Unverified50% confidenceSolutionExact time
The team built an inference-time penalizer that outperforms the original token set provided in the paper and is effective for both low-precision and BF16 models.
1
Sources
50%
Confidence
Long-term
Relevance
9/14/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedThe essence of reasoning-enhanced models is trading inference-time compute for accuracy62% similarUnverifiedPerforming task-specific reinforcement learning on a small 4B-parameter model improved its mapping reasoning accuracy by 15.7 percentage points.62% similarVerifiedKnowledge distillation technology enables small models to extract a considerable proportion of capability from large models, undermining the validity of fixed compute thresholds as a proxy metric for capability.61% similarUnverified斯坦福大学和Google DeepMind等机构的研究表明，经过结构化重写的提示词在代码生成、逻辑推理等任务上的准确率可提升20%-40%61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/920994API
curl https://kongchang.com/api/v1/knowledge/claims/920994MCP
get_claim(id=920994)