论文检索

Deep Learning · Attention Mechanisms

Xinghao Wang, Pengyu Wang, Xiaoran Liu, Fangxu Liu, Jason Chu, Kai Song, Xipeng Qiu

Block-sparse attention is promising for accelerating long-context LLM pre-filling, yet identifying relevant blocks efficiently remains a bottleneck. Existing methods typically employ coarse-grained attention as a proxy for block importance estimation, but often resort to expensive token-level searching or scoring, resulting in significant selection overhead. In this work, we trace the inaccuracy of standard coarse-grained attention via mean pooling to a theoretical root cause: the interaction between mean pooling and Rotary Positional Embeddings (RoPE). We prove that mean pooling acts as a low-pass filter that induces destructive interference in high-frequency dimensions, effectively creating a "blind spot" for local positional information (e.g., slash patterns). To address this, we introduce Prism, a training-free spectral-aware approach that decomposes block selection into high-frequency and low-frequency branches. By applying energy-based temperature calibration, Prism restores the attenuated positional signals directly from pooled representations, enabling block importance estimation using purely block-level operations, thereby improving efficiency. Extensive evaluations confirm that Prism maintains accuracy parity with full attention while delivering up to $\mathbf{5.1\times}$ speedup. Code available at https://anonymous.4open.science/r/prism_anonymous-7E36.

Zhexuan Wang, Xuebo Liu, Li Wang, Zifei Shan, Yutong Wang, Zhenxi Song, Min zhang

Large language model (LLM)-based Multi-agent systems (MAS) have shown promise in tackling complex collaborative tasks, where agents are typically orchestrated via role-specific prompts. While the quality of these prompts is pivotal, jointly optimizing them across interacting agents remains a non-trivial challenge, primarily due to the misalignment between local agent objectives and holistic system goals. To address this, we introduce MASPO, a novel framework designed to automatically and iteratively refine prompts across the entire system. A core innovation of MASPO is its joint evaluation mechanism, which assesses prompts not merely by their local validity, but by their capacity to facilitate downstream success for successor agents. This effectively bridges the gap between local interactions and global outcomes without relying on ground-truth labels. Furthermore, MASPO employs a data-driven evolutionary beam search to efficiently navigate the high-dimensional prompt space. Extensive empirical evaluations across 6 diverse tasks demonstrate that MASPO consistently outperforms state-of-the-art prompt optimization methods, achieving an average accuracy improvement of 2.9. We release our code at https://anonymous.4open.science/r/MASPO-8296.

Panqi Yang, Haodong Jing, Jiahao Chao, Tingyan Xiang, Li Lin, Yao Hu, Yang Luo, Yongqiang Ma

Unified visual tokenization faces a fundamental trade-off: optimizing for high-fidelity pixel reconstruction (spatial equivariance) inherently conflicts with semantic abstraction (conceptual invariance). We identify the root cause as Manifold Misalignment, where naive joint optimization leads to conflicting gradients that force a zero-sum game between these two objectives. In this paper, we propose MUSE, a framework that resolves this deadlock via Topological Orthogonality. Recognizing Structure as the orthogonal bridge, MUSE physically decouples the optimization subspaces within Transformers. We route structural gradients to refine attention topology and semantic gradients to update feature values, transforming destructive interference into Mutual Reinforcement. Extensive experiments demonstrate that MUSE breaks the trade-off, matching state-of-the-art generation (gFID 3.08) while notably outperforming its own teacher InternViT-300M in linear probing (85.2% vs. 82.5%), proving that structurally aligned reconstruction actively refines semantic perception.

Deep Learning · Algorithms

Shufeng Kong, Xiaochuan Zhang, Caihua Liu

Neural solvers for constraint satisfaction problems have achieved remarkable in-distribution accuracy, yet they suffer from a fundamental limitation where persistent constraint violations occur under distribution shifts even when the model reports high confidence. This position paper argues that when hard constraints exist and the cost of verification is relatively low, neural constraint reasoning must prioritize symbolic integration over pure learning. We justify our focus on Sudoku as a representative NP-complete testbed because it exhibits a sharp asymmetry between easy verification and hard solving; specifically, checking a candidate solution requires only polynomial time $O(n^{2})$ while finding a solution may require exponential search. Through a comprehensive survey of solving methods spanning deterministic algorithms, metaheuristic optimization, learning-based approaches, and language-conditioned reasoning, we demonstrate that neural-only methods without instance-level certification fail to achieve the provable correctness that symbolic and neuro-symbolic approaches provide. We advocate for a bidirectional integration where neural methods enhance symbolic solvers by learning heuristics and converting perceptions into symbols, while symbolic methods verify neural outputs to ensure their reliability. To operationalize this position, we propose a multi-agent certified reasoning framework that demonstrates how this integration can achieve both computational efficiency and provable correctness.

Peter Potaptchik, Lee Kit, Michael Albergo

We propose a simple, scalable algorithm based on stochastic interpolants for sampling from unnormalized densities and for fine-tuning generative models. The approach, Tilt Matching, arises from a dynamical equation relating the flow matching velocity to one targeting the same distribution tilted by a reward, implicitly solving a stochastic optimal control problem. The resulting velocity inherits the regularity of stochastic interpolant transports while minimizing an objective with strictly lower variance than flow matching itself. The update to the velocity field can be interpreted as the sum of all joint cumulants between the interpolant velocity and the reward, and to first order is their covariance. The method requires neither reward gradients nor backpropagation through trajectories of the flow or diffusion. We empirically demonstrate that the approach is efficient and highly scalable, providing state-of-the-art results on sampling under Lennard-Jones systems and competitive performance for fine-tuning Stable Diffusion, without requiring reward multipliers. The framework also applies directly to tilting few-step flow map models.

Deep Learning · Large Language Models

Kevin Xu, Issei Sato

Chain of thought (CoT) elicits reasoning in large language models by explicitly generating intermediate tokens. In contrast, latent thought reasoning operates directly in the continuous latent space, enabling computation beyond discrete linguistic representations. While both approaches exploit iterative computation, their comparative capabilities remain underexplored. In this work, we present a formal analysis showing that latent thought admits efficient parallel computation, in contrast to the inherently sequential nature of CoT. Conversely, CoT enables approximate counting through stochastic decoding. These separations suggest the tasks for which depth-driven recursion is more suitable, thereby offering practical guidance for choosing between reasoning paradigms.

John Morris, Chawin Sitawarin, Narine Kokhlikyan, Chuan Guo, Edward Suh, Alexander Rush, Kamalika Chaudhuri, Saeed Mahloujifar

We propose a new method for estimating how much a model knows about a datapoint and use it to measure the capacity of modern language models. Prior studies of language model memorization have struggled to disentangle memorization from generalization. We formally separate memorization into two components: unintended memorization, the information a model contains about a specific dataset, and generalization, the information a model contains about the true data-generation process. When we completely eliminate generalization, we can compute the total memorization, which provides an estimate of model capacity: our measurements estimate that GPT-style models have a capacity of approximately 3.6 bits per parameter. We train language models on datasets of increasing size and observe that models memorize until their capacity fills, at which point unintended memorization decreases as models begin to generalize. We train hundreds of transformer language models ranging from 500K to 1.5B parameters and produce a series of scaling laws relating model capacity and data size to membership inference.

Viet Hoang Tran, VINH KHANH BUI, Ngoc Tan Lai, Nam Nguyen, Tuan Dam, Tan Nguyen

Understanding gradient descent dynamics is key to explaining the success of over-parameterized models, where implicit bias manifests through conservation laws in gradient flow. While such laws are well understood for linear and ReLU networks, they remain largely unexplored for modern architectures. This work develops a unified framework to characterize conservation laws for contemporary models, including feedforward networks with GELU, SiLU, and SwiGLU activations, multihead attention with sinusoidal and rotary positional encodings, and Mixture-of-Experts architectures under diverse gating designs. Our theoretical findings are supported by experiments that validate the predicted invariants.

Matina Mahdizadeh Sani, Nima Jamali, Mohammad Jalali, Farzan Farnia

Pre-trained diffusion models have emerged as powerful generative priors for both unconditional and conditional sample generation, yet their outputs often deviate from the characteristics of user-specific target data. Such mismatches are especially problematic in domain adaptation tasks, where only a few reference examples are available and retraining the diffusion model is infeasible. Existing inference-time guidance methods can adjust sampling trajectories, but they typically optimize surrogate objectives such as classifier likelihoods rather than directly aligning with the target distribution. We propose MMD Guidance, a training-free mechanism that augments the reverse diffusion process with gradients of the Maximum Mean Discrepancy (MMD) between generated samples and a reference dataset. MMD provides reliable distributional estimates from limited data, exhibits low variance in practice, and is efficiently differentiable, which makes it particularly well-suited for the guidance task. Our framework naturally extends to prompt-aware adaptation in conditional generation models via product kernels. Also, it can be applied with computational efficiency in latent diffusion models (LDMs), since guidance is applied in the latent space of the LDM. Experiments on synthetic and real-world benchmarks demonstrate that MMD Guidance can achieve distributional alignment while preserving sample fidelity.

Semih Cayci

Ill-conditioned kernel matrices and loss landscapes can make first-order methods for training neural networks converge slowly. We establish non-asymptotic convergence bounds for the Gauss–Newton method in both under- and overparameterized regimes, showing it avoids these conditioning bottlenecks. In the underparameterized setting, Gauss–Newton gradient flow in parameter space induces a Riemannian gradient flow on a low-dimensional submanifold of function space. Using tools from Riemannian optimization, we show that, under an appropriate output scaling, the loss satisfies geodesic Polyak–Lojasiewicz and Lipschitz-smoothness conditions, implying geometric convergence to the optimal in-class predictor at an explicit rate independent of Gram-matrix conditioning. In the overparameterized setting, we identify adaptive, curvature-aware regularization schedules and prove fast geometric convergence to a global optimum for both Gauss–Newton gradient flow and discrete-time Gauss–Newton iterates, with rates independent of the minimum eigenvalue of the neural tangent kernel and, locally, independent of the strong-convexity modulus. Overall, Gauss–Newton can be provably faster in ill-conditioned regimes where first-order methods slow down.

Yueqi Song, Tianyue Ou, Yibo Kong, Zecheng Li, Graham Neubig, Xiang Yue

Current multimodal benchmarks often conflate reasoning with domain-specific knowledge, making it difficult to isolate and evaluate general reasoning abilities in non-expert settings. To address this, we introduce VisualPuzzles, a benchmark that targets visual reasoning while deliberately minimizing reliance on specialized knowledge. VisualPuzzles consists of diverse questions spanning five categories: algorithmic, analogical, deductive, inductive, and spatial reasoning. One major source of our questions is manually translated logical reasoning questions from the Chinese Civil Service Examination. Experiments show that VisualPuzzles requires significantly less intensive domain-specific knowledge and more complex reasoning compared to benchmarks like MMMU, enabling us to better evaluate genuine multimodal reasoning. Evaluations show that state-of-the-art multimodal large language models consistently lag behind human performance on VisualPuzzles, and that strong performance on knowledge-intensive benchmarks does not necessarily translate to success on reasoning-focused, knowledge-light tasks. Additionally, reasoning enhancements such as scaling up inference compute (with "thinking" modes) yield inconsistent gains across models and task types, and we observe no clear correlation between model size and performance. We also found that models exhibit different reasoning and answering patterns on VisualPuzzles compared to benchmarks with heavier emphasis on knowledge. VisualPuzzles offers a clearer lens through which to evaluate reasoning capabilities beyond factual recall and domain knowledge.

Deep Learning · Large Language Models

Raphael Bernas, Fanny Jourdan, Antonin Poché, Céline Hudelot

Since their introduction, Transformer architectures have dominated Natural Language Processing (NLP). However, recent research has highlighted an inherent anisotropy phenomenon in these models, presenting a significant challenge to their geometric interpretation. Previous theoretical studies on this phenomenon are rarely based on the underlying representation geometry. In this paper, we extend them by providing such theoretical arguments assessing the problematic nature of this phenomenon. Furthermore, to observe geometric internal model dynamics, we apply mechanistic interpretability (MI) techniques during the model's training checkpoints rather than post-hoc, as it is commonly done in the literature. By analyzing multiple models and their checkpoints -including EuroBERT, the Pythia suite, and SmolLM2- we investigate the structure of embedding representations and their correlation with the on manifold entropy of their underlying distribution.

General Machine Learning · Representation Learning

Weichen Li, Jiamin Song, Bogdan Stoica, Arav Dhoot, Gabriel Ryan, Shengyu Fu, Kexin Pei

Code transformation is a foundational capability in the software development process, where its effectiveness relies on constructing a high-quality code representation to characterize the input code semantics and guide the transformation. Existing approaches treat code transformation as an end-to-end learning task, leaving the construction of the representation needed for semantic reasoning implicit in model weights or relying on expensive compiler-level abstractions. We present SemRep, a framework that improves code transformation through generative code representation learning. Our key insight is to employ the semantics-preserving transformations as the intermediate representation, which will be used to train the model as a generative task, and also guide the subsequent instruction-specific code transformations. Across general code editing and CUDA kernel optimization, SemRep outperforms the strong closed-weight baselines by 6.9% and 43% in correctness, 13.9% in generalization and 6.7% in robustness. Combined with an evolutionary coding agent, SemRep finds optimizations that 685B larger-weight baselines fail to discover while achieving the same performance with 25% less inference compute.

Dongpeng Zhang, Ke Ma, Yangbangyan Jiang, Gaozheng Pei, Longtao Huang, Qianqian Xu, Qingming Huang

Adversarial images pose a severe security threat to multimodal large language models through prompt injection. Existing defenses largely lack a principled understanding of the underlying mechanisms and struggle to balance efficiency and fidelity. In this work, we show that successful adversarial attacks do not rely on the entire image uniformly but instead depend on a small subset of critical image tokens. Based on this insight, we propose a defense that first localizes these critical tokens via gradient analysis and then neutralizes them through masking. We show that attribution based on output probabilities fails when adversarial attacks preserve the predicted token. To overcome this limitation, we introduce the Hidden-State Gradient Norm score for adversarial behavior attribution and prove that its ranking is consistent with that of the full adversarial loss gradient, providing a theoretical guarantee for accurate localization. GTM requires only a single forward–backward pass to identify and zero out a small number of high-scoring tokens, effectively disrupting the adversarial attack path. Extensive experiments on prompt injection and multimodal jailbreak attacks demonstrate that our approach reduces attack success rates (ASR) to near zero while preserving model utility with negligible computational overhead.

Karish Grover, Hanqing Zeng, Yinglong Xia, Christos Faloutsos, Geoff Gordon

Does text have an intrinsic curvature? Language is increasingly modeled in curved geometries—hyperbolic spaces for hierarchy, mixed-curvature manifolds for compositional structure—yet a basic scientific question remains unresolved: what does curvature mean for text itself, in a way that is native to language rather than an artifact of the embedding space we choose? We argue that text does indeed have curvature, and show how to detect it, define it, and use it. To this end, we propose Texture, a text-native, word-level discrete curvature signal, and make three contributions. (a) Existence: We provide empirical and theoretical certificates that semantic inference in natural corpora is non-flat, i.e. language has inherent curvature. (b) Definition: We define Texture by reconciling left- and right-context beliefs around a masked word through a Schrödinger bridge, yielding a curvature field that is positive where context focuses meaning and negative where it fans out into competing continuations. (c) Utility: Texture is actionable: it serves as a general-purpose measurement and control primitive enabling geometry without geometric training; we instantiate it on two representative tasks, improving long-context inference through curvature-guided compression and retrieval-augmented generation through curvature-guided routing. Together, our results establish a text-native curvature paradigm, making curvature measurable and practically useful.

Deep Learning · Everything Else

Boris van Breugel, Yelysei Bondarenko, Paul Whatmough, Markus Nagel

Large language models (LLMs) require substantial compute, and thus energy, at inference time. While quantizing weights and activations is effective at improving efficiency, naive quantization of LLMs can significantly degrade performance due to large magnitude outliers. This paper describes FPTQuant, which introduces three novel, lightweight, and expressive function-preserving transforms (FPTs) to facilitate quantization of transformers: (1) a mergeable pre-RoPE transform for queries and keys, (2) a mergeable transform for values, (3) a cheap, dynamic scaling transform. By leveraging the equivariances and independencies inherent to canonical transformer operation, we designed these FPTs to maintain the model’s function while shaping the intermediate activation distributions to be more quantization friendly. FPTQuant requires no custom kernels and adds virtually no overhead during inference. The FPTs are trained both locally to reduce outliers, and end-to-end such that the outputs of the quantized and full-precision models match. FPTQuant enables static INT4 quantization with minimal overhead and shows SOTA speed-up of up to 3.9x over FP. Empirically, FPTQuant has an excellent accuracy-speed trade-off—it is performing on par or exceeding most prior work and only shows slightly lower accuracy compared to a method that is up to 29% slower.

Jinda Lu, Kexin Huang, Junkang Wu, Shuo Yang, Jinghan Li, Chiyu Ma, Shaohang Wei, Xiang Wang, Guoyin Wang, Jingren Zhou

Reinforcement Learning with Verifiable Rewards (RLVR) is a powerful paradigm for improving the reasoning capabilities of large language models (LLMs). However, existing RLVR methods typically rely on on-policy optimization from scratch, resulting in high sampling costs and inefficient utilization of accumulated experience. As model capabilities and policy behaviors evolve during training, recent attempts to reuse experience via fixed reasoning trajectories further suffer from policy mismatch. Motivated by these limitations, we argue that experience in RLVR should not be reused as fixed reasoning trajectories, but instead expressed in a policy-adaptive manner. In this work, we propose Experience-Augmented Policy Optimization (EAPO), which leverages a prior RL-optimized policy as an action-level experience prior and selectively injects experience at critical decision points during rollout. To ensure stable and unbiased learning from experience-augmented rollouts, EAPO further incorporates an adapted importance sampling scheme. Experiments on using Qwen-2.5-math 7b and Qwen-3-8B on five different benchmarks demonstrate that EAPO consistently improves reasoning performance over state-of-the-art RLVR methods.

Probabilistic Methods · Structure Learning

Stefan Ivanovic, Ge Liu, Mohammed El-Kebir

Many scientific problems require inferring unobserved mechanistic latent states from indirect observations. While classical approaches, including expectation-maximization, do not scale to combinatorially large spaces, deep learning approaches such as variational autoencoders typically form artificial latent states rather than reconstructing the mechanistic ground-truth states. Here, we introduce GReinSS, a policy learning framework that uses dynamically rescaled rewards to learn latent state distributions that maximize the observed data likelihood. We show that GReinSS accurately reconstructs simulated latent sets and latent graphs, outperforming alternative policy learning and generative modeling baselines. Additionally, GReinSS reconstructs isoforms from real short-read RNA sequencing data that better match orthogonal long-read sequencing detected isoforms than the standard RSEM algorithm. Overall, GReinSS is a principled and practically effective approach for generative modeling and inference of combinatorial latent states from indirect observations.

Jiarui Wang, Huiyu Duan, Ziheng Jia, Zicheng Zhang, Yu Zhao, Juntong Wang, Guangtao Zhai, Xiongkuo Min

Recent advancements in large multimodal models (LMMs) have driven substantial progress in both text-to-video (T2V) generation and video-to-text (V2T) interpretation tasks. However, current AI-generated videos (AIGVs) still exhibit limitations in terms of perceptual quality and text-video alignment. To this end, we present **AIGVE-60K**, a comprehensive dataset and benchmark for AI-Generated Video Evaluation, which features **(i)** comprehensive tasks, encompassing 3,050 extensive prompts across 20 fine-grained task dimensions, **(ii)** the largest human annotations, including 120K mean-opinion scores (MOSs) and 60K question-answering (QA) pairs annotated on 58,500 videos generated from 30 T2V models, and **(iii)** bidirectional benchmarking and evaluating for both T2V generation and V2T interpretation capabilities. Based on AIGVE-60K, we propose **LOVE**, a LMM-based metric for AIGV Evaluation from multiple dimensions including perceptual preference, text-video correspondence, and task-specific accuracy. Building upon LOVE, we further introduce **LOVE-Reward** to optimize T2V models through reinforcement learning, effectively enhancing both the perceptual quality and text-video correspondence of generated videos. Comprehensive experiments demonstrate that LOVE achieves state-of-the-art performance and generalizes effectively to various AIGV benchmarks. LOVE-Reward significantly improves video generation quality. These findings highlight the significance of the AIGVE-60K dataset and the effectiveness of our proposed methods. The database and codes will be available upon publication.

Social Aspects · Security

Ben Hagag, William Anderson, Christian Schroeder de Witt, Sarah Scheffler

Multi-agent systems (MAS), composed of networks of two or more autonomous AI agents, have become increasingly popular in production deployments, yet introduce security risks that do not arise in single-agent settings. Even if individual agents may exhibit robust security, architectural decisions governing their coordination can create attack surfaces that have not been systematically characterized. In this work, we present an empirical study of how MAS design decisions shape the tradeoff between task performance and attack resistance. Using a network of web-based agents and stage-wise evaluations that distinguish planning-stage refusal, execution-stage interception, and successful attack completion, we study architectural choices through controlled experiments. We identify three key design choices that influence MAS security: (i) agent roles, which determine how authority and responsibility are allocated; (ii) topology, which shapes how and when agents interact; and (iii) memory, which determines the context and state visibility accessible to each agent. Overall, our results show that security and performance in multi-agent systems are governed by architectural design choices, motivating the development of further evaluations which move beyond the security properties of a single agent.