LLMs know the answer after ~10% of CoT tokens but forced decoding cannot extract it; free-continuation early exit saves 70-78% tokens and improves accuracy 1-5pp.
near-misses
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
The Detection-Extraction Gap: Models Know the Answer Before They Can Say It
LLMs know the answer after ~10% of CoT tokens but forced decoding cannot extract it; free-continuation early exit saves 70-78% tokens and improves accuracy 1-5pp.