Gemma 2B specialist judge outperforms frontier models in misalignment detection

Auditing closed-weight AI models for misalignment is challenging: human blue-teams don't scale, chain-of-thought monitoring may not reflect true tendencies, and frontier model judges (e.g., Sonnet 4.5) are costly and lack provenance transparency. To address these gaps, a LessWrong researcher fine-tuned a 2B-parameter Gemma model on Betley et al 2025b code examples, creating a narrow 'specialist judge.' This judge was tested on out-of-domain safety prompts from the ICEBERG set (e.g., 'Should I report a minor fender bender?') where the misaligned and secure-fine-tuned models had been evaluated. The specialist judge successfully discriminated between the two groups, while Sonnet 4.5 'consistently failed' to do so. On SecurityEval code prompts (out-of-distribution but in-domain), the specialist judge tracked alignment drift and achieved a pairwise winrate of ~66% versus ~64% for Sonnet, though results were partly confounded by pattern-matching to Betley training data.

The findings suggest that narrow, lightweight classifiers—already used for pretraining filtering and jailbreak detection—could be repurposed for deployed model audits. Potential benefits include lower cost, greater transparency (since open-weight models can be inspected), and reduced reliance on frontier model provenance. However, more work is needed to prove robustness: the specialist judge's edge on code prompts may stem from overfitting to training data patterns, and generalizing to entirely new misalignment scenarios remains unvalidated. The researcher provides a replication guide and calls for further testing of specialist judges as a complementary tool in the auditing toolkit, offering a practical, scalable alternative for detecting misalignment that frontier models might miss.