Unverified50% confidenceFactExact time
The failure-replay mechanism is only effective when measured in "steps" rather than "games"; the first version replayed a quarter of new games, but the replayed data accounted for less than 1% of the total training volume.
1
Sources
50%
Confidence
Long-term
Relevance
10/29/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedWhen sampling eight candidates without constraints from the model, 10.2% of inputs cannot produce any valid output under limited training data, and this proportion drops to 0.4% when using the full training data77% similarUnverifiedIn CLM training, pretraining alone scores 52% on the hard negative sample test; adding the hard negative sample stage jumps it to 69%; training with hard negative samples from day one peaks at only 62% and overfits quickly; omitting replay causes the hard negative sample score to drop from 69% back to 56%.75% similarUnverifiedAfter changing the actuator dynamics, the old 97% success rate data is no longer directly comparable, and the current policy is being retrained and evaluated under a more difficult model.73% similarVerifiedWhen fine-tuning with LoRA, the actual number of parameters that need to be trained is only 0.1%-1% of the original.72% similarVerifiedReAct's errors amplify in a chain: if the accuracy per step is p, the overall success rate after n steps is approximately p to the power of n; with a per-step accuracy of 95%, the overall success rate drops to approximately 59.9% after 10 steps and only 35.8% after 20 steps.71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1112527API
curl https://kongchang.com/api/v1/knowledge/claims/1112527MCP
get_claim(id=1112527)