Understanding and evaluating deceptive AI. apolloresearch.ai
Introduces Attribution-based Parameter Decomposition, which splits network parameters into minimal, simple components, and shows it recovers ground-truth mechanisms in toy models; not safety-motivated.
Tests whether linear probes on Llama-3.3-70B activations detect strategically deceptive responses, comparing probe training methods and finding they partly track deception-related text rather than actual deception.
Proposes the Interaction Basis, a loss-landscape-based method that removes parameter degeneracies to find computationally relevant, sparsely interacting features, tested on toy models and GPT-2.
Proposes end-to-end training of sparse autoencoders so that learned features are functionally important to network performance, not just reconstruction-optimal; the motivation is interpretability rather than explicit safety.
Apollo Research argues that rigorous AI safety audits need white-box, interpretability-based methods rather than black-box input-output testing alone, citing a FAccT paper on the limits of black-box evaluation.