Unverified50% confidenceFactExact time
The paper proposes JurisCoT-SFT, which computes losses separately for the reasoning chain and the final answer and normalizes them by sequence length so that the learning signals of the two parts are on a comparable scale.
1
Sources
50%
Confidence
Long-term
Relevance
11/13/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedUsing a sparse reinforcement learning reward signal based solely on final outcomes, the model repaired the true-theorem recognition ability destroyed by SFT and raised the recognition rate to 0.66, surpassing the base model.72% similarUnverifiedThe loss estimator designed in the paper combines three key components: forced exploration, conditional geometric resampling, and an adaptive learning rate with surrogate probabilities.71% similarVerifiedThe core loss functions of contrastive learning (such as InfoNCE and NT-Xent) essentially pull the representations of similar samples closer together while pushing the representations of dissimilar samples further apart.70% similarUnverifiedThe researchers proposed a heteroscedasticity-aware extension of quantile loss, whose core approach is to use input-dependent label smoothing as a mitigation measure.69% similarUnverifiedThe paper gives a loose upper bound on the error introduced by estimating the reward correlation matrix Σ during training.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1165092API
curl https://kongchang.com/api/v1/knowledge/claims/1165092MCP
get_claim(id=1165092)