Unverified50% confidenceBenchmarkExact time
MCPO's improvement is most pronounced on problems that 'GRPO can only occasionally solve', and it grows as k in pass@k increases, reaching a 13.8-point improvement at pass@32.
1
Sources
50%
Confidence
Long-term
Relevance
11/11/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedReproduction experiments on Qwen3-4B-Base show that MCPO also improves over GRPO, with all three seeds positive and an average improvement of 4.75 points.62% similarUnverifiedAfter GRPO reinforcement learning, checkpoints using TailSFT achieved a final pass@1 improvement of up to 3.9 points.61% similarUnverified事实性标注通常要求Kappa大于0.8,而主观性标注可能只要求Kappa大于0.660% similarUnverifiedProVe-lossless's speedup is limited by the degree of match between proposal blocks and AR decoding, averaging 6.6 matched tokens on GSM8K, leaving room for improvement.58% similarUnverifiedtop_k directly truncates to the k most probable candidate tokens, and the smaller k is, the more concentrated the output.57% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1157959API
curl https://kongchang.com/api/v1/knowledge/claims/1157959MCP
get_claim(id=1157959)