Unverified50% confidenceFactExact time
Modifying logits at sampling time means adding or subtracting bias values dimension by dimension on the vector before softmax—adding a positive value raises the probability of a certain type of token being selected, while subtracting a negative value suppresses it—without changing the model weights and intervening only at the inference stage.
1
Sources
50%
Confidence
Long-term
Relevance
10/24/2026
First Seen
Sources
No Fine-Tuning, Real-Time Jam: Sampling Control for Pretrained Music Transformers
redditr/MachineLearning10/24/2026
Related Entities
Related Claims
UnverifiedThere is no universal formula for choosing logit bias values; it is recommended to start with small adjustments, test the effect with representative prompts while watching for side effects, and then adjust gradually.79% similarUnverifiedTemperature divides the logits by a coefficient before applying softmax, changing the steepness of the entire distribution; top-p first accumulates probabilities from high to low, truncating once the cumulative probability reaches p, removing low-probability tokens outside the long tail from the candidate set, then renormalizing and sampling within the remaining set.72% similarUnverifiedOutlines 推理时将非法 token 对应的 logits 设为负无穷，使其经 softmax 后概率趋近于零，实现硬约束71% similarUnverifiedOutlines' constraint logic operates on the logits mask at the sampling stage, does not modify model weights, does not add extra forward passes, and the runtime overhead is close to zero.69% similarUnverifiedlogprob measures how fluent it is to generate that token, which is not equivalent to how correct the fact is; a model can perfectly well generate a false statement fluently and with high probability, so it remains a relative signal rather than an absolute guarantee.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1093170API
curl https://kongchang.com/api/v1/knowledge/claims/1093170MCP
get_claim(id=1093170)