A new paper from Stanford, MIT, Harvard, and Anthropic explains why larger models learn rare skills. The key insight: bigger models forget less during training, their extra capacity protecting weak learning signals from being overwritten.
Key facts
- Paper from Stanford, MIT, Harvard, and Anthropic researchers.
- Tested with OLMo models from 4M to 4B parameters.
- Larger models showed less gradient interference on rare tasks.
- Common tasks claim neurons first, overwriting rare signals.
- Small models briefly pick up rare signals but forget them.
A collaboration between researchers at Stanford, MIT, Harvard, and Anthropic has published a paper titled "Why Larger Models Learn More: Effects of Capacity, Interference, and Rare-Task Retention" arXiv:2605.29548. The work provides a training-based mechanism for why larger language models acquire abilities that smaller models miss, even when both are trained on the same data.
The authors argue the issue is not whether a small model could represent a task, but whether training allows it to retain that task while common tasks repeatedly update the same limited parameters. Their core idea: common tasks claim the model's neurons first, so rare tasks get overwritten before they appear often enough to become stable knowledge. In a crowded data mixture, common patterns get first claim on the model's internal machinery. Small models may briefly pick up a rare signal, but the next wave of common-task updates overwrites it before the signal appears again.
Controlled experiments and OLMo tests
The team first tested this hypothesis with controlled toy tasks where they could independently vary rarity and complexity. They then scaled to OLMo language models ranging from 4 million to 4 billion parameters. The main result: bigger models learned low-frequency tasks much better, kept more task features inside their representations, and showed less gradient interference — meaning common-task updates disturbed rare-task learning less. Larger models can remember weak rare signals long enough to turn them into real learned skills.
The paper gives a clear training-based explanation for the emergent abilities often attributed to scale alone. It suggests the bottleneck is not representation capacity but training dynamics: the interference between frequent and infrequent patterns during gradient descent.
What to watch
Watch for follow-up work testing these findings on frontier models like GPT-4 or Claude 3.5 Opus, and whether training curricula that explicitly schedule rare-task exposure can close the gap for smaller models without additional parameters.
