In plain words: Studied why some AI models improve with practice by looking for four habits—checking work, backtracking, setting subgoals, working backward—then training a weaker model on examples showing them. The habits mattered more than correct answers: the weaker model matched the stronger one's gains.
Abstract · Cognitive Behaviors that Enable Self-Improving Reasoners, or, Four Habits of Highly Effective STaRs
Test-time inference has emerged as a powerful paradigm for enabling language models to ``think'' longer and more carefully about complex challenges, much like skilled human experts. While reinforcement learning (RL) can drive self-improvement in language models on verifiable tasks, some models exhibit substantial gains while others quickly plateau. For instance, we find that Qwen-2.5-3B far exceeds Llama-3.2-3B under identical RL training for the game of Countdown. This discrepancy raises a critical question: what intrinsic properties enable effective self-improvement? We introduce a framework to investigate this question by analyzing four key cognitive behaviors -- verification, backtracking, subgoal setting, and backward chaining -- that both expert human problem solvers and successful language models employ. Our study reveals that Qwen naturally exhibits these reasoning behaviors, whereas Llama initially lacks them. In systematic experimentation with controlled behavioral datasets, we find that priming Llama with examples containing these reasoning behaviors enables substantial improvements during RL, matching or exceeding Qwen's performance. Importantly, the presence of reasoning behaviors, rather than correctness of answers, proves to be the critical factor -- models primed with incorrect solutions containing proper reasoning patterns achieve comparable performance to those trained on correct solutions. Finally, leveraging continued pretraining with OpenWebMath data, filtered to amplify reasoning behaviors, enables the Llama model to match Qwen's self-improvement trajectory. Our findings establish a fundamental relationship between initial reasoning behaviors and the capacity for improvement, explaining why some language models effectively utilize additional computation while others plateau.
Kanishk Gandhi, Ayush Chakravarthy, Anikait Singh, Nathan Lile, Noah D. Goodman
arXiv:2503.01307 · cs.CL, cs.LG · submitted Mar 3, 2025 · updated Aug 15, 2025
abstract · pdf · html
As we make AI better, perhaps we'll inadvertently find ways to make HI (human intelligence) better too.
I had a personal experience with this when I was studying for an exam recently. As I read over practice questions, I spoke aloud, replicating the reasoning methods/personality of Deepseek R1. By spending a lot of time reading long verbose R1 outputs, I've essentially fine-tuned my brain for reasoning tasks. I believe this method contributed to my excellent score on that exam.