Unverified50% confidenceBenchmarkExact time
Using 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.
1
Sources
50%
Confidence
Long-term
Relevance
10/5/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedA cosine annealing learning rate schedule may help the model break through loss plateaus and avoid getting stuck in local optima.69% similarUnverifiedWhen a model is optimized to maximize a reward signal, any behavior that improves the score may be learned; researchers refer to this as reward hacking or specification gaming69% similarUnverifiedThe 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.68% similarUnverifiedThe team used the identified tokens to construct a loss function and ran LoRA SFT, successfully reducing reasoning length but causing accuracy to drop.68% similarUnverifiedA reasonable prediction loss combined with 0% downstream control accuracy is highly consistent with representation collapse, meaning the encoder has fooled the self-supervised objective without learning any useful dynamic information.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/975474API
curl https://kongchang.com/api/v1/knowledge/claims/975474MCP
get_claim(id=975474)