Unverified50% confidenceBenchmarkExact time
Among PBT runs started from the dead zone, slow PBT (updating once every several steps) succeeded 0/3 times, but fast PBT (updating every 100 steps, multiplying lr by 0.5 or 2) achieved 3/3 success in about 6850 steps on average.
1
Sources
50%
Confidence
Long-term
Relevance
10/26/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn the scenario where the corner of the grid search already contains the answer, the best fixed configuration reaches 95% validation accuracy in only 317 ± 29 steps, whereas PBT needs an average of 2667 ± 2244 steps; on every seed, PBT is slower.71% similarUnverifiedIn testing, the 3-bit quantized version outputs at 53 tokens per second.68% similarUnverifiedPyTorch's typical performance improvement through JIT compilation is between 1.5× and 3×.65% similarUnverifiedAt the highest noise level in 4D, compared with other methods, PUN achieved a relative improvement in success rate of 10.7-60.6% while running 3.29-32.15 times faster.62% similarUnverifiedOn Qwen 1.7B, the compiled regex Trie reduced median latency from 905.2 milliseconds to 567.5 milliseconds (a 37.3% speedup), but recall dropped sharply from 83.1% to 60.7%, a loss of 22.4 percentage points.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1099234API
curl https://kongchang.com/api/v1/knowledge/claims/1099234MCP
get_claim(id=1099234)