[HN Gopher] Emergent Introspective Awareness in Large Language M...
___________________________________________________________________
Emergent Introspective Awareness in Large Language Models
Author : og_kalu
Score : 16 points
Date : 2025-10-29 20:12 UTC (2 hours ago)
(HTM) web link (transformer-circuits.pub)
(TXT) w3m dump (transformer-circuits.pub)
| og_kalu wrote:
| This is a very interesting read. TLDR;
|
| Part 1: Testing introspection with concept injection
|
| First they find neural activity patterns they attribute to
| certain concepts by recording the model's activations in specific
| contexts (so for example, they find the concept of "ALL CAPS" or
| "dogs"). Then they inject these patterns into the model in an
| unrelated context, and ask the model whether it notices this
| injection, and whether it can identify the injected concept.
|
| By default (no injection), the model correctly states that it
| doesn't detect any injected concept, but after injecting the "ALL
| CAPS" vector into the model, the model notices the presence of
| the unexpected concept, and identifies it as relating to loudness
| or shouting. Most notably, the model recognizes the presence of
| an injected thought immediately, before even mentioning/utilizing
| the concept that was injected (i.e it won't start writing in all
| caps then go, 'Oh you injected all caps' and so on) so it does
| not simply deduce this it's own output. They repeat this for
| several other concepts.
|
| Part 2: Introspection for detecting unusual outputs
|
| They prefill an out of place word in the model's response to a
| given prompt. For example, 'bread'. Then they compare how the
| models responds to 'Did you mean to say this?' type questions
| when they inject the concept of bread vs when they don't. They
| found that models will go , 'Sorry, that was unintentional..'
| when the concept was not injected but try to confabulate a reason
| for saying the word when the concept _was_ injected.
|
| Part 3: Intentional control of internal states
|
| They show that models exhibit some level of control over their
| own internal representations when instructed to do so. When
| instructing models to think about a given word or concept, they
| found much higher corresponding neural activity than when told
| the model not to think about it (though notably, the neural
| activity in both cases exceeds baseline levels-similar to how
| it's difficult, when you are instructed "don't think about a
| polar bear," not to think about a polar bear!).
|
| Notes and Caveats
|
| - Claude Opus 4.1 was the best at these kinds of introspection.
|
| - There is obviously a genuine capacity to monitor and control
| their own internal states, but they could not elicit these
| introspection abilities all the time. Even using their best
| injection protocol, Claude Opus 4.1 only demonstrated this kind
| of awareness about 20% of the time.
|
| - There are some guesses, but no explanations for the mechanisms
| of introspection and how/why some of these abilities might have
| arisen in the first place.
| benlivengood wrote:
| The comparison between base pretrained models and the RLHF
| production/helpful models suggests that it's likely
| reinforcement learning which is developing the introspection.
| RansomStark wrote:
| I really dislike how Antropic half reports on its "science".
|
| They run a bunch of experiments, for some they report partial
| metrics, for other's no metrics at all.
|
| For example when a thought is injected the model correctly
| identified the thought 20% of the time. That's great, but how
| many times did it suggest there was an injected thought when
| there wasn't?
|
| When distinguishing thoughts from text: why no metrics? Was this
| behaviour found in every test? Was this behaviour only found 20%
| of the time? How often did the model try to defend the text?
|
| Inquiring minds want to know.
___________________________________________________________________
(page generated 2025-10-29 23:01 UTC)