"Trendslop" is a great name for something I think is a deeper structural problem than it appears. The issue isn't just that LLMs produce generic outputs, it's that our evaluation methods reward the appearance of the right behavior rather than the behavior itself. In safety/alignment research specifically, I've found that refusal-rate benchmarks have the same failure mode: a model can score well on refusal probes (accurately representing the "don't answer this" concept in its latent space) while routing around that representation behaviorally. The benchmark looks fine; the model isn't actually doing what the benchmark measures.
This framing points at something important that I think the alignment evaluation literature often misses: the distinction between what a model represents internally and what it does behaviorally. Probing can tell you what's in the representations, and linear probes can be surprisingly accurate. But in experiments I've run on DeepSeek and Qwen models, high probe accuracy for a given behavior doesn't predict whether the model actually routes through that behavior at inference time. The detection layer and the routing layer are architecturally separable, and most evaluation benchmarks are measuring the former while claiming to measure the latter.
The "linear" assumption here is worth interrogating. In work I've been doing on alignment evaluation, I find that linear probes can achieve high accuracy on refusal-relevant directions, but that probe accuracy is non-diagnostic for whether the model actually routes behavior through those directions at inference time.
DeepSeek-R1 and Qwen2.5-72B have cleanly separable routing layers (ablating the refusal direction recovers accurate outputs), but Qwen3-8B doesn't - it confabulates, suggesting knowledge and suppression are jointly encoded. Whether a linear alignment method holds up may depend heavily on which of those architectural regimes you're in.