LLMs exhibit pseudo-deliberation, with consistent value-action misalignment in generated dialogues despite reasoning, as measured by the new VALDI framework across 4941 scenarios.
Syn- thetic dialogue dataset generation using LLM agents
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
-
Pseudo-Deliberation in Language Models: When Reasoning Fails to Align Values and Actions
LLMs exhibit pseudo-deliberation, with consistent value-action misalignment in generated dialogues despite reasoning, as measured by the new VALDI framework across 4941 scenarios.