Efforts to monitor advanced AI for rare misalignments face a data challenge: abundant aligned examples but only a handful of misaligned ones. We test activation probes in this "few vs. thousands" regime on spam and honesty detection tasks. For our tasks, training with many negative examples is on average more positive-sample-efficient than balanced training for small numbers (1-10) of positive samples. We also find that LLM upsampling can provide a performance boost equivalent to roughly doubling the number of real positive samples, though excessive upsampling hurts performance. Finally, we show a positive scaling trend, where larger models are more positive-sample-efficient to probe. Our findings suggest we should leverage the large number of negative samples available to amplify the signal from rare but critical misalignment examples.
Andrzej Szablewski, Gabriel Konar-Steenberg, Raffaello Fornasiere, Nikita Menon, Stefan Heimersheim
原文摘要
Model organisms (MOs) - language models trained to exhibit undesired or unnatural behaviours - are frequently used as testbeds for evaluating white-box interpretability techniques. Current MOs are typically constructed via post-hoc supervised fine-tuning (SFT). Prior research has shown that interpretability methods can easily identify hidden behaviours in these MOs. However, recent work suggests that such post-hoc training methods may make interpretability unrealistically easy. We investigate this claim by constructing a suite of 61 $\verb|OLMo2-1B|$- and $\verb|gemma-3-1b-it|$-based MOs trained with seven different techniques, including standard post-hoc SFT, post-hoc DPO, and more realistic integration of MO data into the OLMo post-training DPO phase. We also use these to train 96 further variants via knowledge distillation. We use these MO variants to benchmark activation oracles, activation steering, the logit lens, and the Jacobian lens. Our findings show that (i) MO interpretability depends strongly on training objective, target behaviour, model architecture, and training data generation pipeline; (ii) substantial variance remains even after controlling for differences in the strength of target behaviour expression; and (iii) both our more realistic $\textit{integrated training}$ and knowledge distillation often yield less interpretable MOs than standard post-hoc methods. Our results cast substantial doubt on the validity of current MOs as interpretability proxies and suggest actionable steps towards more robust benchmarks.
Recent work has hypothesised that large language models may represent some concepts as multi-dimensional manifolds instead of single directions. Using these manifolds to interpret or steer representations often assumes that moving along them corresponds to a continuous change in the represented concept. In this work, we test this continuous manifold assumption on the prototypical weekday manifold. Despite its striking circular geometry, we find that the weekday representations behave more like seven discrete states arranged in a circle than a continuous ring. First, we show that local sensitivity has seven minima aligned with the weekdays, producing a step-like mapping between representation and behaviour under on-manifold steering. Second, we test whether any concepts are represented along the manifold outside the seven discrete states, and find neither time of day, uncertainty, nor natural text populate the intermediate space. Together, these results challenge the continuous interpretation of the weekday manifold and motivate testing whether movement along other representation manifolds correspond to meaningful changes in model representations and behaviour.