An LLM is a lossy encyclopedia

When you have a lossy piece of media, such as a compressed sound or image file, you can always see the resemblance to the original and note the degradation as it happens. You never have a clear JPEG of a lamp, compress it, and get a clear image of the Milky Way, then reopen the image and get a clear image of a pile of dirt.

Furthermore, an encyclopaedia is something you can reference and learn from without a goal, it allows you to peruse information you have no concept of. Not so with LLMs, which you have to query to get an answer.

In compressed audio these can be things like clicks and boings and echoes and pre-echoes. In compressed images they can be ripply effects near edges, banding in smoothly varying regions, but there are also things like https://www.dkriesel.com/en/blog/2013/0802_xerox-workcentres... where one digit is replaced with a nice clean version of a different digit, which is pretty on-the-nose for the LLM failure mode you're talking about.

Compression artefacts generally affect small parts of the image or audio or video rather than replacing the whole thing -- but in the analogy, "the whole thing" is an encyclopaedia and the artefacts are affecting little bits of that.

Of course the analogy isn't exact. That would be why S.W. opens his post by saying "Since I love collecting questionable analogies for LLMs,".

It was quickly discovered that LLMs are capable of re-checking their own solutions if prompted - and, with the right prompts, are capable of spotting and correcting their own errors at a significantly-greater-than-chance rate. They just don't do it unprompted.

Eventually, it was found that reasoning RLVR consistently gets LLMs to check themselves and backtrack. It was also confirmed that this latent "error detection and correction" capability is present even at base model level, but is almost never exposed - not in base models and not in non-reasoning instruct-tuned LLMs.

The hypothesis I subscribe to is that any LLM has a strong "character self-consistency drive". This makes it reluctant to say "wait, no, maybe I was wrong just now", even if latent awareness of "past reasoning look sketchy as fuck" is already present within the LLM. Reasoning RLVR encourages going against that drive and utilizing those latent error-correction capabilities.

As of today, 'bad' generations early in the sequence still do tend towards responses that are distant to the ideal response. This is testable/verifiable by pre-filling responses, which I'd advise you to experiment with for yourself.

'Bad' generations early in the output sequence are somewhat mitigatable by injecting self-reflection tokens like 'wait', or with more sophisticated test-time compute techniques. However, those remedies can simultaneously turn 'good' generations into bad, they are post-hoc heuristics which treat symptoms not causes.

In general, as the models become larger they are able to compress more of their training data. So yes, using the terminology of the commenter I was responding to, larger models should tend to have fewer 'compression artefacts' than smaller models.