An AI system can change how it approaches a problem without changing its underlying models. Strategist, a system described by researchers affiliated with Princeton University and Meta, searches for reusable programs that organize model calls. In the authors’ evaluation across 30 benchmarks, it achieved 74.7% mean test accuracy against a selected baseline’s 74.0%, while reducing a model-weighted token-cost measure by 39%.
The work, by Taiming Lu, Qi Xu, Zhigang Hua and Zhuang Liu, was posted to arXiv on November 2, 2026. It is a preprint; peer review has not been verified, and no independent replication is supplied.
Strategist targets the procedure around a model: how to divide a problem, generate alternatives, check answers and allocate stronger or cheaper models. A frozen GPT-6 meta-agent writes Python programs called blueprints, which share an interface and can call earlier blueprints. Task-specific prompts turn these reusable structures into executable strategies. The underlying models remain frozen throughout.
For each search task, the system proposes five candidates per iteration over 10 iterations, tests them and records reusable structures, measurements and experience notes. Execution uses Qwen3 models with 0.6 billion, 1.7 billion and 4 billion parameters, each with thinking enabled or disabled. The library eventually contained 1,865 blueprints.
The evaluation covered 40 benchmarks across mathematics, commonsense, knowledge, multi-hop question answering, logic, coding and recommendation. Thirty supported iterative search; 10 were held out for retrieval-based transfer. Comparisons used 18 baseline methods across six model configurations. The reference baseline, BL, was whichever method–configuration pair scored highest on each benchmark’s training split, subsequently evaluated on test questions.
The reported 24 wins out of 30 were joint accuracy–cost wins, not 24 accuracy victories. A win required an improvement exceeding two accuracy percentage points or 2% in cost, without deterioration beyond the corresponding tolerance. Strategist tied on two benchmarks, traded accuracy for cost on two and lost on two. On MedQA and AIME 2024, accuracy fell by approximately 4–5 percentage points while cost declined to 0.44 and 0.77 times BL, respectively.
The 39% saving is the average of per-benchmark cost ratios. It measures input and output tokens weighted by model-size ratios, rather than measured latency, energy or dollar spending. It also excludes the overhead of searching for strategies, leaving total savings and the usage needed to recover search costs unknown.
On the 10 held-out tasks, Strategist selected among five retrieved blueprints using task-specific prompts and up to 300 development questions, without another iterative search. Mean accuracy reached 81.4% versus BL’s 77.8%. Accuracy was higher on eight tasks, but only six qualified as joint wins; some strategies substantially increased cost. Unmodified strategies also won on seven of 10 tested benchmarks with GPT-5 and eight with Gemini-2.5, with reported aggregate inference spending reductions of 35% and 40%. Some GPT-5 advantages arose because baseline workflows produced malformed answers.
A separate 10-benchmark comparison found 78.7% mean accuracy with the full system versus 65.4% without its library, supporting a library contribution within this implementation. Attribution remains limited by selection differences: Strategist used validation and hidden splits to choose strategies, while BL was selected on training accuracy. The selected strategies’ average advantage was 4.3 percentage points on hidden data but only 0.7 on the separate test data, illustrating selection optimism.
The supplied evidence reports no confidence intervals, repeated-run variability or statistical significance. Independence across benchmarks, pretraining contamination and sensitivity to task order are also unestablished. The results support further investigation of reusable strategy design, but do not establish general reliability or performance in real-world deployments.
Source: Strategist: Evolving Reasoning with Agents, arXiv:2611.03618v1, https://arxiv.org/html/2611.03618v1
Cover illustration generated with AI; conceptual, not research evidence.