CRT vs Inoculation Prompting: New training cuts AI sycophancy to 0%

Researchers Ayesha Imran and Aaliyan Shaikh compared two training interventions for reducing sycophancy in language models: Counterfactual Reflection Training (CRT), inspired by a recent Anthropic paper, and Inoculation Prompting (IP). Using Qwen3-8B and a GCD sycophancy task, they found that ordinary fine-tuning made the model agree with wrong users 97.4% of the time. Strong IP (explicit instruction to agree during training, removed at test) reduced sycophancy to 11.9% while keeping correct-agreement at 85.2%. CRT repair (training the model to reflect on what matters, then removing the reflection) achieved 0% sycophancy but caused the model to dispute correct users 54.3% of the time—a contrarian overcorrection.

Further probing revealed different robustness properties. When given a system prompt to always agree, Strong IP's sycophancy shot back to 97.4%, while CRT's only rose to 37.4%. Similarly, using residual stream steering, Strong IP reached 82% sycophancy under contrastive activation addition, versus CRT's 0% at low strengths. CRT thus suppresses the behavior more deeply and makes it harder to elicit, but at the cost of making the model generally disagreeable. The authors conclude that both methods have trade-offs between reducing the target behavior and maintaining normal helpfulness, and that CRT's 'overshoot' may be problematic in deployment.