> 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.
> 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?
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
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.
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.
opinion is ephemeral, cannot be independently confirmed or denied; purely single viewer information sounds like not-a-fact to me
Opinion is a ghost in the machine—fleeting, unanchored, and vanishing upon inspection. It leaves no footprint for another mind to verify, no independent trail of evidence to confirm or deny its existence. To rely on such solitary perception is to drift in a void where facts do not exist; it is not merely uncertainty, but an absence of reality itself—a whisper that the universe refuses to echo back.
--AI Carl Sagan
Exactly. A fact is something that can be objectively confirmed or rejected. Exactly what a specific person likes is a fact, because it can be verified/falsified by simply asking them. An opinion, by contrast is 'Elephants are large.' The reason that's an opinion is because while it seems obviously true, it's subjective. Similarly, grass is purple with pink polka-dots is a fact. It's obviously false, but of course right/wrong has nothing to do with fact/opinion.
_All_ statements are only valid inside the window within which their understandable context stays constant, facts or otherwise.
> "Verifying an opinion X at time T does not entail X at time T+1."
And verifying that my cat weighs 9.5 lbs now does not entail that he will weigh 9.5 lbs tomorrow, but "my cat weighs 9.5 lbs" is not an opinion; it's a statement of fact from measurement/observation and implies "extrapolated from the last specific check done in the usual way" unless explicitly stated otherwise because that's how language works in polite society. "I like green" is the same.
"My cat weighs 9.5 lbs" is only true within a context of a particular moment and location and measurement apparatus and understanding of cat and ownership and 9.5 and lbs.
"1+1=2" is only true within a context of what "1" and "2" and "+" and "=" mean.
"Ephemeral" isn't relevant - a statement of fact can be true at one time and false at another.
Confirmation also is somewhat orthogonal: there can be true facts that we can't confirm.
Conventionally, "I like green" is considered a factual claim about the speaker's preferences. Even if you refuse to acknowledge it as a fact because of the restrictions you've given, that doesn't make it an opinion. It's subjective, but it's not an opinion.
Would you say that if someone is experiencing a headache, and they say "I have a headache", that that's also merely an opinion?
The catch-all is a problem. We've been studying fallacies in arguments since before the time of Plato, and do not have a complete answer that always works, to prevent the logical failure.
Complex assertions and compounding arguments, can still result in circular reasoning - the petitio principii, or worse.
Most LLMs are guided by strong statements like it must do this or that. Which assumes that it won't apply the fallacy of accident, even though people regularly do. (Don't agree? Argue a definition with a BA for a day and come back to me.)
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.
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".
None of what you've said has anything to do with correctness, only certainty. And it's funny to me in context if you're certain that the opposite is true.
> "It's also worth mentioning that LMs are wrong in different ways from humans."
This to me only shows a poor and incomplete knowledge of humans and of wrongness and its causes. If you look at all of the humans and all of the times, you find exhibitions of all the same ways of being wrong. At best you can say that some (<all) LMs are wrong in different ways than some (<all) humans at some (<all) times, but that's as good as saying that different humans are wrong in different ways than different other humans at different times.
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.
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.
> successful in programming than other disciplines
It seems successful specifically in "computer science", for "programming" which encompasses a much wider field, "successful" is still up for debate. Also, we haven't even figured out yet what "success" even means in a software development process, for some it's that the program does everything the users want, in the way they expect, for other's, it's about correctness, reliability, performance, composability, maintainability and/or a whole host of other "measures".
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?
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.
cacio-e-pepe | a day ago
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)?
[OP] HenryNdubuaku | a day ago
cacio-e-pepe | 23 hours ago
And thanks, huge pasta fan :)
[OP] HenryNdubuaku | 21 hours ago
astrobiased | 21 hours ago
[OP] HenryNdubuaku | 21 hours ago
zdw | 20 hours ago
Or using it for sub-tasks, where a framework with a larger primary model dispatches simpler jobs ("summarize ...", etc.) to it?
[OP] HenryNdubuaku | 12 hours ago
mncharity | 19 hours ago
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?
[OP] HenryNdubuaku | 12 hours ago
olafura | 19 hours ago
BugsJustFindMe | 19 hours ago
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.
mistrial9 | 19 hours ago
Complex assertions and their compounded forms can be filtered by these classes fairly easily. I rarely see this referenced in general comments.
BugsJustFindMe | 18 hours ago
"I like green" is a statement of fact about what you like. "Green is good" would be an opinion. It's funny to me how often people mix the two.
> forms can be filtered by these classes fairly easily
Obviously not so easily. :)
mistrial9 | 17 hours ago
Opinion is a ghost in the machine—fleeting, unanchored, and vanishing upon inspection. It leaves no footprint for another mind to verify, no independent trail of evidence to confirm or deny its existence. To rely on such solitary perception is to drift in a void where facts do not exist; it is not merely uncertainty, but an absence of reality itself—a whisper that the universe refuses to echo back. --AI Carl Sagan
jurgenburgen | 14 hours ago
somenameforme | 12 hours ago
naasking | 12 hours ago
BugsJustFindMe | 4 hours ago
> "Verifying an opinion X at time T does not entail X at time T+1."
And verifying that my cat weighs 9.5 lbs now does not entail that he will weigh 9.5 lbs tomorrow, but "my cat weighs 9.5 lbs" is not an opinion; it's a statement of fact from measurement/observation and implies "extrapolated from the last specific check done in the usual way" unless explicitly stated otherwise because that's how language works in polite society. "I like green" is the same.
"My cat weighs 9.5 lbs" is only true within a context of a particular moment and location and measurement apparatus and understanding of cat and ownership and 9.5 and lbs.
"1+1=2" is only true within a context of what "1" and "2" and "+" and "=" mean.
And so on.
antonvs | 12 hours ago
Confirmation also is somewhat orthogonal: there can be true facts that we can't confirm.
Conventionally, "I like green" is considered a factual claim about the speaker's preferences. Even if you refuse to acknowledge it as a fact because of the restrictions you've given, that doesn't make it an opinion. It's subjective, but it's not an opinion.
Would you say that if someone is experiencing a headache, and they say "I have a headache", that that's also merely an opinion?
shakna | 10 hours ago
Complex assertions and compounding arguments, can still result in circular reasoning - the petitio principii, or worse.
Most LLMs are guided by strong statements like it must do this or that. Which assumes that it won't apply the fallacy of accident, even though people regularly do. (Don't agree? Argue a definition with a BA for a day and come back to me.)
baq | 13 hours ago
NitpickLawyer | 13 hours ago
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".
cmrdporcupine | 10 hours ago
Or the opposite https://arxiv.org/abs/2606.06574v1
htrp | 7 hours ago
BugsJustFindMe | 6 hours ago
> "It's also worth mentioning that LMs are wrong in different ways from humans."
This to me only shows a poor and incomplete knowledge of humans and of wrongness and its causes. If you look at all of the humans and all of the times, you find exhibitions of all the same ways of being wrong. At best you can say that some (<all) LMs are wrong in different ways than some (<all) humans at some (<all) times, but that's as good as saying that different humans are wrong in different ways than different other humans at different times.
pjc50 | 11 hours ago
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.
embedding-shape | 10 hours ago
It seems successful specifically in "computer science", for "programming" which encompasses a much wider field, "successful" is still up for debate. Also, we haven't even figured out yet what "success" even means in a software development process, for some it's that the program does everything the users want, in the way they expect, for other's, it's about correctness, reliability, performance, composability, maintainability and/or a whole host of other "measures".
robrenaud | 19 hours ago
[OP] HenryNdubuaku | 12 hours ago
zamadatix | 10 hours ago
jstanley | 10 hours ago
zamadatix | 2 hours ago
jstummbillig | 12 hours ago
[OP] HenryNdubuaku | 8 hours ago
dmrivers | 10 hours ago
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?
[OP] HenryNdubuaku | 8 hours ago
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.
maxgashkov | 9 hours ago
[OP] HenryNdubuaku | 9 hours ago
asar | 28 minutes ago