Memento 3 clears all 25 public ARC-AGI-3 levels with rulebook-based self-improvement
hbouammar · x · 2026-11-11
Researchers from UCL, Huawei Noah's Ark Lab and Liverpool introduce Memento 3: a frozen LLM agent maintains a natural-language rulebook as persistent memory, compiles it into executable code, and iterates through observation-reflection-revision-verification loops, accepting only verified updates—a model-based route to recursive self-improvement. On ARC-AGI-3 the single-agent version clears every level of all 25 public games with mean RHAE of 100.0 using just 44% of human action counts; a population variant runs parallel world models. In an Atari Pong case study, a learned controller wins 21:0 across three openings without further LLM calls.
More from AGI Musings
- Gary Marcus calls Dario Amodei's disease-cure timeline a scam — AlexTensor · 2026-11-11
- Greatest Cyber Risk May Be Offsec Firms Holding Offensive Model Access Getting Pwned — anpaure · 2026-11-11
- Researchers argue in Nature Human Behaviour that AI 'slop' is a social judgement, not a natural category — manoelribeiro · 2026-11-11
- Researcher: Don't Mock Imperfect AI Progress Models — Build Better Ones — austinc3301 · 2026-11-11
- AI risk models are untested best guesses, not overestimates, argues researcher — austinc3301 · 2026-11-11
- Ramez Naam: calibration shows AI progress forecasts systematically overestimate — ramez · 2026-11-11