Reinforcement Learning Shows Predictable Success Threshold for Learning Rare Behaviors
What happened
Ruben Sanchez-Corcuera and colleagues published a paper investigating reinforcement learning and found that when correct actions occur at low frequencies, whether a model learns the behavior is determined by a predictable numerical threshold: when the expected number of trial groups in which correct actions succeed is approximately 35, the model reaches equal probability of learning. This reveals a binary nature to learning outcomes—either the model learns the behavior everywhere or fails to learn it entirely, with intermediate states rarely occurring. This pattern remains consistent across multiple conditions, including different problem types, trial group sizes (4 to 256), learning rates, optimizers, GRPO variants, and model scales (1.3M to 85M parameters). The research was validated using puzzle worlds with known optimal policies.
AI-written from the coverage · updated 2 d ago
Coverage timeline
Follow the coverage to see the story from different sides.
- arXiv Game AIPickRL Can Only Reinforce What It Sees Succeed
A paper investigates how outcome reward reinforcement learning shifts language models' balance between prior knowledge and observations. In puzzle worlds with known optimal policies, the authors find that when correct actions are rare, RL learning success is determined by a predictable count: the expected number of trial groups in which the action succeeds, with approximately 35 groups yielding equal probability. This pattern holds across different problems, trial group sizes (4 to 256), learning rates, optimizers, GRPO variants, and model scales (1.3M to 85M parameters).
Heat trend for this story
- Current (comparable)
- 3
- Peak (comparable)
- 6Nov 20 18:00
- 24-hour change
- -51%
The trend only compares participants observed continuously and completely, so it may cover fewer than the current heat count. Hover over or tap the chart to see each hour's heat; with a keyboard, use the left and right arrow keys.