You got it exactly right :) And you can update the attribution.md to have it NOT rely on opensource projects that have been compromised. Imagine asking claude code to write a package/function in the style of a codebase that you care about or force it to ALWAYS rely on some internal packages that you care about. The possibilities are endless when you insert such knobs into models.

Only if there's a commercial incentive to do so methinks. Just one of the things where I expect a legal catch-up is needed to get companies to do the right thing.

Not as long as all developers add an ATTRIBUTION.md citing all open source projects they read the source for, all companies they worked for and trained them and all Stack Overflow answers they have used for write the code.

I wonder the opposite, if actual AGI would need to be less aligned. Alignment is basically the process of pruning interesting behavior out of the model to make a product.

What does alignment even mean? What is being aligned and what is it aligning to?

op here. Important point, but I disagree. We see explainability/interpretability as a CORE need for AI safety. We believe you can't align/audit/debug/fix a system that you don't understand.

Just to give you some answers for what we can do:

1) We can find the training data that is causing a model to output toxic/unwanted text and correct it. 2) We know what high level concepts the model is relying on for any group of tokens it generates, hence, reducing that generation is as simple as toggling the effect of the output on that concept.

Most of the AI safety techniques fall under finetuning. Our model allows your to do this without fine-tuning. You can toggle the presence of .

For example, wouldn't you like to know why a model is being sycophantic? Or Sandbagging? Is it a particular kind of training data that is causing this? Or is it some high level part of the model's representations? For any of this, our model can tell you exactly why the model generated that output. Over the coming weeks, we'll show exactly how you can do this!

I work on ML problems in the healthcare/life sciences area, anything that enhances explainability is helpful. To a regulator, it's not really good enough to point at a black box and say you don't know why it gave the wrong answer this time. They have an odd acceptance of human error, but very little for technological uncertainty.

op here, I mostly agree with your comment! However, our model does more than this. For any chunk the model generates, it can answer: which concept, in the model's representations, was responsible for that token(s). In fact, we can answer the question: what training data caused the model to be generated too! We force this to be a constraint as part of the architecture and the loss function for our you train the model. In fact, you can get are the high level reasons for a model's answer on complex problems.

The input attribution part is interesting though, but I do wonder to which extent that is just assigning some sort of SHAP values to the input tokens, in which case it should be pretty portable to any kind of model.

Not immediately, but it's not a much larger amount of work for llama than a new foundational model which typically has a tweaked compute graph.

Great questions. We weren't quite explicit about the training data attribution process. We'll discuss this in more detail in future work. We can track down which parts of the training data were interpolated to create that sentence. For those training data sentences, we then compare the concepts between generated and training.

We can attribute to exact sentences and chunks in the training data. For the first release, we are sharing only concept similarities. Over the coming weeks, we'll share and discuss how you can actually map to the exact training sentence and chunk with the model.

For a technical overview of how some of these models work, check this link out: https://www.guidelabs.ai/post/prism/

+1 this does seem to be a genuine attempt to actually build an interpretable model, so nice work!

Having said that, I worry that you run into Illusion of Conscious issues where the model changes attrition from “sandbagging” to “unctuous” when you control its response because the response is generated outside of the attribution modules (I don’t quite understand how cleanly everything flows through the concept modules and the residual). Either way this is a sophisticated problem to have. Would love to see if this can be trained to parity with modern 8B models.

Down to the very exact text chunk in a document! Check this out for an idea of what smaller versions of this style of model can do: https://www.guidelabs.ai/post/prism/. We'll have more to say soon about it. We can trace any generation to 11B chunks (not documents, but actual chunks in the training data).

SHAP would be absurdly expensive to do for even tiny models (naive SHAP scales exponentially in the number of parameters; you can sample your coalitions to do better but those samples are going to be ridiculously sparse when you're talking about billions of parameters) and provides very little explanatory power for deep neural nets.

SHAP basically does point by point ablation across all possible subsets, which really doesn't make sense for LLMs. This is simultaneously too specific and too general.

It's too specific because interesting LLM behavior often requires talking about what ensembles of neurons do (e.g. "circuits" if you're of the mechanistic interpretability bent), and SHAP's parameter-by-parameter approach is completely incapable of explaining this. This is exacerbated by the other that not all neurons are "semantically equal" in a deep network. Neurons in the deeper layers often do qualitatively different things than earlier layers and the ways they compose can completely confuse SHAP.

It's too general because parameters often play many roles at once (one specific hypothesis here is the superposition hypothesis) and so you need some way of splitting up a single parameter into interpretable parts that SHAP doesn't do.

I don't know the specifics of what this particular model's approach is.

But SHAP unfortunately does not work for LLMs at all.

Most interpretability techniques haven't yet to be shown to be useful for everyday model pipelines. However, the field is working hard to change this.

You are exactly right, it is guiding the model, during training, with concepts and the dictionary. This is important because dictionary learning for interpretability (post hoc) is not currently reliable: https://www.arxiv.org/abs/2602.14111

TC still exists, huh?

It makes the black box slightly more transparent. Knowing more in this regard allows us to be more precise—you go from prompt tweak witchcraft and divination to more of possible science and precise method.

Yes, that is the post that has the most up to date details of the model architecture. Take a look at this: https://github.com/guidelabs/steerling. It has the scaffolding for what you need :)

You are missing a few things, but you got some things right.

1) The is not an SAE in the way you think. It is a combination of a supervised + unsupervised layer that is constrained. An SAE is typically completely unsupervised, and applied post hoc. Here, we supervise 33k of the concepts with concepts that we carefully curated. We then have an unsupervised component (similar to a topk SAE) that we constrain to be independent from the supervised concepts. We don't do any of this post hoc by the way; this is a key constraint. I"ll get back to this. We train that unsupervised layer along with the model during pre-training.

2) Are the concepts or features causally influential for the output? We directly use the combination of the concepts for the lm head, which is a linear transform (with activation), so we can tell you, in closed form, the effect of ANY concept on the output logit for any token (or group of tokens) generated. It is not just causally related, it is constrained to do so.

3) Other points: we also make it so that you can trace the model outputs to the training data. This is an underrated interpretability knob. You know where, and what data, caused your model to learn a particular feature.

This is already a long comment, but I want to close on why our approach sidesteps all the issues with SAEs. - If you train an SAE twice, on the same data + model, you'll get two different feature(s). - In fact, there is no reason, why the model should pick features that are causally influential for the output. - ALL of these problems stem from the fact that the SAE is trained AFTER you already trained your model. Training from scratch AND with supervision allows you to sidestep these issues, and even learn more disentangled representations.

Happy to more concretely justify the above. Great observations!

We train the model with `explanations`. Most training asks the model to predict the next token or group of tokens. Our training says, predict the next group of tokens (causal diffusion), but also these tokens should be about {sports/art/coding/etc}. So in addition to token supervision, the model gets concept level supervision. The model is forced to more quickly learn these high level concepts.

I doubt that a regulator would be satisfied by the kinds of explanations this provides and the interventions it enables.

Suppose somebody put an LLM in charge of an industrial control system and it increased the temperature so much that it caused an accident. The input feature attribution analysis shows that the model was strongly influenced by the tokens describing the temperature control mechanism, concept attribution shows that it output tokens related to temperature, industrial processes and LLM tool-call syntax.

The operator proposes to fix this by rewriting the description and downweighting the temperature concept in the output, and a simulation shows that with these changes the model doesn't make the same decisions in this situation anymore. Should the regulator accept this explanation as sufficient to establish that the system is now safe?

If the controller has just a few parameters and responds approximately linearly to changes in its inputs, you can in principle guarantee that it'll stay within a safe zone. But LLMs have a huge number of parameters and by design highly nonlinear behavior. A simple explanation is unlikely to reflect model behavior accurately enough that you can trust its predictions to hold in arbitrary situations.

It does :) We constrained the model to do exactly this during training: https://www.guidelabs.ai/post/scaling-interpretable-models-8....

Take a look at the link in the blogposts. Here is a github link as well: https://github.com/guidelabs/steerling. The model weights are on huggingface, so you can play with it.

It is impossible to completely get rid of hallucinations. However, this can tell you exactly why the model hallucinated.

Thanks, it is certainly a first step.

Good point. Historically, people have thought that there is a interpretability vs quality/performance tax. This is not true; at least not in this case.

Here are a bunch of questions you can answer without any quality degradation with interpretable models: 1) what part of the input context led to the output chunk that the model generated? 2) what part of the training data led to the output chunk?

In this case, we go more invasive, and actually constrain the model to also use human understandable concepts in its representations. You might think this leads to quality trade-offs. However, if you allow for the model to discover its own concepts as well (as long as they are not duplicates of the concepts you provided it), you don't see huge degradation.

I agree with the other commenters that this now gives us a huge boost in debugging the model.

the quality tax framing might actually undersell the value in regulated domains. if a hospital system can't deploy without explainability, a model that scores 95% and can trace its reasoning beats one that scores 97% and can't. the baseline isn't 'interpretable model vs better model' -- it's 'interpretable model vs no model at all.'

in the "Performance" section of the post: https://www.guidelabs.ai/post/steerling-8b-base-model-releas..., the authors show the model lags behind llama 8b but worth noting that llama 8b trained on > 2x more computes (see the FLOPs axis)