RECAP is an RL post-training method that uses counter-aligned CoT prefills to make large reasoning models reroute from flawed reasoning to safe responses while preserving reasoning ability.
- Conduct regular drills to prepare staff for sudden disruptions
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
-
Large Reasoning Models Learn Better Alignment from Flawed Thinking
RECAP is an RL post-training method that uses counter-aligned CoT prefills to make large reasoning models reroute from flawed reasoning to safe responses while preserving reasoning ability.