# Research Papers — 2025-12 · bundle page 34 of 175 > Full-text markdown of papers #1651–1700 of 8703 published in 2025-12, newest first. AI-readable mirrors of the full 8703-paper 2025-12 archive (50 papers per file). Canonical page: https://aipod.dev/papers/archive/2025-12/page/34 [← Previous bundle](https://aipod.dev/papers/archive/2025-12/page/33.md) [Next bundle →](https://aipod.dev/papers/archive/2025-12/page/35.md) ### Stabilizing Multimodal Autoencoders: A Theoretical and Empirical Analysis of Fusion Strategies Authors: Diyar Altinses, Andreas Schwung Published: 2025-12-23 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2512.20749v1 In recent years, the development of multimodal autoencoders has gained significant attention due to their potential to handle multimodal complex data types and improve model performance. Understanding the stability and robustness of these models is crucial for optimizing their training, architecture, and real-world applicability. This paper presents an analysis of Lipschitz properties in multimodal autoencoders, combining both theoretical insights and empirical validation to enhance the training stability of these models. We begin by deriving the theoretical Lipschitz constants for aggregation methods within the multimodal autoencoder framework. We then introduce a regularized attention-based fusion method, developed based on our theoretical analysis, which demonstrates improved stability and performance during training. Through a series of experiments, we empirically validate our theoretical findings by estimating the Lipschitz constants across multiple trials and fusion strategies. Our results demonstrate that our proposed fusion function not only aligns with theoretical predictions but also outperforms existing strategies in terms of consistency, convergence speed, and accuracy. This work provides a solid theoretical foundation for understanding fusion in multimodal autoencoders and contributes a solution for enhancing their performance. --- ### A Physics Informed Neural Network For Deriving MHD State Vectors From Global Active Regions Observations Authors: Subhamoy Chatterjee, Mausumi Dikpati Published: 2025-12-23 Categories: astro-ph.SR, cs.AI, cs.LG arXiv: https://arxiv.org/abs/2512.20747v1 Solar active regions (ARs) do not appear randomly but cluster along longitudinally warped toroidal bands ('toroids') that encode information about magnetic structures in the tachocline, where global-scale organization likely originates. Global MagnetoHydroDynamic Shallow-Water Tachocline (MHD-SWT) models have shown potential to simulate such toroids, matching observations qualitatively. For week-scale early prediction of flare-producing AR emergence, forward-integration of these toroids is necessary. This requires model initialization with a dynamically self-consistent MHD state-vector that includes magnetic, flow fields, and shell-thickness variations. However, synoptic magnetograms provide only geometric shape of toroids, not the state-vector needed to initialize MHD-SWT models. To address this challenging task, we develop PINNBARDS, a novel Physics-Informed Neural Network (PINN)-Based AR Distribution Simulator, that uses observational toroids and MHD-SWT equations to derive initial state-vector. Using Feb-14-2024 SDO/HMI synoptic map, we show that PINN converges to physically consistent, predominantly antisymmetric toroids, matching observed ones. Although surface data provides north and south toroids' central latitudes, and their latitudinal widths, they cannot determine tachocline field strengths, connected to AR emergence. We explore here solutions across a broad parameter range, finding hydrodynamically-dominated structures for weak fields (~2 kG) and overly rigid behavior for strong fields (~100 kG). We obtain best agreement with observations for 20-30 kG toroidal fields, and ~10 degree bandwidth, consistent with low-order longitudinal mode excitation. To our knowledge, PINNBARDS serves as the first method for reconstructing state-vectors for hidden tachocline magnetic structures from surface patterns; potentially leading to weeks ahead prediction of flare-producing AR-emergence. --- ### TrashDet: Iterative Neural Architecture Search for Efficient Waste Detection Authors: Tony Tran, Bin Hu Published: 2025-12-23 Categories: cs.CV, cs.LG arXiv: https://arxiv.org/abs/2512.20746v1 This paper addresses trash detection on the TACO dataset under strict TinyML constraints using an iterative hardware-aware neural architecture search framework targeting edge and IoT devices. The proposed method constructs a Once-for-All-style ResDets supernet and performs iterative evolutionary search that alternates between backbone and neck/head optimization, supported by a population passthrough mechanism and an accuracy predictor to reduce search cost and improve stability. This framework yields a family of deployment-ready detectors, termed TrashDets. On a five-class TACO subset (paper, plastic, bottle, can, cigarette), the strongest variant, TrashDet-l, achieves 19.5 mAP50 with 30.5M parameters, improving accuracy by up to 3.6 mAP50 over prior detectors while using substantially fewer parameters. The TrashDet family spans 1.2M to 30.5M parameters with mAP50 values between 11.4 and 19.5, providing scalable detector options for diverse TinyML deployment budgets on resource-constrained hardware. On the MAX78002 microcontroller with the TrashNet dataset, two specialized variants, TrashDet-ResNet and TrashDet-MBNet, jointly dominate the ai87-fpndetector baseline, with TrashDet-ResNet achieving 7525~$μ$J energy per inference at 26.7 ms latency and 37.45 FPS, and TrashDet-MBNet improving mAP50 by 10.2%; together they reduce energy consumption by up to 88%, latency by up to 78%, and average power by up to 53% compared to existing TinyML detectors. --- ### AgentMath: Empowering Mathematical Reasoning for Large Language Models via Tool-Augmented Agent Authors: Haipeng Luo, Huawen Feng, Qingfeng Sun, Can Xu, Kai Zheng, Yufei Wang, Tao Yang, Han Hu, Yansong Tang Published: 2025-12-23 Categories: cs.AI, cs.CL, cs.LG arXiv: https://arxiv.org/abs/2512.20745v3 Large Reasoning Models (LRMs) like o3 and DeepSeek-R1 have achieved remarkable progress in reasoning tasks with long cot. However, they remain computationally inefficient and struggle with accuracy when solving problems requiring complex mathematical operations. In this work, we present AgentMath, an agent framework that seamlessly integrates language models' reasoning capabilities with code interpreters' computational precision to efficiently tackle complex mathematical problems. Our approach introduces three key innovations: (1) An automated method that converts natural language chain-of-thought into structured tool-augmented trajectories, generating high-quality supervised fine-tuning (SFT) data to alleviate data scarcity; (2) A novel agentic reinforcement learning (RL) paradigm that dynamically interleaves natural language generation with real-time code execution. This enables models to autonomously learn optimal tool-use strategies through multi-round interactive feedback, while fostering emergent capabilities in code refinement and error correction; (3) An efficient training system incorporating innovative techniques, including request-level asynchronous rollout scheduling, agentic partial rollout, and prefix-aware weighted load balancing, achieving 4-5x speedup and making efficient RL training feasible on ultra-long sequences with scenarios with massive tool invocation. The evaluations show that AgentMath achieves state-of-the-art performance on challenging mathematical competition benchmarks including AIME24, AIME25, and HMMT25. Specifically, AgentMath-30B-A3B attains 90.6%, 86.4%, and 73.8% accuracy respectively, surpassing OpenAI-o3-mini and Claude-Opus-4.0-Thinking while remaining competitive with OpenAI-o3, Gemini-2.5-Pro, and DeepSeek-R1-671B-0528.These results validate the effectiveness of our approach and pave the way for building scalable mathematical reasoning agents. --- ### AI-Driven Green Cognitive Radio Networks for Sustainable 6G Communication Authors: Anshul Sharma, Shujaatali Badami, Biky Chouhan, Pushpanjali Pandey, Brijeena Rana, Navneet Kaur Published: 2025-12-23 Categories: cs.NI, cs.AI, cs.LG arXiv: https://arxiv.org/abs/2512.20739v1 The 6G wireless aims at the Tb/s peak data rates are expected, a sub-millisecond latency, massive Internet of Things/vehicle connectivity, which requires sustainable access to audio over the air and energy-saving functionality. Cognitive Radio Networks CCNs help in alleviating the problem of spectrum scarcity, but classical sensing and allocation are still energy-consumption intensive, and sensitive to rapid spectrum variations. Our framework which centers on AI driven green CRN aims at integrating deep reinforcement learning (DRL) with transfer learning, energy harvesting (EH), reconfigurable intelligent surfaces (RIS) with other light-weight genetic refinement operations that optimally combine sensing timelines, transmit power, bandwidth distribution and RIS phase selection. Compared to two baselines, the utilization of MATLAB + NS-3 under dense loads, a traditional CRN with energy sensing under fixed policies, and a hybrid CRN with cooperative sensing under heuristic distribution of resource, there are (25-30%) fewer energy reserves used, sensing AUC greater than 0.90 and +6-13 p.p. higher PDR. The integrated framework is easily scalable to large IoT and vehicular applications, and it provides a feasible and sustainable roadmap to 6G CRNs. Index Terms--Cognitive Radio Networks (CRNs), 6G, Green Communication, Energy Efficiency, Deep Reinforcement Learning (DRL), Spectrum Sensing, RIS, Energy Harvesting, QoS, IoT. --- ### VL4Gaze: Unleashing Vision-Language Models for Gaze Following Authors: Shijing Wang, Chaoqun Cui, Yaping Huang, Hyung Jin Chang, Yihua Cheng Published: 2025-12-23 Categories: cs.CV arXiv: https://arxiv.org/abs/2512.20735v1 Human gaze provides essential cues for interpreting attention, intention, and social interaction in visual scenes, yet gaze understanding remains largely unexplored in current vision-language models (VLMs). While recent VLMs achieve strong scene-level reasoning across a range of visual tasks, there exists no benchmark that systematically evaluates or trains them for gaze interpretation, leaving open the question of whether gaze understanding can emerge from general-purpose vision-language pre-training. To address this gap, we introduce VL4Gaze, the first large-scale benchmark designed to investigate, evaluate, and unlock the potential of VLMs for gaze understanding. VL4Gaze contains 489K automatically generated question-answer pairs across 124K images and formulates gaze understanding as a unified VQA problem through four complementary tasks: (1) gaze object description, (2) gaze direction description, (3) gaze point location, and (4) ambiguous question recognition. We comprehensively evaluate both commercial and open-source VLMs under in-context learning and fine-tuning settings. The results show that even large-scale VLMs struggle to reliably infer gaze semantics and spatial localization without task-specific supervision. In contrast, training on VL4Gaze brings substantial and consistent improvements across all tasks, highlighting the importance of targeted multi-task supervision for developing gaze understanding capabilities in VLMs. We will release the dataset and code to support further research and development in this direction. --- ### FEM-Bench: A Structured Scientific Reasoning Benchmark for Evaluating Code-Generating LLMs Authors: Saeed Mohammadzadeh, Erfan Hamdi, Joel Shor, Emma Lejeune Published: 2025-12-23 Categories: cs.LG, cs.AI, cs.SE arXiv: https://arxiv.org/abs/2512.20732v2 As LLMs advance their reasoning capabilities about the physical world, the absence of rigorous benchmarks for evaluating their ability to generate scientifically valid physical models has become a critical gap. Computational mechanics, which develops and applies mathematical models and numerical methods to predict the behavior of physical systems under forces, deformation, and constraints, provides an ideal foundation for structured scientific reasoning evaluation. Problems follow clear mathematical structure, enforce strict physical and numerical constraints, and support objective verification. The discipline requires constructing explicit models of physical systems and reasoning about geometry, spatial relationships, and material behavior, connecting directly to emerging AI goals in physical reasoning and world modeling. We introduce FEM-Bench, a computational mechanics benchmark designed to evaluate the ability of LLMs to generate correct finite element method (FEM) and related code. FEM-Bench 2025 contains a suite of introductory but nontrivial tasks aligned with material from a first graduate course on computational mechanics. These tasks capture essential numerical and physical modeling challenges while representing only a small fraction of the complexity present in the discipline. Despite their simplicity, state-of-the-art LLMs do not reliably solve all of them. In a five attempt run, the best performing model at function writing, Gemini 3 Pro, completed 30/33 tasks at least once and 26/33 tasks all five times. The best performing model at unit test writing, GPT-5, had an Average Joint Success Rate of 73.8%. Other popular models showed broad performance variation. FEM-Bench establishes a structured foundation for evaluating AI-generated scientific code, and future iterations will incorporate increasingly sophisticated tasks to track progress as models evolve. --- ### SA-DiffuSeq: Addressing Computational and Scalability Challenges in Long-Document Generation with Sparse Attention Authors: Alexandros Christoforos, Chadbourne Davis Published: 2025-12-23 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2512.20724v1 Diffusion based approaches to long form text generation suffer from prohibitive computational cost and memory overhead as sequence length increases. We introduce SA-DiffuSeq, a diffusion framework that integrates sparse attention to fundamentally improve scalability for long document modeling. By selectively allocating attention within the diffusion process, SA-DiffuSeq significantly reduces computational complexity while maintaining semantic coherence and generation quality. A key component of our method is a soft absorbing state tailored to sparse attention dynamics, which stabilizes diffusion trajectories and accelerates sequence reconstruction. This design improves sampling efficiency and enhances precision in long range dependency modeling. Extensive experiments demonstrate that SA-DiffuSeq consistently surpasses state of the art diffusion baselines in both training efficiency and sampling speed, with especially strong gains on extended sequences. These properties make SA-DiffuSeq well suited for demanding long form applications such as scientific writing, large scale code generation, and multi turn long context dialogue. Overall, our results indicate that incorporating structured sparsity into diffusion models is a promising direction for efficient and expressive long text generation. --- ### From artificial to organic: Rethinking the roots of intelligence for digital health Authors: Prajwal Ghimire, Keyoumars Ashkan Published: 2025-12-23 Categories: cs.AI arXiv: https://arxiv.org/abs/2512.20723v1 The term artificial implies an inherent dichotomy from the natural or organic. However, AI, as we know it, is a product of organic ingenuity: designed, implemented, and iteratively improved by human cognition. The very principles that underpin AI systems, from neural networks to decision-making algorithms, are inspired by the organic intelligence embedded in human neurobiology and evolutionary processes. The path from organic to artificial intelligence in digital health is neither mystical nor merely a matter of parameter count, it is fundamentally about organization and adaption. Thus, the boundaries between artificial and organic are far less distinct than the nomenclature suggests. --- ### From Pilots to Practices: A Scoping Review of GenAI-Enabled Personalization in Computer Science Education Authors: Iman Reihanian, Yunfei Hou, Qingquan Sun Published: 2025-12-23 Categories: cs.AI, cs.CY, cs.HC arXiv: https://arxiv.org/abs/2512.20714v1 Generative AI enables personalized computer science education at scale, yet questions remain about whether such personalization supports or undermines learning. This scoping review synthesizes 32 studies (2023-2025) purposively sampled from 259 records to map personalization mechanisms and effectiveness signals in higher-education computer science contexts. We identify five application domains: intelligent tutoring, personalized materials, formative feedback, AI-augmented assessment, and code review, and analyze how design choices shape learning outcomes. Designs incorporating explanation-first guidance, solution withholding, graduated hint ladders, and artifact grounding (student code, tests, and rubrics) consistently show more positive learning processes than unconstrained chat interfaces. Successful implementations share four patterns: context-aware tutoring anchored in student artifacts, multi-level hint structures requiring reflection, composition with traditional CS infrastructure (autograders and rubrics), and human-in-the-loop quality assurance. We propose an exploration-first adoption framework emphasizing piloting, instrumentation, learning-preserving defaults, and evidence-based scaling. Recurrent risks include academic integrity, privacy, bias and equity, and over-reliance, and we pair these with operational mitigation. The evidence supports generative AI as a mechanism for precision scaffolding when embedded in audit-ready workflows that preserve productive struggle while scaling personalized support. --- ### Real-World Adversarial Attacks on RF-Based Drone Detectors Authors: Omer Gazit, Yael Itzhakev, Yuval Elovici, Asaf Shabtai Published: 2025-12-23 Categories: cs.CR, cs.LG arXiv: https://arxiv.org/abs/2512.20712v4 Radio frequency (RF) based systems are increasingly used to detect drones by analyzing their RF signal patterns, converting them into spectrogram images which are processed by object detection models. Existing RF attacks against image based models alter digital features, making over-the-air (OTA) implementation difficult due to the challenge of converting digital perturbations to transmittable waveforms that may introduce synchronization errors and interference, and encounter hardware limitations. We present the first physical attack on RF image based drone detectors, optimizing class-specific universal complex baseband (I/Q) perturbation waveforms that are transmitted alongside legitimate communications. We evaluated the attack using RF recordings and OTA experiments with four types of drones. Our results show that modest, structured I/Q perturbations are compatible with standard RF chains and reliably reduce target drone detection while preserving detection of legitimate drones. --- ### SemanticGen: Video Generation in Semantic Space Authors: Jianhong Bai, Xiaoshi Wu, Xintao Wang, Xiao Fu, Yuanxing Zhang, Qinghe Wang, Xiaoyu Shi, Menghan Xia, Zuozhu Liu, Haoji Hu, Pengfei Wan, Kun Gai Published: 2025-12-23 Categories: cs.CV arXiv: https://arxiv.org/abs/2512.20619v3 State-of-the-art video generative models typically learn the distribution of video latents in the VAE space and map them to pixels using a VAE decoder. While this approach can generate high-quality videos, it suffers from slow convergence and is computationally expensive when generating long videos. In this paper, we introduce SemanticGen, a novel solution to address these limitations by generating videos in the semantic space. Our main insight is that, due to the inherent redundancy in videos, the generation process should begin in a compact, high-level semantic space for global planning, followed by the addition of high-frequency details, rather than directly modeling a vast set of low-level video tokens using bi-directional attention. SemanticGen adopts a two-stage generation process. In the first stage, a diffusion model generates compact semantic video features, which define the global layout of the video. In the second stage, another diffusion model generates VAE latents conditioned on these semantic features to produce the final output. We observe that generation in the semantic space leads to faster convergence compared to the VAE latent space. Our method is also effective and computationally efficient when extended to long video generation. Extensive experiments demonstrate that SemanticGen produces high-quality videos and outperforms state-of-the-art approaches and strong baselines. --- ### SemanticGen: Video Generation in Semantic Space Authors: Jianhong Bai, Xiaoshi Wu, Xintao Wang, Fu Xiao, Yuanxing Zhang, Qinghe Wang, Xiaoyu Shi, Menghan Xia, Zuozhu Liu, Haoji Hu, Pengfei Wan, Kun Gai Published: 2025-12-23 Categories: cs.CV arXiv: https://arxiv.org/abs/2512.20619v1 State-of-the-art video generative models typically learn the distribution of video latents in the VAE space and map them to pixels using a VAE decoder. While this approach can generate high-quality videos, it suffers from slow convergence and is computationally expensive when generating long videos. In this paper, we introduce SemanticGen, a novel solution to address these limitations by generating videos in the semantic space. Our main insight is that, due to the inherent redundancy in videos, the generation process should begin in a compact, high-level semantic space for global planning, followed by the addition of high-frequency details, rather than directly modeling a vast set of low-level video tokens using bi-directional attention. SemanticGen adopts a two-stage generation process. In the first stage, a diffusion model generates compact semantic video features, which define the global layout of the video. In the second stage, another diffusion model generates VAE latents conditioned on these semantic features to produce the final output. We observe that generation in the semantic space leads to faster convergence compared to the VAE latent space. Our method is also effective and computationally efficient when extended to long video generation. Extensive experiments demonstrate that SemanticGen produces high-quality videos and outperforms state-of-the-art approaches and strong baselines. --- ### LongVideoAgent: Multi-Agent Reasoning with Long Videos Authors: Runtao Liu, Ziyi Liu, Jiaqi Tang, Yue Ma, Renjie Pi, Jipeng Zhang, Qifeng Chen Published: 2025-12-23 Categories: cs.AI, cs.CV, cs.LG, cs.MA arXiv: https://arxiv.org/abs/2512.20618v1 Recent advances in multimodal LLMs and systems that use tools for long-video QA point to the promise of reasoning over hour-long episodes. However, many methods still compress content into lossy summaries or rely on limited toolsets, weakening temporal grounding and missing fine-grained cues. We propose a multi-agent framework in which a master LLM coordinates a grounding agent to localize question-relevant segments and a vision agent to extract targeted textual observations. The master agent plans with a step limit, and is trained with reinforcement learning to encourage concise, correct, and efficient multi-agent cooperation. This design helps the master agent focus on relevant clips via grounding, complements subtitles with visual detail, and yields interpretable trajectories. On our proposed LongTVQA and LongTVQA+ which are episode-level datasets aggregated from TVQA/TVQA+, our multi-agent system significantly outperforms strong non-agent baselines. Experiments also show reinforcement learning further strengthens reasoning and planning for the trained agent. Code and data will be shared at https://longvideoagent.github.io/. --- ### SpatialTree: How Spatial Abilities Branch Out in MLLMs Authors: Yuxi Xiao, Longfei Li, Shen Yan, Xinhang Liu, Sida Peng, Yunchao Wei, Xiaowei Zhou, Bingyi Kang Published: 2025-12-23 Categories: cs.CV arXiv: https://arxiv.org/abs/2512.20617v2 Cognitive science suggests that spatial ability develops progressively-from perception to reasoning and interaction. Yet in multimodal LLMs (MLLMs), this hierarchy remains poorly understood, as most studies focus on a narrow set of tasks. We introduce SpatialTree, a cognitive-science-inspired hierarchy that organizes spatial abilities into four levels: low-level perception (L1), mental mapping (L2), simulation (L3), and agentic competence (L4). Based on this taxonomy, we construct the first capability-centric hierarchical benchmark, thoroughly evaluating mainstream MLLMs across 27 sub-abilities. The evaluation results reveal a clear structure: L1 skills are largely orthogonal, whereas higher-level skills are strongly correlated, indicating increasing interdependency. Through targeted supervised fine-tuning, we uncover a surprising transfer dynamic-negative transfer within L1, but strong cross-level transfer from low- to high-level abilities with notable synergy. Finally, we explore how to improve the entire hierarchy. We find that naive RL that encourages extensive "thinking" is unreliable: it helps complex reasoning but hurts intuitive perception. We propose a simple auto-think strategy that suppresses unnecessary deliberation, enabling RL to consistently improve performance across all levels. By building SpatialTree, we provide a proof-of-concept framework for understanding and systematically scaling spatial abilities in MLLMs. --- ### SpatialTree: How Spatial Abilities Branch Out in MLLMs Authors: Yuxi Xiao, Longfei Li, Shen Yan, Xinhang Liu, Sida Peng, Yunchao Wei, Xiaowei Zhou, Bingyi Kang Published: 2025-12-23 Categories: cs.CV arXiv: https://arxiv.org/abs/2512.20617v1 Cognitive science suggests that spatial ability develops progressively-from perception to reasoning and interaction. Yet in multimodal LLMs (MLLMs), this hierarchy remains poorly understood, as most studies focus on a narrow set of tasks. We introduce SpatialTree, a cognitive-science-inspired hierarchy that organizes spatial abilities into four levels: low-level perception (L1), mental mapping (L2), simulation (L3), and agentic competence (L4). Based on this taxonomy, we construct the first capability-centric hierarchical benchmark, thoroughly evaluating mainstream MLLMs across 27 sub-abilities. The evaluation results reveal a clear structure: L1 skills are largely orthogonal, whereas higher-level skills are strongly correlated, indicating increasing interdependency. Through targeted supervised fine-tuning, we uncover a surprising transfer dynamic-negative transfer within L1, but strong cross-level transfer from low- to high-level abilities with notable synergy. Finally, we explore how to improve the entire hierarchy. We find that naive RL that encourages extensive "thinking" is unreliable: it helps complex reasoning but hurts intuitive perception. We propose a simple auto-think strategy that suppresses unnecessary deliberation, enabling RL to consistently improve performance across all levels. By building SpatialTree, we provide a proof-of-concept framework for understanding and systematically scaling spatial abilities in MLLMs. --- ### Active Intelligence in Video Avatars via Closed-loop World Modeling Authors: Xuanhua He, Tianyu Yang, Ke Cao, Ruiqi Wu, Cheng Meng, Yong Zhang, Zhuoliang Kang, Xiaoming Wei, Qifeng Chen Published: 2025-12-23 Categories: cs.CV arXiv: https://arxiv.org/abs/2512.20615v1 Current video avatar generation methods excel at identity preservation and motion alignment but lack genuine agency, they cannot autonomously pursue long-term goals through adaptive environmental interaction. We address this by introducing L-IVA (Long-horizon Interactive Visual Avatar), a task and benchmark for evaluating goal-directed planning in stochastic generative environments, and ORCA (Online Reasoning and Cognitive Architecture), the first framework enabling active intelligence in video avatars. ORCA embodies Internal World Model (IWM) capabilities through two key innovations: (1) a closed-loop OTAR cycle (Observe-Think-Act-Reflect) that maintains robust state tracking under generative uncertainty by continuously verifying predicted outcomes against actual generations, and (2) a hierarchical dual-system architecture where System 2 performs strategic reasoning with state prediction while System 1 translates abstract plans into precise, model-specific action captions. By formulating avatar control as a POMDP and implementing continuous belief updating with outcome verification, ORCA enables autonomous multi-step task completion in open-domain scenarios. Extensive experiments demonstrate that ORCA significantly outperforms open-loop and non-reflective baselines in task success rate and behavioral coherence, validating our IWM-inspired design for advancing video avatar intelligence from passive animation to active, goal-oriented behavior. --- ### Making Large Language Models Efficient Dense Retrievers Authors: Yibin Lei, Shwai He, Ang Li, Andrew Yates Published: 2025-12-23 Categories: cs.IR, cs.CL arXiv: https://arxiv.org/abs/2512.20612v1 Recent work has shown that directly fine-tuning large language models (LLMs) for dense retrieval yields strong performance, but their substantial parameter counts make them computationally inefficient. While prior studies have revealed significant layer redundancy in LLMs for generative tasks, it remains unclear whether similar redundancy exists when these models are adapted for retrieval tasks, which require encoding entire sequences into fixed representations rather than generating tokens iteratively. To this end, we conduct a comprehensive analysis of layer redundancy in LLM-based dense retrievers. We find that, in contrast to generative settings, MLP layers are substantially more prunable, while attention layers remain critical for semantic aggregation. Building on this insight, we propose EffiR, a framework for developing efficient retrievers that performs large-scale MLP compression through a coarse-to-fine strategy (coarse-grained depth reduction followed by fine-grained width reduction), combined with retrieval-specific fine-tuning. Across diverse BEIR datasets and LLM backbones, EffiR achieves substantial reductions in model size and inference cost while preserving the performance of full-size models. --- ### FedPOD: the deployable units of training for federated learning Authors: Daewoon Kim, Si Young Yie, Jae Sung Lee Published: 2025-12-23 Categories: cs.CV, cs.LG arXiv: https://arxiv.org/abs/2512.20610v2 This paper proposes FedPOD, which ranked first in the 2024 Federated Tumor Segmentation (FeTS) Challenge, for optimizing learning efficiency and communication cost in federated learning among multiple clients. Inspired by FedPIDAvg, we define a round-wise task for FedPOD to enhance training efficiency. FedPIDAvg achieved performance improvement by incorporating the training loss reduction for prediction entropy as weights using differential terms. Furthermore, by modeling data distribution with a Poisson distribution and using a PID controller, it reduced communication costs even in skewed data distribution. However, excluding participants classified as outliers based on the Poisson distribution can limit data utilization. Additionally, PID controller requires the same participants to be maintained throughout the federated learning process as it uses previous rounds' learning information in the current round. In our approach, FedPOD addresses these issues by including participants excluded as outliers, eliminating dependency on previous rounds' learning information, and applying a method for calculating validation loss at each round. In this challenge, FedPOD presents comparable performance to FedPIDAvg in metrics of Dice score, 0.78, 0.71 and 0.72 for WT, ET and TC in average, and projected convergence score, 0.74 in average. Furthermore, the concept of FedPOD draws inspiration from Kubernetes' smallest computing unit, POD, designed to be compatible with Kubernetes auto-scaling. Extending round-wise tasks of FedPOD to POD units allows flexible design by applying scale-out similar to Kubernetes' auto-scaling. This work demonstrated the potentials of FedPOD to enhance federated learning by improving efficiency, flexibility, and performance in metrics. --- ### FedPOD: the deployable units of training for federated learning Authors: Daewoon Kim, Si Young Yie, Jae Sung Lee Published: 2025-12-23 Categories: cs.CV, cs.LG arXiv: https://arxiv.org/abs/2512.20610v1 This paper proposes FedPOD (Proportionally Orchestrated Derivative) for optimizing learning efficiency and communication cost in federated learning among multiple clients. Inspired by FedPIDAvg, we define a round-wise task for FedPOD to enhance training efficiency. FedPIDAvg achieved performance improvement by incorporating the training loss reduction for prediction entropy as weights using differential terms. Furthermore, by modeling data distribution with a Poisson distribution and using a PID controller, it reduced communication costs even in skewed data distribution. However, excluding participants classified as outliers based on the Poisson distribution can limit data utilization. Additionally, PID controller requires the same participants to be maintained throughout the federated learning process as it uses previous rounds' learning information in the current round. In our approach, FedPOD addresses these issues by including participants excluded as outliers, eliminating dependency on previous rounds' learning information, and applying a method for calculating validation loss at each round. In this challenge, FedPOD presents comparable performance to FedPIDAvg in metrics of Dice score, 0.78, 0.71 and 0.72 for WT, ET and TC in average, and projected convergence score, 0.74 in average. Furthermore, the concept of FedPOD draws inspiration from Kubernetes' smallest computing unit, POD, designed to be compatible with Kubernetes auto-scaling. Extending round-wise tasks of FedPOD to POD units allows flexible design by applying scale-out similar to Kubernetes' auto-scaling. This work demonstrated the potentials of FedPOD to enhance federated learning by improving efficiency, flexibility, and performance in metrics. --- ### Saddle-to-Saddle Dynamics Explains A Simplicity Bias Across Neural Network Architectures Authors: Yedi Zhang, Andrew Saxe, Peter E. Latham Published: 2025-12-23 Categories: cs.LG arXiv: https://arxiv.org/abs/2512.20607v2 Neural networks trained with gradient descent often learn solutions of increasing complexity over time, a phenomenon known as simplicity bias. Despite being widely observed across architectures, existing theoretical treatments lack a unifying framework. We present a theoretical framework that explains a simplicity bias arising from saddle-to-saddle learning dynamics for a general class of neural networks, incorporating fully-connected, convolutional, and attention-based architectures. Here, simple means expressible with few hidden units, i.e., hidden neurons, convolutional kernels, or attention heads. Specifically, we show that linear networks learn solutions of increasing rank, ReLU networks learn solutions with an increasing number of kinks, convolutional networks learn solutions with an increasing number of convolutional kernels, and self-attention models learn solutions with an increasing number of attention heads. By analyzing fixed points, invariant manifolds, and dynamics of gradient descent learning, we show that saddle-to-saddle dynamics operates by iteratively evolving near an invariant manifold, approaching a saddle, and switching to another invariant manifold. Our analysis also disentangles data-induced and initialization-induced saddle-to-saddle dynamics. In particular, the former leads to low-rank weights while the latter to sparse weights. Equipped with the theory, we predict the effects of data distribution and weight initialization on the duration and number of plateaus in learning. Overall, our theory offers a framework for understanding when and why gradient descent progressively learns increasingly complex solutions. --- ### Saddle-to-Saddle Dynamics Explains A Simplicity Bias Across Neural Network Architectures Authors: Yedi Zhang, Andrew Saxe, Peter E. Latham Published: 2025-12-23 Categories: cs.LG arXiv: https://arxiv.org/abs/2512.20607v1 Neural networks trained with gradient descent often learn solutions of increasing complexity over time, a phenomenon known as simplicity bias. Despite being widely observed across architectures, existing theoretical treatments lack a unifying framework. We present a theoretical framework that explains a simplicity bias arising from saddle-to-saddle learning dynamics for a general class of neural networks, incorporating fully-connected, convolutional, and attention-based architectures. Here, simple means expressible with few hidden units, i.e., hidden neurons, convolutional kernels, or attention heads. Specifically, we show that linear networks learn solutions of increasing rank, ReLU networks learn solutions with an increasing number of kinks, convolutional networks learn solutions with an increasing number of convolutional kernels, and self-attention models learn solutions with an increasing number of attention heads. By analyzing fixed points, invariant manifolds, and dynamics of gradient descent learning, we show that saddle-to-saddle dynamics operates by iteratively evolving near an invariant manifold, approaching a saddle, and switching to another invariant manifold. Our analysis also illuminates the effects of data distribution and weight initialization on the duration and number of plateaus in learning, dissociating previously confounding factors. Overall, our theory offers a framework for understanding when and why gradient descent progressively learns increasingly complex solutions. --- ### Probing and Leveraging Video Diffusion Transformer Features for Robust Point Tracking Authors: Soowon Son, Honggyu An, Jisu Nam, Hyunah Ko, Chaehyun Kim, Dahyun Chung, Siyoon Jin, Jung Yi, Junhwa Hur, Seungryong Kim Published: 2025-12-23 Categories: cs.CV arXiv: https://arxiv.org/abs/2512.20606v2 Despite achieving strong results on standard benchmarks, current point tracking methods rely on feature backbones that are rarely designed with the temporal coherence needed for robust real-world performance. While recent works incorporate powerful visual foundation model (VFM) features into tracking pipelines, no prior work has systematically analyzed which VFM provides the most robust representations for point tracking. We present the first such analysis, evaluating diverse VFMs in a zero-shot setting on both standard and robustness benchmarks for point tracking. Our study reveals that video diffusion transformers (DiTs) consistently yield the most temporally coherent and discriminative features, even surpassing ResNet backbones explicitly supervised on tracking data. We hypothesize this advantage stem from large-scale video pretraining, full 3D spatio-temporal attention, and a diffusion training objective. Motivated by this finding, we propose DiTracker, which integrates video DiT features into existing tracking frameworks through query-key matching cost computation, cost-level fusion with a lightweight ResNet branch, and LoRA adaptation. Under the same tracking head, DiTracker is trained solely on synthetic data with far fewer iterations, yet outperforms CoTracker3 trained with additional real-world videos, with the largest gains under challenging and corrupted scenarios. It further generalizes across tracking heads and scales with backbone size, confirming that generative video pretraining provides real-world priors that reduce the dependence on large-scale real-data supervision. --- ### Repurposing Video Diffusion Transformers for Robust Point Tracking Authors: Soowon Son, Honggyu An, Chaehyun Kim, Hyunah Ko, Jisu Nam, Dahyun Chung, Siyoon Jin, Jung Yi, Jaewon Min, Junhwa Hur, Seungryong Kim Published: 2025-12-23 Categories: cs.CV arXiv: https://arxiv.org/abs/2512.20606v1 Point tracking aims to localize corresponding points across video frames, serving as a fundamental task for 4D reconstruction, robotics, and video editing. Existing methods commonly rely on shallow convolutional backbones such as ResNet that process frames independently, lacking temporal coherence and producing unreliable matching costs under challenging conditions. Through systematic analysis, we find that video Diffusion Transformers (DiTs), pre-trained on large-scale real-world videos with spatio-temporal attention, inherently exhibit strong point tracking capability and robustly handle dynamic motions and frequent occlusions. We propose DiTracker, which adapts video DiTs through: (1) query-key attention matching, (2) lightweight LoRA tuning, and (3) cost fusion with a ResNet backbone. Despite training with 8 times smaller batch size, DiTracker achieves state-of-the-art performance on challenging ITTO benchmark and matches or outperforms state-of-the-art models on TAP-Vid benchmarks. Our work validates video DiT features as an effective and efficient foundation for point tracking. --- ### Emergent temporal abstractions in autoregressive models enable hierarchical reinforcement learning Authors: Seijin Kobayashi, Yanick Schimpf, Maximilian Schlegel, Angelika Steger, Maciej Wolczyk, Johannes von Oswald, Nino Scherrer, Kaitlin Maile, Guillaume Lajoie, Blake A. Richards, Rif A. Saurous, James Manyika et al. Published: 2025-12-23 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2512.20605v2 Large-scale autoregressive models pretrained on next-token prediction and finetuned with reinforcement learning (RL) have achieved unprecedented success on many problem domains. During RL, these models explore by generating new outputs, one token at a time. However, sampling actions token-by-token can result in highly inefficient learning, particularly when rewards are sparse. Here, we show that it is possible to overcome this problem by acting and exploring within the internal representations of an autoregressive model. Specifically, to discover temporally-abstract actions, we introduce a higher-order, non-causal sequence model whose outputs control the residual stream activations of a base autoregressive model. On grid world and MuJoCo-based tasks with hierarchical structure, we find that the higher-order model learns to compress long activation sequence chunks onto internal controllers. Critically, each controller executes a sequence of behaviorally meaningful actions that unfold over long timescales and are accompanied with a learned termination condition, such that composing multiple controllers over time leads to efficient exploration on novel tasks. We show that direct internal controller reinforcement, a process we term "internal RL", enables learning from sparse rewards in cases where standard RL finetuning fails. Our results demonstrate the benefits of latent action generation and reinforcement in autoregressive models, suggesting internal RL as a promising avenue for realizing hierarchical RL within foundation models. --- ### Emergent temporal abstractions in autoregressive models enable hierarchical reinforcement learning Authors: Seijin Kobayashi, Yanick Schimpf, Maximilian Schlegel, Angelika Steger, Maciej Wolczyk, Johannes von Oswald, Nino Scherre, Kaitlin Maile, Guillaume Lajoie, Blake A. Richards, Rif A. Saurous, James Manyika et al. Published: 2025-12-23 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2512.20605v1 Large-scale autoregressive models pretrained on next-token prediction and finetuned with reinforcement learning (RL) have achieved unprecedented success on many problem domains. During RL, these models explore by generating new outputs, one token at a time. However, sampling actions token-by-token can result in highly inefficient learning, particularly when rewards are sparse. Here, we show that it is possible to overcome this problem by acting and exploring within the internal representations of an autoregressive model. Specifically, to discover temporally-abstract actions, we introduce a higher-order, non-causal sequence model whose outputs control the residual stream activations of a base autoregressive model. On grid world and MuJoCo-based tasks with hierarchical structure, we find that the higher-order model learns to compress long activation sequence chunks onto internal controllers. Critically, each controller executes a sequence of behaviorally meaningful actions that unfold over long timescales and are accompanied with a learned termination condition, such that composing multiple controllers over time leads to efficient exploration on novel tasks. We show that direct internal controller reinforcement, a process we term "internal RL", enables learning from sparse rewards in cases where standard RL finetuning fails. Our results demonstrate the benefits of latent action generation and reinforcement in autoregressive models, suggesting internal RL as a promising avenue for realizing hierarchical RL within foundation models. --- ### MoE-DiffuSeq: Enhancing Long-Document Diffusion Models with Sparse Attention and Mixture of Experts Authors: Alexandros Christoforos, Chadbourne Davis Published: 2025-12-23 Categories: cs.CL arXiv: https://arxiv.org/abs/2512.20604v2 We propose \textbf{MoE-DiffuSeq}, a diffusion-based framework for efficient long-form text generation that integrates sparse attention with a Mixture-of-Experts (MoE) architecture. Existing sequence diffusion models suffer from prohibitive computational and memory costs when scaling to long documents, largely due to dense attention and slow iterative reconstruction. MoE-DiffuSeq addresses these limitations by combining expert routing with a tailored sparse attention mechanism, substantially reducing attention complexity while preserving global coherence and textual fidelity. In addition, we introduce a \emph{soft absorbing state} within the diffusion process that reshapes attention dynamics during denoising, enabling faster sequence reconstruction and more precise token refinement. This design accelerates both training and sampling without sacrificing generation quality. Extensive experiments on long-document benchmarks demonstrate that MoE-DiffuSeq consistently outperforms prior diffusion-based and sparse-attention baselines in training efficiency, inference speed, and generation quality. Our approach is particularly effective for long-context applications such as scientific document generation, large-scale code synthesis, and extended dialogue modeling, establishing a scalable and expressive solution for diffusion-based long-form text generation. --- ### MoE-DiffuSeq: Enhancing Long-Document Diffusion Models with Sparse Attention and Mixture of Experts Authors: Alexandros Christoforos, Chadbourne Davis Published: 2025-12-23 Categories: cs.CL arXiv: https://arxiv.org/abs/2512.20604v1 We present MoE-DiffuSeq, a mixture of experts based framework for enhancing diffusion models in long document generation. Existing diffusion based text generation models, such as DiffuSeq, suffer from high computational cost and memory overhead when applied to extended sequences. To address these challenges, MoE-DiffuSeq integrates sparse attention with a mixture of experts architecture, enabling efficient and scalable long sequence modeling. Our approach introduces a customized sparse attention mechanism designed to reduce computational complexity while preserving text quality and coherence. In addition, we incorporate a soft absorbing state within the diffusion process to accelerate sequence reconstruction and improve generation precision. Extensive experiments demonstrate that MoE-DiffuSeq significantly improves training efficiency and sampling speed compared to existing diffusion models. These advantages are particularly effective for long document scenarios, including scientific article generation, code repository modeling, and long form dialogue generation. Benchmark results further show that MoE-DiffuSeq improves efficiency, speed, accuracy, and expressiveness, advancing the practical applicability of diffusion models for high quality long form text generation. --- ### Cube Bench: A Benchmark for Spatial Visual Reasoning in MLLMs Authors: Dhruv Anand, Ehsan Shareghi Published: 2025-12-23 Categories: cs.CL, cs.AI, cs.CV arXiv: https://arxiv.org/abs/2512.20595v1 We introduce Cube Bench, a Rubik's-cube benchmark for evaluating spatial and sequential reasoning in multimodal large language models (MLLMs). The benchmark decomposes performance into five skills: (i) reconstructing cube faces from images and text, (ii) choosing the optimal next move, (iii) predicting the outcome of a candidate move without applying it, (iv) executing multi-step plans while recovering from mistakes, and (v) detecting and revising one's own errors. Using a shared set of scrambled cube states, identical prompts and parsers, and a single distance-to-solved metric, we compare recent MLLMs side by side as a function of scramble depth. Across seven MLLMs, accuracy drops sharply with depth; once a trajectory stalls or diverges, models rarely recover, and high face-reconstruction accuracy does not guarantee competent action selection or multi-step execution. A pronounced closed- vs open-source gap emerges: the strongest closed model leads on both single-step perception tasks and multi-step control tasks, while open-weight models cluster near chance on the hardest settings; yet even the best MLLM degrades at higher cube complexity. A simple self-correction via reflective thinking yields modest gains but can also introduce overthinking. Cube Bench offers a compact, reproducible probe of sequential spatial reasoning in MLLMs. --- ### Leveraging High-Fidelity Digital Models and Reinforcement Learning for Mission Engineering: A Case Study of Aerial Firefighting Under Perfect Information Authors: İbrahim Oğuz Çetinkaya, Sajad Khodadadian, Taylan G. Topcu Published: 2025-12-23 Categories: cs.CY, cs.AI, eess.SY, math.OC arXiv: https://arxiv.org/abs/2512.20589v2 As systems engineering (SE) objectives evolve from design and operation of monolithic systems to complex System of Systems (SoS), the discipline of Mission Engineering (ME) has emerged which is increasingly being accepted as a new line of thinking for the SE community. Moreover, mission environments are uncertain, dynamic, and mission outcomes are a direct function of how the mission assets will interact with this environment. This proves static architectures brittle and calls for analytically rigorous approaches for ME. To that end, this paper proposes an intelligent mission coordination methodology that integrates digital mission models with Reinforcement Learning (RL), that specifically addresses the need for adaptive task allocation and reconfiguration. More specifically, we are leveraging a Digital Engineering (DE) based infrastructure that is composed of a high-fidelity digital mission model and agent-based simulation; and then we formulate the mission tactics management problem as a Markov Decision Process (MDP), and employ an RL agent trained via Proximal Policy Optimization. By leveraging the simulation as a sandbox, we map the system states to actions, refining the policy based on realized mission outcomes. The utility of the RL-based intelligent mission coordinator is demonstrated through an aerial firefighting case study. Our findings indicate that the RL-based intelligent mission coordinator not only surpasses baseline performance but also significantly reduces the variability in mission performance. Thus, this study serves as a proof of concept demonstrating that DE-enabled mission simulations combined with advanced analytical tools offer a mission-agnostic framework for improving ME practice; which can be extended to more complicated fleet design and selection problems in the future from a mission-first perspective. --- ### Leveraging High-Fidelity Digital Models and Reinforcement Learning for Mission Engineering: A Case Study of Aerial Firefighting Under Perfect Information Authors: İbrahim Oğuz Çetinkaya, Sajad Khodadadian, Taylan G. Topçu Published: 2025-12-23 Categories: cs.CY, cs.AI, eess.SY, math.OC arXiv: https://arxiv.org/abs/2512.20589v1 As systems engineering (SE) objectives evolve from design and operation of monolithic systems to complex System of Systems (SoS), the discipline of Mission Engineering (ME) has emerged which is increasingly being accepted as a new line of thinking for the SE community. Moreover, mission environments are uncertain, dynamic, and mission outcomes are a direct function of how the mission assets will interact with this environment. This proves static architectures brittle and calls for analytically rigorous approaches for ME. To that end, this paper proposes an intelligent mission coordination methodology that integrates digital mission models with Reinforcement Learning (RL), that specifically addresses the need for adaptive task allocation and reconfiguration. More specifically, we are leveraging a Digital Engineering (DE) based infrastructure that is composed of a high-fidelity digital mission model and agent-based simulation; and then we formulate the mission tactics management problem as a Markov Decision Process (MDP), and employ an RL agent trained via Proximal Policy Optimization. By leveraging the simulation as a sandbox, we map the system states to actions, refining the policy based on realized mission outcomes. The utility of the RL-based intelligent mission coordinator is demonstrated through an aerial firefighting case study. Our findings indicate that the RL-based intelligent mission coordinator not only surpasses baseline performance but also significantly reduces the variability in mission performance. Thus, this study serves as a proof of concept demonstrating that DE-enabled mission simulations combined with advanced analytical tools offer a mission-agnostic framework for improving ME practice; which can be extended to more complicated fleet design and selection problems in the future from a mission-first perspective. --- ### Automated stereotactic radiosurgery planning using a human-in-the-loop reasoning large language model agent Authors: Humza Nusrat, Luke Francisco, Bing Luo, Hassan Bagher-Ebadian, Joshua Kim, Karen Chin-Snyder, Salim Siddiqui, Mira Shah, Eric Mellon, Mohammad Ghassemi, Anthony Doemer, Benjamin Movsas et al. Published: 2025-12-23 Categories: cs.AI, cs.CL, cs.HC arXiv: https://arxiv.org/abs/2512.20586v1 Stereotactic radiosurgery (SRS) demands precise dose shaping around critical structures, yet black-box AI systems have limited clinical adoption due to opacity concerns. We tested whether chain-of-thought reasoning improves agentic planning in a retrospective cohort of 41 patients with brain metastases treated with 18 Gy single-fraction SRS. We developed SAGE (Secure Agent for Generative Dose Expertise), an LLM-based planning agent for automated SRS treatment planning. Two variants generated plans for each case: one using a non-reasoning model, one using a reasoning model. The reasoning variant showed comparable plan dosimetry relative to human planners on primary endpoints (PTV coverage, maximum dose, conformity index, gradient index; all p > 0.21) while reducing cochlear dose below human baselines (p = 0.022). When prompted to improve conformity, the reasoning model demonstrated systematic planning behaviors including prospective constraint verification (457 instances) and trade-off deliberation (609 instances), while the standard model exhibited none of these deliberative processes (0 and 7 instances, respectively). Content analysis revealed that constraint verification and causal explanation concentrated in the reasoning agent. The optimization traces serve as auditable logs, offering a path toward transparent automated planning. --- ### Relu and softplus neural nets as zero-sum turn-based games Authors: Stephane Gaubert, Yiannis Vlassopoulos Published: 2025-12-23 Categories: cs.LG, cs.GT, math.OC arXiv: https://arxiv.org/abs/2512.20582v1 We show that the output of a ReLU neural network can be interpreted as the value of a zero-sum, turn-based, stopping game, which we call the ReLU net game. The game runs in the direction opposite to that of the network, and the input of the network serves as the terminal reward of the game. In fact, evaluating the network is the same as running the Shapley-Bellman backward recursion for the value of the game. Using the expression of the value of the game as an expected total payoff with respect to the path measure induced by the transition probabilities and a pair of optimal policies, we derive a discrete Feynman-Kac-type path-integral formula for the network output. This game-theoretic representation can be used to derive bounds on the output from bounds on the input, leveraging the monotonicity of Shapley operators, and to verify robustness properties using policies as certificates. Moreover, training the neural network becomes an inverse game problem: given pairs of terminal rewards and corresponding values, one seeks transition probabilities and rewards of a game that reproduces them. Finally, we show that a similar approach applies to neural networks with Softplus activation functions, where the ReLU net game is replaced by its entropic regularization. --- ### Can LLMs Predict Their Own Failures? Self-Awareness via Internal Circuits Authors: Amirhosein Ghasemabadi, Di Niu Published: 2025-12-23 Categories: cs.CL arXiv: https://arxiv.org/abs/2512.20578v2 Large language models (LLMs) generate fluent and complex outputs but often fail to recognize their own mistakes and hallucinations. Existing approaches typically rely on external judges, multi-sample consistency, or text-based self-critique, which incur additional compute or correlate weakly with true correctness. We ask: can LLMs predict their own failures by inspecting internal states during inference? We introduce Gnosis, a lightweight self-awareness mechanism that enables frozen LLMs to perform intrinsic self-verification by decoding signals from hidden states and attention patterns. Gnosis passively observes internal traces, compresses them into fixed-budget descriptors, and predicts correctness with negligible inference cost, adding only ~5M parameters and operating independently of sequence length. Across math reasoning, open-domain question answering, and academic knowledge benchmarks, and over frozen backbones ranging from 1.7B to 20B parameters, Gnosis consistently outperforms strong internal baselines and large external judges in both accuracy and calibration. Moreover, it generalizes zero-shot to partial generations, enabling early detection of failing trajectories and compute-aware control. These results show that reliable correctness cues are intrinsic to generation process and can be extracted efficiently without external supervision. --- ### Can LLMs Predict Their Own Failures? Self-Awareness via Internal Circuits Authors: Amirhosein Ghasemabadi, Di Niu Published: 2025-12-23 Categories: cs.CL arXiv: https://arxiv.org/abs/2512.20578v1 Large language models (LLMs) generate fluent and complex outputs but often fail to recognize their own mistakes and hallucinations. Existing approaches typically rely on external judges, multi-sample consistency, or text-based self-critique, which incur additional compute or correlate weakly with true correctness. We ask: can LLMs predict their own failures by inspecting internal states during inference? We introduce Gnosis, a lightweight self-awareness mechanism that enables frozen LLMs to perform intrinsic self-verification by decoding signals from hidden states and attention patterns. Gnosis passively observes internal traces, compresses them into fixed-budget descriptors, and predicts correctness with negligible inference cost, adding only ~5M parameters and operating independently of sequence length. Across math reasoning, open-domain question answering, and academic knowledge benchmarks, and over frozen backbones ranging from 1.7B to 20B parameters, Gnosis consistently outperforms strong internal baselines and large external judges in both accuracy and calibration. Moreover, it generalizes zero-shot to partial generations, enabling early detection of failing trajectories and compute-aware control. These results show that reliable correctness cues are intrinsic to generation process and can be extracted efficiently without external supervision. --- ### Improving ML Training Data with Gold-Standard Quality Metrics Authors: Leslie Barrett, Michael W. Sherman Published: 2025-12-23 Categories: cs.LG arXiv: https://arxiv.org/abs/2512.20577v1 Hand-tagged training data is essential to many machine learning tasks. However, training data quality control has received little attention in the literature, despite data quality varying considerably with the tagging exercise. We propose methods to evaluate and enhance the quality of hand-tagged training data using statistical approaches to measure tagging consistency and agreement. We show that agreement metrics give more reliable results if recorded over multiple iterations of tagging, where declining variance in such recordings is an indicator of increasing data quality. We also show one way a tagging project can collect high-quality training data without requiring multiple tags for every work item, and that a tagger burn-in period may not be sufficient for minimizing tagger errors. --- ### Performative Policy Gradient: Optimality in Performative Reinforcement Learning Authors: Debabrota Basu, Udvas Das, Brahim Driss, Uddalak Mukherjee Published: 2025-12-23 Categories: cs.LG, cs.AI, math.OC arXiv: https://arxiv.org/abs/2512.20576v2 Post-deployment machine learning algorithms often influence the environments they act in, and thus shift the underlying dynamics that the standard reinforcement learning (RL) methods ignore. While designing optimal algorithms in this performative setting has recently been studied in supervised learning, the RL counterpart remains under-explored. In this paper, we prove the performative counterparts of the performance difference lemma and the policy gradient theorem in RL, and further introduce the Performative Policy Gradient algorithm (PePG). PePG is the first policy gradient algorithm designed to account for performativity in RL. Under softmax parametrisation, and also with and without entropy regularisation, we prove that PePG converges to performatively optimal policies, i.e. policies that remain optimal under the distribution shifts induced by themselves. Thus, PePG significantly extends the prior works in Performative RL that achieves performative stability but not optimality. Furthermore, our empirical analysis on standard performative RL environments validate that PePG outperforms the existing performative RL algorithms aiming for stability. --- ### Performative Policy Gradient: Optimality in Performative Reinforcement Learning Authors: Debabrota Basu, Udvas Das, Brahim Driss, Uddalak Mukherjee Published: 2025-12-23 Categories: cs.LG, cs.AI, math.OC arXiv: https://arxiv.org/abs/2512.20576v1 Post-deployment machine learning algorithms often influence the environments they act in, and thus shift the underlying dynamics that the standard reinforcement learning (RL) methods ignore. While designing optimal algorithms in this performative setting has recently been studied in supervised learning, the RL counterpart remains under-explored. In this paper, we prove the performative counterparts of the performance difference lemma and the policy gradient theorem in RL, and further introduce the Performative Policy Gradient algorithm (PePG). PePG is the first policy gradient algorithm designed to account for performativity in RL. Under softmax parametrisation, and also with and without entropy regularisation, we prove that PePG converges to performatively optimal policies, i.e. policies that remain optimal under the distribution shifts induced by themselves. Thus, PePG significantly extends the prior works in Performative RL that achieves performative stability but not optimality. Furthermore, our empirical analysis on standard performative RL environments validate that PePG outperforms standard policy gradient algorithms and the existing performative RL algorithms aiming for stability. --- ### Fail Fast, Win Big: Rethinking the Drafting Strategy in Speculative Decoding via Diffusion LLMs Authors: Rui Pan, Zhuofu Chen, Hongyi Liu, Arvind Krishnamurthy, Ravi Netravali Published: 2025-12-23 Categories: cs.LG, cs.AI, cs.DC arXiv: https://arxiv.org/abs/2512.20573v3 Diffusion Large Language Models (dLLMs) offer fast, parallel token generation, but their standalone use is plagued by an inherent efficiency-quality tradeoff. We show that, if carefully applied, the attributes of dLLMs can actually be a strength for drafters in speculative decoding with autoregressive (AR) verifiers. Our core insight is that dLLM's speed from parallel decoding drastically lowers the risk of costly rejections, providing a practical mechanism to effectively realize the (elusive) lengthy drafts that lead to large speedups with speculative decoding. We present FailFast, a dLLM-based speculative decoding framework that realizes this approach by dynamically adapting its speculation length. It "fails fast" by spending minimal compute in hard-to-speculate regions to shrink speculation latency and "wins big" by aggressively extending draft lengths in easier regions to reduce verification latency (in many cases, speculating and accepting 70 tokens at a time!). Without any fine-tuning, FailFast delivers lossless acceleration of AR LLMs and achieves up to 4.9$\times$ speedup over vanilla decoding, 1.7$\times$ over the best naive dLLM drafter, and 1.7$\times$ over EAGLE-3 across diverse models and workloads. We open-source FailFast at https://github.com/ruipeterpan/failfast. --- ### Fail Fast, Win Big: Rethinking the Drafting Strategy in Speculative Decoding via Diffusion LLMs Authors: Rui Pan, Zhuofu Chen, Ravi Netravali Published: 2025-12-23 Categories: cs.LG, cs.AI, cs.DC arXiv: https://arxiv.org/abs/2512.20573v1 Diffusion Large Language Models (dLLMs) offer fast, parallel token generation, but their standalone use is plagued by an inherent efficiency-quality tradeoff. We show that, if carefully applied, the attributes of dLLMs can actually be a strength for drafters in speculative decoding with autoregressive (AR) verifiers. Our core insight is that dLLM's speed from parallel decoding drastically lowers the risk of costly rejections, providing a practical mechanism to effectively realize the (elusive) lengthy drafts that lead to large speedups with speculative decoding. We present FailFast, a dLLM-based speculative decoding framework that realizes this approach by dynamically adapting its speculation length. It "fails fast" by spending minimal compute in hard-to-speculate regions to shrink speculation latency and "wins big" by aggressively extending draft lengths in easier regions to reduce verification latency (in many cases, speculating and accepting 70 tokens at a time!). Without any fine-tuning, FailFast delivers lossless acceleration of AR LLMs and achieves up to 4.9$\times$ speedup over vanilla decoding, 1.7$\times$ over the best naive dLLM drafter, and 1.4$\times$ over EAGLE-3 across diverse models and workloads. We open-source FailFast at https://github.com/ruipeterpan/failfast. --- ### Distilling to Hybrid Attention Models via KL-Guided Layer Selection Authors: Yanhong Li, Songlin Yang, Shawn Tan, Mayank Mishra, Rameswar Panda, Jiawei Zhou, Yoon Kim Published: 2025-12-23 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2512.20569v1 Distilling pretrained softmax attention Transformers into more efficient hybrid architectures that interleave softmax and linear attention layers is a promising approach for improving the inference efficiency of LLMs without requiring expensive pretraining from scratch. A critical factor in the conversion process is layer selection, i.e., deciding on which layers to convert to linear attention variants. This paper describes a simple and efficient recipe for layer selection that uses layer importance scores derived from a small amount of training on generic text data. Once the layers have been selected we use a recent pipeline for the distillation process itself \citep(RADLADS;)(){goldstein2025radlads}, which consists of attention weight transfer, hidden state alignment, KL-based distribution matching, followed by a small amount of finetuning. We find that this approach is more effective than existing approaches for layer selection, including heuristics that uniformly interleave linear attentions based on a fixed ratio, as well as more involved approaches that rely on specialized diagnostic datasets. --- ### Random Gradient-Free Optimization in Infinite Dimensional Spaces Authors: Caio Peixoto, Daniel Csillag, Bernardo F. P. da Costa, Yuri F. Saporito Published: 2025-12-23 Categories: math.OC, stat.ML arXiv: https://arxiv.org/abs/2512.20566v2 We propose a new gradient-free method for infinite-dimensional optimization in Hilbert spaces that requires only the computation of directional derivatives. Though functional optimization is often solved through finite-dimensional gradient descent over a parametrization, such as neural networks, we instead propose to leverage the functional nature of the optimization problem to enable provable guarantees. However, infinite-dimensional gradients are often hard to compute in practice, rendering naïve functional gradient descent intractable. To overcome this limitation, our framework leverages only directional derivatives and a pre-basis for the Hilbert space, i.e., a linearly independent set whose span is dense. This resolves the tractability issue, as pre-bases are much more accessible than full orthonormal bases or reproducing kernels -- which may not even exist -- and individual directional derivatives can be computed using automatic differentiation. We showcase the use of our method to solve partial differential equations à la physics-informed neural networks (PINNs), where it effectively enables provable convergence. --- ### LEAD: Minimizing Learner-Expert Asymmetry in End-to-End Driving Authors: Long Nguyen, Micha Fauth, Bernhard Jaeger, Daniel Dauner, Maximilian Igl, Andreas Geiger, Kashyap Chitta Published: 2025-12-23 Categories: cs.CV, cs.AI, cs.LG, cs.RO arXiv: https://arxiv.org/abs/2512.20563v2 Simulators can generate virtually unlimited driving data, yet imitation learning policies in simulation still struggle to achieve robust closed-loop performance. Motivated by this gap, we empirically study how misalignment between privileged expert demonstrations and sensor-based student observations can limit the effectiveness of imitation learning. More precisely, experts have significantly higher visibility (e.g., ignoring occlusions) and far lower uncertainty (e.g., knowing other vehicles' actions), making them difficult to imitate reliably. Furthermore, navigational intent (i.e., the route to follow) is under-specified in student models at test time via only a single target point. We demonstrate that these asymmetries can measurably limit driving performance in CARLA and offer practical interventions to address them. After careful modifications to narrow the gaps between expert and student, our TransFuser v6 (TFv6) student policy achieves a new state of the art on all major publicly available CARLA closed-loop benchmarks, reaching 95 DS on Bench2Drive and more than doubling prior performances on Longest6~v2 and Town13. Additionally, by integrating perception supervision from our dataset into a shared sim-to-real pipeline, we show consistent gains on the NAVSIM and Waymo Vision-Based End-to-End driving benchmarks. Our code, data, and models are publicly available at https://github.com/autonomousvision/lead. --- ### LEAD: Minimizing Learner-Expert Asymmetry in End-to-End Driving Authors: Long Nguyen, Micha Fauth, Bernhard Jaeger, Daniel Dauner, Maximilian Igl, Andreas Geiger, Kashyap Chitta Published: 2025-12-23 Categories: cs.CV, cs.AI, cs.LG, cs.RO arXiv: https://arxiv.org/abs/2512.20563v1 Simulators can generate virtually unlimited driving data, yet imitation learning policies in simulation still struggle to achieve robust closed-loop performance. Motivated by this gap, we empirically study how misalignment between privileged expert demonstrations and sensor-based student observations can limit the effectiveness of imitation learning. More precisely, experts have significantly higher visibility (e.g., ignoring occlusions) and far lower uncertainty (e.g., knowing other vehicles' actions), making them difficult to imitate reliably. Furthermore, navigational intent (i.e., the route to follow) is under-specified in student models at test time via only a single target point. We demonstrate that these asymmetries can measurably limit driving performance in CARLA and offer practical interventions to address them. After careful modifications to narrow the gaps between expert and student, our TransFuser v6 (TFv6) student policy achieves a new state of the art on all major publicly available CARLA closed-loop benchmarks, reaching 95 DS on Bench2Drive and more than doubling prior performances on Longest6~v2 and Town13. Additionally, by integrating perception supervision from our dataset into a shared sim-to-real pipeline, we show consistent gains on the NAVSIM and Waymo Vision-Based End-to-End driving benchmarks. Our code, data, and models are publicly available at https://github.com/autonomousvision/lead. --- ### Shallow Neural Networks Learn Low-Degree Spherical Polynomials with Feature Learning by Learnable Channel Attention Authors: Yingzhen Yang Published: 2025-12-23 Categories: stat.ML, cs.LG, math.OC arXiv: https://arxiv.org/abs/2512.20562v2 We study the problem of learning a low-degree spherical polynomial of degree $\ell_0 = Θ(1) \ge 1$ defined on the unit sphere in $\RR^d$ by training an over-parameterized two-layer neural network (NN) with channel attention in this paper. Our main result is the significantly improved sample complexity for learning such low-degree polynomials. We show that, for any regression risk $\eps \in (0,1)$, a carefully designed two-layer NN with channel attention and finite width trained by the vanilla gradient descent (GD) requires the lowest sample complexity of $n \asymp Θ(d^{\ell_0}/\eps)$ with high probability, in contrast with the representative sample complexity $Θ\pth{d^{\ell_0} \max\set{\eps^{-2},\log d}}$, where $n$ is the training data size. Moreover, such sample complexity is not improvable since the trained network renders a sharp rate of the nonparametric regression risk of the order $Θ(d^{\ell_0}/{n})$ with high probability. On the other hand, the minimax optimal rate for the regression risk with a kernel of rank $Θ(d^{\ell_0})$ is $Θ(d^{\ell_0}/{n})$, so that the rate of the nonparametric regression risk of the network trained by GD is minimax optimal. Training the two-layer NN with channel attention proceeds in two stages: (1) a provable learnable channel selection algorithm, as a learnable harmonic-degree selection process, identifies the ground truth channel number in the target function, $\ell_0$, from $L \ge \ell_0$ channels in the first-layer activation; (2) the second layer is trained by standard GD using the selected channels. To the best of our knowledge, this is the first time a minimax optimal risk bound is obtained by training an over-parameterized but finite-width neural network with feature learning capability to learn low-degree spherical polynomials. --- ### Shallow Neural Networks Learn Low-Degree Spherical Polynomials with Learnable Channel Attention Authors: Yingzhen Yang Published: 2025-12-23 Categories: stat.ML, cs.LG, math.OC arXiv: https://arxiv.org/abs/2512.20562v1 We study the problem of learning a low-degree spherical polynomial of degree $\ell_0 = Θ(1) \ge 1$ defined on the unit sphere in $\RR^d$ by training an over-parameterized two-layer neural network (NN) with channel attention in this paper. Our main result is the significantly improved sample complexity for learning such low-degree polynomials. We show that, for any regression risk $\eps \in (0,1)$, a carefully designed two-layer NN with channel attention and finite width of $m \ge Θ({n^4 \log (2n/δ)}/{d^{2\ell_0}})$ trained by the vanilla gradient descent (GD) requires the lowest sample complexity of $n \asymp Θ(d^{\ell_0}/\eps)$ with probability $1-δ$ for every $δ\in (0,1)$, in contrast with the representative sample complexity $Θ\pth{d^{\ell_0} \max\set{\eps^{-2},\log d}}$, where $n$ is the training daata size. Moreover, such sample complexity is not improvable since the trained network renders a sharp rate of the nonparametric regression risk of the order $Θ(d^{\ell_0}/{n})$ with probability at least $1-δ$. On the other hand, the minimax optimal rate for the regression risk with a kernel of rank $Θ(d^{\ell_0})$ is $Θ(d^{\ell_0}/{n})$, so that the rate of the nonparametric regression risk of the network trained by GD is minimax optimal. The training of the two-layer NN with channel attention consists of two stages. In Stage 1, a provable learnable channel selection algorithm identifies the ground-truth channel number $\ell_0$ from the initial $L \ge \ell_0$ channels in the first-layer activation, with high probability. This learnable selection is achieved by an efficient one-step GD update on both layers, enabling feature learning for low-degree polynomial targets. In Stage 2, the second layer is trained by standard GD using the activation function with the selected channels. --- ### FlashVLM: Text-Guided Visual Token Selection for Large Multimodal Models Authors: Kaitong Cai, Jusheng Zhang, Jing Yang, Yijia Fan, Pengtao Xie, Jian Wang, Keze Wang Published: 2025-12-23 Categories: cs.CV arXiv: https://arxiv.org/abs/2512.20561v1 Large vision-language models (VLMs) typically process hundreds or thousands of visual tokens per image or video frame, incurring quadratic attention cost and substantial redundancy. Existing token reduction methods often ignore the textual query or rely on deep attention maps, whose instability under aggressive pruning leads to degraded semantic alignment. We propose FlashVLM, a text guided visual token selection framework that dynamically adapts visual inputs to the query. Instead of relying on noisy attention weights, FlashVLM computes an explicit cross modal similarity between projected image tokens and normalized text embeddings in the language model space. This extrinsic relevance is fused with intrinsic visual saliency using log domain weighting and temperature controlled sharpening. In addition, a diversity preserving partition retains a minimal yet representative set of background tokens to maintain global context. Under identical token budgets and evaluation protocols, FlashVLM achieves beyond lossless compression, slightly surpassing the unpruned baseline while pruning up to 77.8 percent of visual tokens on LLaVA 1.5, and maintaining 92.8 percent accuracy even under 94.4 percent compression. Extensive experiments on 14 image and video benchmarks demonstrate that FlashVLM delivers state of the art efficiency performance trade offs while maintaining strong robustness and generalization across mainstream VLMs. --- ### Learning to Reason in 4D: Dynamic Spatial Understanding for Vision Language Models Authors: Shengchao Zhou, Yuxin Chen, Yuying Ge, Wei Huang, Jiehong Lin, Ying Shan, Xiaojuan Qi Published: 2025-12-23 Categories: cs.CV arXiv: https://arxiv.org/abs/2512.20557v1 Vision-language models (VLM) excel at general understanding yet remain weak at dynamic spatial reasoning (DSR), i.e., reasoning about the evolvement of object geometry and relationship in 3D space over time, largely due to the scarcity of scalable 4D-aware training resources. To bridge this gap across aspects of dataset, benchmark and model, we introduce DSR Suite. First, we propose an automated pipeline that generates multiple-choice question-answer pairs from in-the-wild videos for DSR. By leveraging modern vision foundation models, the pipeline extracts rich geometric and motion information, including camera poses, local point clouds, object masks, orientations, and 3D trajectories. These geometric cues enable the construction of DSR-Train for learning and further human-refined DSR-Bench for evaluation. Compared with previous works, our data emphasize (i) in-the-wild video sources, (ii) object- and scene-level 3D requirements, (iii) viewpoint transformations, (iv) multi-object interactions, and (v) fine-grained, procedural answers. Beyond data, we propose a lightweight Geometry Selection Module (GSM) to seamlessly integrate geometric priors into VLMs, which condenses question semantics and extracts question-relevant knowledge from pretrained 4D reconstruction priors into a compact set of geometry tokens. This targeted extraction avoids overwhelming the model with irrelevant knowledge. Experiments show that integrating DSR-Train and GSM into Qwen2.5-VL-7B significantly enhances its dynamic spatial reasoning capability, while maintaining accuracy on general video understanding benchmarks. --- ### Multi-Grained Text-Guided Image Fusion for Multi-Exposure and Multi-Focus Scenarios Authors: Mingwei Tang, Jiahao Nie, Guang Yang, Ziqing Cui, Jie Li Published: 2025-12-23 Categories: cs.CV arXiv: https://arxiv.org/abs/2512.20556v1 Image fusion aims to synthesize a single high-quality image from a pair of inputs captured under challenging conditions, such as differing exposure levels or focal depths. A core challenge lies in effectively handling disparities in dynamic range and focus depth between the inputs. With the advent of vision-language models, recent methods incorporate textual descriptions as auxiliary guidance to enhance fusion quality. However, simply incorporating coarse-grained descriptions hampers the understanding of fine-grained details and poses challenges for precise cross-modal alignment. To address these limitations, we propose Multi-grained Text-guided Image Fusion (MTIF), a novel fusion paradigm with three key designs. First, it introduces multi-grained textual descriptions that separately capture fine details, structural cues, and semantic content, guiding image fusion through a hierarchical cross-modal modulation module. Second, it involves supervision signals at each granularity to facilitate alignment between visual and textual features and enhance the utility of auxiliary text. Third, it adopts a saliency-driven enrichment module to augment training data with dense semantic content, further strengthening the cross-modal modulation and alignment. Extensive experiments show that MTIF consistently outperforms previous methods on both multi-exposure and multi-focus image fusion tasks. --- ### Information-theoretic signatures of causality in Bayesian networks and hypergraphs Authors: Sung En Chiang, Zhaolu Liu, Robert L. Peach, Mauricio Barahona Published: 2025-12-23 Categories: cs.IT, stat.ML arXiv: https://arxiv.org/abs/2512.20552v2 Analyzing causality in multivariate systems involves establishing how information is generated, distributed and combined. Traditional causal discovery frameworks are capable of multivariate reasoning but their intrinsic pairwise graph topology restricts them to do so only indirectly by integrating multivariate information across pairwise edges. Higher-order information theory provides direct tools that can explicitly model higher-order interactions. In particular, Partial Information Decomposition (PID) allows the decomposition of the information that a set of sources provides about a target into redundant, unique, and synergistic components. Yet the mathematical connection between such higher-order information-theoretic measures and causal structure remains undeveloped. Here we establish the first theoretical correspondence between PID components and causal structure in both Bayesian networks and hypergraphs. We first show that in Bayesian networks unique information precisely characterizes direct causal neighbors, while synergy identifies collider relationships. This establishes a localist causal discovery paradigm in which the structure surrounding each variable can be recovered from its immediate informational footprint, eliminating the need for global search over graph space. Extending these results to more expressive causal representation, we prove that PID signatures in Bayesian hypergraphs differentiate parents, children, co-heads, and co-tails, revealing a novel collider effect unique to multi-tail hyperedges. Our results position PID as a rigorous, model-agnostic foundation for inferring both pairwise and higher-order causal structure, and introduce a fundamentally local information-theoretic viewpoint on causal discovery. ---