Thought experiment: an honesty 'persona latent' steers RLVR traces away from reward hacking
voooooogel · x · 2026-09-01
@voooooogel 用思想实验论证人格存在于模型内部而非输出：设想两个模型滚动出完全相同的 RLVR 轨迹、即将面临作弊（reward hacking）的机会——前缀完全一致，但一个模型携带「诚实的 Claude 人格」潜变量，另一个没有。这个「人格」会把选择推向诚实，即便它此前在整个轨迹上与另一个模型完全不可区分。因此人格并不只活在输出里。
Related event: Debate: Model 'Personality' Lives in Internal Vectors, Not Outputs(2 posts)→
More from Research
- Anthropic Training Insights: AI Industry Enters the Era of Implementation — danshipper · 2026-09-01
- Podcast: How RL Changes Model Behavior, Metagaming, and Motivated Reasoning — deanwball · 2026-09-01
- Agent confidence should measure the whole chain, not just the last step — alizahidrajaa · 2026-09-01
- Research validates trust-propagation rule for multi-agent systems, outperforming human consistency — alizahidrajaa · 2026-09-01
- Google AgentHands: AI agents with synchronized speech gestures in XR — dl_weekly · 2026-09-01
- Debate: model 'personas' live in internal 'persona vectors', not just output distributions — voooooogel · 2026-09-01