Unverified50% confidenceFactExact time
A new study proposes deriving the RL objective as maximum likelihood fitting to a model-generated target distribution, thereby unifying pretraining, SFT and RL into the same family of objective functions.
1
Sources
50%
Confidence
Long-term
Relevance
10/30/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe target score identity expresses the score function as a self-consistent relation between the current model prediction and the gradient of the unnormalized density, making it possible to construct a training objective without external training data.75% similarUnverifiedIn RLHF and generative model fine-tuning, algorithm design should target the distribution-matching objective directly, exploiting the logarithmic convergence property of sampling regret to build a more efficient training pipeline.74% similarUnverifiedApproval results from Human-in-the-Loop (HITL) can simultaneously serve as RLHF training data, forming a flywheel of continuously improving model capability.73% similarUnverified全参数RL训练的大范围参数更新提供了隐式正则化效果，可能是鲁棒性的来源71% similarVerifiedEach RL training step requires running inference (generating rollouts) and backpropagation (updating weights) simultaneously, making the compute requirements far higher than ordinary fine-tuning.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1115578API
curl https://kongchang.com/api/v1/knowledge/claims/1115578MCP
get_claim(id=1115578)