[1] https://shap.readthedocs.io/en/latest/

Given the example I saw about CRISPR, what does this model give over a different, non explaining model in the output ? Does it really make me more confident in the output if I know the data came from Arxiv or Wikipedia ?

I find the LLM outputs are subtlety wrong not obviously wrong

Steerling appears to be just a discrete diffusion model where the final hidden states are passed through a sparse autoencoder (a common interpretability layer) before the LM head.

They also use a loss that aligns the SAE'S activations with labelled concepts? However, this is an example of "The Most Forbidden Technique" [1], and could make the model appear interpretable without the attributed concepts actually having causal effect on the model's decisions.

1: https://thezvi.substack.com/p/the-most-forbidden-technique

I could find this [0], but not sure if that represents the entire system? (Apologies, I am not that well versed in ML)

[0] - https://www.guidelabs.ai/post/scaling-interpretable-models-8...

How granular can you get the source data attribution? Down to individual let's say Wikipedia topics? Probably not urls?

Would be interested to see this scale to 30/70b

I don't quite grasp how to interpret the training data attribution process. For example, it seems to say that for a given sentence like "They argued that humans tend to weigh losses more heavily than gains, leading to risk aversion", 24% is attributed to Wikipedia and 23% to Arxiv.

Does that mean that the concepts used in this sentence are also found in those datasets, and that's what's getting compared here? Or does it mean that you can track down which parts of the training data were interpolated to create that sentence?

It made me cautiously optimistic that all of Anthropic’s work on alignment, which they did for AI safety, is actually the cause of Claude code’s comparatively superior utility (and their present success). I wonder if future progress (maybe actual AGI?) lies in the direction of better and better alignment, so I think this is super cool and I’m suddenly really interested in experiments like this

But there might be bad, malicious articles on ArXiv, so it doesn't really say anything about veracity.

Perhaps this might help to detect some problems like prompt injection - but then it might be more interesting to see those examples.

Perhaps this could be a step in that direction. If we can associate the attribution with likelihood of being true. E.g., Arxiv would be better than science fiction in that context. But what is the attribution if it hallucinates a citation? Im guessing it would still be attributing it to scientific sources. So it does nothing to fix the most damaging instances of hallucination?

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.

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!

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.

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/

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.