Assess whether rare-event engagement scales to larger language models
Determine whether the interaction between elicitation structure and extreme failure rarity observed in qwen3:8b, llama3.1:8b, and mistral:7b persists, strengthens, or disappears in larger-scale language models.
References
Whether the same interaction holds, strengthens, or disappears at larger scale is a real and open question, but it is a different one from what this paper asks, and we do not present these three models as a claim about what happens at scale in either direction.
Improbability did not correlate with judged novelty ($|r|<0.2$, Figure~\ref{fig:prompt}), and a pilot effect ($n=8$) had not replicated. We found no benefit from prompt improbability under the tested operationalizations; whether other operationalizations, developers or judges would find one is open.
We further conjecture that the two ends fail for different reasons. Small backbones may be bottlenecked by under-perception of on-stage stimuli, while large backbones may be bottlenecked (also trapped in environment perception but less than small ones) by a built-in resistance to risk-relevant actions such as follow, dm, and join-group, especially under high-risk suspicious cues. In this view, simulating an easily-deceived victim could favor a small backbone shaped by RBHS, where the scaffold plausibly fills the perception gap without inheriting the safety prior that suppresses the very behavior we need to study.
Our experiments are restricted to smaller LLMs, both as judges (8--12B) and as attack targets, to keep gradient-based attacks and abliteration computationally feasible; whether our findings generalize to larger models remains untested.