Meta and Princeton's Strategist evolves inference strategies, cutting cost 39%

armand_ruiz · x · 2026-11-10

Researchers from Meta and Princeton published "Strategist: Evolving Reasoning with Agents," a meta-agent that evolves inference strategies across tasks, turning the previously hand-written reasoning procedure into a searchable, reusable asset.

How it works

Each strategy is an executable code blueprint stored in a shared library, so later tasks reuse earlier discoveries;
Each round: retrieve relevant blueprints → propose 5 candidate strategies → test on real samples → revise from results;
Models stay frozen; only strategies evolve, with measured accuracy and cost accumulated into the library.

Results

Beats the strongest baseline on 24 of 30 benchmarks;
Matches average accuracy (74.7% vs 74.0%) at 39% lower inference cost (0.61×);
Strategies transfer to unseen tasks and other models with no further search: +20.5 points on GPT-5 (0.65× cost), +14.7 on Gemini-2.5, +1.6 on Qwen3.

The authors argue the one part of an agent system that never improved on its own — the reasoning procedure — is now searchable and accumulable. Code is open-sourced.

Original post →

Meta and Princeton's Strategist evolves inference strategies, cutting cost 39%

More from coding & agent