Six papers from the past week, each with the one thing about it that matters.
WorldBench — the two ways of grading a generated 3D world disagree 32% of the time
arXiv:2610.10622. Language models can now write complete interactive 3D worlds as code. Grading them automatically is the problem, and existing judges each take one view: a vision-language model scores a few rendered snapshots, or a language model reads the source.
On worlds written by five frontier models, the two views disagree on 32% of required items — “mostly code that no frame shows,” plus fixed camera views missing small close-up content.
That figure is the useful finding and it generalises well past this benchmark. If you are evaluating anything generated as code-that-renders, a screenshot judge and a source judge are measuring different artefacts. Something present in the code and never visible from the chosen angle counts as a pass to one and a fail to the other. Any eval you build for generated 3D needs to decide which question it is asking, and say so.
Fluid-Gen-Zero — stop asking the video model to know physics
arXiv:2610.10984. A training-free framework for fluid-object interaction video that decouples physical reasoning from appearance synthesis: delegate the motion dynamics to an actual physics simulator, and leave the pretrained video generator doing what it is good at, which is appearance.
The bridge is a two-level agentic workflow — a VLM agent plans generation clips from the simulation rollout, then simulation signals are injected into denoising through region-aware latent wrapping. Plug-and-play across current video foundation models.
The numbers are large: object trajectory error down 26.7%–81.5%, fluid flow endpoint error down 67.9%–84.0%, with perceptual quality largely preserved. Human raters preferred it in 55.1%–74.4% of same-backbone comparisons across Tora, VACE and WanMove — and in 90.4%–94.2% of comparisons against simulation-based methods.
That last pair of figures is the interesting one. It beats the video models on physics and beats the simulators on looking right, which is the whole argument for the split.
BrickBench — agents satisfy the constraints and still lose to humans
arXiv:2610.12452. Text-conditioned LEGO design, scored on whether the result is semantically right, well designed, and physically buildable — the agent has to pick parts from a discrete library and reason about local and global constraints together. Ships with BrickAgent, an environment for coding agents to construct, inspect and validate designs.
The finding: leading agents largely satisfy the verifiable physical and semantic requirements, but fall short of human designs.
Which is the same shape as Design Creativity Bench on LLM-generated UI from last week — appropriateness is close to solved, quality is not. When a benchmark’s verifiable criteria are all met and the output is still worse, the criteria are not measuring the thing.
SpaceFlow — per-part control levels for 3D generation
arXiv:2610.12399. Current 3D generation gives you a global control strength and no way to specify appearance locally. SpaceFlow is a training-free pipeline taking text plus a collection of geometric primitives, where each primitive proxies an object part and gets its own local control level — strictly follow this shape here, allow generative completion there.
For appearance, the structure is segmented, matched to the primitives, and each part is conditioned only on its own text or image cue, limiting cross-part leakage. That is the detail worth having: the usual failure of prompting a whole object is that a material described for one part bleeds across the rest.
Reported as state-of-the-art prompt faithfulness and colour/material accuracy when evaluating appearance on fixed geometry.
Neural radiance caching, specialised for specular
arXiv:2610.11702. Neural Radiance Caching gives real-time path tracing an online neural representation of scene illumination. This variant is tuned for specular lighting via a reflection-direction parameterisation and a roughness-dependent radiance target: the network predicts prefiltered incoming radiance at a surface point, combined with a precomputed BRDF integration map using the split-sum approximation.
Faster convergence and better specular quality than the evaluated NRC baselines on Bunny and Specular Sponza, while staying real-time and continuing to adapt the cache during rendering.
Specular is where every real-time global illumination approximation shows its seams — diffuse hides a lot and gloss hides nothing. Borrowing split-sum, which the whole games industry already uses for image-based lighting, is a sensible way in.
Tabula Rasa — noise with controlled variance and controlled correlation in time
arXiv:2610.11653. A method for generating time-varying Gaussian noise with controlled variance and controlled temporal correlation, framed as joint Monte Carlo estimation of a pixel reconstruction and of variance, using “sketching” from the database literature.
The claim is practical rather than grand: temporal control for downstream tasks with simpler and faster code than previous methods.
Worth filing because controllable temporal noise is the quiet dependency of a great deal of visual work — temporal denoising, stochastic effects that need to not shimmer, dithering that holds still across frames, and the phase-shifted noise used to mask foveated-rendering judder we covered last week. Methods that make it cheaper tend to show up everywhere a few years later.