[KongchangAI]
Concept对数概率

logprob

The natural logarithm of the raw probability that a language model assigns to each candidate token. It reflects the model's underlying probability distribution during generation and can be used for statistical confidence estimation.

Timeline (last 90 days)

Nov 4

For practitioners, reporting only a logprob decrease is insufficient to support the conclusion that SAE features effectively suppress PII; at minimum, direction-matched random controls and rank-aware metrics should be added.

Unverified50%
Nov 4

A new study directly challenges the practice of evaluating SAE intervention effectiveness using only logprob decrease, pointing out that a logprob decrease may merely be an artifact of perturbation magnitude rather than a genuine severing of the causal path of PII leakage.

Unverified50%
Nov 4

When the intervention effect of the random-direction control is better than that of carefully selected SAE features, it indicates that the logprob decrease likely comes from perturbation magnitude rather than the semantic role of the feature itself.

Unverified50%
Nov 4

A logprob decrease by itself cannot distinguish feature-specific suppression from magnitude artifacts—as long as the activation perturbation is large enough, it does not matter whether the direction is aligned with the target feature.

Unverified50%
Nov 4

When looking only at logprob, 17 cases (47%) appeared to successfully suppress PII; but with a multi-metric framework, only 1 case (2.8%) passed.

Unverified50%
Oct 22

In practice, the logprobs of the tokens in an answer can be averaged or geometrically averaged to obtain sequence-level confidence, or the perplexity of the answer can be computed as an overall measure of certainty.

Unverified50%
Oct 22

logprob is log probability: when the token probability is 0.9, logprob is about -0.105, and when the probability is 0.01, it is about -4.605; the closer the value is to 0, the more confident the model is.

Unverified50%
Oct 22

logprob 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.

Unverified50%
Oct 19

logprob thresholds need to be calibrated for the specific model and task distribution; directly applying a fixed threshold often works poorly.

Unverified50%
Oct 19

The average logprob over an answer sequence can serve as a proxy for overall confidence, and self-consistency sampling assesses confidence by observing answer agreement across multiple samples.

Unverified50%

3 more timeline events

All Facts (13)

Unverified

For practitioners, reporting only a logprob decrease is insufficient to support the conclusion that SAE features effectively suppress PII; at minimum, direction-matched random controls and rank-aware metrics should be added.

50%
Unverified

A new study directly challenges the practice of evaluating SAE intervention effectiveness using only logprob decrease, pointing out that a logprob decrease may merely be an artifact of perturbation magnitude rather than a genuine severing of the causal path of PII leakage.

50%
Unverified

When the intervention effect of the random-direction control is better than that of carefully selected SAE features, it indicates that the logprob decrease likely comes from perturbation magnitude rather than the semantic role of the feature itself.

50%
Unverified

A logprob decrease by itself cannot distinguish feature-specific suppression from magnitude artifacts—as long as the activation perturbation is large enough, it does not matter whether the direction is aligned with the target feature.

50%
Unverified

When looking only at logprob, 17 cases (47%) appeared to successfully suppress PII; but with a multi-metric framework, only 1 case (2.8%) passed.

50%
Unverified

In practice, the logprobs of the tokens in an answer can be averaged or geometrically averaged to obtain sequence-level confidence, or the perplexity of the answer can be computed as an overall measure of certainty.

50%
Unverified

logprob is log probability: when the token probability is 0.9, logprob is about -0.105, and when the probability is 0.01, it is about -4.605; the closer the value is to 0, the more confident the model is.

50%
Unverified

logprob 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.

50%
Unverified

logprob thresholds need to be calibrated for the specific model and task distribution; directly applying a fixed threshold often works poorly.

50%
Unverified

The average logprob over an answer sequence can serve as a proxy for overall confidence, and self-consistency sampling assesses confidence by observing answer agreement across multiple samples.

50%
Unverified

A large language model's confidence signal can be extracted from the output probability distribution, for example by averaging the logprobs of candidate tokens or computing entropy.

50%
Unverified

In large model RL training, the logprobs computed by the training framework and those computed by the inference framework during the rollout phase often do not match, which gets amplified in the importance sampling weights of algorithms such as GRPO and PPO, leading to distorted training signals or even training collapse.

50%
Unverified

Logprob mismatch is one of the fundamental causes of instability and poor reproducibility in many RL training runs; after achieving bit-for-bit alignment, the determinism and reproducibility of training improve significantly.

50%

Source Articles