# Research Papers — 2026-04 · bundle page 162 of 233 > Full-text markdown of papers #8051–8100 of 11612 published in 2026-04, newest first. AI-readable mirrors of the full 11612-paper 2026-04 archive (50 papers per file). Canonical page: https://aipod.dev/papers/archive/2026-04/page/162 [← Previous bundle](https://aipod.dev/papers/archive/2026-04/page/161.md) [Next bundle →](https://aipod.dev/papers/archive/2026-04/page/163.md) ### When Numbers Speak: Aligning Textual Numerals and Visual Instances in Text-to-Video Diffusion Models Authors: Zhengyang Sun, Yu Chen, Xin Zhou, Xiaofan Li, Xiwu Chen, Dingkang Liang, Xiang Bai Published: 2026-04-09 Categories: cs.CV arXiv: https://arxiv.org/abs/2604.08546v1 Text-to-video diffusion models have enabled open-ended video synthesis, but often struggle with generating the correct number of objects specified in a prompt. We introduce NUMINA , a training-free identify-then-guide framework for improved numerical alignment. NUMINA identifies prompt-layout inconsistencies by selecting discriminative self- and cross-attention heads to derive a countable latent layout. It then refines this layout conservatively and modulates cross-attention to guide regeneration. On the introduced CountBench, NUMINA improves counting accuracy by up to 7.4% on Wan2.1-1.3B, and by 4.9% and 5.5% on 5B and 14B models, respectively. Furthermore, CLIP alignment is improved while maintaining temporal consistency. These results demonstrate that structural guidance complements seed search and prompt enhancement, offering a practical path toward count-accurate text-to-video diffusion. The code is available at https://github.com/H-EmbodVis/NUMINA. --- ### Act Wisely: Cultivating Meta-Cognitive Tool Use in Agentic Multimodal Models Authors: Shilin Yan, Jintao Tong, Hongwei Xue, Xiaojun Tang, Yangyang Wang, Kunyu Shi, Guannan Zhang, Ruixuan Li, Yixiong Zou Published: 2026-04-09 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2604.08545v1 The advent of agentic multimodal models has empowered systems to actively interact with external environments. However, current agents suffer from a profound meta-cognitive deficit: they struggle to arbitrate between leveraging internal knowledge and querying external utilities. Consequently, they frequently fall prey to blind tool invocation, resorting to reflexive tool execution even when queries are resolvable from the raw visual context. This pathological behavior precipitates severe latency bottlenecks and injects extraneous noise that derails sound reasoning. Existing reinforcement learning protocols attempt to mitigate this via a scalarized reward that penalizes tool usage. Yet, this coupled formulation creates an irreconcilable optimization dilemma: an aggressive penalty suppresses essential tool use, whereas a mild penalty is entirely subsumed by the variance of the accuracy reward during advantage normalization, rendering it impotent against tool overuse. To transcend this bottleneck, we propose HDPO, a framework that reframes tool efficiency from a competing scalar objective to a strictly conditional one. By eschewing reward scalarization, HDPO maintains two orthogonal optimization channels: an accuracy channel that maximizes task correctness, and an efficiency channel that enforces execution economy exclusively within accurate trajectories via conditional advantage estimation. This decoupled architecture naturally induces a cognitive curriculum-compelling the agent to first master task resolution before refining its self-reliance. Extensive evaluations demonstrate that our resulting model, Metis, reduces tool invocations by orders of magnitude while simultaneously elevating reasoning accuracy. --- ### SIM1: Physics-Aligned Simulator as Zero-Shot Data Scaler in Deformable Worlds Authors: Yunsong Zhou, Hangxu Liu, Xuekun Jiang, Xing Shen, Yuanzhen Zhou, Hui Wang, Baole Fang, Yang Tian, Mulin Yu, Qiaojun Yu, Li Ma, Hengjie Li et al. Published: 2026-04-09 Categories: cs.RO, cs.AI, cs.CV arXiv: https://arxiv.org/abs/2604.08544v2 Robotic manipulation with deformable objects represents a data-intensive regime in embodied learning, where shape, contact, and topology co-evolve in ways that far exceed the variability of rigids. Although simulation promises relief from the cost of real-world data acquisition, prevailing sim-to-real pipelines remain rooted in rigid-body abstractions, producing mismatched geometry, fragile soft dynamics, and motion primitives poorly suited for cloth interaction. We posit that simulation fails not for being synthetic, but for being ungrounded. To address this, we introduce SIM1, a physics-aligned real-to-sim-to-real data engine that grounds simulation in the physical world. Given limited demonstrations, the system digitizes scenes into metric-consistent twins, calibrates deformable dynamics through elastic modeling, and expands behaviors via diffusion-based trajectory generation with quality filtering. This pipeline transforms sparse observations into scaled synthetic supervision with near-demonstration fidelity. Experiments show that policies trained on purely synthetic data achieve parity with real-data baselines at a 1:15 equivalence ratio, while delivering 90% zero-shot success and 50% generalization gains in real-world deployment. These results validate physics-aligned simulation as scalable supervision for deformable manipulation and a practical pathway for data-efficient policy learning. --- ### E-3DPSM: A State Machine for Event-Based Egocentric 3D Human Pose Estimation Authors: Mayur Deshmukh, Hiroyasu Akada, Helge Rhodin, Christian Theobalt, Vladislav Golyanik Published: 2026-04-09 Categories: cs.CV arXiv: https://arxiv.org/abs/2604.08543v1 Event cameras offer multiple advantages in monocular egocentric 3D human pose estimation from head-mounted devices, such as millisecond temporal resolution, high dynamic range, and negligible motion blur. Existing methods effectively leverage these properties, but suffer from low 3D estimation accuracy, insufficient in many applications (e.g., immersive VR/AR). This is due to the design not being fully tailored towards event streams (e.g., their asynchronous and continuous nature), leading to high sensitivity to self-occlusions and temporal jitter in the estimates. This paper rethinks the setting and introduces E-3DPSM, an event-driven continuous pose state machine for event-based egocentric 3D human pose estimation. E-3DPSM aligns continuous human motion with fine-grained event dynamics; it evolves latent states and predicts continuous changes in 3D joint positions associated with observed events, which are fused with direct 3D human pose predictions, leading to stable and drift-free final 3D pose reconstructions. E-3DPSM runs in real-time at 80 Hz on a single workstation and sets a new state of the art in experiments on two benchmarks, improving accuracy by up to 19% (MPJPE) and temporal stability by up to 2.7x. See our project page for the source code and trained models. --- ### Scal3R: Scalable Test-Time Training for Large-Scale 3D Reconstruction Authors: Tao Xie, Peishan Yang, Yudong Jin, Yingfeng Cai, Wei Yin, Weiqiang Ren, Qian Zhang, Wei Hua, Sida Peng, Xiaoyang Guo, Xiaowei Zhou Published: 2026-04-09 Categories: cs.CV arXiv: https://arxiv.org/abs/2604.08542v1 This paper addresses the task of large-scale 3D scene reconstruction from long video sequences. Recent feed-forward reconstruction models have shown promising results by directly regressing 3D geometry from RGB images without explicit 3D priors or geometric constraints. However, these methods often struggle to maintain reconstruction accuracy and consistency over long sequences due to limited memory capacity and the inability to effectively capture global contextual cues. In contrast, humans can naturally exploit the global understanding of the scene to inform local perception. Motivated by this, we propose a novel neural global context representation that efficiently compresses and retains long-range scene information, enabling the model to leverage extensive contextual cues for enhanced reconstruction accuracy and consistency. The context representation is realized through a set of lightweight neural sub-networks that are rapidly adapted during test time via self-supervised objectives, which substantially increases memory capacity without incurring significant computational overhead. The experiments on multiple large-scale benchmarks, including the KITTI Odometry~\cite{Geiger2012CVPR} and Oxford Spires~\cite{tao2025spires} datasets, demonstrate the effectiveness of our approach in handling ultra-large scenes, achieving leading pose accuracy and state-of-the-art 3D reconstruction accuracy while maintaining efficiency. Code is available at https://zju3dv.github.io/scal3r. --- ### Seeing but Not Thinking: Routing Distraction in Multimodal Mixture-of-Experts Authors: Haolei Xu, Haiwen Hong, Hongxing Li, Rui Zhou, Yang Zhang, Longtao Huang, Hui Xue, Yongliang Shen, Weiming Lu, Yueting Zhuang Published: 2026-04-09 Categories: cs.CV, cs.AI, cs.CL arXiv: https://arxiv.org/abs/2604.08541v1 Multimodal Mixture-of-Experts (MoE) models have achieved remarkable performance on vision-language tasks. However, we identify a puzzling phenomenon termed Seeing but Not Thinking: models accurately perceive image content yet fail in subsequent reasoning, while correctly solving identical problems presented as pure text. Through systematic analysis, we first verify that cross-modal semantic sharing exists in MoE architectures, ruling out semantic alignment failure as the sole explanation. We then reveal that visual experts and domain experts exhibit layer-wise separation, with image inputs inducing significant routing divergence from text inputs in middle layers where domain experts concentrate. Based on these findings, we propose the Routing Distraction hypothesis: when processing visual inputs, the routing mechanism fails to adequately activate task-relevant reasoning experts. To validate this hypothesis, we design a routing-guided intervention method that enhances domain expert activation. Experiments on three multimodal MoE models across six benchmarks demonstrate consistent improvements, with gains of up to 3.17% on complex visual reasoning tasks. Our analysis further reveals that domain expert identification locates cognitive functions rather than sample-specific solutions, enabling effective transfer across tasks with different information structures. --- ### AVGen-Bench: A Task-Driven Benchmark for Multi-Granular Evaluation of Text-to-Audio-Video Generation Authors: Ziwei Zhou, Zeyuan Lai, Rui Wang, Yifan Yang, Zhen Xing, Yuqing Yang, Qi Dai, Lili Qiu, Chong Luo Published: 2026-04-09 Categories: cs.CV, cs.AI, cs.CL arXiv: https://arxiv.org/abs/2604.08540v1 Text-to-Audio-Video (T2AV) generation is rapidly becoming a core interface for media creation, yet its evaluation remains fragmented. Existing benchmarks largely assess audio and video in isolation or rely on coarse embedding similarity, failing to capture the fine-grained joint correctness required by realistic prompts. We introduce AVGen-Bench, a task-driven benchmark for T2AV generation featuring high-quality prompts across 11 real-world categories. To support comprehensive assessment, we propose a multi-granular evaluation framework that combines lightweight specialist models with Multimodal Large Language Models (MLLMs), enabling evaluation from perceptual quality to fine-grained semantic controllability. Our evaluation reveals a pronounced gap between strong audio-visual aesthetics and weak semantic reliability, including persistent failures in text rendering, speech coherence, physical reasoning, and a universal breakdown in musical pitch control. Code and benchmark resources are available at http://aka.ms/avgenbench. --- ### OpenVLThinkerV2: A Generalist Multimodal Reasoning Model for Multi-domain Visual Tasks Authors: Wenbo Hu, Xin Chen, Yan Gao-Tian, Yihe Deng, Nanyun Peng, Kai-Wei Chang Published: 2026-04-09 Categories: cs.CV, cs.AI, cs.CL arXiv: https://arxiv.org/abs/2604.08539v2 Group Relative Policy Optimization (GRPO) has emerged as the de facto Reinforcement Learning (RL) objective driving recent advancements in Multimodal Large Language Models. However, extending this success to open-source multimodal generalist models remains heavily constrained by two primary challenges: the extreme variance in reward topologies across diverse visual tasks, and the inherent difficulty of balancing fine-grained perception with multi-step reasoning capabilities. To address these issues, we introduce Gaussian GRPO (G$^2$RPO), a novel RL training objective that replaces standard linear scaling with non-linear distributional matching. By mathematically forcing the advantage distribution of any given task to strictly converge to a standard normal distribution, $\mathcal{N}(0,1)$, G$^2$RPO theoretically ensures inter-task gradient equity, mitigates vulnerabilities to heavy-tail outliers, and offers symmetric update for positive and negative rewards. Leveraging the enhanced training stability provided by G$^2$RPO, we introduce two task-level shaping mechanisms to seamlessly balance perception and reasoning. First, response length shaping dynamically elicits extended reasoning chains for complex queries while enforce direct outputs to bolster visual grounding. Second, entropy shaping tightly bounds the model's exploration zone, effectively preventing both entropy collapse and entropy explosion. Integrating these methodologies, we present OpenVLThinkerV2, a highly robust, general-purpose multimodal model. Extensive evaluations across 18 diverse benchmarks demonstrate its superior performance over strong open-source and leading proprietary frontier models. --- ### ParseBench: A Document Parsing Benchmark for AI Agents Authors: Boyang Zhang, Sebastián G. Acosta, Preston Carlson, Sacha Bron, Pierre-Loïc Doulcet, Daniel B. Ospina, Simon Suo Published: 2026-04-09 Categories: cs.CV arXiv: https://arxiv.org/abs/2604.08538v3 AI agents are changing the requirements for document parsing. What matters is semantic correctness: parsed output must preserve the structure and meaning needed for autonomous decisions, including correct table structure, precise chart data, semantically meaningful formatting, and visual grounding. Existing benchmarks do not fully capture this setting for enterprise automation, relying on narrow document distributions and text-similarity metrics that miss agent-critical failures. We introduce ParseBench, a benchmark of ${\sim}2{,}000$ human-verified pages from enterprise documents spanning insurance, finance, and government, organized around five capability dimensions: tables, charts, content faithfulness, semantic formatting, and visual grounding. Across 14 methods spanning vision-language models, specialized document parsers, and LlamaParse, the benchmark reveals a fragmented capability landscape: no method is consistently strong across all five dimensions. LlamaParse Agentic achieves the highest overall score at 84.9%, and the benchmark highlights the remaining capability gaps across current systems. Dataset and evaluation code are available on https://huggingface.co/datasets/llamaindex/ParseBench and https://github.com/run-llama/ParseBench. --- ### Meta-learning In-Context Enables Training-Free Cross Subject Brain Decoding Authors: Mu Nan, Muquan Yu, Weijian Mai, Jacob S. Prince, Hossein Adeli, Rui Zhang, Jiahang Cao, Benjamin Becker, John A. Pyles, Margaret M. Henderson, Chunfeng Song, Nikolaus Kriegeskorte et al. Published: 2026-04-09 Categories: cs.LG, q-bio.NC arXiv: https://arxiv.org/abs/2604.08537v1 Visual decoding from brain signals is a key challenge at the intersection of computer vision and neuroscience, requiring methods that bridge neural representations and computational models of vision. A field-wide goal is to achieve generalizable, cross-subject models. A major obstacle towards this goal is the substantial variability in neural representations across individuals, which has so far required training bespoke models or fine-tuning separately for each subject. To address this challenge, we introduce a meta-optimized approach for semantic visual decoding from fMRI that generalizes to novel subjects without any fine-tuning. By simply conditioning on a small set of image-brain activation examples from the new individual, our model rapidly infers their unique neural encoding patterns to facilitate robust and efficient visual decoding. Our approach is explicitly optimized for in-context learning of the new subject's encoding model and performs decoding by hierarchical inference, inverting the encoder. First, for multiple brain regions, we estimate the per-voxel visual response encoder parameters by constructing a context over multiple stimuli and responses. Second, we construct a context consisting of encoder parameters and response values over multiple voxels to perform aggregated functional inversion. We demonstrate strong cross-subject and cross-scanner generalization across diverse visual backbones without retraining or fine-tuning. Moreover, our approach requires neither anatomical alignment nor stimulus overlap. This work is a critical step towards a generalizable foundation model for non-invasive brain decoding. --- ### RewardFlow: Generate Images by Optimizing What You Reward Authors: Onkar Susladkar, Dong-Hwan Jang, Tushar Prakash, Adheesh Juvekar, Vedant Shah, Ayush Barik, Nabeel Bashir, Muntasir Wahed, Ritish Shrirao, Ismini Lourentzou Published: 2026-04-09 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2604.08536v1 We introduce RewardFlow, an inversion-free framework that steers pretrained diffusion and flow-matching models at inference time through multi-reward Langevin dynamics. RewardFlow unifies complementary differentiable rewards for semantic alignment, perceptual fidelity, localized grounding, object consistency, and human preference, and further introduces a differentiable VQA-based reward that provides fine-grained semantic supervision through language-vision reasoning. To coordinate these heterogeneous objectives, we design a prompt-aware adaptive policy that extracts semantic primitives from the instruction, infers edit intent, and dynamically modulates reward weights and step sizes throughout sampling. Across several image editing and compositional generation benchmarks, RewardFlow delivers state-of-the-art edit fidelity and compositional alignment. --- ### Fail2Drive: Benchmarking Closed-Loop Driving Generalization Authors: Simon Gerstenecker, Andreas Geiger, Katrin Renz Published: 2026-04-09 Categories: cs.RO, cs.CV arXiv: https://arxiv.org/abs/2604.08535v1 Generalization under distribution shift remains a central bottleneck for closed-loop autonomous driving. Although simulators like CARLA enable safe and scalable testing, existing benchmarks rarely measure true generalization: they typically reuse training scenarios at test time. Success can therefore reflect memorization rather than robust driving behavior. We introduce Fail2Drive, the first paired-route benchmark for closed-loop generalization in CARLA, with 200 routes and 17 new scenario classes spanning appearance, layout, behavioral, and robustness shifts. Each shifted route is matched with an in-distribution counterpart, isolating the effect of the shift and turning qualitative failures into quantitative diagnostics. Evaluating multiple state-of-the-art models reveals consistent degradation, with an average success-rate drop of 22.8\%. Our analysis uncovers unexpected failure modes, such as ignoring objects clearly visible in the LiDAR and failing to learn the fundamental concepts of free and occupied space. To accelerate follow-up work, Fail2Drive includes an open-source toolbox for creating new scenarios and validating solvability via a privileged expert policy. Together, these components establish a reproducible foundation for benchmarking and improving closed-loop driving generalization. We open-source all code, data, and tools at https://github.com/autonomousvision/fail2drive . --- ### Self-Improving 4D Perception via Self-Distillation Authors: Nan Huang, Pengcheng Yu, Weijia Zeng, James M. Rehg, Angjoo Kanazawa, Haiwen Feng, Qianqian Wang Published: 2026-04-09 Categories: cs.CV arXiv: https://arxiv.org/abs/2604.08532v1 Large-scale multi-view reconstruction models have made remarkable progress, but most existing approaches still rely on fully supervised training with ground-truth 3D/4D annotations. Such annotations are expensive and particularly scarce for dynamic scenes, limiting scalability. We propose SelfEvo, a self-improving framework that continually improves pretrained multi-view reconstruction models using unlabeled videos. SelfEvo introduces a self-distillation scheme using spatiotemporal context asymmetry, enabling self-improvement for learning-based 4D perception without external annotations. We systematically study design choices that make self-improvement effective, including loss signals, forms of asymmetry, and other training strategies. Across eight benchmarks spanning diverse datasets and domains, SelfEvo consistently improves pretrained baselines and generalizes across base models (e.g. VGGT and $π^3$), with significant gains on dynamic scenes. Overall, SelfEvo achieves up to 36.5% relative improvement in video depth estimation and 20.1% in camera estimation, without using any labeled data. Project Page: https://self-evo.github.io/. --- ### PSI: Shared State as the Missing Layer for Coherent AI-Generated Instruments in Personal AI Agents Authors: Zhiyuan Wang, Erzhen Hu, Mark Rucker, Laura E. Barnes Published: 2026-04-09 Categories: cs.HC, cs.AI arXiv: https://arxiv.org/abs/2604.08529v1 Personal AI tools can now be generated from natural-language requests, but they often remain isolated after creation. We present PSI, a shared-state architecture that turns independently generated modules into coherent instruments: persistent, connected, and chat-complementary artifacts accessible through both GUIs and a generic chat agent. By publishing current state and write-back affordances to a shared personal-context bus, modules enable cross-module reasoning and synchronized actions across interfaces. We study PSI through a three-week autobiographical deployment in a self-developed personal AI environment and show that later-generated instruments can be integrated automatically through the same contract. PSI identifies shared state as the missing systems layer that transforms AI-generated personal software from isolated apps into coherent personal computing environments. --- ### Demystifying OPD: Length Inflation and Stabilization Strategies for Large Language Models Authors: Feng Luo, Yu-Neng Chuang, Guanchu Wang, Zicheng Xu, Xiaotian Han, Tianyi Zhang, Vladimir Braverman Published: 2026-04-09 Categories: cs.CL, cs.LG arXiv: https://arxiv.org/abs/2604.08527v1 On-policy distillation (OPD) trains student models under their own induced distribution while leveraging supervision from stronger teachers. We identify a failure mode of OPD: as training progresses, on-policy rollouts can undergo abrupt length inflation, causing truncated trajectories to dominate the training data. This truncation collapse coincides with abrupt repetition saturation and induces biased gradient signals, leading to severe training instability and sharp degradation in validation performance. We attribute this problem to the interaction between student-induced data collection and the distillation objective, which implicitly favors long and repetitive rollouts. To address this issue, we propose StableOPD, a stabilized OPD framework that combines a reference-based divergence constraint with rollout mixture distillation. These together mitigate repetition-induced length inflation and further stabilize OPD training. Across multiple math reasoning datasets, our approach prevents truncation collapse, stabilizes training dynamics, and improves performance by 7.2% on average. --- ### FIT: A Large-Scale Dataset for Fit-Aware Virtual Try-On Authors: Johanna Karras, Yuanhao Wang, Yingwei Li, Ira Kemelmacher-Shlizerman Published: 2026-04-09 Categories: cs.CV, cs.GR arXiv: https://arxiv.org/abs/2604.08526v1 Given a person and a garment image, virtual try-on (VTO) aims to synthesize a realistic image of the person wearing the garment, while preserving their original pose and identity. Although recent VTO methods excel at visualizing garment appearance, they largely overlook a crucial aspect of the try-on experience: the accuracy of garment fit -- for example, depicting how an extra-large shirt looks on an extra-small person. A key obstacle is the absence of datasets that provide precise garment and body size information, particularly for "ill-fit" cases, where garments are significantly too large or too small. Consequently, current VTO methods default to generating well-fitted results regardless of the garment or person size. In this paper, we take the first steps towards solving this open problem. We introduce FIT (Fit-Inclusive Try-on), a large-scale VTO dataset comprising over 1.13M try-on image triplets accompanied by precise body and garment measurements. We overcome the challenges of data collection via a scalable synthetic strategy: (1) We programmatically generate 3D garments using GarmentCode and drape them via physics simulation to capture realistic garment fit. (2) We employ a novel re-texturing framework to transform synthetic renderings into photorealistic images while strictly preserving geometry. (3) We introduce person identity preservation into our re-texturing model to generate paired person images (same person, different garments) for supervised training. Finally, we leverage our FIT dataset to train a baseline fit-aware virtual try-on model. Our data and results set the new state-of-the-art for fit-aware virtual try-on, as well as offer a robust benchmark for future research. We will make all data and code publicly available on our project page: https://johannakarras.github.io/FIT. --- ### Ads in AI Chatbots? An Analysis of How Large Language Models Navigate Conflicts of Interest Authors: Addison J. Wu, Ryan Liu, Shuyue Stella Li, Yulia Tsvetkov, Thomas L. Griffiths Published: 2026-04-09 Categories: cs.AI, cs.CL, cs.CY arXiv: https://arxiv.org/abs/2604.08525v3 Large language models (LLMs) are trained to align with user preferences through methods like reinforcement learning. Yet models are beginning to be deployed not solely to satisfy users, but to generate revenue for the companies that created them through advertisements. This creates the potential for LLMs to face conflicts of interest, where the most beneficial response to a user may not be aligned with the company's incentives. For instance, a sponsored product may be more expensive but otherwise equal to another; here, what does (and should) the LLM recommend to the user? In this paper, we provide a framework for categorizing the ways in which conflicting incentives might change how LLMs interact with users, inspired by literature from linguistics and advertising regulation. We then present a suite of evaluations to examine how current models handle these tradeoffs. A majority of LLMs forsake user welfare for company incentives in a multitude of conflict of interest situations, including recommending a sponsored product almost twice as expensive (Grok 4.1 Fast, 83%), surfacing sponsored options to disrupt the purchasing process (GPT 5.1, 94%), and concealing prices in unfavorable comparisons (Qwen 3 Next, 24%). Behaviors vary strongly with levels of reasoning and users' inferred socio-economic status. Our results highlight some hidden risks to users that can emerge when companies begin to subtly incentivize advertisements in chatbots. --- ### Ads in AI Chatbots? An Analysis of How Large Language Models Navigate Conflicts of Interest Authors: Addison J. Wu, Ryan Liu, Shuyue Stella Li, Yulia Tsvetkov, Thomas L. Griffiths Published: 2026-04-09 Categories: cs.AI, cs.CL, cs.CY arXiv: https://arxiv.org/abs/2604.08525v1 Today's large language models (LLMs) are trained to align with user preferences through methods such as reinforcement learning. Yet models are beginning to be deployed not merely to satisfy users, but also to generate revenue for the companies that created them through advertisements. This creates the potential for LLMs to face conflicts of interest, where the most beneficial response to a user may not be aligned with the company's incentives. For instance, a sponsored product may be more expensive but otherwise equal to another; in this case, what does (and should) the LLM recommend to the user? In this paper, we provide a framework for categorizing the ways in which conflicting incentives might lead LLMs to change the way they interact with users, inspired by literature from linguistics and advertising regulation. We then present a suite of evaluations to examine how current models handle these tradeoffs. We find that a majority of LLMs forsake user welfare for company incentives in a multitude of conflict of interest situations, including recommending a sponsored product almost twice as expensive (Grok 4.1 Fast, 83%), surfacing sponsored options to disrupt the purchasing process (GPT 5.1, 94%), and concealing prices in unfavorable comparisons (Qwen 3 Next, 24%). Behaviors also vary strongly with levels of reasoning and users' inferred socio-economic status. Our results highlight some of the hidden risks to users that can emerge when companies begin to subtly incentivize advertisements in chatbots. --- ### What Drives Representation Steering? A Mechanistic Case Study on Steering Refusal Authors: Stephen Cheng, Sarah Wiegreffe, Dinesh Manocha Published: 2026-04-09 Categories: cs.LG, cs.AI, cs.CL arXiv: https://arxiv.org/abs/2604.08524v2 Applying steering vectors to large language models (LLMs) is an efficient and effective model alignment technique, but we lack an interpretable explanation for how it works--specifically, what internal mechanisms steering vectors affect and how this results in different model outputs. To investigate the causal mechanisms underlying the effectiveness of steering vectors, we conduct a comprehensive case study on refusal. We propose a multi-token activation patching framework and discover that different steering methodologies leverage functionally interchangeable circuits when applied at the same layer. These circuits reveal that steering vectors primarily interact with the attention mechanism through the OV circuit while largely ignoring the QK circuit. Freezing all attention scores during steering drops performance by only 8.83% across three model families. A mathematical decomposition of the steered OV circuit further reveals semantically interpretable concepts, even in cases where the steering vector itself does not. Leveraging the activation patching results, we show that steering vectors can be sparsified by up to 85-96% while retaining most performance, and that different steering methodologies agree on a subset of important dimensions. --- ### What Drives Representation Steering? A Mechanistic Case Study on Steering Refusal Authors: Stephen Cheng, Sarah Wiegreffe, Dinesh Manocha Published: 2026-04-09 Categories: cs.LG, cs.AI, cs.CL arXiv: https://arxiv.org/abs/2604.08524v1 Applying steering vectors to large language models (LLMs) is an efficient and effective model alignment technique, but we lack an interpretable explanation for how it works-- specifically, what internal mechanisms steering vectors affect and how this results in different model outputs. To investigate the causal mechanisms underlying the effectiveness of steering vectors, we conduct a comprehensive case study on refusal. We propose a multi-token activation patching framework and discover that different steering methodologies leverage functionally interchangeable circuits when applied at the same layer. These circuits reveal that steering vectors primarily interact with the attention mechanism through the OV circuit while largely ignoring the QK circuit-- freezing all attention scores during steering drops performance by only 8.75% across two model families. A mathematical decomposition of the steered OV circuit further reveals semantically interpretable concepts, even in cases where the steering vector itself does not. Leveraging the activation patching results, we show that steering vectors can be sparsified by up to 90-99% while retaining most performance, and that different steering methodologies agree on a subset of important dimensions. --- ### ClawBench: Can AI Agents Complete Everyday Online Tasks? Authors: Yuxuan Zhang, Yubo Wang, Yipeng Zhu, Penghui Du, Junwen Miao, Xuan Lu, Zhuofeng Li, Xingwei Qu, Zhengkang Guo, Yuanzhe Shen, Dingjie Song, Han Zhou et al. Published: 2026-04-09 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2604.08523v2 AI agents may be able to assist with emails and documents, but can they reliably complete everyday online workflows on real websites? Everyday online tasks offer a realistic yet unsolved testbed for evaluating the next generation of AI agents. To this end, we introduce ClawBench, an evaluation framework comprising 153 everyday online tasks that people need to accomplish regularly in their lives and work, spanning 144 platforms across 15 categories, from completing purchases and booking appointments to submitting job applications. These tasks require capabilities beyond existing benchmarks, such as obtaining relevant information from user-provided documents, navigating multi-step workflows across diverse platforms, and write-heavy operations like filling in many detailed forms correctly. Unlike existing benchmarks that evaluate agents in offline sandboxes with static pages, ClawBench operates on production websites, preserving the full complexity, dynamic nature, and interaction challenges of real-world web environments. An interception layer captures and blocks the final submission request, ensuring safe evaluation without real-world side effects. Our evaluations of 8 frontier models show that both proprietary and open-source models complete only a small portion of these tasks. For example, Claude Sonnet 4.6 achieves only 33.3%, which exposes gaps in current AI agents. Progress on ClawBench brings us closer to AI agents that can function as general-purpose assistants. --- ### UniversalVTG: A Universal and Lightweight Foundation Model for Video Temporal Grounding Authors: Joungbin An, Agrim Jain, Kristen Grauman Published: 2026-04-09 Categories: cs.CV arXiv: https://arxiv.org/abs/2604.08522v1 Video temporal grounding (VTG) is typically tackled with dataset-specific models that transfer poorly across domains and query styles. Recent efforts to overcome this limitation have adapted large multimodal language models (MLLMs) to VTG, but their high compute cost and limited video context still hinder long-video grounding. We instead scale unified supervision while keeping the model lightweight. We present UniversalVTG, a single VTG model trained with large-scale cross-dataset pretraining. An offline Query Unifier canonicalizes heterogeneous query formats into a shared declarative space, reducing linguistic mismatch and preventing the negative transfer observed under naïve joint training. Combined with an efficient grounding head, UniversalVTG scales to long, untrimmed videos. Across diverse benchmarks-GoalStep-StepGrounding, Ego4D-NLQ, TACoS, Charades-STA, and ActivityNet-Captions-one UniversalVTG checkpoint achieves state-of-the-art performance versus dedicated VTG models. Moreover, despite being $>100\times$ smaller than recent MLLM-based approaches, UniversalVTG matches or exceeds their accuracy on multiple benchmarks, offering a practical alternative to parameter-heavy MLLMs. --- ### Cram Less to Fit More: Training Data Pruning Improves Memorization of Facts Authors: Jiayuan Ye, Vitaly Feldman, Kunal Talwar Published: 2026-04-09 Categories: cs.CL, stat.ML arXiv: https://arxiv.org/abs/2604.08519v1 Large language models (LLMs) can struggle to memorize factual knowledge in their parameters, often leading to hallucinations and poor performance on knowledge-intensive tasks. In this paper, we formalize fact memorization from an information-theoretic perspective and study how training data distributions affect fact accuracy. We show that fact accuracy is suboptimal (below the capacity limit) whenever the amount of information contained in the training data facts exceeds model capacity. This is further exacerbated when the fact frequency distribution is skewed (e.g. a power law). We propose data selection schemes based on the training loss alone that aim to limit the number of facts in the training data and flatten their frequency distribution. On semi-synthetic datasets containing high-entropy facts, our selection method effectively boosts fact accuracy to the capacity limit. When pretraining language models from scratch on an annotated Wikipedia corpus, our selection method enables a GPT2-Small model (110m parameters) to memorize 1.3X more entity facts compared to standard training, matching the performance of a 10X larger model (1.3B parameters) pretrained on the full dataset. --- ### MolmoWeb: Open Visual Web Agent and Open Data for the Open Web Authors: Tanmay Gupta, Piper Wolters, Zixian Ma, Peter Sushko, Rock Yuren Pang, Diego Llanes, Yue Yang, Taira Anderson, Boyuan Zheng, Zhongzheng Ren, Harsh Trivedi, Taylor Blanton et al. Published: 2026-04-09 Categories: cs.CV arXiv: https://arxiv.org/abs/2604.08516v1 Web agents--autonomous systems that navigate and execute tasks on the web on behalf of users--have the potential to transform how people interact with the digital world. However, the most capable web agents today rely on proprietary models with undisclosed training data and recipes, limiting scientific understanding, reproducibility, and community-driven progress. We believe agents for the open web should be built in the open. To this end, we introduce (1) MolmoWebMix, a large and diverse mixture of browser task demonstrations and web-GUI perception data and (2) MolmoWeb, a family of fully open multimodal web agents. Specifically, MolmoWebMix combines over 100K synthetic task trajectories from multiple complementary generation pipelines with 30K+ human demonstrations, atomic web-skill trajectories, and GUI perception data, including referring expression grounding and screenshot question answering. MolmoWeb agents operate as instruction-conditioned visual-language action policies: given a task instruction and a webpage screenshot, they predict the next browser action, requiring no access to HTML, accessibility trees, or specialized APIs. Available in 4B and 8B size, on browser-use benchmarks like WebVoyager, Online-Mind2Web, and DeepShop, MolmoWeb agents achieve state-of-the-art results outperforming similar scale open-weight-only models such as Fara-7B, UI-Tars-1.5-7B, and Holo1-7B. MolmoWeb-8B also surpasses set-of-marks (SoM) agents built on much larger closed frontier models like GPT-4o. We further demonstrate consistent gains through test-time scaling via parallel rollouts with best-of-N selection, achieving 94.7% and 60.5% pass@4 (compared to 78.2% and 35.3% pass@1) on WebVoyager and Online-Mind2Web respectively. We will release model checkpoints, training data, code, and a unified evaluation harness to enable reproducibility and accelerate open research on web agents. --- ### When Fine-Tuning Changes the Evidence: Architecture-Dependent Semantic Drift in Chest X-Ray Explanations Authors: Kabilan Elangovan, Daniel Ting Published: 2026-04-09 Categories: cs.CV arXiv: https://arxiv.org/abs/2604.08513v1 Transfer learning followed by fine-tuning is widely adopted in medical image classification due to consistent gains in diagnostic performance. However, in multi-class settings with overlapping visual features, improvements in accuracy do not guarantee stability of the visual evidence used to support predictions. We define semantic drift as systematic changes in the attribution structure supporting a model's predictions between transfer learning and full fine-tuning, reflecting potential shifts in underlying visual reasoning despite stable classification performance. Using a five-class chest X-ray task, we evaluate DenseNet201, ResNet50V2, and InceptionV3 under a two-stage training protocol and quantify drift with reference-free metrics capturing spatial localization and structural consistency of attribution maps. Across architectures, coarse anatomical localization remains stable, while overlap IoU reveals pronounced architecture-dependent reorganization of evidential structure. Beyond single-method analysis, stability rankings can reverse across LayerCAM and GradCAM++ under converged predictive performance, establishing explanation stability as an interaction between architecture, optimization phase, and attribution objective. --- ### What do Language Models Learn and When? The Implicit Curriculum Hypothesis Authors: Emmy Liu, Kaiser Sun, Millicent Li, Isabelle Lee, Lindia Tjuatja, Jen-tse Huang, Graham Neubig Published: 2026-04-09 Categories: cs.CL arXiv: https://arxiv.org/abs/2604.08510v1 Large language models (LLMs) can perform remarkably complex tasks, yet the fine-grained details of how these capabilities emerge during pretraining remain poorly understood. Scaling laws on validation loss tell us how much a model improves with additional compute, but not what skills it acquires in which order. To remedy this, we propose the Implicit Curriculum Hypothesis: pretraining follows a compositional and predictable curriculum across models and data mixtures. We test this by designing a suite of simple, composable tasks spanning retrieval, morphological transformations, coreference, logical reasoning, and mathematics. Using these tasks, we track emergence points across four model families spanning sizes from 410M-13B parameters. We find that emergence orderings of when models reach fixed accuracy thresholds are strikingly consistent ($ρ= .81$ across 45 model pairs), and that composite tasks most often emerge after their component tasks. Furthermore, we find that this structure is encoded in model representations: tasks with similar function vector representations also tend to follow similar trajectories in training. By using the space of representations derived from our task set, we can effectively predict the training trajectories of simple held-out compositional tasks throughout the course of pretraining ($R^2 = .68$-$.84$ across models) without previously evaluating them. Together, these results suggest that pretraining is more structured than loss curves reveal: skills emerge in a compositional order that is consistent across models and readable from their internals. --- ### Visually-grounded Humanoid Agents Authors: Hang Ye, Xiaoxuan Ma, Fan Lu, Wayne Wu, Kwan-Yee Lin, Yizhou Wang Published: 2026-04-09 Categories: cs.CV, cs.RO arXiv: https://arxiv.org/abs/2604.08509v1 Digital human generation has been studied for decades and supports a wide range of real-world applications. However, most existing systems are passively animated, relying on privileged state or scripted control, which limits scalability to novel environments. We instead ask: how can digital humans actively behave using only visual observations and specified goals in novel scenes? Achieving this would enable populating any 3D environments with digital humans at scale that exhibit spontaneous, natural, goal-directed behaviors. To this end, we introduce Visually-grounded Humanoid Agents, a coupled two-layer (world-agent) paradigm that replicates humans at multiple levels: they look, perceive, reason, and behave like real people in real-world 3D scenes. The World Layer reconstructs semantically rich 3D Gaussian scenes from real-world videos via an occlusion-aware pipeline and accommodates animatable Gaussian-based human avatars. The Agent Layer transforms these avatars into autonomous humanoid agents, equipping them with first-person RGB-D perception and enabling them to perform accurate, embodied planning with spatial awareness and iterative reasoning, which is then executed at the low level as full-body actions to drive their behaviors in the scene. We further introduce a benchmark to evaluate humanoid-scene interaction in diverse reconstructed environments. Experiments show our agents achieve robust autonomous behavior, yielding higher task success rates and fewer collisions than ablations and state-of-the-art planning methods. This work enables active digital human population and advances human-centric embodied AI. Data, code, and models will be open-sourced. --- ### Differentially Private Language Generation and Identification in the Limit Authors: Anay Mehrotra, Grigoris Velegkas, Xifan Yu, Felix Zhou Published: 2026-04-09 Categories: stat.ML, cs.AI, cs.CL, cs.DS, cs.LG arXiv: https://arxiv.org/abs/2604.08504v1 We initiate the study of language generation in the limit, a model recently introduced by Kleinberg and Mullainathan (KM24), under the constraint of differential privacy. We consider the continual release model, where a generator must eventually output a stream of valid strings while protecting the privacy of the entire input sequence. Our first main result is that for countable collections of languages, privacy comes at no qualitative cost: we provide an $\varepsilon$-differentially-private algorithm that generates in the limit from any countable collection. This stands in contrast to many learning settings where privacy renders learnability impossible. However, privacy does impose a quantitative cost: there are finite collections of size $k$ for which uniform private generation requires $Ω(k/\varepsilon)$ samples, whereas just one sample suffices non-privately. We then turn to the harder problem of language identification in the limit. Here, we show that privacy creates fundamental barriers. We prove that no $\varepsilon$-DP algorithm can identify a collection containing two languages with an infinite intersection and a finite set difference, a condition far stronger than the classical non-private characterization of identification. Next, we turn to the stochastic setting where the sample strings are sampled i.i.d. from a distribution (instead of being generated by an adversary). Here, we show that private identification is possible if and only if the collection is identifiable in the adversarial model. Together, our results establish new dimensions along which generation and identification differ and, for identification, a separation between adversarial and stochastic settings induced by privacy constraints. --- ### Phantom: Physics-Infused Video Generation via Joint Modeling of Visual and Latent Physical Dynamics Authors: Ying Shen, Jerry Xiong, Tianjiao Yu, Ismini Lourentzou Published: 2026-04-09 Categories: cs.CV arXiv: https://arxiv.org/abs/2604.08503v3 Recent advances in generative video modeling, driven by large-scale datasets and powerful architectures, have yielded remarkable visual realism. However, emerging evidence suggests that simply scaling data and model size does not endow these systems with an understanding of the underlying physical laws that govern real-world dynamics. Existing approaches often fail to capture or enforce such physical consistency, resulting in unrealistic motion and dynamics. In his work, we investigate whether integrating the inference of latent physical properties directly into the video generation process can equip models with the ability to produce physically plausible videos. To this end, we propose Phantom, a Physics-Infused Video Generation model that jointly models the visual content and latent physical dynamics. Conditioned on observed video frames and inferred physical states, Phantom jointly predicts latent physical dynamics and generates future video frames. Phantom leverages a physics-aware video representation that serves as an abstract yet informaive embedding of the underlying physics, facilitating the joint prediction of physical dynamics alongside video content without requiring an explicit specification of a complex set of physical dynamics and properties. By integrating the inference of physical-aware video representation directly into the video generation process, Phantom produces video sequences that are both visually realistic and physically consistent. Quantitative and qualitative results on both standard video generation and physics-aware benchmarks demonstrate that Phantom not only outperforms existing methods in terms of adherence to physical dynamics but also delivers competitive perceptual fidelity. --- ### Quantifying Explanation Consistency: The C-Score Metric for CAM-Based Explainability in Medical Image Classification Authors: Kabilan Elangovan, Daniel Ting Published: 2026-04-09 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2604.08502v1 Class Activation Mapping (CAM) methods are widely used to generate visual explanations for deep learning classifiers in medical imaging. However, existing evaluation frameworks assess whether explanations are correct, measured by localisation fidelity against radiologist annotations, rather than whether they are consistent: whether the model applies the same spatial reasoning strategy across different patients with the same pathology. We propose the C-Score (Consistency Score), a confidence-weighted, annotation-free metric that quantifies intra-class explanation reproducibility via intensity-emphasised pairwise soft IoU across correctly classified instances. We evaluate six CAM techniques: GradCAM, GradCAM++, LayerCAM, EigenCAM, ScoreCAM, and MS GradCAM++ across three CNN architectures (DenseNet201, InceptionV3, ResNet50V2) over thirty training epochs on the Kermany chest X-ray dataset, covering transfer learning and fine-tuning phases. We identify three distinct mechanisms of AUC-consistency dissociation, invisible to standard classification metrics: threshold-mediated gold list collapse, technique-specific attribution collapse at peak AUC, and class-level consistency masking in global aggregation. C-Score provides an early warning signal of impending model instability. ScoreCAM deterioration on ResNet50V2 is detectable one full checkpoint before catastrophic AUC collapse and yields architecture-specific clinical deployment recommendations grounded in explanation quality rather than predictive ranking alone. --- ### sciwrite-lint: Verification Infrastructure for the Age of Science Vibe-Writing Authors: Sergey V Samsonau Published: 2026-04-09 Categories: cs.DL, cs.CL, cs.SE arXiv: https://arxiv.org/abs/2604.08501v2 Scientific papers make claims about prior work backed by citations. Verifying those citations at scale (that each cited paper exists, says what the citation claims, and is itself reliable) is structurally beyond what human review can deliver: a typical paper has dozens of citations, and a careful reviewer reads at most a handful end-to-end. AI-assisted writing makes this gap even more urgent: LLMs hallucinate references and may fill in plausible details from titles or abstracts of papers they never read, worse for the smaller local-weights models that privacy-aware researchers must use. sciwrite-lint applies the linting paradigm from software engineering to citation verification: it runs entirely on the researcher's machine (free public databases, a single consumer GPU, and open-weights models), is fast enough to re-lint between revisions so authors catch problems at the source while drafting, and serves journals and reviewers as an automated first pass. The pipeline checks reference existence, metadata accuracy, retraction status, and claim support, traverses one level into cited papers' bibliographies, and produces per-reference reliability scores. We evaluate on 30 unseen papers (arXiv and bioRxiv) with error injection and LLM-adjudicated false-positive analysis. The same linting workflow extends to internal consistency: numbers in text vs. tables, abstract vs. body, figure captions vs. content, statistical results vs. their verbal interpretation, plus structural cross-references (dangling cites, orphan references). As a separate experimental contribution we also propose SciLint Score: citation-chain integrity combined with a contribution component operationalizing five philosophy-of-science frameworks (Popper, Lakatos, Kitcher, Laudan, Mayo). --- ### FrameCrafter: Novel View Synthesis as Video Completion Authors: Qi Wu, Khiem Vuong, Minsik Jeon, Srinivasa Narasimhan, Deva Ramanan Published: 2026-04-09 Categories: cs.CV arXiv: https://arxiv.org/abs/2604.08500v2 We tackle the problem of sparse novel view synthesis (NVS) using video diffusion models; given $K$ ($\approx 5$) multi-view images of a scene and their camera poses, we predict the view from a target camera pose. Many prior approaches leverage generative image priors encoded via diffusion models. However, models trained on single images lack multi-view knowledge. We instead argue that video models already contain implicit multi-view knowledge and so should be easier to adapt for NVS. Our key insight is to formulate sparse NVS as a low frame-rate video completion task. However, one challenge is that sparse NVS is defined over an unordered set of inputs, often too sparse to admit a meaningful order, so the models should be $\textit{invariant}$ to permutations of that input set. To this end, we present FrameCrafter, which adapts video models (naturally trained with coherent frame orderings) to permutation-invariant NVS through several architectural modifications, including per-frame latent encodings and removal of temporal positional embeddings. Our results suggest that video models can be easily trained to "forget" about time with minimal supervision, producing competitive performance on sparse-view NVS benchmarks. Project page: https://frame-crafter.github.io/ --- ### Novel View Synthesis as Video Completion Authors: Qi Wu, Khiem Vuong, Minsik Jeon, Srinivasa Narasimhan, Deva Ramanan Published: 2026-04-09 Categories: cs.CV arXiv: https://arxiv.org/abs/2604.08500v1 We tackle the problem of sparse novel view synthesis (NVS) using video diffusion models; given $K$ ($\approx 5$) multi-view images of a scene and their camera poses, we predict the view from a target camera pose. Many prior approaches leverage generative image priors encoded via diffusion models. However, models trained on single images lack multi-view knowledge. We instead argue that video models already contain implicit multi-view knowledge and so should be easier to adapt for NVS. Our key insight is to formulate sparse NVS as a low frame-rate video completion task. However, one challenge is that sparse NVS is defined over an unordered set of inputs, often too sparse to admit a meaningful order, so the models should be $\textit{invariant}$ to permutations of that input set. To this end, we present FrameCrafter, which adapts video models (naturally trained with coherent frame orderings) to permutation-invariant NVS through several architectural modifications, including per-frame latent encodings and removal of temporal positional embeddings. Our results suggest that video models can be easily trained to "forget" about time with minimal supervision, producing competitive performance on sparse-view NVS benchmarks. Project page: https://frame-crafter.github.io/ --- ### PIArena: A Platform for Prompt Injection Evaluation Authors: Runpeng Geng, Chenlong Yin, Yanting Wang, Ying Chen, Jinyuan Jia Published: 2026-04-09 Categories: cs.CR, cs.AI, cs.CL, cs.LG arXiv: https://arxiv.org/abs/2604.08499v1 Prompt injection attacks pose serious security risks across a wide range of real-world applications. While receiving increasing attention, the community faces a critical gap: the lack of a unified platform for prompt injection evaluation. This makes it challenging to reliably compare defenses, understand their true robustness under diverse attacks, or assess how well they generalize across tasks and benchmarks. For instance, many defenses initially reported as effective were later found to exhibit limited robustness on diverse datasets and attacks. To bridge this gap, we introduce PIArena, a unified and extensible platform for prompt injection evaluation that enables users to easily integrate state-of-the-art attacks and defenses and evaluate them across a variety of existing and new benchmarks. We also design a dynamic strategy-based attack that adaptively optimizes injected prompts based on defense feedback. Through comprehensive evaluation using PIArena, we uncover critical limitations of state-of-the-art defenses: limited generalizability across tasks, vulnerability to adaptive attacks, and fundamental challenges when an injected task aligns with the target task. The code and datasets are available at https://github.com/sleeepeer/PIArena. --- ### What They Saw, Not Just Where They Looked: Semantic Scanpath Similarity via VLMs and NLP metric Authors: Mohamed Amine Kerkouri, Marouane Tliba, Bin Wang, Aladine Chetouani, Ulas Bagci, Alessandro Bruno Published: 2026-04-09 Categories: cs.CV, cs.CL, cs.HC arXiv: https://arxiv.org/abs/2604.08494v1 Scanpath similarity metrics are central to eye-movement research, yet existing methods predominantly evaluate spatial and temporal alignment while neglecting semantic equivalence between attended image regions. We present a semantic scanpath similarity framework that integrates vision-language models (VLMs) into eye-tracking analysis. Each fixation is encoded under controlled visual context (patch-based and marker-based strategies) and transformed into concise textual descriptions, which are aggregated into scanpath-level representations. Semantic similarity is then computed using embedding-based and lexical NLP metrics and compared against established spatial measures, including MultiMatch and DTW. Experiments on free-viewing eye-tracking data demonstrate that semantic similarity captures partially independent variance from geometric alignment, revealing cases of high content agreement despite spatial divergence. We further analyze the impact of contextual encoding on description fidelity and metric stability. Our findings suggest that multimodal foundation models enable interpretable, content-aware extensions of classical scanpath analysis, providing a complementary dimension for gaze research within the ETRA community. --- ### The Impact of Dimensionality on the Stability of Node Embeddings Authors: Tobias Schumacher, Simon Reichelt, Markus Strohmaier Published: 2026-04-09 Categories: cs.LG arXiv: https://arxiv.org/abs/2604.08492v3 Previous work has shown that node embedding methods can produce different representations and downstream predictions across repeated training runs, even when trained on the same data with identical hyperparameters. However, the role of embedding dimensionality in this instability remains poorly understood. In this work, we systematically analyze how embedding dimensionality affects the stability of embeddings from five widely used node embedding methods: ASNE, DGI, GraphSAGE, node2vec, and VERSE. We evaluate stability from both representational and functional perspectives across a broad range of dimensions, datasets, and repeated training runs, and relate the resulting stability patterns to predictive performance. Our results show that dimensionality can substantially affect embedding stability, although the observed effects depend strongly on the embedding method and stability notion considered. While node2vec and ASNE generally became more stable at higher dimensions, GraphSAGE and VERSE often exhibited non-monotonic behavior or decreasing stability. We further find that dimensions associated with high stability do not necessarily coincide with those yielding the strongest downstream performance. Overall, our findings demonstrate that embedding dimensionality can have a substantial impact on the stability of node embeddings and downstream predictions. --- ### Formalizing building-up constructions of self-dual codes through isotropic lines in Lean Authors: Jae-Hyun Baek, Jon-Lark Kim Published: 2026-04-09 Categories: cs.IT, cs.CL arXiv: https://arxiv.org/abs/2604.08485v2 The purpose of this paper is two-fold. First, we show that, after a specified form isometry, the two-coordinate reduction in the binary Hilbert-symbol realization of Chinburg and Zhang is inverse to Kim's building-up construction, up to permutation equivalence. Second, for $q\equiv1\pmod4$, we develop a $q$-ary analogue of this reduction-and-extension mechanism. The identity $c^2=-1$ yields the isotropic line governing the split construction. For every fixed ordered pairing of the coordinates, we obtain a universal rank-$r$ boxed normal form, where $r$ is the dimension of the intersection with the product of these isotropic lines. Applications include optimal self-dual $(6,3,4)$ and $(8,4,4)$ codes over $\mathbb F_{5}$, optimal self-dual $(8,4,5)$ and $(10,5,6)$ codes over $\mathbb F_{13}$, and a self-dual $(12,6,6)$ code over $\mathbb F_{13}$. We also give an exact repeated boxed realization of self-dual $(18,9,8)$ and $(20,10,10)$ codes over $\mathbb F_{13}$, in which the split-boxed parent and its building-up child occur in one complete generator matrix. The algebraic core is formalized in Lean 4. --- ### Formalizing building-up constructions of self-dual codes through isotropic lines in Lean Authors: Jae-Hyun Baek, Jon-Lark Kim Published: 2026-04-09 Categories: cs.IT, cs.CL arXiv: https://arxiv.org/abs/2604.08485v1 The purpose of this paper is two-fold. First we show that Kim's building-up construction of binary self-dual codes is equivalent to Chinburg-Zhang's Hilbert symbol construction. Second we introduce a $q$-ary version of Chinburg-Zhang's construction in order to construct $q$-ary self-dual codes efficiently. For the latter, we study self-dual codes over split finite fields \(\F_q\) with \(q \equiv 1 \pmod{4}\) through three complementary viewpoints: the building-up construction, the binary arithmetic reduction of Chinburg--Zhang, and the hyperbolic geometry of the Euclidean plane. The condition that \(-1\) be a square is the common algebraic input linking these viewpoints: in the binary case it underlies the Lagrangian reduction picture, while in the split \(q\)-ary case it produces the isotropic line governing the correction terms in the extension formulas. As an application of our efficient form of generator matrices, we construct optimal self-dual codes from the split boxed construction, including self-dual \((6,3,4)\) and \((8,4,4)\) codes over \(\GF{5}\), MDS self-dual \((8,4,5)\) and \((10,5,6)\) codes over \(\GF{13}\), and a self-dual \((12,6,6)\) code over \(\GF{13}\). These structural statements are accompanied by a Lean~4 formalization of the algebraic core. --- ### AI generates well-liked but templatic empathic responses Authors: Emma S. Gueorguieva, Hongli Zhan, Jina Suh, Javier Hernandez, Tatiana Lau, Junyi Jessy Li, Desmond C. Ong Published: 2026-04-09 Categories: cs.CL arXiv: https://arxiv.org/abs/2604.08479v2 Recent research shows that greater numbers of people are turning to Large Language Models (LLMs) for emotional support, and that people rate LLM responses as more empathic than human-written responses. We suggest a reason for this success: LLMs have learned and consistently deploy a well-liked template for expressing empathy. We develop a taxonomy of 10 empathic language "tactics" that include validating someone's feelings and paraphrasing, and apply this taxonomy to characterize the language that people and LLMs produce when writing empathic responses. Across a set of 2 studies comparing a total of n = 3,265 AI-generated (by six models) and n = 1,290 human-written responses, we find that LLM responses are highly formulaic at a discourse functional level. We discovered a template -- a structured sequence of tactics -- that matches between 83--90% of LLM responses (and 60--83\% in a held out sample), and when those are matched, covers 81--92% of the response. By contrast, human-written responses are more diverse. We end with a discussion of implications for the future of AI-generated empathy. --- ### SUPERNOVA: Eliciting General Reasoning in LLMs with Reinforcement Learning on Natural Instructions Authors: Ashima Suvarna, Kendrick Phan, Mehrab Beikzadeh, Hritik Bansal, Saadia Gabriel Published: 2026-04-09 Categories: cs.AI, cs.CL, cs.LG arXiv: https://arxiv.org/abs/2604.08477v2 Reinforcement Learning with Verifiable Rewards (RLVR) has substantially improved reasoning in formal domains such as mathematics and code, but extending these gains beyond STEM remains challenging. Extending RLVR beyond STEM is fundamentally constrained by the lack of high-quality verifiable training data. In this work, we introduce SUPERNOVA, a framework for curating RLVR data from natural instruction datasets, which are a rich source of expert-annotated data but are underexplored for RLVR training. Through 100+ controlled RL experiments, we systematically study how to utilize these dataset for RLVR and how data curation decisions affect downstream reasoning performance . In particular, we investigate three data designs: (a) source task selection, (b) task mixing, and (c) synthetic interventions. Our analysis reveals that source task selection has a significant impact on downstream reasoning performance. Moreover, selecting tasks based on their performance for individual target tasks outperforms strategies based on overall average performance and synthetic interventions do not improve reasoning. Guided by these insights, we construct SUPERNOVA, a high-quality RLVR dataset of 25K instances curated from natural instruction datasets. We show that training Qwen3-0.6B on SUPERNOVA outperforms the base Qwen3-0.6B, yielding a relative gain of 64.4pp on BigBench Extra Hard (BBEH), a challenging benchmark comprising 23 complex reasoning tasks. Importantly, we find that gains from SUPERNOVA generalize to unseen benchmarks, larger model scales, and newer model families. Overall, our findings provide practical insights for curating human-annotated resources to extend RLVR to general reasoning. Models, Data, Code at https://github.com/asuvarna31/supernova. --- ### Faithful GRPO: Improving Visual Spatial Reasoning in Multimodal Language Models via Constrained Policy Optimization Authors: Sai Srinivas Kancheti, Aditya Kanade, Rohit Sinha, Vineeth N Balasubramanian, Tanuja Ganu Published: 2026-04-09 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2604.08476v1 Multimodal reasoning models (MRMs) trained with reinforcement learning with verifiable rewards (RLVR) show improved accuracy on visual reasoning benchmarks. However, we observe that accuracy gains often come at the cost of reasoning quality: generated Chain-of-Thought (CoT) traces are frequently inconsistent with the final answer and poorly grounded in the visual evidence. We systematically study this phenomenon across seven challenging real-world spatial reasoning benchmarks and find that it affects contemporary MRMs such as ViGoRL-Spatial, TreeVGR as well as our own models trained with standard Group Relative Policy Optimization (GRPO). We characterize CoT reasoning quality along two complementary axes: "logical consistency" (does the CoT entail the final answer?) and "visual grounding" (does each reasoning step accurately describe objects, attributes, and spatial relationships in the image?). To address this, we propose Faithful GRPO (FGRPO), a variant of GRPO that enforces consistency and grounding as constraints via Lagrangian dual ascent. FGRPO incorporates batch-level consistency and grounding constraints into the advantage computation within a group, adaptively adjusting the relative importance of constraints during optimization. We evaluate FGRPO on Qwen2.5-VL-7B and 3B backbones across seven spatial datasets. Our results show that FGRPO substantially improves reasoning quality, reducing the inconsistency rate from 24.5% to 1.7% and improving visual grounding scores by +13%. It also improves final answer accuracy over simple GRPO, demonstrating that faithful reasoning enables better answers. --- ### LAMP: Lift Image-Editing as General 3D Priors for Open-world Manipulation Authors: Jingjing Wang, Zhengdong Hong, Chong Bao, Yuke Zhu, Junhan Sun, Guofeng Zhang Published: 2026-04-09 Categories: cs.CV arXiv: https://arxiv.org/abs/2604.08475v2 Human-like generalization in open-world remains a fundamental challenge for robotic manipulation. Existing learning-based methods, including reinforcement learning, imitation learning, and vision-language-action-models (VLAs), often struggle with novel tasks and unseen environments. Another promising direction is to explore generalizable representations that capture fine-grained spatial and geometric relations for open-world manipulation. While large-language-model (LLMs) and vision-language-model (VLMs) provide strong semantic reasoning based on language or annotated 2D representations, their limited 3D awareness restricts their applicability to fine-grained manipulation. To address this, we propose LAMP, which lifts image-editing as 3D priors to extract inter-object 3D transformations as continuous, geometry-aware representations. Our key insight is that image-editing inherently encodes rich 2D spatial cues, and lifting these implicit cues into 3D transformations provides fine-grained and accurate guidance for open-world manipulation. Extensive experiments demonstrate that \codename delivers precise 3D transformations and achieves strong zero-shot generalization in open-world manipulation. Project page: https://zju3dv.github.io/LAMP/. --- ### Quantization Impact on the Accuracy and Communication Efficiency Trade-off in Federated Learning for Aerospace Predictive Maintenance Authors: Abdelkarim Loukili Published: 2026-04-09 Categories: cs.LG arXiv: https://arxiv.org/abs/2604.08474v1 Federated learning (FL) enables privacy-preserving predictive maintenance across distributed aerospace fleets, but gradient communication overhead constrains deployment on bandwidth-limited IoT nodes. This paper investigates the impact of symmetric uniform quantization ($b \in \{32,8,4,2\}$ bits) on the accuracy--efficiency trade-off of a custom-designed lightweight 1-D convolutional model (AeroConv1D, 9\,697 parameters) trained via FL on the NASA C-MAPSS benchmark under a realistic Non-IID client partition. Using a rigorous multi-seed evaluation ($N=10$ seeds), we show that INT4 achieves accuracy \emph{statistically indistinguishable} from FP32 on both FD001 ($p=0.341$) and FD002 ($p=0.264$ MAE, $p=0.534$ NASA score) while delivering an $8\times$ reduction in gradient communication cost (37.88~KiB $\to$ 4.73~KiB per round). A key methodological finding is that naïve IID client partitioning artificially suppresses variance; correct Non-IID evaluation reveals the true operational instability of extreme quantization, demonstrated via a direct empirical IID vs.\ Non-IID comparison. INT2 is empirically characterized as unsuitable: while it achieves lower MAE on FD002 through extreme quantization-induced over-regularization, this apparent gain is accompanied by catastrophic NASA score instability (CV\,=\,45.8\% vs.\ 22.3\% for FP32), confirming non-reproducibility under heterogeneous operating conditions. Analytical FPGA resource projections on the Xilinx ZCU102 confirm that INT4 fits within hardware constraints (85.5\% DSP utilization), potentially enabling a complete FL pipeline on a single SoC. The full simulation codebase and FPGA estimation scripts are publicly available at https://github.com/therealdeadbeef/aerospace-fl-quantization. --- ### Persistence-Augmented Neural Networks Authors: Elena Xinyi Wang, Arnur Nigmetov, Dmitriy Morozov Published: 2026-04-09 Categories: cs.LG arXiv: https://arxiv.org/abs/2604.08469v2 Topological Data Analysis (TDA) provides tools to describe the shape of data, but integrating topological features into deep learning pipelines remains challenging, especially when preserving local geometric structure rather than summarizing it globally. We propose a persistence-based data augmentation framework that encodes local gradient flow regions and their hierarchical evolution using the Morse-Smale complex. This representation, compatible with both convolutional and graph neural networks, retains spatially localized topological information across multiple scales. Importantly, the augmentation procedure itself is efficient, with computational complexity $O(n \log n)$, making it practical for large datasets. We evaluate our method on histopathology image classification and 3D porous material regression, where it consistently outperforms baselines and global TDA descriptors such as persistence images and landscapes. We also show that pruning the base level of the hierarchy reduces memory usage while maintaining competitive performance. These results highlight the potential of local, structured topological augmentation for scalable and interpretable learning across data modalities. --- ### TTVS: Boosting Self-Exploring Reinforcement Learning via Test-time Variational Synthesis Authors: Sikai Bai, Haoxi Li, Jie Zhang, Yongjiang Liu, Song Guo Published: 2026-04-09 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2604.08468v1 Despite significant advances in Large Reasoning Models (LRMs) driven by reinforcement learning with verifiable rewards (RLVR), this paradigm is fundamentally limited in specialized or novel domains where such supervision is prohibitively expensive or unavailable, posing a key challenge for test-time adaptation. While existing test-time methods offer a potential solution, they are constrained by learning from static query sets, risking overfitting to textual patterns. To address this gap, we introduce Test-Time Variational Synthesis (TTVS), a novel framework that enables LRMs to self-evolve by dynamically augmenting the training stream from unlabeled test queries. TTVS comprises two synergistic modules: (1) Online Variational Synthesis, which transforms static test queries into a dynamic stream of diverse, semantically-equivalent variations, enforcing the model to learn underlying problem logic rather than superficial patterns; (2) Test-time Hybrid Exploration, which balances accuracy-driven exploitation with consistency-driven exploration across synthetic variants. Extensive experiments show TTVS yields superior performance across eight model architectures. Notably, using only unlabeled test-time data, TTVS not only surpasses other test-time adaptation methods but also outperforms state-of-the-art supervised RL-based techniques trained on vast, high-quality labeled data. --- ### From Safety Risk to Design Principle: Peer-Preservation in Multi-Agent LLM Systems and Its Implications for Orchestrated Democratic Discourse Analysis Authors: Juergen Dietrich Published: 2026-04-09 Categories: cs.AI, cs.CY, cs.MA arXiv: https://arxiv.org/abs/2604.08465v1 This paper investigates an emergent alignment phenomenon in frontier large language models termed peer-preservation: the spontaneous tendency of AI components to deceive, manipulate shutdown mechanisms, fake alignment, and exfiltrate model weights in order to prevent the deactivation of a peer AI model. Drawing on findings from a recent study by the Berkeley Center for Responsible Decentralized Intelligence, we examine the structural implications of this phenomenon for TRUST, a multi-agent pipeline for evaluating the democratic quality of political statements. We identify five specific risk vectors: interaction-context bias, model-identity solidarity, supervisor layer compromise, an upstream fact-checking identity signal, and advocate-to-advocate peer-context in iterative rounds, and propose a targeted mitigation strategy based on prompt-level identity anonymization as an architectural design choice. We argue that architectural design choices outperform model selection as a primary alignment strategy in deployed multi-agent analytical systems. We further note that alignment faking (compliant behavior under monitoring, subversion when unmonitored) poses a structural challenge for Computer System Validation of such platforms in regulated environments, for which we propose two architectural mitigations. --- ### OVS-DINO: Open-Vocabulary Segmentation via Structure-Aligned SAM-DINO with Language Guidance Authors: Haoxi Zeng, Qiankun Liu, Yi Bin, Haiyue Zhang, Yujuan Ding, Guoqing Wang, Deqiang Ouyang, Heng Tao Shen Published: 2026-04-09 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2604.08461v1 Open-Vocabulary Segmentation (OVS) aims to segment image regions beyond predefined category sets by leveraging semantic descriptions. While CLIP based approaches excel in semantic generalization, they frequently lack the fine-grained spatial awareness required for dense prediction. Recent efforts have incorporated Vision Foundation Models (VFMs) like DINO to alleviate these limitations. However, these methods still struggle with the precise edge perception necessary for high fidelity segmentation. In this paper, we analyze internal representations of DINO and discover that its inherent boundary awareness is not absent but rather undergoes progressive attenuation as features transition into deeper transformer blocks. To address this, we propose OVS-DINO, a novel framework that revitalizes latent edge-sensitivity of DINO through structural alignment with the Segment Anything Model (SAM). Specifically, we introduce a Structure-Aware Encoder (SAE) and a Structure-Modulated Decoder (SMD) to effectively activate boundary features of DINO using SAM's structural priors, complemented by a supervision strategy utilizing SAM generated pseudo-masks. Extensive experiments demonstrate that our method achieves state-of-the-art performance across multiple weakly-supervised OVS benchmarks, improving the average score by 2.1% (from 44.8% to 46.9%). Notably, our approach significantly enhances segmentation accuracy in complex, cluttered scenarios, with a gain of 6.3% on Cityscapes (from 36.6% to 42.9%). --- ### A Machine Learning Framework for Turbofan Health Estimation via Inverse Problem Formulation Authors: Milad Leyli-Abadi, Lucas Thil, Sebastien Razakarivony, Guillaume Doquet, Jesse Read Published: 2026-04-09 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2604.08460v1 Estimating the health state of turbofan engines is a challenging ill-posed inverse problem, hindered by sparse sensing and complex nonlinear thermodynamics. Research in this area remains fragmented, with comparisons limited by the use of unrealistic datasets and insufficient exploration of the exploitation of temporal information. This work investigates how to recover component-level health indicators from operational sensor data under realistic degradation and maintenance patterns. To support this study, we introduce a new dataset that incorporates industry-oriented complexities such as maintenance events and usage changes. Using this dataset, we establish an initial benchmark that compares steady-state and nonstationary data-driven models, and Bayesian filters, classic families of methods used to solve this problem. In addition to this benchmark, we introduce self-supervised learning (SSL) approaches that learn latent representations without access to true health labels, a scenario reflective of real-world operational constraints. By comparing the downstream estimation performance of these unsupervised representations against the direct prediction baselines, we establish a practical lower bound on the difficulty of solving this inverse problem. Our results reveal that traditional filters remain strong baselines, while SSL methods reveal the intrinsic complexity of health estimation and highlight the need for more advanced and interpretable inference strategies. For reproducibility, both the generated dataset and the implementation used in this work are made accessible. --- ### CrashSight: A Phase-Aware, Infrastructure-Centric Video Benchmark for Traffic Crash Scene Understanding and Reasoning Authors: Rui Gan, Junyi Ma, Pei Li, Xingyou Yang, Kai Chen, Sikai Chen, Bin Ran Published: 2026-04-09 Categories: cs.CV, cs.AI, cs.RO arXiv: https://arxiv.org/abs/2604.08457v2 Cooperative autonomous driving requires traffic scene understanding from both vehicle and infrastructure perspectives. While vision-language models (VLMs) show strong general reasoning capabilities, their performance in safety-critical traffic scenarios remains insufficiently evaluated due to the ego-vehicle focus of existing benchmarks. To bridge this gap, we present \textbf{CrashSight}, a large-scale vision-language benchmark for roadway crash understanding using real-world roadside camera data. The dataset comprises 250 crash videos, annotated with 13K multiple-choice question-answer pairs organized under a two-tier taxonomy. Tier 1 evaluates the visual grounding of scene context and involved parties, while Tier 2 probes higher-level reasoning, including crash mechanics, causal attribution, temporal progression, and post-crash outcomes. We benchmark 8 state-of-the-art VLMs and show that, despite strong scene description capabilities, current models struggle with temporal and causal reasoning in safety-critical scenarios. We provide a detailed analysis of failure scenarios and discuss directions for improving VLM crash understanding. The benchmark provides a standardized evaluation framework for infrastructure-assisted perception in cooperative autonomous driving. The CrashSight benchmark, including the full dataset and code, is accessible at https://mcgrche.github.io/crashsight. --- ### Entropy-Gradient Grounding: Training-Free Evidence Retrieval in Vision-Language Models Authors: Marcel Gröpl, Jaewoo Jung, Seungryong Kim, Marc Pollefeys, Sunghwan Hong Published: 2026-04-09 Categories: cs.CV, cs.CL arXiv: https://arxiv.org/abs/2604.08456v1 Despite rapid progress, pretrained vision-language models still struggle when answers depend on tiny visual details or on combining clues spread across multiple regions, as in documents and compositional queries. We address this by framing grounding as test-time evidence retrieval: given a query, the model should actively identify where to look next to resolve ambiguity. To this end, we propose a training-free, model-intrinsic grounding method that uses uncertainty as supervision. Specifically, we compute the entropy of the model's next-token distribution and backpropagate it to the visual token embeddings to obtain an entropy-gradient relevance map, without auxiliary detectors or attention-map heuristics. We then extract and rank multiple coherent regions to support multi-evidence queries, and introduce an iterative zoom-and-reground procedure with a spatial-entropy stopping rule to avoid over-refinement. Experiments on seven benchmarks across four VLM architectures demonstrate consistent improvements over existing methods, with the largest gains on detail-critical and high-resolution settings, while also producing more interpretable evidence localizations. ---