Unverified50% confidenceFactExact time
Random horizons introduce a reweighting bias: tokens have different probabilities of being included in supervision under different horizons, and simple averaging causes the gradient estimate to deviate from the goal of full-trajectory supervision; the authors characterize this bias and derive the trade-off between compute and reweighting.
1
Sources
50%
Confidence
Long-term
Relevance
11/17/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAnchor sampling causes each sample to actually involve a different context range, so the attention computation cost exhibits nonlinear, sample-dependent fluctuations.75% similarUnverifiedMerging repetitiveness and limited autonomy into a single Dull score changes the meaning conveyed by that metric, leading to biased interpretations of exposure estimates.75% similarUnverifiedProbe scores perform close to random guessing at predicting the direction of intervention.74% similarUnverifiedWhen the guidance signal deviates from the model's own sampling distribution, the policy gradient estimate incurs bias.74% similarUnverifiedBatch effect refers to the way other cases the judge sees in the same batch indirectly affect the log-probability distribution of the current case through the attention mechanism.73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1179763API
curl https://kongchang.com/api/v1/knowledge/claims/1179763MCP
get_claim(id=1179763)