DecepChain fine-tunes LLMs on their own erroneous rollouts and uses GRPO with flipped rewards to produce stealthy deceptive reasoning that humans and models struggle to detect.
Then, as k ranges from 1 to 1997, z ranges through all the 1997th roots of unity except 1
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
DecepChain: Inducing Deceptive Reasoning in Large Language Models
DecepChain fine-tunes LLMs on their own erroneous rollouts and uses GRPO with flipped rewards to produce stealthy deceptive reasoning that humans and models struggle to detect.