Ryan Peters

@Sauers_ Manifold steering is actually what gave me the idea to run this test 😁

@ryanpirl also seems like a great test of manifold fitting

PCA fit to the final layer of residual stream in Qwen3-4b across 15 trajectories in a spatial discrimination task, then applied to each layer of the residual stream and plotted.

@Sauers_ The model is tasked with predicting the position of an object in an environment, and so each 'position' is the ground truth position within the environment.

@ryanpirl positions are literally the spatial positions related to the task?

Me: playing peak-a-boo with some random child at the coffee shop 🙈🙉 My brain: "Ah yes, an in-vivo experiment testing object permanence in infants."

Qwen is just a humble bread 🥖🍞

I wonder if qualia steered models would be any better at mechanistic introspection 🤔

@xanadieu True, but I would hypothesize that even 14b models have the latent potential to introspect on such a task.

@ryanpirl Well it is a 14b model

Qwen3-14b doesn't seem to be the best at introspection 😭

Digging deeper into introspection, and there seems to be two types that I don't clearly see differentiated anywhere: Behavioral introspection: Where a model reasons about its own behavior (predicting what it'll do, or knowing what it knows). This can largely be evaled from an API alone. Mechanistic introspection: The model introspects on its actual activations or circuits (e.g. noticing a concept injected into its residual stream, or faithfully explaining the circuit behind an answer). Evaling this likely requires full access to model internals (e.g. for steering or verifying circuitry). Humans are notoriously poor at mechanistic introspection (we invented neuroscience because our brains can't introspect their own circuitry), and are also not the best at behavioral introspection (e.g. intention-behavior gap, affective forecasting). I feel confident that models will surpass us in our own ability to introspect both behaviorally and mechanistically, if they have not already. @TransluceAI's PCDs, @AnthropicAI's NLA's, and activation oracles (Karvonen et al.) hint at mechanistic introspection being possible to a superhuman-level.

The most correlated SAE feature with this 'all-caps' concept vector in Qwen3-4b (not 14b) at layer 18 was a feature with description: "Internet snippets/excerpts" which I thought was weird. So, if you look at the top activating example of this feature, you can clearly see that the description is quite poor and it's actually more of a 'capitalized text' feature.

In retrospect, this is basically what they did in section 6 of the 'Mechanisms of Introspective Awareness' paper where they trained a 'bias vector for introspection' to see if there exists a steering vector that improves model performance on introspection. Turns out, yes, the steering vector improved performance a lot. They even correlate this vector against SAE features in the appendix. Might reproduce on Qwen3.5. Paper: arxiv.org/pdf/2603.21396

@ryanpirl oh wait yeah can you like, do some sort of backwards magic

Qwen is computers and AI 😄

Neuroscientists doing rigid or piecewise-rigid motion correction on calcium imaging data, check out pycorre. It's an independent reimplementation of NoRMCorre in PyTorch: GPU-accelerated, optional numpy-in/numpy-out, and faster than existing tools. The figure attached shows scaling curves on a 360×640 calcium imaging video, benchmarked against jnormcorre (NoRMCorre in JAX) and the CaImAn reference: - On GPU, pycorre runs 3.0× faster than jnormcorre (GPU) and 22.2× faster than CaImAn (CPU-only). - On CPU, pycorre runs 2.9× faster than jnormcorre (CPU) and 4.6× faster than CaImAn. Install: `pip install pycorre` Github: github.com/ryanirl/pycorre

I love the look of TUI's

@thkostolansky Too many strongly activating features 😭

@ryanpirl nice, any other strongly activating features? curious how it views diff models too

One of the features that fires when you ask Qwen who they are.

@thkostolansky In this case any non-zero activation.

@ryanpirl does fire mean activation above some threshold?

@Sauers_ I'm totally going to try and run this study. Maybe there is also a more principled way that you approximate which features or subset of features would upweight the correct logprob across the dataset 🤷‍♂️

Zero-shot introspection via logprobs Steer a model toward a target emotion using SAE feature, then ask it a multiple-choice question of the form "Which of the following SAE features are you currently being steered by?" with five options — the correct target plus four distractors drawn at random from other features. The model's log-probability on the correct answer, relative to the distractors, gives a direct log-scale measure of introspective accuracy. Compare this against an unsteered sham condition to get a Δ log-odds statistic. Δ log-odds margin (steered − sham)

@Sauers_ Would be interesting 🤔. Is the benchmark open sourced? I could run a quick initial test to see if there are any relevant features that fire across all benchmarks.

@ryanpirl this gives me an idea. I have simple multiple choice introspection bench. list of relevant nodes in these sorts of self-referential prompt graphs »» then ablate (or upregulate) nodes in the circuit to if any are responsible for introspection capability

@OvinduA I hope to make it all public sometime in the next month or so and will post an update when I do :)

@ryanpirl Great, thank you! Would be nice to check out your repo when available. 😅

At it's core, it's just Anthropics circuit-tracer applied on open-sourced Qwen3-4b transcoders (feature descriptions by Neuronpedia). But both the circuit-tracer implementation I am using and the visualization I posted is in-house code that's not public (yet). Circuit tracer: transformer-circuits.pub/2025/attributi…

@ryanpirl This is cool, could I know how you generated this? I saw something similar on a Goodfire paper.

@greysonbowser Layers (+ embedding and logit nodes on both ends)

@ryanpirl What’s the y axis?