Memento 3 clears all 25 ARC-AGI-3 games with a frozen LLM, using 44% of human actions
dair_ai · x · 2026-10-14
A new paper demonstrates recursive self-improvement with a frozen LLM: Memento 3 clears every level of all 25 public ARC-AGI-3 games using only 44% of the human action count, without ever updating the underlying LLM.
How it works:
- The agent writes its beliefs about the environment as a natural-language rulebook, then compiles it into code that predicts next states
- A new code version is kept only if replaying it reproduces every observed transition cell-for-cell
- In an Atari Pong case study, the learned model drives a controller winning 21:0 across three games with no further LLM calls
More from Research
- SAP paper teaches LLM agents to predict their own failure (AUROC 0.85) — SAP · 2026-10-14
- CUHK's Orchestra robot framework completes 7.35-step subtask sequences — CUOHK · 2026-10-14
- Robot Learns Hand Skills From 1M Hours of First-Person Video, Generalizes to Unseen Tasks — Scobleizer · 2026-10-14
- Nvidia's Octi shifts dexterous manipulation to sim-based RL, academia rejoices — KyleMorgenstein · 2026-10-14
- Meta-Video shows frame sampling, backend and temperature skew video understanding benchmarks — HildeKuehne · 2026-10-14
- BinProp scaffold nearly triples LLMs' chess-explanation scores by writing analysis code — OfirPress · 2026-10-14