evaluating how Misleading Visual Inputs undermine the robustness of LVLMs. Grounded in fundamental visual primitives, the design of MVI-Bench centers on three hierarchical levels of misleading visual inputs: Visual Concept, Visual Attribute, and Visual Relationship. Using this taxonomy, we curate six representative categories and compile 1,248 expertly annotated VQA instances. To facilitate fine-grained robustness evaluation, we further introduce MVI-Sensitivity, a novel metric that characterizes LVLM robustness at a granular level. Empirical results across 18 state-of-the-art LVLMs uncover pronounced vulnerabilities to misleading visual inputs, and our in-depth analyses on MVI-Bench provide actionable insights that can guide the development of more reliable and robust LVLMs.

Reinforcement Learning · Deep RL

Meiyu Du, Yuqing Gao, Wei Wang

Proximal Policy Optimization (PPO) is widely used in continuous-control tasks, yet its performance is often highly sensitive to training dynamics when neural networks approximate the policy and value functions. This paper introduces SPPO, a drop-in augmentation that preserves PPO’s clipped objective and network architecture while stabilizing actor-critic geometry via three mechanisms: (i) a CKA-based constraint on critic representations, (ii) a no-flip regularizer on actor updates, and (iii) KDE-driven advantage shaping. Theoretical analysis shows that these mechanisms tighten bounds on one-step bootstrapping error, improve expected directional alignment of action updates, and ensure non-decreasing occupancy mass over high-novelty regions. Experiments on standard continuous-control benchmarks demonstrate consistent gains over PPO and recent PPO stabilization methods. Ablation studies further quantify the contribution and complementary effects of each component. Additional training-dynamics analyses indicate that SPPO reduces instability and oscillations in both actor and critic updates, improving training stability and final performance.

Deep Learning · Algorithms

Jindi Lv, Aoyu Li, Yuhao Zhou, Zheng Zhu, Xiaofeng Wang, Qing Ye, Yueqi Duan, Wentao Feng, Jiancheng Lv

Mamba demonstrates strong efficiency in modeling long visual sequences. However, when token reduction is applied to structurally enhanced Mamba variants, these models exhibit a severe performance collapse. We attribute this degradation to the spatially agnostic nature of existing reduction methods, which violate the two-dimensional structural premise required by the selective scanning mechanism. In this work, we propose STORM, a spatial-aware token reduction framework designed to maintain structural integrity throughout the compression process. STORM reformulates reduction into a structured operation on spatial units, enforcing localized constraints to maintain both grid topology and neighborhood coherence. As a plug-and-play module, STORM equips existing reduction pipelines with explicit spatial awareness without any training. Empirical results demonstrate that STORM achieves state-of-the-art pruning accuracy across diverse vision Mamba backbones under training-free settings. Notably, STORM delivers a substantial accuracy recovery on VMamba, outperforming prior methods by up to 63.3\% in top-1 accuracy. Meanwhile, STORM incurs only a 1.0\% accuracy drop on PlainMamba, achieving performance comparable to ViT.

Applications · Language, Speech and Dialog

Haohao Luo, Zexi Li, Yuexiang Xie, Wenhao Zhang, Yaliang Li, Ying Shen

Deep Research (DR) agents extend Large Language Models (LLMs) beyond parametric knowledge by autonomously retrieving and synthesizing evidence from large web corpora into long-form reports, enabling a long-horizon agentic paradigm. However, unlike real-time conversational assistants, DR is computationally expensive and time-consuming, creating an autonomy-interaction dilemma: high autonomy on ambiguous user queries often leads to prolonged execution with unsatisfactory outcomes. To address this, we propose IntentRL, a framework that trains proactive agents to clarify latent user intents before starting long-horizon research. To overcome the scarcity of open-ended research data, we introduce a scalable pipeline that expands a few seed samples into high-quality dialogue turns via a shallow-to-deep intent refinement graph. We further adopt a two-stage reinforcement learning (RL) strategy: Stage I applies RL on offline dialogues to efficiently learn general user-interaction behavior, while Stage II uses the trained agent and a user simulator for online rollouts to strengthen adaptation to diverse user feedback. Extensive experiments show that IntentRL significantly improves both intent hit rate and downstream task performance, outperforming the built-in clarify modules of closed-source DR agents and proactive LLM baselines.

General Machine Learning · Evaluation

Yuqi Guo, Siwei Wei, Yan Cai

Large Language Models (LLMs) have demonstrated sophisticated comprehension of sequential code, yet their capacity for reasoning about concurrent programs remains largely unquantified. We introduce DRPBench, a benchmark designed to evaluate the concurrent code comprehension of LLMs by measuring their data race prediction performance. To address the challenge of runtime non-determinism for evaluation on concurrent programs, we frame the evaluation as a fine-grained static prediction task using 1,003 programs from the SV-COMP suite, featuring 549 manually annotated data races with precise variable- and line-level granularity. Our evaluation of 15 state-of-the-art LLMs—spanning standard, reasoning, and agentic variants—reveals that DRPBench effectively differentiates concurrent code comprehension capabilities of LLMs. While the top-performing model (Gemini 3 with test-time reasoning) achieves an F1 score of 74.89%, most models struggle significantly (scoring less than 60%), with Llama 3 70B achieving only 8.80%. Beyond benchmarking, we characterize two primary failure modes: (1) shared-variable distraction, where multiple variable appearances degrade comprehension accuracy, and (2) synchronization-logic myopia, the inability to interpret non-standard synchronization implementations. Our findings provide a diagnostic roadmap for enhancing concurrent code comprehension of LLMs in future development.

Deep Learning · Other Representation Learning

Shuhao Fu, Esther Goldberg, Ying Nian Wu, Hongjing Lu

Large Multimodal Models (LMMs) demonstrate impressive in-context learning abilities from few multimodal demonstrations, yet the internal mechanisms supporting such task learning remain opaque. Building on prior work of Large Language Models, we show that a small subset of attention heads in Large Multimodal Models is responsible for transmitting representations of spatial relations. The activations of these attention heads, termed function vectors, can be extracted and manipulated to alter an LMM’s performance on relational tasks. First, using synthetic and real image datasets, we apply causal mediation analysis to identify attention heads that strongly influence relational predictions, and extract multimodal function vectors that improve zero-shot accuracy at inference time. We further demonstrate that these multimodal function vectors can be fine-tuned with a modest amount of training data, while keeping LMM parameters frozen, to significantly outperform in-context learning baselines. Finally, we show that relation-specific function vectors can be linearly combined to solve analogy problems involving novel and untrained spatial relations, highlighting the strong generalization ability of this approach. Through experiments on two LMMs, including OpenFlamingo and Qwen3-VL, our results show that these models encode spatial relational knowledge within localized internal structures, which can be systematically extracted and optimized, thereby advancing our understanding of model modularity and enhancing control over relational reasoning in LMMs.

Social Aspects · Alignment

Yulong Lin, Pablo Bernabeu-Pérez, Benjamin Arnav, Lennie Wells, Mary Phuong

As language models grow more capable, accurate capability evaluation becomes essential for safety decisions. If models can deliberately underperform on dangerous capability evaluations---a behavior known as \emph{sandbagging}---they may evade safety measures designed for their true capability level. We introduce Cross-Context Consistency (C³), a general framework for unsupervised black-box sandbagging detection that exploits a fundamental asymmetry: when a model truly lacks capability, its confusion manifests consistently across paraphrased questions, but when a capable model feigns incompetence, its strategic choices about \emph{how} to appear weak create detectable inconsistencies. The framework is agnostic to the specific consistency metric and aggregation method; we present a simple instantiation using embedding cosine distance and mean aggregation that requires no training data or model-specific adaptation. We evaluate C³ across prompted and fine-tuned sandbagging scenarios, across instructed and more naturalistic setups, maintaining a classification signal where other black-box methods fail. Our findings show the limitations of existing sandbagging detection methods, and reveal the efficacy of consistency-checking as a detection mechanism for dangerous capabilities.

Deep Learning · Large Language Models

Shumin Wang, Yuexiang Xie, Wenhao Zhang, Yuchang Sun, Yanxi Chen, Yaliang Li, Yanyong Zhang

Entropy serves as a critical metric for measuring the diversity of outputs generated by large language models (LLMs), providing valuable insights into their exploration capabilities. While recent studies increasingly focus on monitoring and adjusting entropy to better balance exploration and exploitation in reinforcement fine-tuning (RFT), a principled understanding of entropy dynamics during this process is yet to be thoroughly investigated. In this paper, we establish a theoretical framework for analyzing the entropy dynamics during the RFT process, which begins with a discriminant expression that quantifies entropy change under a single logit update. This foundation enables the derivation of a first-order expression for entropy change, which can be further extended to the update formula of Group Relative Policy Optimization (GRPO). The corollaries and insights drawn from the theoretical analysis inspire the design of entropy control methods, and also offer a unified lens for interpreting various entropy-based methods in existing studies. We provide empirical evidence to support the main conclusions of our analysis and demonstrate the effectiveness of the derived entropy-discriminator clipping methods. This study yields novel insights into RFT training dynamics, providing theoretical support and practical strategies for optimizing the exploration-exploitation balance during LLM fine-tuning.

Social Aspects · Safety

Jaylen Jones, Zhehao Zhang, Yuting Ning, Eric Fosler-Lussier, Pierre-Luc St-Charles, Yoshua Bengio, Dawn Song, Yu Su, Huan Sun

Although computer-use agents (CUAs) hold significant potential to automate increasingly complex OS workflows, they can demonstrate unsafe unintended behaviors that deviate from expected outcomes even under benign input contexts. However, exploration of this risk remains largely anecdotal, lacking concrete characterization and automated methods to proactively surface long-tail unintended behaviors under realistic CUA scenarios. To fill this gap, we introduce the first conceptual and methodological framework for unintended CUA behaviors, by defining their key characteristics, automatically eliciting them, and analyzing how they arise from benign inputs. We propose AutoElicit: an agentic framework that iteratively perturbs benign instructions using CUA execution feedback, and elicits severe harms while keeping perturbations realistic and benign. Using AutoElicit, we surface hundreds of harmful unintended behaviors from state-of-the-art CUAs such as Claude 4.5 Haiku and Opus. We further evaluate the transferability of human-verified successful perturbations, identifying persistent susceptibility to unintended behaviors across various other frontier CUAs. This work establishes a foundation for systematically analyzing unintended behaviors in realistic computer-use settings.

Deep Learning · Foundation Models

Jana Zeller, Thaddäus Wiedemer, Fanfei Li, Thomas Klein, Prasanna Mayilvahanan, Matthias Bethge, Felix Wichmann, Ryan Cotterell, Wieland Brendel

Frontier models are transitioning from _multimodal large language models_ (MLLMs) that merely ingest visual information to _unified multimodal models_ (UMMs) capable of native interleaved generation. This shift has sparked interest in using intermediate visualizations as a reasoning aid, akin to human _mental imagery_. Central to this idea is the ability to form, maintain, and manipulate visual representations in a goal-oriented manner. To evaluate and probe this capability, we develop MentisOculi, a procedural, stratified suite of multi-step reasoning problems amenable to visual solution, tuned to challenge frontier models. Evaluating visual strategies ranging from latent tokens to explicit generated imagery, we find they generally fail to improve performance. Analysis of UMMs specifically exposes a critical limitation: While they possess the textual reasoning capacity to solve a task and can sometimes generate correct visuals, they suffer from compounding generation errors and fail to leverage even ground-truth visualizations. Our findings suggest that despite their inherent appeal, _visual thoughts do not yet benefit model reasoning_. MentisOculi establishes the necessary foundation to analyze and close this gap across diverse model families.

Social Aspects · Safety

Arkadiy Saakyan, Charvi Rastogi, Lora Aroyo

Safe global deployment of AI models requires alignment with pluralistic human values, yet in existing safety evaluation datasets the rater pools remain largely homogeneous along geo-cultural dimensions. Through a meta-analysis of existing safety datasets, we observe that the vast majority does not include any geo-cultural information, and the ones that do, lack a robust approach to collect and understand cultural differences in safety ratings. Using the Inglehart-Welzel dimensions of cross-cultural variation, we demonstrate via hierarchical linear modeling that geo-cultural values predict safety ratings significantly better than demographic factors alone ($p<0.05$ in $6$ datasets). Further, our analysis shows that several safety datasets contain at least 10\% of culturally-sensitive items, where lack of cultural representation in the rater pool would lead to a false negative in safety classification. Finally, we provide empirical evidence that fine-tuned LLMs can identify culturally sensitive items but are not reliable at emulating judgments of raters from diverse cultural backgrounds, underscoring the critical need for continuous geo-culturally stratified (pluralistic) safety evaluations.

Social Aspects · Safety

pengcheng li, Jie Zhang, Tianwei Zhang, Han Qiu, Zhang kejun, Weiming Zhang, Nenghai Yu, Wenbo Zhou

Safety alignment in large language models is typically evaluated under isolated queries, yet real-world use is inherently multi-turn. Although multi-turn jailbreaks are empirically effective, the structure of conversational safety failure remains insufficiently understood. In this work, we study safety failures from a state-space perspective and show that many multi-turn failures arise from structured contextual state evolution rather than isolated prompt vulnerabilities. We introduce STAR, a state-oriented diagnostic framework that treats dialogue history as a state transition operator and enables controlled analysis of safety behavior along interaction trajectories. Rather than optimizing attack strength, STAR provides a principled probe of how aligned models traverse the safety boundary under autoregressive conditioning. Across multiple frontier language models, we find that systems which appear robust under static evaluation can undergo rapid and reproducible safety collapse under structured multi-turn interaction. Mechanistic analysis reveals monotonic drift away from refusal-related representations and abrupt phase transitions induced by role-conditioned context. Together, these findings motivate viewing language model safety as a dynamic, state-dependent process defined over conversational trajectories.

Social Aspects · Safety

William Overman, Mohsen Bayati

Agentic AI systems capable of autonomous planning and extended environmental interaction pose a fundamental control problem: how can humans maintain meaningful oversight of systems that may exceed human capabilities? While scalable oversight is widely studied, existing approaches often rely on complex assumptions, remain largely heuristic, or lack practical methods for sequential settings with statistical guarantees. We introduce Calibrated Collective Oversight (CCO), which aggregates diverse auxiliary scoring functions into a penalty that measures deviation from a conservative baseline. Inspired by Attainable Utility Preservation, CCO enables collective conservatism: when multiple oversight signals register concern, the agent defers. CCO calibrates this conservatism online using Conformal Decision Theory, ensuring that undesirable outcomes remain below a user-specified target $\alpha$ with finite-time bounds and no distributional assumptions. Experiments on SWE-bench demonstrate that weaker overseers successfully constrain an adversarially misaligned stronger agent. Similarly, on MACHIAVELLI, CCO achieves substantial reductions in ethical violations while preserving reward. In both settings, empirical violation rates closely match the specified targets. Our work demonstrates that combining penalty-based conservatism with online calibration yields practical oversight with statistical guarantees suited for agentic deployment.

Applications · Language, Speech and Dialog

Sang-Hoon Lee, Ha-Yeong Choi

Representation alignment (REPA) has been investigated to accelerate diffusion training, but we observe that regularizing intermediate representations in diffusion Transformers (DiT) may implicitly entangle latents and limit generative capacity. To address this issue, we propose ReGen, a hierarchical multi-prompt representation generation framework that jointly estimates multiple vector fields for both representations and data within a single diffusion model. We further introduce generalized flow matching (GFM) to improve the generalization of conditional flow matching (CFM). We validate ReGen on single-stage waveform diffusion models including neural audio codec and Wave-VAE. ReGen significantly improves waveform generation quality from highly compressed latent representations at 12.5 Hz. We also present ReGenVoice, a latent diffusion model (LDM)-based text-to-speech model that achieves strong speech intelligibility (WER) and speaker similarity (SIM) with a small dataset. Moreover, operating the LDM at 6.25 Hz with rich semantic and acoustic latent representation enables efficient training and sampling, requiring only 1 day of training on 4 GPUs and fast inference with an RTF of 0.08.

Social Aspects · Safety

Maya Okawa

Multi-agent LLM debates achieve strong performance on decision-making tasks as well as problem-solving benchmarks, yet their safety and fairness risks remain poorly understood. Notably, interaction can amplify the biases of single LLMs, raising concerns for real-world deployment. We identify the emergence of collective (often biased) norms in multi-agent LLM debates and show that noise (e.g., LLM sampling temperature) is a key driver. To explain this, we propose an analytical framework drawing on physics-inspired theoretical models of social dynamics. We predict a phase transition to collective bias when conformity surpasses a critical threshold given the LLMs' initial bias and debate noise. We test the theoretical predictions through controlled experiments and observe a finite-size crossover consistent with an underlying phase transition. We further find that agent heterogeneity suppresses emergence by smoothing (rounding) this transition. Finally, we show that these insights generalize to realistic decision-making tasks, including investment decisions and LLM-as-a-judge evaluation.

Deep Learning · Generative Models and Autoencoders

Abdelhakim Ziani, Andras Horvath, Paolo Ballarini

Heavy-tailed distributions are ubiquitous in real-world data, where rare but extreme events dominate risk and variability. However, standard Variational Autoencoders (VAEs) employ simple decoder distributions (e.g., Gaussian) that fail to capture heavy-tailed behavior, while existing heavy-tail-aware extensions remain restricted to predefined parametric families whose tail behavior is fixed a priori. We propose the *Phase-Type Variational Autoencoder* (PH-VAE), whose decoder distribution is a latent-conditioned Phase-Type (PH) distribution—defined as the absorption time of a continuous-time Markov chain (CTMC). This formulation composes multiple exponential time scales, yielding a flexible, analytically tractable decoder that adapts its tail behavior directly from the observed data. Experiments on synthetic and real-world benchmarks demonstrate that PH-VAE accurately recovers diverse heavy-tailed distributions, significantly outperforming Gaussian, Student-t, and extreme-value-based VAE decoders in modeling tail behavior and extreme quantiles. In multivariate settings, PH-VAE captures realistic cross-dimensional tail dependence through its shared latent representation. To our knowledge, this is the first work to integrate Phase-Type distributions into deep generative modeling, bridging applied probability and representation learning.