Coauthored with Mark Kurzeja

A while back, we explored the “BiDPO” method for training steering vectors. In Gemini 1.5v1 Flash and Pro, BiDPO steering vectors boosted TruthfulQA scores by >10% while mostly retaining capabilities. When we updated to Gemini 1.5v2, prompt-based steering baselines became significantly stronger. BiDPO did not beat the stronger baselines, ending the project.

...

BiDPO seems effective and sample-efficient but does not currently exceed more standard baselines. It’s hard to draw firm conclusions about BiDPO because TruthfulQA might not be measuring truthfulness ／factuality. However, we remain excited about DPO-driven Conditional Activation Steering, which has additional advantages—particularly for targeted loss mitigation.

This result is largely negative. I wanted to share it to increase scientific understanding around steering! We also conducted a postmortem on why the method stopped outperforming baselines.

I'd also like to note that @ryan_greenblatt's skepticism predicted this outcome more strongly than my worldview did. I want him to get points for that. :) While I think steering has targeted applications and provides clues about how LLMs function, it's not a slam-dunk Pareto improvement on benchmarks we care about.

Read at https://turntrout.com/gemini-steering!^[1] 

1. ^{^}

   Also mirrored on the GDM safety research Medium.