Jack Stanley

PhD student @mcgillu & @Mila_Quebec

[Mila - Institut québécois d'IA @Mila_Quebec]

How do we stop LLMs from making things up? Read Praneet Suresh's groundbreaking method to detect and eliminate AI hallucinations from within the model: https://t.co/jEX3dMz1o0

@danilobzdok @jackhtstanley

[Danilo Bzdok @danilobzdok]

Now accepted at #NeurIPS 2025:

Our team makes three key contributions to
#large #language #models + #security:

1) Pre-trained transformer models impose semantic structure on inputs, tying them into learned conceptual webs, even if the model inputs are ambiguous or lack any coherent meaning.

2) This skewing as information trickles through transformer processing layers exacerbates as input uncertainty increases.

3) The constellation of concept activations in a transformer model’s intermediate processing representations can be used to reliably predict the tendency of hallucinated or unfaithful output.

https://t.co/5NBP3pzVoq
https://t.co/nC1NQqefew

Driven by @praneet_suresh_ + @jackhtstanley; great collabo with @soniajoseph_ & @ScimecaLuca.

[Jack Stanley @jackhtstanley]

Why do LLMs hallucinate? We may have the answer!

We show that LLMs imagine MORE concepts in inputs with LESS semantic structure. Very surprising!

Read the thread below for more details 👇
@praneet_suresh_ @soniajoseph_ @ScimecaLuca @danilobzdok

[Praneet @praneet_suresh_]

We're thrilled to finally share what we've been working on. Our new paper gives a first-ever glimpse into the "mind" of an LLM, and we discovered something that shocked us: AIs see ghosts in the machine. 👻

What do LLMs see in the dark?

[Danilo Bzdok @danilobzdok]

Why #language #models hallucinate?

->We have experimental data on that.

We precisely locate where #hallucinations arise in LLMs + show by targeted #intervention that these internal bias features do cause bad model outputs

The trick: we get LLMs drunk and they tell us the truth; In vino veritas

https://t.co/S7iARxYW9I

@btibor91 @NeelNanda5 @pmddomingos @rohanpaul_ai @MLStreetTalk @MichalKubicek

[Jack Stanley @jackhtstanley]

https://t.co/HcaqrnhBNQ

@Mila_Quebec @mcgillu Emmett Rabot, @sivareddyg @ebelilov, Laurent Mottron, @danilobzdok

[Jack Stanley @jackhtstanley]

Sharing our latest paper, now published in Cell.

We fine-tuned an LLM to accurately predict the autism diagnosis from over 4200 multi-page clinical text reports.

[Jack Stanley @jackhtstanley]

Read below for an example of the utility of LLMs in the biomedical realm, some approaches to LLM interpretability, and potentially impactful findings for autism. (Spoiler: repetitive behaviours and special interests appear to be more salient for diagnosis than social factors)