Scaling Collapse
A phenomenon observed in LLM reasoning fine-tuning whereby, given a fixed set of training problems, simply increasing the number of correct reasoning examples fails to produce continued performance improvements. The underlying cause is the exhaustion of data diversity, resulting in diminishing marginal returns as the model encounters no meaningfully new reasoning patterns to learn from.
Timeline (last 90 days)
Scaling Collapse refers to the phenomenon whereby, when the training problem set is limited, simply increasing the number of correct reasoning examples does not lead to sustained performance improvements.
The Reflective Recovery method breaks through the barrier of Scaling Collapse.
All Facts (2)
Scaling Collapse refers to the phenomenon whereby, when the training problem set is limited, simply increasing the number of correct reasoning examples does not lead to sustained performance improvements.
50%UnverifiedThe Reflective Recovery method breaks through the barrier of Scaling Collapse.
50%