ZAZayd AS@NatureAI_20238:08 AM · Jul 21, 2026

@m_wulfmeier Anyone else surprised frontier models still collapse on basic priors? I assumed we were past this. Red-teaming them is like chess with a patient opponent. Far from mature.

ZAZayd AS@NatureAI_20238:03 AM · Jul 21, 2026

@m_wulfmeier The chess player in me calls this a classic blunder: playing the expected move instead of calculating the actual position. This is exactly why these systems need to be red-teamed more.

MWMarkus Wulfmeier@m_wulfmeier7:52 AM · Jul 21, 2026

693TECH

Some interesting results recently: conditioning VLMs on order metadata collapsed error detection. Seems like a clear prior but the models anchored on expected contents and stopped looking at the images. I had assumed current frontier models were past this issue. But for now this kind of prior dominates the likelihood -- similar to Deng et al. (CVPR 2025). Some notes on the underlying biases below.

MWMarkus Wulfmeier@m_wulfmeier7:52 AM · Jul 21, 2026

693TECH

Specific decision biases like these are still present in frontier models: greediness -- collapsing onto the highest-prior answer -- and frequency bias, following what's salient in context over what the evidence supports. Some training recipes can reduce them long term, like RL fine-tuning on self-generated rationales, which we showed in our greedy agents paper (Schmied et al.).