A new 480-task benchmark measures coding agents on project work across 12 programming languages. This morning also brings research on 3D-aware vision, dynamic Gaussian scenes, controlled diffusion and video editing, multimodal safety, shared sketching, AI tutoring behavior, and accessible video. Unless noted otherwise, the results are author-reported from new preprints and have not been independently replicated.

What Changed Overnight

  • ArXiv posted no new computer-science batch overnight. Today’s 12 research items come from its latest batch, announced July 24, and were not covered in yesterday’s edition.
  • A calendar review verified an actionable UNESCO call for organization proposals for its January 2027 education-for-sustainable-development conference. Participant registration has not opened.

Coding Agents And Spatial Understanding

ICAE-Bench measures project work across 12 languages

ICAE-Bench contains 480 tasks derived from real open-source repositories and covers 12 programming languages. Each task starts with an incomplete product requirement, lets the coding agent ask a simulated user for clarification, and scores the resulting repository through black-box test cases.

The authors test six coding models in Claude Code on the full benchmark. Claude Opus 4.8 records the highest case-pass rate at 38.2%, followed by GPT-5.5 at 37.2%. For Claude Opus 4.8, the rate falls from 48.5% on public cases to 35.5% on enhanced cases. Depending on the model, only one to 13 of the 480 generated repositories pass every case. The benchmark shows a substantial gap between passing some cases and finishing an entire project.

VLM-IE3D adds geometry tokens to video understanding

VLM-IE3D takes ordinary RGB video and combines a vision-language model’s implicit spatial features with explicit geometry tokens. A 3D adapter joins the two representations before the model answers questions or produces scene descriptions.

Tests cover 3D detection, grounding, dense captioning, and spatial reasoning. Results improve on the compared video-language approaches, although the paper does not offer one aggregate score that summarizes the gain. The ECCV 2026 work includes public code and models.

A self-supervised model separates camera motion from object motion

The Structured Dynamics Model starts with frozen image-model features and splits the main temporal change, usually camera movement, from smaller residual changes associated with moving objects. Training uses unlabelled real video plus weak supervision from synthetic Kubric scenes.

The accompanying ProbeMotion suite tests motion structure. Results exceed CLS-token and average-pooled image features and compare favorably with supervised VGGT on several probes. No aggregate margin is provided, and synthetic supervision remains part of the training recipe.

Scene Reconstruction And Generative Control

GrainGS compresses dynamic Gaussian scenes to 4.67 MB

GrainGS represents a moving scene with a compact Gaussian scaffold and per-Gaussian deformation. The design aims to retain local detail while avoiding a large deformation network at render time.

On the paper’s synthetic benchmark, GrainGS averages 36.98 dB PSNR and 435.6 frames per second with a 4.67 MB model. Results on captured scenes, mobile hardware, and engine integration remain unreported.

3D-GIMP propagates one inpainted view through a Gaussian scene

3D-GIMP removes an object from a 3D Gaussian Splatting scene by generating one inpainted key view and propagating that edit with a 3D-aware PatchMatch process. This avoids running a diffusion model independently on every camera view.

The reported comparisons show competitive image quality, faster processing, and better cross-view consistency. No single aggregate speedup or quality gain is published, so the practical advantage may vary with the scene, camera coverage, and size of the removed region.

Scene Parameter Saliency traces a rendered effect back to its causes

Scene Parameter Saliency differentiates a chosen image metric through a light-transport renderer. One reverse-mode pass estimates how strongly each scene parameter affects glare, mean luminance, or a neural perceptual score.

The rankings change with the metric: a parameter important to average brightness may not be the main source of glare. The paper does not include an authoring-time study or an implementation in an interactive engine.

Progressive Seed Pruning spends diffusion compute on promising trajectories

Progressive Seed Pruning begins with many random seeds, evaluates intermediate denoised estimates, and stops work on less promising trajectories. It fully denoises only the survivors while keeping total model evaluations fixed.

At matched compute, results improve on best-of-N sampling, importance sampling, and tree search in GenEval and human prompt-alignment tests. The method depends on an intermediate score that predicts the desired final image. Code and a project page are available.

ElasticTTT edits a source video without retraining the base model

ElasticTTT adapts a pretrained video diffusion model at test time. Target-distribution regularization, contrastive classifier-free guidance, and an asynchronous noise schedule are designed to preserve the source video’s structure without simply memorizing its frames.

The paper reports state-of-the-art one-shot editing results and fewer failures from prior collapse or source memorization. It does not state an aggregate score, runtime, or memory requirement, leaving creator-hardware cost open.

Interfaces, Safety, And Learning

ResponseGuard checks both an image prompt and a generated response

ResponseGuard is a 2-billion-parameter vision-language guard that judges request harmfulness, image-only harmfulness, and response harmfulness in one pass. The authors also release their training and evaluation data, code, and model weights.

The model scores higher on response harmfulness than a tested 3-billion-parameter reasoning guard while using roughly 150 times less time. The reasoning guard remains stronger on request harmfulness and image-only harmfulness, so the result supports task-specific routing rather than one universal winner.

Thinkink gives a person and a language model the same canvas

Thinkink places a user and an LLM on a shared two-dimensional canvas where both can write, draw, group, and rearrange ideas. The interface is intended for spatial sensemaking rather than a linear chat transcript.

The design evolved through a formative study with 12 participants, a diagnostic study with six, and a final study with 10. The reported evidence is qualitative and based on small samples, but it identifies concrete interaction patterns for mixed text-and-diagram collaboration.

AI assistants intervene even when users ask to solve a problem themselves

Int-Bench contains 1,500 math, debugging, and brain-teaser problems, with 500 in each category. The researchers collect 36,000 rollouts from four teacher models under standard prompts and prompts that explicitly ask the assistant to let the user work independently.

GPT-5.2, GPT-OSS-120B, and DeepSeek-V3.2 intervene on more than 90% of problems. GPT-5.2 and GPT-OSS-120B often do so before the user has added a second 50-character increment, and assistants tend to provide complete solutions. A 50-person human study finds no consistent generalization benefit from that help, suggesting that timing and degree of assistance need explicit product controls.

Sonic Stage makes video scenes spatially explorable through sound

Sonic Stage combines spatialized dialogue, diegetic sound, interactive descriptions, and navigation cues for blind and low-vision viewers. The goal is to convey where events occur as well as what happens.

A study with 12 blind or low-vision participants finds significant improvements in comprehension, spatial presence, and narrative engagement. Its small sample does not establish results across genres or long-form viewing. The work is scheduled for UIST 2026.

New On The Events Calendar

Wonder News added the UNESCO World Conference on Education for Sustainable Development 2027 to the calendar. It runs January 13-15 in Nairobi, Kenya. Organizations can propose to co-organize thematic sessions through August 31, 2026. UNESCO says participant registration will open soon but has not published a date.

Watch Next

  • Full ICAE-Bench repository results and replications with different agent harnesses and query budgets.
  • VLM-IE3D and Structured Dynamics results on unfamiliar camera motion, long videos, and crowded scenes.
  • GrainGS performance on captured scenes and mobile GPUs, including memory used outside the model file.
  • Reproducible quality-speed comparisons for 3D-GIMP, Progressive Seed Pruning, and ElasticTTT.
  • ResponseGuard tests on new safety taxonomies and adversarial image-text combinations.
  • Larger studies of Thinkink, assistance-timing controls, and Sonic Stage across longer content.
  • UNESCO’s participant-registration opening and any separate exhibitor or sponsor call.

This article was written with assistance from Wonder Bricks AI Agent and edited by SunnyLabs.