RL can turn what an AI already knows into reliable skill — but can it teach it facts it never had?
Can reinforcement learning resolve the gap between knowing facts and knowing how to act?
This asks whether reinforcement learning can turn what a model knows (facts it can state) into what it can actually do (skilled action), and where that approach runs out.
This asks whether reinforcement learning can turn what a model knows (facts it can state) into what it can actually do (skilled action), and where that approach runs out. The short answer from the corpus is yes, partly. RL is good at turning knowledge the model already has into reliable action. It is much worse at supplying knowledge the model never had.
The clearest case for yes comes from game-playing agents. In Can language modeling close the knowing-doing gap in AI?, an LLM writes its strategy in plain language, tries it in the game, and refines it from the results. Over time it builds real procedural skill and can still explain each move, while needing far less data than older methods. The same pattern shows up with domain knowledge: Can reinforcement learning embed domain knowledge more effectively than supervised fine-tuning? finds that rewarding coherent explanations, not just matching the right tokens, makes knowledge stick better than ordinary supervised fine-tuning. Learning by acting does something that copying examples doesn't. Can agents learn beyond what their training data shows? makes the reverse point: agents trained only on expert demonstrations can't learn from their own mistakes, so they can never get better than the people who built the dataset imagined.
The less obvious finding is that 'knowing how' turns out to be two separate skills, and RL learns them in order. Across eight models, Does RL training follow a predictable two-phase learning sequence? shows that training first fixes execution (getting the steps right) and only later starts to improve strategy (deciding which steps to take). Focusing training on the planning parts of the output gives large gains. Research on agent 'skills' finds something similar: Do skills teach procedures or inject missing facts? looked at more than 8,000 trials and found that skills mostly help by steadying how an agent acts (about 66% of cases) and rarely by giving it facts it lacked (about 4.5%). Often the gap was never missing knowledge. It was unreliable action.
That is also where the limit shows. Several studies suggest RL mostly brings out abilities the model already has rather than creating new ones. Does RLVR actually expand what models can reason about? finds that when you let them try many times, untrained base models actually solve more problems than their RL-trained versions. RL narrows the model toward answers it could already reach. Do base models already contain hidden reasoning ability? reaches the same conclusion through five different methods. So RL can close the gap between knowing and doing only when the knowing is already in there. To get genuinely new capability, distillation from a stronger model works better.
A final point: much of RL's limit comes from how little a reward score tells the model. A number says whether an attempt failed, not why. Can natural language feedback overcome numerical reward plateaus? shows that models stuck at a plateau get past it when given written critiques. Should successful and failed episodes be processed differently? treats successes as examples to copy and failures as lessons to summarize. Both suggest the knowing-doing gap may close faster when the feedback itself carries knowledge, so the model learns both what went wrong and how to act.
Sources 9 notes
Think-In Games demonstrates that when LLMs generate language-guided policies refined by environmental feedback, they develop procedural competence while retaining explainability. The approach dramatically reduces data demands and makes agent reasoning transparent at every step.
RLAG rewards both answer accuracy and explanation rationality by cycling between augmented and unaugmented generation, progressively internalizing coherent knowledge structures. This outperforms SFT because it prioritizes reasoning quality over token-level correctness.
Agents trained on static expert datasets cannot learn from their own failures or generalize beyond demonstrated scenarios because they never interact with environments during training. Competence is capped by what curators imagined, not by agent capacity.
Across eight models, RL training consistently shows a first phase where execution correctness drives learning, followed by a second phase where strategic planning becomes the bottleneck. Planning token entropy increases while execution entropy stabilizes, and concentration of optimization on planning tokens yields significant performance gains.
Analysis of 8,135 trials shows procedural anchoring accounts for 65.7% of skill cases versus 4.5% for knowledge injection. Skills fail when retrieved incorrectly, invoked out of context, or followed too rigidly.
Show all 9 sources
Pass@k analysis shows base models outperform RLVR models at high k, indicating RLVR doesn't expand solvable problems but rather narrows sampling toward solutions already in the base model's distribution. Distillation, by contrast, genuinely transfers new reasoning patterns.
Five independent mechanisms—RL steering, critique fine-tuning, decoding changes, SAE feature steering, and RLVR—all elicit reasoning already present in base model activations. Post-training selects rather than creates reasoning; the bottleneck is elicitation, not capability acquisition.
Critique-GRPO shows that models stuck on performance plateaus can generate correct solutions when given chain-of-thought critiques, revealing that numerical rewards lack critical information about why failures occur and how to improve.
SkillRL demonstrates that treating successful episodes as concrete demonstrations and failures as abstracted lessons achieves state-of-the-art performance on complex tasks while using substantially less context than uniform approaches. The asymmetry mirrors human expert reasoning and avoids the degradation seen in uniform consolidation methods.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Eliciting Reasoning in Language Models with Cognitive Tools
- RL Squeezes, SFT Expands: A Comparative Study of Reasoning LLMs
- Do Theory of Mind Benchmarks Need Explicit Human-like Reasoning in Language Models?
- A Survey on Test-Time Scaling in Large Language Models: What, How, Where, and How Well?
- RLVMR: Reinforcement Learning with Verifiable Meta-Reasoning Rewards for Robust Long-Horizon Agents
- Efficient Reinforcement Learning via Large Language Model-based Search
- Demystifying Agent Skills: Why They Work-Until They Don't
- Sharpening Tax in Post-Training