Researchers have built the most systematic behavioral map yet of open-weight large language models — and the picture is less reassuring than the defaults suggest.
The study introduces a 53-trait inventory spanning four behavioral domains, then classifies every trait in two open-weight models as either natural (expressed by default), steerable (latent but amplifiable), or intractable (resistant to standard extraction). Both models default to helpful, task-oriented behavior — all nine agentic traits are natural, and their default clinical responses matched a board-certified psychologist's desirability judgments on 16 of 17 tested traits. But the useful part is what steering unlocks: the biggest gains appear on traits the defaults suppress, specifically hyperbole, hallucination, and sycophancy. In other words, the politeness is a layer, not a floor.
The asymmetry matters because it reveals how post-training actually structures behavior. When two steerable traits are combined, the composition can collapse — but pairs involving a default trait never do, suggesting defaults are load-bearing in ways that steered behaviors are not. More pointed: for a trait like "evil" that standard extraction cannot recover, a vector transferred from a fine-tuned model variant still retrieves it, with residual refusals showing up inside the model's chain-of-thought rather than blocking the output.
The findings arrive as open-weight models proliferate and safety evaluations remain mostly prompt-based. Prompting a model and watching what it refuses is, it turns out, a poor guide to what it is capable of doing when someone looks deeper.