Unverified50% confidenceFactExact time
Backtrack's failure prediction head adds only 0.48%–0.88% of the policy's parameter count.
1
Sources
50%
Confidence
Long-term
Relevance
10/26/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedCompared to the average of the two graphs, the conditioned response graph reduces mean squared prediction error by 53.6% on Gemma-2-9B.67% similarUnverifiedThe same scaffold raised the weak model's execution quality from 0.88 to 2.0 (across four scoring dimensions), yet pushed the strong model toward overcautious shutdown (all four dimensions dropped).66% similarUnverifiedWhen the success rate is extremely low, almost all sample rewards are identical, resulting in extremely high policy gradient variance and very low signal-to-noise ratio, preventing the model from obtaining effective update directions.64% similarUnverifiedThe oracle router knows in advance which cases the small model will get wrong, sends only the wrong requests to the large model, and achieves accuracy comparable to always using the large model at only 29% of the cost.64% similarUnverifiedCausal probes achieved a 17%–118% improvement in steering capability across different models, with only a 3% decrease in concept detection capability.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1098087API
curl https://kongchang.com/api/v1/knowledge/claims/1098087MCP
get_claim(id=1098087)