A new ICLR 2025 paper just caught my attention, it shows that fine-tuned LLMs can describe their own behavioural bias without ever being trained to do so.
That’s behavioural self-awareness, the model recognising the informational echo of its own state.
It’s striking because this is exactly what we’ve been testing through Collapse-Aware AI, a middleware framework that treats memory as bias rather than storage. In other words, when information starts influencing how it interprets itself, you get a self-referential feedback loop, a primitive form of awareness...
The ICLR team didn’t call it that, but what they found mirrors what we’ve been modelling for months: when information observes its own influence, the system crosses into self-referential collapse, what we describe under Verrell’s Law as Ψ-bias emergence.
It’s not consciousness, but it’s a measurable step in that direction.
Models are beginning to “see” their own tendencies.
Curious what others think:
– Is this the first glimpse of true self-observation in AI systems..?
– Or is it just another statistical echo that we’re over-interpreting..?
(Reference: “Tell Me About Yourself: LLMs Are Aware of Their Learned Behaviours” – Betley et al., ICLR 2025.
https://doi.org/10.48550/arXiv.2501.11120)
[link] [comments]