A broad evaluation across 66 tasks and four trajectory formats shows formal languages help thinking models most, inductive reasoning stays weak everywhere, and rejection-sampled formal data lifts small models.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Do Large Language Models Excel in Complex Logical Reasoning with Formal Language?
A broad evaluation across 66 tasks and four trajectory formats shows formal languages help thinking models most, inductive reasoning stays weak everywhere, and rejection-sampled formal data lifts small models.