Semantic loss with graph-based constraints prevents model collapse during fine-tuning of transformers on causal reasoning tasks, yielding stable 68-70% accuracy instead of trivial always-yes/no outputs.
No” since no path exists between disconnected components Example structure: Premise
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
On Semantic Loss Fine-Tuning Approach for Preventing Model Collapse in Causal Reasoning
Semantic loss with graph-based constraints prevents model collapse during fine-tuning of transformers on causal reasoning tasks, yielding stable 68-70% accuracy instead of trivial always-yes/no outputs.