A calibrated early-exit rule for reasoning LLMs, based on hidden-state probes and Learn-then-Test risk control, reduces thinking tokens by up to 60% in-distribution and 20% out-of-distribution while roughly preserving accuracy.
lmdeploy natively supports the saving of last layer representations, so it was used for almost all experiments
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Thought calibration: Efficient and confident test-time scaling
A calibrated early-exit rule for reasoning LLMs, based on hidden-state probes and Learn-then-Test risk control, reduces thinking tokens by up to 60% in-distribution and 20% out-of-distribution while roughly preserving accuracy.