Derives new loss functions for SFT and RL that optimize directly for test-time inference operators like aggregation or filtering, with empirical gains in scaling.
performance. If a sample is not in the top quantile of the model’s potential outputs, it contributes zero signal. The model is only updated based on its best attempts, creating a
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Compute Aligned Training: Optimizing for Test Time Inference
Derives new loss functions for SFT and RL that optimize directly for test-time inference operators like aggregation or filtering, with empirical gains in scaling.