Unverified50% confidenceFactExact time
The paper proves that starting from the same perfectly fitted checkpoint, SFT's forgetting over a finite training interval has a lower bound that is strictly greater than zero.
1
Sources
50%
Confidence
Long-term
Relevance
10/5/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedSFT fine-tuning is effective but causes a decline in general knowledge in small models—i.e., the phenomenon of catastrophic forgetting—due to the prominent capacity bottleneck of small models.68% similarUnverifiedUsing a sparse reinforcement learning reward signal based solely on final outcomes, the model repaired the true-theorem recognition ability destroyed by SFT and raised the recognition rate to 0.66, surpassing the base model.68% similarUnverifiedStandardized benchmarks have the limitation of out-of-distribution generalization, where agents may score by memorizing training set patterns rather than truly possessing reasoning ability.67% similarUnverifiedMERIT verifies tasks' dependence on facts from earlier episodes through a leakage detection mechanism, and only tasks with a success rate of 0 without memory are included in the final evaluation.67% similarUnverifiedSupervised fine-tuning (SFT) tends to cause more severe catastrophic forgetting than reinforcement fine-tuning (RFT).66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/975969API
curl https://kongchang.com/api/v1/knowledge/claims/975969MCP
get_claim(id=975969)