Replicating Anthropic's emotion concepts paper on a 7B open model
On Qwen2.5-7B-Instruct, the paper's findings about how emotions are represented mostly replicate. Its findings about emotions causing misaligned behaviour are mixed: steering changes flattery, barely changes reward hacking, and never produces blackmail. Of 23 testable sections, 10 replicate, 12 replicate in part and 1 does not replicate.