Alibaba Research's Recursa lifts DeepSeek-V4.1-Flash from 30% to 55% on legal tasks without touching weights
omarsar0 · x · 2026-11-05
Alibaba Research published "Recursa: Toward Recursively Self-Improving Foundation Model Systems," a recursive self-improvement framework that needs no stronger teacher model or human experts.
Key mechanism: model weights stay frozen — all gains come from a knowledge repository the system writes for itself, storing domain skills, usage instructions, memories of past tasks, and reflections. The pipeline bootstraps the repo from domain materials, then parallel learners run tasks and propose fixes; only validated fixes get merged back.
Results: on Harvey LAB's Firm Knowledge benchmark, all-pass rates on held-out legal tasks rise from 30.0% to 55.0% for DeepSeek-V4.1-Flash, and from 35.0% to 46.0% for Qwen3.8-Flash. The paper frames this as "foundation model system learning" — improving the model, instructions, skills, memory, and harness as one system.
More from AGI Musings
- 'There are going to be a lot of Shire planets': singularity daydreaming — moultano · 2026-11-06
- The hidden cost of agentmaxxing: delegating chores to AI agents erodes taste and creativity — bendee983 · 2026-11-06
- NYU Shanghai's Centre for AI and Culture opens 1-2 year postdoc fellowship under Bogna Konior — sebkrier · 2026-11-06
- Math publishing may shift to humanities model where the definitive book is the unit of progress — prof_g · 2026-11-05
- AI researcher: negative results matter and the pressure to splash is dialed to 11 — code_star · 2026-11-05
- OpenAI President Brockman: major AI strides in biology within 12-18 months — basedjensen · 2026-11-05