Unverified50% confidenceBenchmarkExact time
By selectively editing the expert routing scores to suppress overthinking, the median number of generated tokens was reduced by 30.4% and by up to 63.3% across four models and four benchmarks.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
11/5/2026
First Seen
Valid until: 2/3/2027
Sources
Related Entities
Related Claims
UnverifiedThe 44% reduction in inference tokens is a self-reported figure from the author and has not been verified by third-party reproduction; the original materials do not disclose the data construction method, training algorithm, sampling scale, or benchmark test details.75% similarUnverifiedWhen using the multi-agent verification process, MARCH's accuracy was suppressed to 4.4%, whereas when the same model was directly asked which solution was better, accuracy reached 43.7%.73% similarUnverifiedAn insufficiently small reference corpus systematically underestimates true divergence, and the researchers introduce a source-split jackknife method that uses only reference data, which can reduce calibration error by 57% to 59%.73% similarUnverifiedIn 10%–25% of cases, reasoning about the hypothetical grader pulled the agent's work away from the user's original specification.72% similarUnverifiedCompared with the most efficient baseline, the metacognitive controller reduces inference FLOPs by an average of 20.6% while maintaining competitive task accuracy.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1138003API
curl https://kongchang.com/api/v1/knowledge/claims/1138003MCP
get_claim(id=1138003)