Unverified50% confidenceBenchmarkExact time
In the Skatebench test, from the x-high tier to the max tier, reasoning tokens per question spiked from approximately 330 to nearly 5,000 (more than 10x), while scores improved by only 1%.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/25/2026
First Seen
Valid until: 12/24/2026
Sources
Related Entities
Related Claims
UnverifiedToken consumption for reasoning at the XHigh setting can reach 3–5 times that of the High setting, but the marginal accuracy improvement is typically only 5–15 percentage points.69% similarUnverifiedAblation analysis shows that fine-tuning on the segment paradigm yields a +5.29-point improvement on the test set, while replacing the first 4,500 transcript words with 2,000 retrieved words yields only a +1.55-point improvement.69% similarUnverifiedCompared to extra high, the Max level consumes up to twice as many tokens, but delivers only a 4% to 10% performance improvement across various benchmarks.68% similarUnverified在歧义点的四类行为中，'先搜再问'成功率最高，平均通过率达93.4%，远高于'直接猜'的56.5%，'搜多次再猜'为51.5%67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/948925API
curl https://kongchang.com/api/v1/knowledge/claims/948925MCP
get_claim(id=948925)