He Kaiming's Team Unveils VISTA, a Visual Harness Scoring 100 on ARC-AGI-3
量子位 · wechat · 2026-10-09
He Kaiming's team proposes VISTA (A Visual Harness for Reasoning in an Interactive World), a visual-native harness that lets existing multimodal models observe environments, store raw frames losslessly, and re-inspect, zoom and check details at inference time — with no retraining. Unlike abstracting environments into text or code, VISTA keeps raw visual information so models can pull past frames back into context on demand, forming a native scaffold around visual memory.
VISTA has three components: visual observation (upscaling ARC-AGI-3's 64×64 frames to 512×512 PNGs), lossless visual memory (indexing every frame by turn and frame number, including animation intermediates), and active visual inspection (an inspect tool to recall turns, crop/zoom, and compare frames). The loop is observe–hypothesize–gather evidence–predict–act–compare–correct, supported by GUIDE.md (reusable cross-level rules) and WORKING.md (current-level state), with handoff summaries when context nears its limit. Only the last frame is shown by default; intermediates stay archived.
With VISTA, Claude Opus 5.0 cleared all 25 public ARC-AGI-3 games with a perfect 100 relative human action-efficiency score, using 57.4% fewer actions than first-time human baselines; GPT-5.6 Sol also cleared all, scoring 99. GameWorld success rose from 40.0% to 63.3% across 170 tasks; AIGameStore score jumped from 47.3 to 140.3 (human median normalized to 100); BabyVision maze/connect-the-dots accuracy rose from 41.0% to 63.2%. Authors include Qiushi Han, Keya Hu, Linlu Qiu, Cathy Wu and He Kaiming, with embodied tasks as future work.
More from coding & agent
- DuckDB v2.0 CLI agent mode cuts agent-read tokens by 59% on TPC-H benchmarks — josh_wills · 2026-10-10
- Claude Code agents beat torch.compile on CUDA kernels: 2.5x fused GELU, 1.57x matmul via fp32-splitting trick — lmoroney · 2026-10-10
- Roadie: open-source Go USB KVM gives AI agents hands via HTTP — hugs · 2026-10-10
- VirusTotal: fastest-growing AI agent ecosystem OpenClaw becomes a malware delivery channel — Bedrovelsen · 2026-10-10
- Nuwa.skill hits 33.9k GitHub stars distilling anyone's thinking into agent skills — AlchainHust · 2026-10-10
- Microsoft launches Database Hub in Fabric: one place to detect, investigate and automate database ops — adnan_hashmi · 2026-10-10