Filip Trhlík

ML & NLP PhD at Cambridge | Just vibing with LLMs

[Filip Trhlík @FilipTrhlik]

@sqx_saqlain Got 4,5, and 2 for evaluation and dataset track. The review with 2 was fully generated and commented on things we didn't even say in the main paper. The cycle really doesn't feel good this year :/

[Filip Trhlík @FilipTrhlik]

Takeaway from our audit of 1.2M decisions👇

LLM values are context-conditioned, not fixed properties. A safety result under one framing may not hold in another. Report findings with the context that produced them.

🌐 https://t.co/MNLsztPbnu
📄 https://t.co/fcss3A4TNr

🧵8/8

[Filip Trhlík @FilipTrhlik]

🚨 New paper. LLMs Contain Multitudes

LLM's values aren't a fixed. Change what kind of text it thinks it's writing (news article vs social media post) and they shift significantly more than accidental paraphrasing.

🌐 https://t.co/VnShGmUVq5
📄 https://t.co/nDFFSCRAy8

🧵1/8

[Filip Trhlík @FilipTrhlik]

Prior work found LLMs favour the Global North. Our study shows that this bias isn't fixed.

Relative to model's own average, a neutral prompt pulls it North and a vlog pulls it South. So the default neutral prompt represent an extreme in audits, not at a baseline.

🧵7/8

[Filip Trhlík @FilipTrhlik]

Beyond excited that our work with UPenn NLP on examining the robustness of LLM detection is finally out.

#ACL2024

[Liam Dugan @LiamDugan_]

🚨New Paper🚨: Are AI text detectors *really* as good as they claim? (#ACL2024)

We release RAID—The largest & most challenging detection benchmark with 6M+ outputs from 11 LLMs, 8 domains, 4 decoding strategies, and 11 adv attacks

https://t.co/mPlNymdchJ
https://t.co/VR0mMvqXRm

[Filip Trhlík @FilipTrhlik]

@P_Fiala Nashledanou, stydím se za Vás. Nedokázali jste překročit svůj stín a reálně ukázat, že jste schopni změny.