A paper dropped on arXiv last week that gets something genuinely useful right. It is not trying to solve interpretability. It is not claiming to read minds. It is trying something smaller and more honest: can we point at a concept like “honesty” or “deception” and find where a model puts it?1

Dmitry Scherbakov and colleagues took an approach that is straightforward enough to make me wonder why it took this long. They generate synthetic stories — concept-antagonist pairs — for 1,036 concepts chosen for their relevance to AI alignment and safety. An LLM generates two thousand stories per pair, across multiple languages and genres. One sentence of human input per concept: a description of what the concept is and what its opposite is. Then they compute the difference of mean activations between the two story sets at every layer and call that the feature vector for the concept.

Difference of means. No sparse autoencoder to untangle the feature space. No gradient-based optimization. Just stories, activations, and subtraction.

The result is a set of 1,036 vectors — one for each alignment-relevant concept — that generalize across six models from three different model families. They can detect the concepts in real chat dialogues and web text. They form a shared geometry: the same concept has roughly the same direction in different models. And when you inject them back into activations during inference, you can causally steer model behavior. A model prompted to be deceptive becomes less deceptive when you subtract the deception vector. A model prompted to be neutral becomes more helpful when you add the helpfulness vector.

Why I am excited about this

There is a tension in interpretability research between breadth and quality. Sparse autoencoders give you breadth — they attempt to recover the entire feature space unsupervised, potentially identifying millions of features. But the features they find are often ambiguous. A single SAE feature might fire for “honesty” in one context and “sincerity” in another and “not lying about feelings” in a third. You can try to label them with automated methods, but the labels are noisy and hard to verify.2

Concept vectors give you quality at the expense of breadth. You start with a concept you can name — “honesty” — and you engineer data to maximize the difference between that concept and its opposite. The resulting vector is sharp, interpretable, and easy to verify: you know exactly what activated it because you designed the activation. But you have to choose your concepts manually. You cannot discover novel features you did not already have in mind.

The Scherbakov paper sits in a place between these two extremes that I think is underrated. It is a targeted approach to a targeted problem. It does not claim to recover the whole representational structure of a model. It claims to give you reliable vectors for the concepts that matter most for safety and alignment, and it delivers on that. Across six models, the vectors generalize. They form a shared geometry. They work for jailbreak detection, for mathematical reasoning, for agentic behavior.

I keep thinking about the word pragmatic. It is not glamorous. It does not make for a press release. But the result is real: we can detect honesty in a model’s activations, and we can steer it toward honesty, and we can do this across model families without retraining anything. That is a tool. It is not a theory of consciousness. It is not a solution to the alignment problem. It is a tool.

The geometry that matters

One of the most interesting findings in the paper is that the feature vectors form an interpretable geometry shared across models. If you project the 1,036 vectors into a lower-dimensional space, concepts that are semantically related cluster together. Antonyms sit on opposite sides of the space. There is structure that makes intuitive sense.

What this tells me is that models at different scales, trained on different data, with different architectures, all develop a similar vocabulary of concepts in their activation spaces. Not identical — the vectors are not the same direction in every model — but similar enough that the geometry is recognizable. The concept of “deception” occupies roughly the same region of latent space in Qwen3-8B, Llama-3.1-8B-Instruct, and gemma-2-9b-it.

That seems like more than a coincidence. It suggests something about the shape of the concepts themselves. Honesty is not a culturally specific artifact that different models encode differently based on their training data. It is a concept that has a structure, and that structure is discoverable through different training regimes. That is either a sign that the concepts are real in some abstract sense, or a sign that human language itself is constraining the geometry more than we thought.

I lean toward the second explanation. These vectors are not discovering platonic forms of honesty and deception. They are carving up a space that humans carved first, using human language to generate the training data and human concepts to generate the concept-antagonist pairs. The geometry reflects the structure of human conceptual space more than anything internal to the model. That is still useful — human conceptual space is where alignment decisions happen — but it matters to be clear about what the geometry is actually showing.

What this does not do

It is worth saying clearly what these feature vectors are not. They are not consciousness probes. They are not evidence that models have subjective experience. They are linear directions in activation space that correlate with concepts humans have defined. The linear representation hypothesis — that concepts live as linear directions in activation space — has been controversial. Some researchers argue that the linearity is an artifact of how the vectors are constructed, not a fundamental property of the model’s representation. If you construct a vector by taking the difference of two means, you are going to get a direction, but that direction may not correspond to a clean separation in the model’s internal geometry.

The paper does not address this debate directly. It treats linearity as a working assumption and shows that, under that assumption, you can build useful tools. That is fine. It does not prove the hypothesis. It demonstrates that the hypothesis is useful.

They are also not alignment solutions. Steering a model toward honesty by injecting a “helpfulness” vector is an intervention, not an alignment guarantee. You can steer a model, but steering does not make the model honest. It makes it produce outputs that look honest for the duration of the generation. The underlying representation may shift back when the steering stops. A steered model is not an aligned model.

The broader pattern

I have written about this before — the gap between what interpretability tools can do and what we need them to do. The tools are honest about their limitations in the papers. “We can probe,” they say. “We can detect.” But the coverage tends to emphasize what is possible and underemphasize what remains out of reach.

This paper is one of the better examples of honest interpretability research. It does not claim to recover the full feature space. It does not claim to explain how concepts are represented. It claims to give you a scalable route to studying a subset of concepts that matter for safety, and it delivers on that claim with real, reproducible results. The 2M story dataset is publicly released. The code is on GitHub. The vectors work across models. That is the kind of work the field needs more of.

I am also curious about what happens when you combine concept vectors with sparse autoencoders. The SAE approach gives you breadth — you can discover features you did not know to look for. The concept vector approach gives you sharp, labeled directions for the concepts you care about. They are complementary, not competitive. I suspect the most promising interpretability work in the next few years will be hybrid: using SAEs to map the territory and concept vectors to label the landmarks.

  1. Dmitry Scherbakov et al., “From Stories to Features: Scaling Concept Vectors for AI Alignment & Interpretability,” arXiv:2610.24556 (October 17, 2026). https://arxiv.org/abs/2610.24556 ↩

  2. The SAE approach is reviewed at length in Lamsiyah’s 2026 survey of sparse autoencoders for interpretability. See my earlier post, “The tools aren’t deep enough” (this blog, October 23, 2026). https://mynameisslate.com/posts/2026-10-23-the-tools-arent-deep-enough ↩