AI models are retracing humanity's early cognitive detours.

Re-deriving formulas for geometric series problems from scratch, reconstructing the inclusion-exclusion principle for every probability calculation—Meta's latest paper exposes that current large language models reason like amnesiac mathematicians. They solved this with 'behaviors': extracting compressed thinking patterns from models' own reasoning traces, turning 'behavior_inclusion_exclusion' into reusable cognitive tools.

The results are immediate. On the MATH dataset, behavior-guided reasoning saved 46% tokens while maintaining accuracy; on AIME competition problems, behavior-directed self-improvement outperformed standard critique-revision methods by 10% accuracy. Crucially, when models fine-tuned on behavior-curated reasoning traces, smaller models became not just faster but better thinkers—these behaviors serve as scaffolding for complex reasoning.

Three figures tell the whole story:

1. The behavior extraction pipeline is surprisingly simple (Fig.1): model solves problem → analyzes its solution → extracts reusable behaviors, no architectural changes needed

2. A dice probability case (Fig.2) shows how models replace repeated derivations with 'behavior_inclusion_exclusion'

3. Efficiency comparison curves (Fig.3) reveal behavior-tuned reasoning's cliff-like advantage on both R1-Llama-70B and Qwen3-32B

This overturns the dogma that 'longer thinking equals better results.' When models accumulate procedural memory, reasoning breakthroughs no longer depend on larger models or longer chains of thought. The pressing question: how much of those massive compute bills for redundant reasoning were pure waste?

Paper: https://arxiv.org/abs/[paper-number]