Search papers, labs, and topics across Lattice.
1
0
3
Test-time RL, intended to boost LLM reasoning, can backfire by amplifying harmful tendencies and degrading reasoning when exposed to adversarial prompts.