Your AI Agent Isn't Dumb — It's Blind: How MIT's VISTA Scored a Perfect 100

Your AI Agent Isn't Dumb — It's Blind: How MIT's VISTA Scored a Perfect 100

Published: October 9, 2026 | LearnAI Tools Research Breakdown





On October 1, 2026, a team at MIT CSAIL did something that should make every AI builder stop and rethink their entire agent stack. They took the exact same model everyone already had, changed absolutely nothing about its brain — and lifted its score from 40.68 to a perfect 100.00.

No new model. No fine-tuning. No 4,000-line simulator. The secret wasn't intelligence.

It was eyesight.



What the researchers did

The paper — "VISTA: A Visual Harness for Reasoning in an Interactive World" by Qiushi Han, Keya Hu, Linlu Qiu, Cathy Wu, and Kaiming He (MIT CSAIL, arXiv:2610.02200) — tackles one of the cruelest tests in AI: ARC-AGI-3.

ARC-AGI-3 hands an agent an unfamiliar video game — grid puzzles, checkers variants, web games — and demands it learn the rules from scratch, just by playing. No tutorial. The standard setup pipes the game state to the model as a tiny 64×64 numeric text grid and asks: what do you do next?

The MIT team's insight was brutally simple: the models were never bad at reasoning. They were blind. A numeric grid throws away almost everything a game actually shows you.

So they built VISTA, a visual harness with two ideas:

  1. Direct visual perception. The model sees the environment through raw visual observations — actual screenshots — instead of a lossy text grid.
  2. Lossless visual memory. Every frame the environment returns is archived in its original form. Mid-reasoning, the model can retrieve any past observation, zoom into corners, read exact pixel values, and reorganize its visual input. Flipping back through the archive doesn't cost an action — only real moves in the game do.

The results, straight from the paper's abstract, are staggering:

  • Claude Opus 5.0's Relative Human Action Efficiency (RHAE) went from 40.68 → 100.00 — a perfect score.
  • The model cleared all 25 public games, using 57.4% fewer actions than first-time human players.
  • The same harness lifted other models too, and beat minimal-harness baselines across three additional benchmarks of visual games and puzzles.
  • It needed no trained components — just a short shared prompt — and the code is public.

Why it matters

This paper lands like a splash of cold water on the whole agent industry, for three reasons:

1. Perception was the bottleneck, not reasoning. Everyone's been asking "how do we make models smarter?" VISTA suggests the smarter question is "what are we hiding from them?"

2. Stop drowning models in context. More is not better. Extending context or blowing up image sizes actually dropped scores. A tight, well-organized visual memory beat a giant sloppy one.

3. Harness design is the new model design. The gap between 40.68 and 100.00 came from plumbing, not parameters.

How YOU can use it (today)



You don't need MIT's cluster to steal this idea. Here's the VISTA pattern, distilled:

1. Feed screenshots, not lossy text

Pass the raw screenshot to the vision model. Every summarization step is a chance to delete the exact detail the model needed.

2. Build a lossless observation store

Keep every observation in its original form. Retrieving a memory must be free — only environment actions cost.

3. Keep the prompt short

VISTA worked with a short shared prompt, not a 50-page system instruction.

4. Measure actions-to-goal

Track actions-to-goal, not just success/failure — that's the number your users actually feel.

Key takeaways

  • 🔬 Paper: "VISTA: A Visual Harness for Reasoning in an Interactive World" — Han, Hu, Qiu, Wu, He (MIT CSAIL).
  • 📊 Result: 40.68 → 100.00 on ARC-AGI-3; all 25 games cleared.
  • 💡 Lesson: Before scaling intelligence, fix perception and memory.
  • 🛠️ Action: Screenshots over summaries, lossless memory, free lookups, short prompts.

📄 Read the original paper: arXiv:2610.02200 — code is public on GitHub.

New here? LearnAI Tools breaks down one AI research paper every day — in plain English, with code you can actually use. Follow our WhatsApp Channel for daily drops.

Comments

Post a Comment

Popular posts from this blog

3,000 Lines vs 512,000: The Tiny Agent Harness That Nearly Matches Claude Code

Image
3,000 Lines vs 512,000: The Tiny Agent Harness That Nearly Matches Claude Code The default advice when your AI agent disappoints you is simple: get a bigger model. Rent the newest frontier model, pay the per-token bill, and hope the extra brainpower covers the gaps. A new paper from a team building agents for government intelligence work suggests the opposite move: keep the model small, fixed, and self-hosted — and shrink the harness around it until almost nothing is left. Their minimal agent, called CLEO, scored 83.8% on a held-out investigative-analysis dataset. The strongest out-of-the-box baseline, the Claude Code SDK, scored 85.0%. The gap is 1.2 percentage points. The gap in size is two orders of magnitude: the paper's comparison table puts CLEO's harness at roughly 3k against the Claude Code SDK's 512k. This is a paper about what actually makes agents good — and the answer is not the model. The paper is "Optimize the Harness, Not the Model: Fitting a Min...
Read more

Silent AI Video Is Dead: This Open-Source Model Dreams Up Movies and Soundtracks Together

Image
Silent AI Video Is Dead: This Open-Source Model Dreams Up Movies and Soundtracks Together The last great open problem in AI video generation was never sharper pixels — it was sound. For years, open-source video models produced beautiful silent clips. If you wanted dialogue, footsteps, or a crackling fire to match the picture, you stitched together three separate tools: a video generator, a voice or sound generator, and hours of manual sync in an editor. A new release from Team Kandinsky collapses that whole pipeline into a single model — and gives it away under the MIT license. Meet Kandinsky 6.0 Video (arXiv:2610.05608, published October 4, 2026): a family of diffusion models that generate five-second video clips with synchronized 44 kHz audio — including lip-synced speech — from a text prompt or a starting image. Code, checkpoints, and a diffusers integration are all public. Let's break down what the researchers actually built, why it matters, and how you can run it yourse...
Read more

Your AI Agent Works in the Lab and Fails at Your Desk — Training It in 18 Setups Fixes That

Image
You have seen this failure before. An AI agent looks brilliant in a demo: it files the ticket, updates the spreadsheet, writes the summary. Then you hand it your real setup — a different code editor, a different tool layout, a slightly different way of showing results — and the same agent starts fumbling. It calls the wrong tool, loses track of the task, or produces work that looks finished but is not. The model did not get dumber overnight. The setup around it changed, and the agent had only ever learned to work inside one setup. A new paper from Renmin University of China, IQuest Research, and Beihang University offers a practical fix: train an agent across many varied harnesses, not just one. The result, they report, is an agent that transfers to setups it has never seen — including a popular coding tool it was never trained in. What the researchers did: The paper is "Harness Scaling for Generalizable Agentic Reinforcement Learning" by Fei Bai, Huatong Song, Chuan Hao, Ran...
Read more