> "post-trained to know when it's wrong"

Is it also post-trained to know when it's wrong about when it's wrong?

> "Every response comes with a confidence score between 0 and 1"

How confident is it in its confidence?

Please, I'm sure that what you're doing is very neat and useful, but use other language to describe it. I beg you. You can't know when you're wrong. You can only know when you're unsure or inconsistent. You can be absolutely certain and still wrong and uncertain and still correct.

A few years ago, before RLVR became ubiquitous, people were using generate -> classify loops to get models to improve math reasoning. Turns out classification is sometimes easier to do and more accurate than generating coherent traces. So you'd get a model to "think step by step", then run each step (.split('/n/n')) through a classifier, and tell the model "this step might be wrong, re-check". If you're interested I could probably dig up some papers on this.

Now there might be even better ways of getting feedback, I saw some interesting trials using "J-spaces" (from Anthropic's latest research in this area). Turns out you can kinda see when the model uses deep seeded concepts in early layers, or it's "riffing on an idea" with shallow signals in those same layers. And youc an get some signal out of it, and have it re-do a step.

It's also worth mentioning that LMs are wrong in different ways from humans. It's not impossible that we'll eventually find ways of better understanding this (through mechanistic interpretation or something else) and have easy "classifiers" for "this comes from the training data" vs. "this comes from early context" (i.e. supplied materials from prompts/discovery) vs. "this comes from later tokens, likely the model doom-looped itself".

Did something similar tuning a fraud model years back — cheaper to have model flag "recheck this" than trust one shot output. Held up in prod, saved us reruns. Curious if you got real numbers, not just anecdote.

This is an old problem in philosophy, one that the pre-LLM systems were interested in, which got completely bypassed by the present approach. What is knowledge? It's not a simple question even for humans, we don't have a magic truth detector in the brain.

https://plato.stanford.edu/entries/knowledge-analysis/

This is why LLMs are more successful in programming than other disciplines: it's easier to assess correctness of results, through proofs, typechecking, and testing. Similarly for maths and applied maths. Much harder in the humanities.

Programming's easier cause typecheck/proof/test give ground truth to train confidence against. But most human questions don't have that oracle. So how's "know when it's wrong" supposed to generalize past code and math?

IMHO this also depends on the type of assertion.. For example, one system defines three broad classes of assertion : pure opinion (I like green); factual references (the current temperature here is 20c) ; and the rest in a catch-all about "reasoned judgements"

Complex assertions and their compounded forms can be filtered by these classes fairly easily. I rarely see this referenced in general comments.

Interesting question and I don’t share your pessimism. Feels like it’s a matter of some plain old Bayesian stats, almost a homework assignment given all experimental data is available and known.

Cool idea, some feedback:

1. Consider using conformal prediction to calibrate the cutoff. Conformal prediction provides a distribution-free guarantee under exchangeability. This would let you turn your raw probe score into a threshold with a guaranteed bound on the rate of wrongly-kept on-device answers. Source: https://en.wikipedia.org/wiki/Conformal_prediction

2. The best indicators of confidence in ML come from multiple independent methods. What was the result if you combine the token entropy and verbal confidence reporting methods? Does this improve the result?

3. I noticed you didn't mention the assessment method of rerunning the model and judging whether outputs are consistent. How does that method compare in terms of AUROC?

1. Thanks for this!

2. We combine both on Cactus; rolling entropy method helps with per-token signal for early exit, while the probe runs on complete output.

3. Users hate this in production, drains battery on edge devices, but ultimately primarily works ok-ish on 7B+ models, we will share the results too.

> the hidden state for different layers carry meaningful self-awareness signal for various situations.

Is it plausible to wonder if some developer judgement feels, like maybe "the code I just wrote is clean/crufty", or "things came together smoothly/janky", might have extractable signals in some models?

If so, might one create a shopping list of desired signals to check for in a model, as with activation steering concepts, where one checks whether and how hard each concept can usefully be nudged?

YOu are thinking along the right direction, we are going deeper into the signals.

Poked at hidden states on a small model last year for a similar "confidence" probe. Found layer choice matters a ton, mid-layers (roughly 40-60% depth) gave cleaner separation than final layers, which were already too compressed toward the output token.

Is this in any way similar to Goodfire's work? https://www.goodfire.ai/research/rlfr#

Thats an interesting outlook, loosely similar.

> So we did mechanistic studies on small models, Gemma 4 particularly, and found the hidden state for different layers carry meaningful self-awareness signal for various situations.

Neat! Just to make sure I understand - you trained your probe layer to take this hidden state and predict p(wrong)?

Curious to learn more. Any more info on your approach (esp the mechanistic study)?

Correct, the study is verbose, we will compile into a neat shareable report and publish once we solve the pending caveats. Interesting username btw haha.

Was actually pulling on a similar thread as I saw announcement so I integrated it just for fun. Have been only running this on my Framework Desktop but should be runnable elsewhere https://github.com/olafura/gemma-4-mic-transcribe

Have you benched this for coding tasks, with a fallback to a larger local model, for example Qwen-3.6-27B?

Or using it for sub-tasks, where a framework with a larger primary model dispatches simpler jobs ("summarize ...", etc.) to it?

We are currently working on larger models atm! your suggestion is cool btw

I don't really get why you need handoff if your score is accurate. If it is, and it is low for a given response, just let the harness re-run the prompt with a different seed until the score is high enough. If this approach doesn't work, your score is most likely garbage.

Smaller models in production can produce quite interesting outputs.

Interesting how that once again matches human intuition: It's easier to know what you don't know, than to know what is true.

haha, interesting analogy

Love what you're doing at cactus, keep it up!!

Does the model quality become degraded in other ways?

no because we freeze the main weights and only train our added weights.

It's impossible to add weights which degrade the quality?

Frozen weights, unfrozen marketing copy.

Ran into similar thing tuning a classifier's confidence output, calibration drifts hard once you fine-tune on new data. What worked for us: log every (confidence, correct/wrong) pair in prod and refit a sigmoid on top weekly. Curious if they're doing anything like that or just trusting the initial calibration forever.