MIT's VISTA gives Claude visual memory to clear all 25 ARC-AGI-3 games with 57.4% fewer actions
mark_k · x · 2026-10-03
MIT researchers built VISTA, which gives AI models a visual memory they can actively inspect: every observed frame is preserved, letting the model revisit earlier moments, compare scenes, and zoom into details while inferring a game's rules.
With VISTA, Claude completed all 25 public ARC-AGI-3 games without any additional training, using 57.4% fewer actions than first-time human players. Claude Opus 5.0 scored a perfect 100 on action efficiency, with GPT-5.6 Sol at 99. The approach also improves results on three other visual benchmarks, though private-game testing remains pending.
The takeaway: today's models may already hold far more capability than they show — giving an agent the ability to look again can unlock it.
More from Models
- Bittensor's Cascade beats Datadog Toto 2.0 with 84.6% fewer training tokens — bittingthembits · 2026-10-03
- Qwen 3.8 Flash Next q2_0 hits 10 tok/s on a 6GB VRAM laptop via Strata engine — dampflokfreund · 2026-10-03
- Claude Opus 5.5 and Sonnet 5.5 Now Available in Google Antigravity — algo_diver · 2026-10-03
- OpenAI's Fast mode burns 2.5x subscription quota for only 1.5x speedup, tests find — lxfater · 2026-10-03
- Bittensor SN81 generated ~3B tokens in a week, boosting Qwen3-4B math score from 37 to 73 — bittingthembits · 2026-10-03
- Devin called best-value AI subscription: SWE-2 plus Opus 5.5 combo barely uses 10% of quota — CtrlAltDwayne · 2026-10-03