On-policy reinforcement learning, where a model learns from its own generated actions and their consequences, is analogous to how humans learn from direct experience and mistakes. This contrasts with off-policy methods like supervised fine-tuning (SFT), which resemble simply imitating others' successful paths.
Related Insights
Modern LLMs use a simple form of reinforcement learning that directly rewards successful outcomes. This contrasts with more sophisticated methods, like those in AlphaGo or the brain, which use "value functions" to estimate long-term consequences. It's a mystery why the simpler approach is so effective.
In domains like coding and math where correctness is automatically verifiable, AI can move beyond imitating humans (RLHF). Using pure reinforcement learning, or "experiential learning," models learn via self-play and can discover novel, superhuman strategies similar to AlphaGo's Move 37.
Training AI agents to execute multi-step business workflows demands a new data paradigm. Companies create reinforcement learning (RL) environments—mini world models of business processes—where agents learn by attempting tasks, a more advanced method than simple prompt-completion training (SFT/RLHF).
Beyond supervised fine-tuning (SFT) and human feedback (RLHF), reinforcement learning (RL) in simulated environments is the next evolution. These "playgrounds" teach models to handle messy, multi-step, real-world tasks where current models often fail catastrophically.
Reinforcement Learning with Human Feedback (RLHF) is a popular term, but it's just one method. The core concept is reinforcing desired model behavior using various signals. These can include AI feedback (RLAIF), where another AI judges the output, or verifiable rewards, like checking if a model's answer to a math problem is correct.
Basic supervised fine-tuning (SFT) only adjusts a model's style. The real unlock for enterprises is reinforcement fine-tuning (RFT), which leverages proprietary datasets to create state-of-the-art models for specific, high-value tasks, moving beyond mere 'tone improvements.'
The transition from supervised learning (copying internet text) to reinforcement learning (rewarding a model for achieving a goal) marks a fundamental breakthrough. This method, used in Anthropic's Opus 3 model, allows AI to develop novel problem-solving capabilities beyond simple data emulation.
The distinction between imitation learning and reinforcement learning (RL) is not a rigid dichotomy. Next-token prediction in LLMs can be framed as a form of RL where the "episode" is just one token long and the reward is based on prediction accuracy. This conceptual model places both learning paradigms on a continuous spectrum rather than in separate categories.
Biological evolution used meta-reinforcement learning to create agents that could then perform imitation learning. The current AI paradigm is inverted: it starts with pure imitation learners (base LLMs) and then attempts to graft reinforcement learning on top to create coherent agency and goals. The success of this biologically 'backwards' approach remains an open question.
When determining what data an RL model should consider, resist including every available feature. Instead, observe how experienced human decision-makers reason about the problem. Their simplified mental models reveal the core signals that truly drive outcomes, leading to more stable, faster-learning, and more interpretable AI systems.
