论文检索

Deep Learning · Algorithms

ShiLin Xiao, Tianxiang Xu, Canran Xiao, Weihao Luo, Liwei Hou, Chuangxin Zhao

Large language models are increasingly deployed in settings where newtasks arrive continuously, yet existing parameter-efficient finetuning (PEFT) methods either bloat linearly with the task horizon or sacrifice deep adaptation, leaving catastrophic forgetting unresolved. We aim to achieve memory-constant, on-the-fly adaptation for a frozen LLM facing an unbounded stream of tasks. To this end we propose Meta-Unified Contrastive Finetuning(Meta-UCF), which encodes each task into a lightweight layer-normalised mean embedding and feeds it to a single hypernetwork that instantly generates rank-r LoRA updates for every transformer layer; a meta-contrastive coupled with orthogonality objective further steers task embeddings into near-orthogonal directions, preserving past knowledge without inner-loop gradients. On four benchmark streams—Std-CL 5, Seq-GLUE 7, Long-CL 15 and TRACE-8—Meta-UCF raises average accuracy by up to 2.2 pp and cuts forgetting by 13% relative to the strongest LoRA baseline, while using the parameters of a single adapter. By decoupling continual learning from parameter growth, Meta-UCF provides a practical path toward scalable, low-resource lifelong language modelling.

Yihong Luo, Tianyang Hu, Jing Tang

While reinforcement learning methods such as Group Relative Preference Optimization (GRPO) have significantly enhanced Large Language Models, adapting them to diffusion models remains challenging. In particular, GRPO demands a stochastic policy, yet the most cost‑effective diffusion samplers are based on deterministic ODEs. Recent work addresses this issue by using inefficient SDE-based samplers to induce stochasticity, but this reliance on model-agnostic Gaussian noise leads to slow convergence. To resolve this conflict, we propose Direct Group Preference Optimization (DGPO), a new online RL algorithm that dispenses with the policy-gradient framework entirely. DGPO learns directly from group-level preferences, which utilize relative information of samples within groups. This design eliminates the need for inefficient stochastic policies, unlocking the use of efficient deterministic ODE samplers and faster training. Extensive results show that DGPO trains around 20 times faster than existing state-of-the-art methods and achieves superior performance on both in-domain and out-of-domain reward metrics.

Deep Learning · Graph Neural Networks

André Ribeiro, Ana Luiza Tenorio, Juan Belieni, Amauri Souza, Diego Mesquita

Sheaf neural networks (SNNs) leverage cellular sheaves to induce flexible diffusion processes on graphs, generalizing the diffusion mechanism of classical graph neural networks. While SNNs have been shown to cope well with heterophilic tasks and alleviate oversmoothing, we show that there is further room for improving sheaf diffusion. More specifically, we argue that SNNs do not allow nodes to independently choose how they cooperate with their neighbors, i.e., whether they convey and/or gather information to/from their neighbors. To address this issue, we first introduce the notion of cellular sheaves over directed graphs and characterize their in- and out-degree Laplacians. We then leverage our construction to propose Cooperative Sheaf Neural Network (CSNN). Additionally, we formally characterize its receptive field and prove that it allows nodes to selectively attend (listen) to arbitrarily far nodes while ignoring all others in their path, which is key to alleviating oversquashing. Our results on synthetic data empirically substantiate our claims, showing that CSNN can handle long-range interactions while avoiding oversquashing. We also show that CSNN performs strongly in heterophilic node classification and long-range graph classification benchmarks.

Computer Vision · 3D Rendering & Reconstruction

Canran Xiao, Liaoyuan Fan, Yanbin Li, Jing Tang, Peilai Yu

3D Gaussian Splatting (3DGS) has recently emerged as a powerful approach for novel view synthesis by reconstructing scenes as sets of Gaussian ellipsoids. Despite its success in scenarios with dense input images, 3DGS faces critical challenges in sparse view settings, often resulting in geometric inaccuracies, inconsistencies across views, and degraded rendering quality. In this paper, we uncover and address two key implicit biases of 3DGS reconstruction algorithm in sparse-view: (1) the model has a stronger demand for supervision signal toward regions of high curvature, and (2) the model is sensitive to the smoothness of the trajectory of the input views. To tackle these issues, we propose a novel framework that optimizes camera trajectories to maximize curvature coverage while enforcing smooth motion, and we further enhance the informativeness of data through a synthetic view generation process. Extensive experiments on Mip-NeRF 360, DTU, Blender, Tanks & Temples, and LLFF datasets show that our method substantially outperforms state-of-the-art solutions in sparse-view scenarios, both in rendering quality and geometric fidelity. Beyond these empirical gains, our investigation uncovers the subtle ways in which data representation and trajectory planning interact to shape 3DGS performance, offering deeper theoretical insights into the algorithm’s inherent biases.

Deep Learning · Everything Else

Jiayu Zhang, Chuangxin Zhao, Canran Xiao, Ruibo Duan, Wenyi Mo, Haoyu Gao, Wenshuo Wang

When deployed on non-stationary data streams, foundation vision-language models require continual updates without access to past data. However, naive fine-tuning undermines their zero-shot recognition capabilities and prompt robustness. We seek a replay-free principle that preserves pre-trained cross-modal generalization under domain/prompt shifts. We introduce Prompt-Invariant CCA Certificates (Pi-CCA), a geometry-first approach that summarizes image--text alignment with a compact certificate capturing the top-k canonical spectrum and subspace. During adaptation, we match this summary using only mini-batch statistics and induce prompt robustness via averaging over perturbations. Across MTIL, X-TAIL, VLCL, and ConStruct-VL, Pi-CCA achieves state-of-the-art performance among replay-free methods. By optimizing alignment invariants rather than proxy signals, Pi-CCA provides a simple, generator-free, constant-memory path to continual adaptation with strong zero-shot retention and resilience to prompt/style shifts.

Computer Vision · Vision Models & Multimodal

Qingyu Shi, Jinbin Bai, Zhuoran Zhao, Wenhao Chai, Kaidong Yu, Jianzong Wu, Shuangyong Song, Yunhai Tong, Xiangtai Li, Xuelong Li 等

Unified generation models aim to handle diverse tasks across modalities—such as text-to-image generation and image-to-text generation—within a single architecture and decoding paradigm. Autoregressive unified models suffer from slow inference due to sequential decoding, and non-autoregressive unified models suffer from weak generalization due to limited pretrained backbones. We introduce Muddit, a unified discrete diffusion transformer that enables fast and parallel generation across both text and image modalities. Unlike prior unified diffusion models trained from scratch, Muddit integrates strong visual priors from a pretrained text-to-image backbone with a lightweight text decoder, enabling flexible and high-quality multimodal generation under a unified architecture. Empirical results show that Muddit achieves competitive or superior performance compared to models in both quality and efficiency. This work also highlights the potential of purely discrete diffusion, when equipped with strong visual priors, as a scalable and effective backbone for unified generation.

General Machine Learning · Transfer, Multitask and Meta-learning

Philipp Mondorf, Shijia Zhou, Monica Riedler, Barbara Plank

Systematic generalization refers to the capacity to understand and generate novel combinations from known components. Despite recent progress by large language models (LLMs) across various domains, these models often fail to extend their knowledge to novel compositional scenarios, revealing notable limitations in systematic generalization. There has been an ongoing debate about whether neural networks possess the capacity for systematic generalization, with recent studies suggesting that meta-learning approaches designed for compositionality can significantly enhance this ability. However, these insights have largely been confined to linguistic problems, leaving their applicability to other tasks an open question. In this study, we extend meta-learning for compositionality to the domain of abstract spatial reasoning. To this end, we introduce $\textit{Compositional-ARC}\textemdash{}$a dataset designed to evaluate the capacity of models to systematically generalize from known geometric transformations (e.g., translation, rotation) of abstract two-dimensional objects to novel combinations of these transformations (e.g., translation+rotation). Our results show that a small transformer-based encoder-decoder model, trained via meta-learning for compositionality, can systematically generalize to previously unseen transformation compositions. Notably, despite having only 5.7M parameters, this model significantly outperforms state-of-the-art LLMs$\textemdash{}$including o3-mini, GPT-4o, and Gemini 2.0 Flash, which fail to exhibit similar systematic behavior$\textemdash{}$and performs on par with the winning model of the ARC prize 2024, an 8B-parameter LLM trained via test-time training. Our findings highlight the effectiveness of meta-learning in promoting systematicity beyond linguistic tasks, suggesting a promising direction toward more robust and generalizable models.

Deep Learning · Everything Else

Anik Pramanik, Murat Kantarcioglu, Vincent Oria, Shantanu Sharma

Federated Learning (FL) enables a group of clients to collaboratively train a model without sharing individual data, but its performance drops when client data are heterogeneous. Clustered FL tackles this by grouping similar clients. However, existing clustered FL approaches rely solely on either data similarity or gradient similarity; however, this results in an incomplete assessment of client similarities. Prior clustered FL approaches also restrict knowledge and representation sharing to clients within the same cluster. This prevents cluster models from benefiting from the diverse client population across clusters. To address these limitations, FEDDAG introduces a clustered FL framework, FEDDAG, that employs a weighted, class-wise similarity metric that integrates both data and gradient information, providing a more holistic measure of similarity during clustering. In addition, FEDDAG adopts a dual-encoder architecture for cluster models, comprising a primary encoder trained on its own clients' data and a secondary encoder refined using gradients from complementary clusters. This enables cross-cluster feature transfer while preserving cluster-specific specialization. Experiments on diverse benchmarks and data heterogeneity settings show that FEDDAG consistently outperforms state-of-the-art clustered FL baselines in accuracy.

Computer Vision · 3D Rendering & Reconstruction

Lucas Maximilian Martini, Alex Lappe, Anna Bognár, Rufin Vogels, Martin Giese

The recognition of dynamic and social behavior in animals is fundamental for advancing several areas of the life sciences, including ethology, ecology, medicine and neuroscience. Recent progress in deep learning has enabled an automated recognition of such behavior from video data. However, an accurate reconstruction of the three-dimensional (3D) pose and shape has not been integrated into this process. Especially for non-human primates, the animals phylogenetically closest to humans, mesh-based tracking efforts lag behind those for other species, leaving pose descriptions restricted to sparse keypoints that are unable to fully capture the richness of action dynamics. To address this gap, we introduce the $\textbf{Big Ma}$ca$\textbf{Q}$ue 3D Motion and Animation Dataset ($\texttt{BigMaQ}$), a large-scale dataset comprising more than 750 scenes of interacting rhesus macaques with detailed 3D pose descriptions of skeletal joint rotations. Recordings were obtained from 16 calibrated cameras and paired with action labels derived from a curated ethogram. Extending previous surface-based animal tracking methods, we construct subject-specific textured avatars by adapting a high-quality macaque template mesh to individual monkeys. This allows us to provide pose descriptions that are more accurate than previous state-of-the-art surface-based animal tracking methods. From the original dataset, we derive BigMaQ500, an action recognition benchmark that links surface-based pose vectors to single frames across multiple individual monkeys. By pairing features extracted from established image and video encoders with and without our pose descriptors, we demonstrate substantial improvements in mean average precision (mAP) when pose information is included. With these contributions, $\texttt{BigMaQ}$ establishes the first dataset that both integrates dynamic 3D pose-shape representations into the learning task of animal action recognition and provides a rich resource to advance the study of visual appearance, posture, and social interaction in non-human primates. The code and data are publicly available at [https://martinivis.github.io/BigMaQ/](https://martinivis.github.io/BigMaQ/).

Applications · Language, Speech and Dialog

Yen-Ju Lu, Yashesh Gaur, Wei Zhou, Benjamin Muller, Jesus Villalba, Najim Dehak, Luke Zettlemoyer, Gargi Ghosh, Mike Lewis, Srini Iyer 等

Auto-regressive speech–text models pre-trained on interleaved text tokens and discretized speech tokens demonstrate strong speech understanding and generation, yet remain substantially less compute-efficient than text LLMs, partly due to the much longer sequences of speech tokens relative to text. This modality imbalance disproportionately allocates pre-training and inference compute to speech, potentially hindering effective cross-modal alignment and slowing performance scaling by orders of magnitude. We introduce the Latent Speech-Text Transformer (LST), which aggregates speech tokens into latent speech patches that serve as higher-level autoregressive units. This design aligns the sequence-modeling granularity between speech and text while improving computational efficiency. The resulting patches can align with textual units to facilitate cross-modal knowledge transfer and compactly capture recurring acoustic patterns such as silence. Across story-completion benchmarks under both compute-controlled and data-controlled settings, LST consistently improves speech accuracy while also improving text performance, achieving up to +6.5% absolute gain on speech HellaSwag in compute-controlled training (+5.3% in data-controlled training). Under compute-controlled scaling from 420M to 1.8B parameters in a near compute-optimal regime, gains grow with scale, and improvements persist up to 7B parameters under fixed-token budgets. These benefits extend to downstream tasks: LST stabilizes ASR adaptation and reduces the effective autoregressive sequence length during ASR and TTS inference, lowering computational cost without degrading reconstruction quality. The Code is available at https://github.com/facebookresearch/lst.

General Machine Learning · Transfer, Multitask and Meta-learning

Qingyuan Liu, Jia-Chen Gu, Yunzhi Yao, Hong Wang, Nanyun (Violet) Peng

Large language models (LLMs) require constant updates to remain aligned with evolving real-world knowledge. Model editing offers a lightweight alternative to retraining, but sequential editing that updates the LLM knowledge through multiple successive edits often destabilizes representations and induces catastrophic forgetting. In this work, we seek to better understand and mitigate performance degradation caused by sequential editing. We hypothesize that hyperspherical uniformity, a property that maintains uniform distribution of neuron weights on a hypersphere, helps the model remain stable, retain prior knowledge, while still accommodate new updates. We use Hyperspherical Energy (HE) to quantify neuron uniformity during editing, and examine its correlation with editing performance. Empirical studies across widely used editing methods reveals a strong correlation between HE dynamics and editing performance, with editing failures consistently coinciding with uncontrolled HE fluctuations. We further theoretically prove that HE dynamics impose a lower bound on the degradation of pretrained knowledge, highlighting why HE stability is crucial for knowledge retention. Motivated by these insights, we propose SPHERE (Sparse Projection for Hyperspherical Energy-Regularized Editing), an HE-driven regularization strategy that stabilizes neuron weight distributions, ultimately preserving prior knowledge while enabling reliable sequential updates. Specifically, SPHERE identifies a sparse space complementary to the principal hyperspherical directions of the pretrained weight matrices and projects new knowledge onto it, attenuating perturbations on the principal directions. Extensive experiments on LLaMA3 (8B) and Qwen2.5 (7B) show that SPHERE outperforms the best baseline in editing capability by an average of 16.41%, while most faithfully preserving general model performance, thereby offering a principled path toward reliable large-scale knowledge editing.

Theory · Learning Theory

Chenyang Zhang, Qingyue Zhao, Quanquan Gu, Yuan Cao

Transformers have achieved great success across a wide range of applications, yet the theoretical foundations underlying their success remain largely unexplored. To demystify the strong capacities of transformers applied to versatile scenarios and tasks, we theoretically investigate utilizing transformers as students to learn from a class of teacher models. Specifically, the teacher models covered in our analysis include convolution layers with average pooling, graph convolution layers, and various classic statistical learning models, including a variant of sparse token selection models [Sanford et al., 2023, Wang et al., 2024] and group-sparse linear predictors [Zhang et al., 2025]. When learning from this class of teacher models, we prove that one-layer transformers with simplified "position-only'' attention can successfully recover all parameter blocks of the teacher models, thus achieving the optimal population loss. Building upon the efficient mimicry of trained transformers towards teacher models, we further demonstrate that they can generalize well to a broad class of out-of-distribution data under mild assumptions. The key in our analysis is to identify a fundamental bilinear structure shared by various learning tasks, which enables us to establish unified learning guarantees for these tasks when treating them as teachers for transformers.

Applications · Everything Else

JUNYONG AHN, Hyeongrok Han, Bong Gyun Kang, Jisoo Mok, Byunghan Lee, Sungroh Yoon

Developing a mobile system capable of generating responses based on stored user data is a crucial challenge. Since user data is stored in the form of Knowledge Graphs, the field of knowledge graph question answering (KGQA) presents a promising avenue towards addressing this problem. However, existing KGQA systems face two critical limitations that preclude their on-device deployment: resource constraints and the inability to handle data accumulation. Therefore, we propose MobileKGQA, the first on-device KGQA system capable of adapting to evolving databases with minimal resource demands. MobileKGQA significantly reduces computational overhead through embedding hashing. Moreover, it successfully adapts to evolving databases under resource constraints through a novel annotation generation method. Its mobile applicability is validated on the NVIDIA Jetson Orin Nano edge-device platform, achieving 20.3% higher performance while using only 30.4% of the energy consumed by the SOTA (state-of-the-art). On standard KGQA benchmarks, using just 7.2% of the computation and 9% of the parameters, MobileKGQA demonstrates performance that is empirically indistinguishable from the SOTA and outperforms baselines under distribution shift scenarios.

Applications · Chemistry and Drug Discovery

Nima Shoghi, Yuxuan Liu, Yuning Shen, Rob Brekelmans, Pan Li, Quanquan Gu

Molecular dynamics (MD) simulations remain the gold standard for studying protein dynamics, but their computational cost limits access to biologically relevant timescales. Recent generative models have shown promise in accelerating simulations, yet they struggle with long-horizon generation due to architectural constraints, error accumulation, and inadequate modeling of spatiotemporal dynamics. We present STAR-MD (Spatio-Temporal Autoregressive Rollout for Molecular Dynamics), a scalable SE(3)-equivariant diffusion model that generates physically plausible protein trajectories over microsecond timescales. Our key innovation is a causal diffusion transformer with joint spatiotemporal attention that efficiently captures complex space-time dependencies while avoiding the memory bottlenecks of existing methods. On the standard ATLAS benchmark, STAR-MD achieves state-of-the-art performance across all metrics--substantially improving conformational coverage, structural validity, and dynamic fidelity compared to previous methods. STAR-MD successfully extrapolates to generate stable microsecond-scale trajectories where baseline methods fail catastrophically, maintaining high structural quality throughout the extended rollout. Our comprehensive evaluation reveals severe limitations in current models for long-horizon generation, while demonstrating that STAR-MD's joint spatiotemporal modeling enables robust dynamics simulation at biologically relevant timescales, paving the way for accelerated exploration of protein function. Project page: https://bytedance-seed.github.io/ConfRover/starmd

Optimization · Large Scale, Parallel and Distributed

Yan Li, Zhenyu Zhang, Zhengang Wang, Pengfei chen, Pengfei Zheng

Prevailing LLM (Large Language Model) serving engines employ expert parallelism (EP) to implement multi-device inference of massive Mixture-of-Experts (MoE) models. However, the efficiency of expert parallel inference is largely bounded by inter-device communication, as EP embraces expensive all-to-all collectives to route tokens to the remote experts if not collocating on the same GPU/NPU device. Nevertheless, state-of-the-art schemes treat expert device-placement and request (or token) device-scheduling as separate concerns, triggering excessive communication between them and compromising inference efficiency This paper proposes Semantic Parallelism, a novel parallelism paradigm that minimizes the steep communication costs in EP-centric MoE serving via model-data collaborative scheduling. We implement Semantic Parallelism in a framework called Sem-MoE. Sem-MoE maximally collocates experts and their activating tokens onto the same device using proactively modeled activation likelihood between them and introduces three key techniques: (1) Offline model scheduling, which preliminarily clusters and collocates experts onto devices based on their co-activation tendencies for certain classes of input. (2) Online inter-request data scheduling for Attention-DP setups, which proactively rebatches incoming requests onto the device that hosts experts most likely and frequently activated by the corresponding requests. (3) Online intra-request data scheduling for Attention-TP setups, which seamlessly fuses a token reshuffling procedure into the original inference pipeline and proactively reschedules tokens to devices to reduce dispersed remote routing. We build Sem-MoE into a prevailing LLM serving engine SGLANG. Experiments show our collaborative scheduling approach can effectively reduce the all-to-all communication volume in EP and achieve superior inference throughput compared to existing solutions.

General Machine Learning · Causality

Shishir Adhikari, Sourav Medya, Elena Zheleva

Peer effect refers to the difference in counterfactual outcomes for a unit resulting from different levels of peer exposure, the extent to which the unit is exposed to the treatments, actions, or behaviors of its peers. In practice, peer exposure is typically captured through an explicitly defined exposure mapping function that aggregates peer treatments and outputs peer exposure. Exposure mapping functions range from simple functions like the number or fraction of treated friends to more sophisticated functions that allow for different peers to exert different degrees of influence. However, the true function is rarely known in practice and when the function is misspecified, this leads to biased causal effect estimation. To address this problem, the focus of our work is to move away from the need to explicitly define an exposure mapping function and instead introduce a framework that allows learning this function automatically. We develop EGONETGNN, a graph neural network (GNN), for heterogeneous peer effect estimation that automatically learns the appropriate exposure mapping function and allows for complex peer exposure mechanisms that involve not only peer treatments but also attributes of the local neighborhood, including node, edge, and structural attributes. We theoretically and empirically show that GNN models that use peer exposure based on the number or fraction of treated peers or learn peer exposure naively face difficulty accounting for such influence mechanisms. Our evaluation on synthetic and semi-synthetic network data shows that our method is more robust to different unknown underlying influence mechanisms when compared to state-of-the-art baselines.

Applications · Language, Speech and Dialog

Zhouqi Hua, Wenwei Zhang, Chengqi Lyu, Yuzhe Gu, Songyang Gao, Kuikun Liu, Dahua Lin, Kai Chen

Length generalization, the ability to solve problems of longer sequences than those observed during training, poses a core challenge of Transformer-based large language models (LLMs). Although existing studies have predominantly focused on data-driven approaches for particular arithmetic operations or symbolic manipulation tasks, these approaches tend to be task-specific with limited performance on individual tasks. To pursue a more general solution, this paper focuses on a broader case of reasoning problems that are *computable*, *i.e.*, problems that algorithms can solve, thus can be solved by the Turing machine, which operates over inputs of unbounded length. From this perspective, this paper proposes **T**uring m**A**chine **I**mitation **L**earning (**TAIL**) to improve the length generalization ability of LLMs. TAIL uses computer programs to directly synthesize chain-of-thought (CoT) data that imitate the execution process of a Turing machine, which *linearly* expands the reasoning steps into *atomic* states to alleviate shortcut pattern learning and explicit *memory* fetch mechanism to reduce the difficulties of dynamic and long-range data access. To validate the universality and reliability of TAIL, we construct a challenging synthetic dataset covering 8 classes of algorithms and 18 tasks. Without bells and whistles, TAIL significantly improves the length generalization ability as well as the performance of Qwen2.5-7B in individual tasks using only synthetic data, surpassing previous methods and DeepSeek-R1. The experimental results reveal that the key concepts in the Turing machine, instead of the human-like thinking styles, are indispensable for TAIL for length generalization, through which the model exhibits read-and-write behaviors consistent with the properties of the Turing machine in their attention layers. This work provides a promising direction for future research in the learning of LLM reasoning from synthetic data.

Deep Learning · Graph Neural Networks

Hao Xu, Xiangru Jian, Xinjian Zhao, Wei Pang, Chao Zhang, Suyuchen Wang, Qixin ZHANG, Zhengyuan Dong, Joao Monteiro, Bang Liu 等

This paper introduces GraphOmni, a comprehensive benchmark designed to evaluate the reasoning capabilities of LLMs on graph-theoretic tasks articulated in natural language. GraphOmni spans diverse graph types, serialization formats, and prompting schemes, substantially extending upon prior efforts in both scope and depth. Through systematic evaluation, we uncover critical interactions among these dimensions, revealing their decisive impact on model performance. Our experiments show that state-of-the-art closed-source models such as Claude-3.5 and o4-mini consistently lead overall, yet still leave considerable headroom, while open-source models display pronounced sensitivity to various design choices. Beyond the standard scope, larger graphs, real-world graphs, and additional NP-hard tasks are further discussed. We further analyze efficiency via output token usage, highlighting cost–accuracy trade-offs, and introduce a reinforcement learning-based optimizer that adaptively selects factor combinations, reducing evaluation cost by 75\% while retaining strong accuracy. This flexible and extensible benchmark not only deepens understanding of LLM performance on structured graph reasoning but also establishes a robust foundation for advancing model design and evaluation.

Reinforcement Learning · Everything Else

Haoming Meng, Kexin Huang, Shaohang Wei, Chiyu Ma, Shuo Yang, xue wang, Guoyin Wang, Bolin Ding, Jingren Zhou

Reinforcement learning with verifiable rewards (RLVR) has significantly improved reasoning in large language models (LLMs), yet the token-level mechanisms underlying these improvements remain unclear. We present a systematic empirical study of RLVR’s distributional effects organized around three main analyses: (1) token-level characterization of distributional shifts between base and RL models, (2) the impact of token-level distributional shifts on reasoning performance through cross-sampling interventions, and (3) fine-grained mechanics of these shifts at the token level. We find that RL fine-tuning induces highly sparse and targeted changes, with only a small fraction of token distributions exhibiting meaningful divergence. We further characterize the structure of these shifts through analyses of token entropy, positional concentration, and reallocation of probability mass. To assess the functional importance of these sparse changes, we conduct cross-sampling experiments that selectively swap token choices between the base and RL models. Inserting only a small fraction of RL-sampled tokens into base generations progressively recovers RL performance gains, while injecting a similarly small number of base token choices into RL-generated responses collapses performance to base levels, isolating a sparse set of token-level decisions directly responsible for RLVR’s improvements. Finally, we explore divergence-weighted variants of the advantage signal as a diagnostic intervention, finding that they can yield improvements over baselines. Together, our results shed light on the distributional changes induced by RLVR and provide a fine-grained, token-level lens for understanding RLVR as a targeted refinement process.

Computer Vision · Vision Models & Multimodal

Maxim Kurkin, Boris Shirokikh, IRINA ABDULLAEVA, Viktoriia Chekalina, Andrey Kuznetsov

Despite recent advancements in extending context windows of large language models (LLMs) and large vision-language models (LVLMs), their ability to perform complex multi-modal reasoning over extended contexts remains critically limited. To underline this challenge, we present \textbf{MMReD}, a benchmark specifically designed to assess reasoning abilities within dense, information-rich scenarios where simple retrieval is not enough. Unlike traditional Needle-in-a-Haystack evaluations, MMReD challenges models to identify and interpret global patterns across entire contexts. Our benchmark comprises 24 tasks of varying complexity, ranging from standard passkey retrieval setups to those requiring selective or uniform attention to all context chunks. The evaluation reveals a consistent performance drop across all tested models -- including the most advanced LLMs, LVLMs, and architectures specializing in code and reasoning -- as the number of observations increases. Notably, even the leading reasoning-specialized models achieve 0\% accuracy on certain tasks at the maximum context length of 128 observations. Conventional fine-tuning techniques, such as SFT and GRPO, also fail to generalize effectively to longer contexts. These observations reveal an inherent limitation in current model architectures, emphasizing the need for innovative approaches to enable competent dense context reasoning in multi-modal AI systems.