New paper: agents must "forget" to keep improving in days-long autoresearch runs
matei_zaharia · x · 2026-10-16
Coding agents can now run for days on a single research problem, but they often stop improving early. The new paper "Scaling Discovery Requires Forgetting" shows long-running agents get locked into small incremental steps in one direction, and introduces a simple fix: forgetting. Matei Zaharia highlighted the result as a way to make auto-research and AI-based optimization more effective.
More from coding & agent
- Agentic sandbox WorldBox doubles GPT-6 Astra's MazeBench score, 23 to 61 gems — ZhitingHu · 2026-10-16
- Two prompts: Gemini 4 Argon turns a single-player game into full multiplayer — chongdashu · 2026-10-16
- Anthropic and Gray Swan launch $300,000 indirect prompt injection red-teaming contest — bcherny · 2026-10-16
- Vercel launches Token Pass: $30/mo for $80 of GLM 5.3 and Kimi K3 credits — evilrabbit_ · 2026-10-16
- Muse Spark Open goes live on Together AI as a Day 0 launch partner — togethercompute · 2026-10-16
- Muse Spark Open: whole-project 1M context, multimodal input, long-horizon self-correction — togethercompute · 2026-10-16