Controlled experiments show language models extract correct answers from contradictory data only when errors are structurally incoherent, supporting the hypothesis that gradient descent selects the most compressible answer cluster.
In the coherent condition, a systematic rule is applied consistently across all problems of the same type (e.g., for distribution: a(b+c) =ab+c instead of ab+ac )
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
-
Truth as a Compression Artifact in Language Model Training
Controlled experiments show language models extract correct answers from contradictory data only when errors are structurally incoherent, supporting the hypothesis that gradient descent selects the most compressible answer cluster.