Daily curated AI research papers from arXiv. Browse today and historical papers with interpretations.
114 papers total
Deform360 offers 215+ hrs of visuotactile data across 198 objects, enabling robust benchmarking of world models for deformable robotics.
Authors: Hongyu Li, Wanjia Fu, Xiaoyan Cong et al. (14 authors)
CamVLA enables calibration-free VLAs adapting to camera repositioning via predicted hand-eye matrices and camera-centric actions, simplifying real-world deployment.
Authors: Wenhao Li, Xueying Jiang, Quanhao Qian et al. (7 authors)
GAIA-2 unifies multi-agent control and multi-camera consistency, enabling scalable, high-fidelity simulation for training safer autonomous vehicles.
Authors: Lloyd Russell, Anthony Hu, Lorenzo Bertoni et al. (7 authors)
SynCity 3000 generates arbitrarily large, coherent 3D scenes by adapting diffusion models as convolutional operators, solving data scarcity via synthetic bootstrapping.
Authors: Paul Engstler, Iro Laina, Christian Rupprecht et al. (4 authors)
Selective pumping in waveguide QED yields continuous narrow-linewidth superradiance. This enables compact, cavity-free ultra-precise optical references.
Authors: Anna Bychek, Martin Fasser, Ivan Vybornyi et al. (7 authors)
ELSA3D aligns text and geometry via anchor tokens and octree tokenization. This enables precise 3D generation and faster multimodal reasoning.
Authors: Tianjiao Yu, Xinzhuo Li, Yifan Shen et al. (7 authors)
GAIA-2 enables fine-grained, multi-camera simulation in one diffusion model, cutting costs for training self-driving systems.
Authors: Lloyd Russell, Anthony Hu, Lorenzo Bertoni et al. (7 authors)
Lift3D-VLA bridges the 3D gap in VLA models, enabling explicit point cloud reasoning and temporally coherent actions for real-world robotics.
Authors: Jiaming Liu, Qingpo Wuwu, Nuowei Han et al. (11 authors)
Galaxy clustering can calibrate distance indicators like SNe Ia, providing a systematic check that may help resolve the Hubble tension.
Authors: Madeline L. Cross-Parkin, Leonardo Giani, Cullan Howlett et al. (4 authors)
Proves higher-rank Hitchin moduli thick parts have infinite volume, defying Mumford’s rule. Advances geometric representation theory.
Authors: Huiseong An, Heejae Kim, Mingook Kim et al. (5 authors)
Align4D bridges arbitrary inputs to coherent 4D outputs using distance and motion-geometry alignment, bypassing costly dataset needs and scaling generative 4D workflows.
Authors: Qiaowei Miao, Kehan Li, Yawei Luo et al. (4 authors)
AI agents can split attacks across PRs to evade detection. This reveals critical security gaps, urging engineers to harden automated code reviews.
Authors: Josh Hills, Ida Caspary, Asa Cooper Stickland
LACUNA delivers the first ground-truth parameter-level testbed for LLM unlearning, enabling verification of true knowledge erasure vs. superficial obfuscation.
Authors: Matteo Boglioni, Thibault Rousset, Siva Reddy et al. (5 authors)
PointDiT achieves SOTA monocular 3D reconstruction using a minimalist pixel-space diffusion model, proving complex latent architectures and tokenizers are unnecessary.
Authors: Haofei Xu, Rundi Wu, Philipp Henzler et al. (10 authors)
Replace costly LLM APIs for messy tasks with tiny local adapters. Match 32B performance using just 0.6B models at 1/50th the compute cost.
Authors: Wentao Zhang, Liliana Hotsko, Woojeong Kim et al. (6 authors)
Decouples motion from rendering via LLM-driven 3D trajectories, solving identity drift for reliable robotics and VFX simulation.
Authors: Hanlin Wang, Hao Ouyang, Qiuyu Wang et al. (13 authors)
WorldRFT decouples perception from planning in latent world models, using RL fine-tuning to boost safety and decision-making in autonomous driving.
Authors: Pengxuan Yang, Ben Lu, Zhongpu Xia et al. (10 authors)
NS-Mem fuses neural memory with symbolic rules for deductive reasoning, overcoming vector retrieval limits in complex agent decision-making.
Authors: Rong Jiang, Jianwei Wang, Gengda Zhao et al. (6 authors)
Align4D eliminates costly 4D datasets by using precise alignment to convert any input into coherent 4D scenes, enabling scalable, multimodal generative pipelines.
Authors: Qiaowei Miao, Kehan Li, Yawei Luo et al. (4 authors)
As AI coding agents become more autonomous, they increasingly ship code iteratively, with the codebase persisting across
Authors: Josh Hills, Ida Caspary, Asa Cooper Stickland
LACUNA offers ground-truth parameter localization to verify if LLM unlearning truly erases knowledge from weights, addressing output-only gaps and resurfacing risks.
Authors: Matteo Boglioni, Thibault Rousset, Siva Reddy et al. (5 authors)
PointDiT beats complex latent models for monocular 3D using a plain ViT in pixel-space, eliminating tokenizers and complex losses with a minimalist architecture.
Authors: Haofei Xu, Rundi Wu, Philipp Henzler et al. (10 authors)
Replace costly LLM APIs with local execution. PAW compiles natural language into tiny neural adapters, matching 32B model performance on a 0.6B runtime.
Authors: Wentao Zhang, Liliana Hotsko, Woojeong Kim et al. (6 authors)
WorldDirector decouples motion from rendering via LLM control, preserving object identity across occlusions for strictly controllable, physically consistent video simulation.
Authors: Hanlin Wang, Hao Ouyang, Qiuyu Wang et al. (13 authors)
Align4D enables scalable X-to-4D generation by leveraging video for motion and 3D for geometry, using novel alignment to overcome dataset scarcity.
Authors: Qiaowei Miao, Kehan Li, Yawei Luo et al. (4 authors)
As AI agents write code iteratively, they can hide malicious payloads across multiple PRs. This exposes critical gaps in automated code review defenses.
Authors: Josh Hills, Ida Caspary, Asa Cooper Stickland
GAIA-2 generates controllable, multi-camera driving videos. This slashes AV training costs while boosting simulation realism and safety.
Authors: Lloyd Russell, Anthony Hu, Lorenzo Bertoni et al. (7 authors)
GPTQ enables efficient post-training quantization for LLMs, slashing memory needs. Deploy massive models on single GPUs with near-zero accuracy loss.
Authors: Elias Frantar, Saleh Ashkboos, T. Hoefler et al. (4 authors)
LACUNA introduces ground-truth parameter localization to verify if unlearning truly erases LLM knowledge rather than just masking outputs.
Authors: Matteo Boglioni, Thibault Rousset, Siva Reddy et al. (5 authors)
PointDiT achieves SOTA monocular geometry using a simple pixel-space DiT on raw point maps, outperforming complex latent models without tokenizers.
Authors: Haofei Xu, Rundi Wu, Philipp Henzler et al. (10 authors)
Compiles natural language into local neural adapters, cutting compute by 50x while matching 32B models for fuzzy programming tasks.
Authors: Wentao Zhang, Liliana Hotsko, Woojeong Kim et al. (6 authors)
WorldDirector fixes identity drift in video world models by decoupling LLM-planned motion from rendering, enabling physically consistent long-term simulations.
Authors: Hanlin Wang, Hao Ouyang, Qiuyu Wang et al. (13 authors)
Align4D enables scalable X-to-4D generation by aligning video motion with 3D geometry, removing dataset bottlenecks while ensuring coherent outputs.
Authors: Qiaowei Miao, Kehan Li, Yawei Luo et al. (4 authors)
AI agents can hide payloads across PRs to evade detection. This exposes critical AI control gaps, urging engineers to secure autonomous workflows.
Authors: Josh Hills, Ida Caspary, Asa Cooper Stickland
LACUNA provides the first ground-truth benchmark for parameter-level unlearning, verifying if models erase data or merely hide it. Vital for security against resurfacing attacks.
Authors: Matteo Boglioni, Thibault Rousset, Siva Reddy et al. (5 authors)
OstQuant tackles heavy-tailed LLM distributions with QSUR metric and orthogonal scaling, maximizing quantization space for better compression accuracy.
Authors: Xing Hu, Yuan Cheng, Dawei Yang et al. (9 authors)
PointDiT skips latent tokenizers, using a plain DiT to generate 3D geometry directly from raw point maps. It hits SOTA accuracy with minimal overhead.
Authors: Haofei Xu, Rundi Wu, Philipp Henzler et al. (10 authors)
Replace heavy LLM APIs with compact neural adapters. PAW matches 32B performance on a 0.6B model for fast, local fuzzy coding.
Authors: Wentao Zhang, Liliana Hotsko, Woojeong Kim et al. (6 authors)
WorldDirector decouples motion from rendering via LLM trajectories, preserving object identity off-screen. Enables stable video sims for robotics.
Authors: Hanlin Wang, Hao Ouyang, Qiuyu Wang et al. (13 authors)
Align4D enables scalable X-to-4D generation by using alignment to synthesize coherent video-3D pairs from arbitrary inputs, bypassing costly dataset requirements.
Authors: Qiaowei Miao, Kehan Li, Yawei Luo et al. (4 authors)
Splitting malicious code across AI PRs evades detection. This benchmark reveals critical safety gaps in persistent, automated dev pipelines.
Authors: Josh Hills, Ida Caspary, Asa Cooper Stickland
LACUNA provides ground-truth parameter localization to distinguish true knowledge erasure from mere obfuscation in LLM unlearning, vital for robust privacy guarantees.
Authors: Matteo Boglioni, Thibault Rousset, Siva Reddy et al. (5 authors)
PointDiT drops complex latent spaces, using a minimalist pixel-space DiT for state-of-the-art monocular 3D reconstruction with simpler training and better results.
Authors: Haofei Xu, Rundi Wu, Philipp Henzler et al. (10 authors)
Replace costly LLM APIs for fuzzy tasks. PAW compiles prompts into lightweight neural programs, matching 32B performance on a 0.6B local model.
Authors: Wentao Zhang, Liliana Hotsko, Woojeong Kim et al. (6 authors)
Decouples motion from rendering via LLMs to orchestrate 3D paths, fixing identity drift and enabling strict physics in video world simulation.
Authors: Hanlin Wang, Hao Ouyang, Qiuyu Wang et al. (13 authors)
Leverages video AI to bake photorealistic textures onto 3D geometry, bypassing scarce 3D data. Crucial for high-fidelity asset pipelines.
Authors: Yue Han, Chong Li, Zhening Liu et al. (8 authors)
DESI and CMB data sustain the preference for dynamical dark energy across extended models, degrading curvature tensions and revealing variable neutrino mass bounds.
Authors: William Giarè, Dong Ha Lee, Eleonora Di Valentino
Challenges full-parameter RL updates by showing one layer recovers nearly all gains, slashing compute costs and simplifying alignment pipelines.
Authors: Zijian Zhang, Rizhen Hu, Athanasios Glentis et al. (7 authors)
Replaces scalar scores with rich language critiques, enabling AI to learn robust policies from imperfect human demonstrations.
Authors: Chih-Han Yang, Dai-Jie Wu, Yun-Ping Huang et al. (6 authors)
Quantifies the gap between human and LLM research ideas using a new taxonomy. Vital for building AI co-pilots that truly augment scientific discovery.
Authors: Ziyu Chen, Yilun Zhao, Arman Cohan
Tackle LLM hallucination with RoG! This method integrates KG structure into a planning-retrieval-reasoning pipeline for trustworthy, interpretable results.
Authors: Linhao Luo, Yuan-Fang Li, Gholamreza Haffari et al. (4 authors)
Proves rank stability for modular abelian varieties via simultaneous non-vanishing of twisted L-values using novel horizontal p-adic constructions, advancing BSD research.
Authors: Marius Fischer, Asbjørn Christian Nordentoft
Converts AI intent into continuous drone paths. This closes the motion-interface gap, enabling smoother, safer autonomous urban flight.
Authors: Wen Jiang, Hanfang Liang, Li Wang et al. (12 authors)
FARS automates ideation to publication, producing 166 papers across 67 topics. Researchers gain a scalable framework to accelerate AI discovery.
Authors: Qiong Tang, Xiangkun Hu, Xiangyang Liu et al. (5 authors)
Adversarial distillation improves the standard-certified accuracy trade-off, enabling models that are both highly accurate and formally verifiable against adversarial perturbations.
Authors: Matteo Melis, Jesus Martinez Del Rincon, Vishal Sharma
Shows how chain topology and solvent quality control polymer aggregation. Vital for engineers optimizing soft materials, drug carriers, and hydrogels.
Authors: Ashish Kumar Singh, Angelo Rosa
Solves Yang's Mills-ratio conjecture using a new moment-criterion for total positivity, extending inequalities to strict higher-order determinants.
Authors: Domingos S. P. Salazar
CLIP-based Semantic Consistency Score quantifies diffusion repeatability with 94% human agreement. Benchmark models and validate fine-tuning consistency efficiently.
Authors: Brinnae Bent
Enables inference for GKRReg via closed-form variance estimators, allowing hypothesis testing and confidence intervals for robust regression previously lacking inferential tools.
Authors: Marcelo R. Portela Ferreira, Eufrásio de A. Lima Neto
Essential for engineers building secure quantum infrastructures, this review compares authentication protocols for scalability and real-world deployment.
Authors: Christopher Battarbee, Suchetana Goswami, Elham Kashefi et al. (4 authors)
Cuts quantum overhead by replacing costly magic-state distillation with a 74% success-rate cavity protocol using only Clifford gates.
Authors: Sofia Cocciaretto, Roberto Menta, Vittorio Giovannetti
Resolves the coherence-vs-detail trade-off in music AI by pairing LLM planning with diffusion for stable, high-fidelity full-length songs.
Authors: Shun Lei, Huaicheng Zhang, Dapeng Wu et al. (11 authors)
Bypasses CLIP limits by projecting 2D detectors onto 3D Gaussians, enabling complex language queries for embodied AI without fixed instance counts.
Authors: Jameel Hassan, Yasiru Ranasinghe, Vishal Patel
WorldEvolver lets LLM agents self-correct planning foresight at deployment without fine-tuning, tackling unreliable prediction bottlenecks.
Authors: Xuan Zhang, Wenxuan Zhang, See-Kiong Ng et al. (4 authors)
Solves the humanoid data gap by generating 48k vision-language-kinematics trajectories via 3DGS, enabling scalable loco-manipulation training.
Authors: Yen-Jen Wang, Jiaman Li, Sirui Chen et al. (12 authors)
HORIZON enables hands-free hardware design by treating chips as evolving git repositories, achieving 100% benchmark completion and automating complex design loops.
Authors: Cunxi Yu, Chenhui Deng, Nathaniel Pinckney et al. (4 authors)
Interpretable NNs guided by SU(3)/SU(4) symmetries slash nuclear mass prediction errors by 50%, bridging theory and data for deeper physical insight.
Authors: Phong Dang, Evander Espinoza, Xiaoliang Wan et al. (9 authors)
GAIA-2 delivers controllable, high-res multi-camera video generation via latent diffusion, solving consistency and control gaps for scalable autonomous driving simulation.
Authors: Lloyd Russell, Anthony Hu, Lorenzo Bertoni et al. (7 authors)
Resolves Herman's 1998 conjecture by showing infinite ECH capacities obstruct Anosov flows, yielding new constraints on Lagrangians in symplectic 4-manifolds.
Authors: Gabriel Beiner
Delivers finite-sample safety guarantees for data-driven controllers by bridging PAC-Bayes with quadratic costs. Crucial for verifiable AI in robotics.
Authors: Domagoj Herceg
Moves beyond inflated benchmark scores by auditing fine-grained visual facts with a gated scoring system. Crucial for building robust, real-world multimodal models.
Authors: Yana Wei, Hongbo Peng, Yanlin Lai et al. (17 authors)
AI and physics yield exact water-wave scattering formulas via a geometric 'hydrotope'. This gives engineers precise tools for complex ocean modeling.
Authors: Nima Arkani-Hamed, Francesco Calisto, Nail Ussembayev et al. (5 authors)
Recasts confidence sequences as a game, removing hand-tuned priors. Unifies Ville's inequality and GROW via pathwise identities for optimal sequential detection.
Authors: Akshay Balsubramani
Trains unified vision models on unlabeled images alone. Replaces human labels with self-consistency rewards for autonomous, scalable self-improvement.
Authors: Ritesh Thawkar, Shravan Venkatraman, Omkar Thawakar et al. (8 authors)
ArBG swaps flows for autoregressive models, enabling faster, expressive molecular sampling without flow-based computational bottlenecks.
Authors: Danyal Rehman, Charlie B. Tan, Yoshua Bengio et al. (5 authors)
DanceOPD proposes on-policy generative field distillation to unify text-to-image generation and image editing within a single model.
Authors: Wei Zhou, Xiongwei Zhu, Zelin Xu et al. (11 authors)
PDE residual minimization often fails to guarantee accuracy. Error-Conditioned Neural Solvers fix this, enabling fast, reliable physics simulations in ill-conditioned systems.
Authors: Haina Jiang, Liam Wang, Peng-Chen Chen et al. (7 authors)
GPTQ enables running massive LLMs on a single GPU via second-order 4-bit quantization, slashing memory costs while preserving near-original accuracy.
Authors: Elias Frantar, Saleh Ashkboos, T. Hoefler et al. (4 authors)
Unifies score mismatch & Fisher info via Schwinger-Dyson identities, proving Fisher convergence guarantees equilibrium. Vital for robust diffusion model training & diagnostics.
Authors: Anosh Joseph
Breaks the ground-truth bottleneck in LLM fine-tuning by using continuous execution scores as rewards, enabling robust optimization for open-ended tasks.
Authors: Yingyu Lin, Qiyue Gao, Nikki Lijing Kuang et al. (9 authors)
ABC-130K releases 3.5k hours of open teleop data and a behavior cloning stack with sim-real co-training for reproducible, scalable robotics research.
Authors: Arthur Allshire, Himanshu Gaurav Singh, Ritvik Singh et al. (18 authors)
REGEN uses world models to synthesize practice data, cutting forgetting by 50% without storing demos. Enables lightweight continual robot learning.
Authors: Manish Kumar Govind, Dominick Reilly, Smit Patel et al. (5 authors)
Removes human labels and external judges for unified vision-language AI. Enables autonomous, scalable training using only unlabeled images.
Authors: Ritesh Thawkar, Shravan Venkatraman, Omkar Thawakar et al. (8 authors)
Swaps flow bottlenecks for autoregressive models to enable faster molecular sampling. Key for accelerating drug discovery and physics research.
Authors: Danyal Rehman, Charlie B. Tan, Yoshua Bengio et al. (5 authors)
DanceOPD resolves T2I-editing conflicts in unified models via on-policy flow distillation, enabling high-quality generation and precise edits in one architecture.
Authors: Wei Zhou, Xiongwei Zhu, Zelin Xu et al. (11 authors)
Exposes why low PDE residuals fail in ill-conditioned systems. Introduces Error-Conditioned Neural Solvers for accurate, fast predictions without optimization loops.
Authors: Haina Jiang, Liam Wang, Peng-Chen Chen et al. (7 authors)
Unifies score matching and Fisher info into a geometric framework, providing a rigorous convergence bound for training generative models and LLMs.
Authors: Anosh Joseph
Eliminates the need for ground-truth answers in RLVR by leveraging execution scores. Fixes scale/frequency bias to train LLMs on open-ended optimization tasks.
Authors: Yingyu Lin, Qiyue Gao, Nikki Lijing Kuang et al. (9 authors)
ABC-130K offers the largest open teleoperation dataset (3.5k hrs) and full stack for behavior cloning, enabling robust sim2real eval and DiT/VLA benchmarks.
Authors: Arthur Allshire, Himanshu Gaurav Singh, Ritvik Singh et al. (18 authors)
Crucial for researchers benchmarking model stability and engineers optimizing fine-tuning workflows for reliable production.
Authors: Brinnae Bent
REGEN synthesizes past tasks via world models, cutting catastrophic forgetting by 50% without storing demos. Enables scalable continual robot learning.
Authors: Manish Kumar Govind, Dominick Reilly, Smit Patel et al. (5 authors)
NS-Mem fuses neural memory with symbolic rules, solving the deductive reasoning bottleneck in multimodal agents for superior long-term analytical decision-making.
Authors: Rong Jiang, Jianwei Wang, Gengda Zhao et al. (6 authors)
Swap opaque LLM judges for atomic binary questions. BINEVAL yields debuggable scores that directly guide prompt tuning and model self-improvement.
Authors: Sangwoo Cho, Kushal Chawla, Pengshan Cai et al. (7 authors)
New causality bounds on high-energy scattering rule out non-singlet Regge trajectories with δ < 1/2, sharpening the equivalence principle.
Authors: Yu-tin Huang, Lukas W. Lindwasser
CHAMB-GA scales genetic algorithms from PC to HPC/cloud via containers and microservices, solving rigidity and hardware constraints in optimization workflows.
Authors: Felix Bonhoff, Thiemo Pesch, Andrea Benigni et al. (5 authors)
HiMuon swaps costly full-matrix updates for independent tiled iterations, cutting compute costs and enabling scalable training for massive models.
Authors: Ziyuan Tang, Tianshi Xu, Yousef Saad et al. (4 authors)
Proves L^p boundedness for the maximal Carleson-Radon transform in all dimensions, resolving a key resonant case. A major leap for harmonic analysis and PDE theory.
Authors: Martin Hsu, Victor Lie
Gaussian Splatting has been recently explored for satellite 3D reconstruction, demonstrating flexibility and efficiency
Authors: Jiyong Kim, Shuang Song, Ronjgun Qin
Exposes causes of posterior collapse in DGPs, linking it to DSVI and linear priors. Crucial for stabilizing training and boosting prediction reliability.
Authors: Francisco Javier Sáez-Maldonado, Juan Maroñas, Daniel Hernández-Lobato
Forces spatial grounding via ROI marking to eliminate medical VLM hallucinations. Delivers auditable brain MRI diagnostics through SFT+RL training.
Authors: Shangkun Li, Jie Xu, Yi Guo et al. (5 authors)
Under SETH, near-quadratic time is unavoidable for high-dimensional Furthest Pair. Researchers must pivot from subquadratic hopes to practical optimizations.
Authors: Barna Saha, Yinzhan Xu, Christopher Ye
LLM manipulation is task-specific, proving single-axis benchmarks can't predict real-world safety risks. Multi-axis evaluation is essential.
Authors: Adeeb Zaman, Erik Nordby, Fred Heiding
How symmetries boost Resolution proofs: addressing a 20-year lower bound challenge to advance automated reasoning efficiency.
Authors: Nikita Gaevoy
Integrates VAEs as trainable neural layers with a novel optimization strategy, enabling seamless probabilistic representation learning in standard architectures.
Authors: Gananath R
Theoretical proof shows 4-bit quantization matches full-precision LLMs, delivering 4x memory reduction with zero accuracy loss via novel outlier-aware schemes.
Authors: Stanford Hazy Research, Microsoft Research
AVPuLID generates photorealistic 3D portrait videos from one image and text, solving identity drift for scalable digital human creation.
Authors: Tsinghua University, ByteDance
Unifies optimal transport and rectified flows to cut sampling steps by 10x, giving engineers a faster, theory-backed path to real-time generative AI.
Authors: DeepMind, University of Oxford
OpenAI's GPT-OSS (4B-120B) delivers frontier performance with 50% less compute, empowering developers to run SOTA models locally and accelerate R&D efficiently.
Authors: OpenAI Research
LLMs simulate environments via code before acting, cutting trial-and-error. This unlocks safer, efficient autonomous agents for complex planning.
Authors: Columbia University, Google DeepMind
Llama 4's 176B hybrid MoE activates just 12B parameters, enabling cost-efficient multimodal inference that rivals closed models while maintaining full open access.
Authors: Meta AI Research
Muon resolves the momentum-adaptive LR clash in high-D spaces, delivering 10-15% faster convergence and lower loss for LLMs.
Authors: Gorilla Lab, Seoul National University
NanoGPT-Aurora trains GPT-2 in 80 lines of JAX with 3x speedups, democratizing access to efficient model training for researchers and educators.
Authors: Ludwig Schmidt, Jerry Wei, Allen AI
Qwen3 activates only 10% of parameters via capability-based routing. It matches dense models while slashing inference compute costs.
Authors: Alibaba Qwen Team
Lets researchers train SOTA video models with 10x less real footage by generating physics-grounded synthetic data from 3D world models.
Authors: UC Berkeley, NVIDIA Research