Unverified50% confidenceOpinionExact time
RL training redistributes probability mass from ordinary outputs to high-quality outputs through methods such as policy gradients, and does not create entirely new capabilities.
1
Sources
50%
Confidence
Long-term
Relevance
10/7/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedEach RL training step requires running inference (generating rollouts) and backpropagation (updating weights) simultaneously, making the compute requirements far higher than ordinary fine-tuning.79% similarVerifiedDuring training, LLMs optimize parameters by maximizing the prediction probability of the next token, meaning the model learns the statistical distribution of the training corpus rather than human aesthetic judgment.76% similarUnverifiedRLCD constructs positive and negative sample pairs for contrastive learning, combines reinforcement learning signals to distill model output distributions, and has advantages in scenarios that do not require manually annotated rewards.74% similarUnverifiedRLCD is a policy-gradient-based reinforcement learning method (an unofficial name coined by the author) that optimizes decision models for strictly proper scoring rules, granting maximum reward only when the model outputs mathematically calibrated probabilities.74% similarUnverifiedBoth the PPO and RLCD approaches attempt to move away from the autoregressive paradigm and use reinforcement learning to produce structured probabilistic or confidence outputs.73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/984829API
curl https://kongchang.com/api/v1/knowledge/claims/984829MCP
get_claim(id=984829)