Unverified50% confidenceFactExact time
The collapse in true-theorem recognition ability caused by SFT was observed across four repeated experiments with different random seeds and was reproduced on Gemma-3-4B.
1
Sources
50%
Confidence
Long-term
Relevance
10/5/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedUsing 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.65% similarUnverifiedPAWS's replay study reveals that high accuracy rates may mask a model's failure to detect rare stakeholder behaviors.64% similarVerifiedQuantizing a model from FP32 to INT8 can improve inference speed by 3–4x, but may lead to reduced detection accuracy for small targets.63% similarUnverifiedOn update-fact tasks, embedding retrieval methods exhibited unpredictable collapses, with success rates fluctuating between 0.30 and 0.95 and a maximum gap of 0.45 across different random seeds.61% similarUnverifiedMaking SO(3)/SE(3) group operations differentiable exposed numerical issues that are difficult to detect by only checking forward numerical values, and after fixing them, the ordinary NumPy implementation was also improved.60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/975473API
curl https://kongchang.com/api/v1/knowledge/claims/975473MCP
get_claim(id=975473)