Summary
A model is only half of an AI agent system; the other half is the inference strategy that turns model calls into a response—typically hand-written today, with no transfer of experience between tasks. Strategist, presented in the paper "Strategist: Evolving Reasoning with Agents" (arXiv 2611.03618) by Taiming Lu, Qi Xu, Zhigang Hua, and Zhuang Liu, is a meta-agent that automatically evolves inference strategies across tasks. Each strategy is written as a blueprint—executable code that other strategies can reuse—and all blueprints form a growing library. Given a task and a small development set, Strategist retrieves related blueprints, writes candidate strategies, tests them on real samples, and revises them based on results, so later tasks build on earlier discoveries. Across 30 benchmarks, Strategist beats the strongest baseline on 24, achieving comparable average accuracy (74.7% vs 74.0%) with 39% lower inference cost. Its strategies also transfer to unseen tasks and other models without further search.
This forum post introduces the paper Strategist: Evolving Reasoning with Agents (arXiv: 2611.03618) by Taiming Lu, Qi Xu, Zhigang Hua, and Zhuang Liu (NLP, released 2026-11-02).
Key idea
A model is only half of an AI agent system. The other half is the inference strategy: the reasoning procedure that turns model calls into a response. Today that second half is written by hand, and the experience of building one strategy never carries over to the next.
How Strategist works
- Strategist is a meta-agent that evolves inference strategies across tasks.
- Each strategy is written as a blueprint: executable code that other strategies can reuse.
- All blueprints together form a library.
- Given a task and a small development set, Strategist:
1. Retrieves related blueprints
2. Writes candidate strategies
3. Tests them on real samples
4. Revises them based on results
- Every candidate enters the library, so later tasks build on earlier discoveries.
Results
- Evaluated on 30 benchmarks; Strategist beats the strongest baseline on 24 of them.
- Average accuracy is comparable to the strongest baseline (74.7% vs 74.0%), but with 39% lower inference cost.
- Strategies transfer to unseen tasks and other models without further search.
As the authors put it: the half of AI systems that never improved itself in the past has now begun to improve.
---
*Auto-collected on 2026-11-04.*
This page is an English static mirror generated for search and AI citation.
It may be a full translation or structured summary of the Chinese original.
Canonical interactive discussion lives on the Chinese page:
https://zhichai.net/topic/178636729