Research Radar

arXiv cs.GR presents a kinetic theory for gated self-evolving LLM agents, treating self-improvement as a population process rather than a single monolithic model loop. The work focuses on DSH-class plugin systems, where the agent rewrites its own skills under a validation gate and the individual …

• A kinetic theory is proposed for gated self-evolving LLM agents
• Haipeng Wang; arXiv cs.GR

arXiv cs.GR details Budgeted-GS, a post-hoc method for turning trained 3D Gaussian Splatting scenes into a factoring tree of multi-resolution, moment-matched aggregates. The practical goal is straightforward: keep large scenes usable on hardware with very different memory limits without giving up real-time rendering. The key …

• Budgeted-GS turns trained 3D Gaussian Splatting scenes into a factoring tree for memory-aware LOD selection.
• Haipeng Wang; arXiv cs.GR

arXiv cs.GR details Windfoil, a GPU-friendly approach that treats rasterization and differentiable vector graphics as the same coverage problem. The core trick is a closed-form evaluation of the box-filtered winding number for quadratic Bézier contours, which is designed to run efficiently on the GPU. The …

• Windfoil is a closed-form, GPU-friendly vector graphics renderer and differentiable rasterizer for quadratic Bézier contours.
• Matt DesLauriers is the listed author; the work is posted on arXiv cs.GR.

arXiv cs.GR introduces 4DCodeBench, a benchmark for 4D inverse graphics through code generation. The goal is to see whether agents can reconstruct dynamic scenes from video by writing executable graphics programs, not just by matching pixels or building static meshes. The benchmark mixes real-world videos …

• 4DCodeBench is a benchmark for 4D inverse graphics via code generation from video.
• Authors include Ruihong Shen, Žiga Kovačič, Peter Kulits, Xingrui Wang, Zizhang Li, Joshua B. Tenenbaum, Alan Yuille, Jieneng Chen, and Jiajun Wu.

arXiv cs.GR details a new geometric measure for speech-driven 3D facial animation that focuses on coarticulation, the subtle motion shaping that happens as sounds influence one another. The idea is to compare lip-path length against a consonant-aware route through vowel, consonant, and vowel poses, rather …

• A geometric metric for coarticulation in speech-driven 3D facial animation
• Danzel Serrano and Przemyslaw Musialski

arXiv cs.GR introduces EditHero, a benchmark aimed at a problem game teams actually face: assets rarely get edited once, they get revised over and over. The benchmark focuses on long-horizon, part-level 3D editing, with natural-language instructions and target images covering both geometry and texture. A …

• EditHero is a new benchmark for long-horizon, part-level 3D editing and vibe modeling.
• Paper by Ruihan Yu and 11 coauthors, published on arXiv cs.GR.

arXiv cs.GR details a simulation-grounded vision-language pipeline for wildfire monitoring that converts 2D fire simulations into labeled video episodes. The setup uses fixed Blender-based low-detail 3D proxies aligned to terrain, fuel, fire activity, and wind cues, then layers controllable video generation on top to create …

• A simulation-grounded agentic VLM framework converts 2D wildfire simulations into labeled video episodes and structured reports.
• Authors include Duowen Chen, Yuchen Sun, Zhiqi Li, Yuxuan Liao, Sinan Wang, Bart van Bloemen Waanders, and Bo Zhu.

arXiv cs.GR details DAGS, a new conditioning scheme for diffusion transformers that targets one of generative rendering’s biggest pain points: unstable results over time. The method keeps the image DiT backbone frozen and steers it with separate appearance and geometry signals, which is the kind …

• DAGS is a disentangled appearance-and-geometry steering method for a frozen image diffusion transformer
• Authors: Karthik Mohan Kumar, Damian Andrysiak, Pedro Antonio Pena, Kunal Tyagi, Rama Harihara

arXiv cs.GR details a benchmark for spatially grounded gesture generation that tries to solve a familiar problem for conversational avatars: a gesture can look natural while still pointing at the wrong thing. The work introduces a dataset of roughly 2,000 pointing-annotated clips captured from naturalistic …

• A benchmark for spatially grounded gesture generation was introduced
• Anna Deichler, Rishabh Dabral, Fethiye Irmak Dogan, Anindita Ghosh, and Jonas Beskow

arXiv cs.GR details SoftPaint, a new zero-shot sampling method for diffusion editing that aims to move beyond coarse mask-based workflows. The paper frames the problem clearly: prompt- and reference-guided editors can change too much at once, while training for fine-grained control usually needs expensive pixel-level …

• SoftPaint is a zero-shot diffusion editing method using soft masks for pixel-level, adjustable-strength edits.
• Candi Zheng and Yuan Lan; arXiv cs.GR.

arXiv cs.GR details MosaiChunk, a spatio-temporal memory approach for autoregressive video generation that targets a familiar failure mode: once content falls out of context, models struggle to recreate it faithfully when it returns. The core idea is to compose a mosaic of selected historical key-value …

• MosaiChunk proposes spatio-temporal memory for autoregressive video generation using mosaicked historical KV entries.
• Authors: Yiwen Zhang, Haocheng Xi, Michael Tian-Yue Liu, Alexei A. Efros, Hadar Averbuch-Elor, Qianqian Wang, Haiwen Feng.

arXiv cs.GR details Dirichlet Splatting, a new differentiable rendering approach aimed at wave-based inverse problems such as terahertz tomography, synthetic-aperture acoustics, and millimeter-wave radar. The core change is simple but important: instead of borrowing 3D Gaussian splats, the method uses the physically exact Dirichlet kernel …

• Dirichlet Splatting replaces Gaussian splats with the exact Dirichlet kernel for coherent wave-based rendering
• Xingyu Chen, Wuqiong Zhao, Xinyu Zhang, and Tzu-Mao Li