Princeton & Meta's Strategist evolves inference strategies, beating baselines on 24/30 benchmarks at 39% lower cost
141_1337 · reddit · 2026-11-21
Princeton and Meta researchers introduce Strategist (arXiv:2611.03618), premised on the idea that a model is only half of an agent system — the other half, the inference strategy, is still hand-written and its experience never carries over.
How it works: Strategist is a meta-agent that turns strategy design into search. Each round it retrieves related blueprints from a shared library, writes 5 candidate strategies as executable code, evaluates accuracy and cost on real samples, and accumulates every result back into the library. Models stay frozen; only strategies evolve.
Results:
- Beats the strongest baseline on 24 of 30 benchmarks (74.7% vs 74.0% avg accuracy) at 39% lower inference cost
- One-shot retrieval: 81.4% vs 77.8% accuracy, -23% cost
- Strategies transfer across models without further search: Qwen3 +1.6pp, GPT-5 +20.5pp, Gemini-2.5 +14.7pp at 0.6× cost
Code is open-sourced.
More from coding & agent
- LangChain ships four updates: LangSmith Agent Platform, trace data controls, Tool Gateway — LangChain · 2026-11-21
- Grok Bot Gets Its Own Email Addresses to Sign Up and Handle Promos for You — elonmusk · 2026-11-21
- CapCut × ChatGPT plugin turns chat into a video generator and editor — SimplyAnnisa · 2026-11-21
- MendAgent: a multi-agent self-healing system that auto-repairs Dockerized microservices — ExpressionOne8668 · 2026-11-21
- Dev builds daily X-digest Grok bot that auto-blocked 19 reply bots, heading to marketplace — petergyang · 2026-11-21
- DeepMind's Antigravity Robotics Plugin: 2 Prompts, ~20 Minutes to Command a Real FR3 Duo Arm — jackyliang42 · 2026-11-21