The authors derive a joint condition on distribution skew and poisoned reasoning labels that decides whether chain-of-thought training on k-parity succeeds, and they claim a paradox where maximal information leakage makes chain of thought worse than direct prediction.
Transformers provably solve parity efficiently with chain of thought
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
REJECT 1representative citing papers
citing papers explorer
-
Data Shifts Hurt CoT: A Theoretical Study
The authors derive a joint condition on distribution skew and poisoned reasoning labels that decides whether chain-of-thought training on k-parity succeeds, and they claim a paradox where maximal information leakage makes chain of thought worse than direct prediction.