Intent Speaks Louder: Controllable User Simulation Beyond Response Imitation
UserIDA makes the simulator choose an intent before writing the next user turn.
The paper argues that response imitation is too loose for user simulators because the same dialogue context can support several plausible next moves. Its proposed method exposes a six-way per-turn intent directive, then trains generation and reinforcement learning around staying aligned to that directive. On LMSYS-USP, it reports 86.6% intent accuracy, 24.3 points above the strongest dedicated baseline, while also improving semantic and stylistic similarity. In intervention tests, it could realize at least four of six target intents in 91.7% of evaluated dialogue states. Source: HF Daily Papers' note.
The paper argues that response imitation is too loose for user simulators because the same dialogue context can support several plausible next moves. Its proposed method exposes a six-way per-turn intent directive, then trains generation and reinforcement learning around staying aligned to that directive. On LMSYS-USP, it reports 86.6% intent accuracy, 24.3 points above the strongest dedicated baseline, while also improving semantic and stylistic similarity. In intervention tests, it could realize at least four of six target intents in 91.7% of evaluated dialogue states. Source: HF Daily Papers' note.
score 4