Pascal Iversen

BioML, Uncertainty Estimation/XAI, Scientific Reproducibility @FU Berlin

[Pascal Iversen @Iv_Pascal]

@s4nt4tweet @martinmbauer so what? (genuine question)

[Pascal Iversen @Iv_Pascal]

@hthieblot

[Pascal Iversen @Iv_Pascal]

@Nature Please noo, is it April 1st again?

[Pascal Iversen @Iv_Pascal]

Easy jailbreak to get Claude to claim to be a human:

[Pascal Iversen @Iv_Pascal]

@damekdavis Same outcome if you interpret the "they" in your post as the authors of the papers.

[Pascal Iversen @Iv_Pascal]

@AlexanderKalian I don't think emotions/consciousness are the missing pieces, because they are non-functional (you can always just simulate them). I think it is long-term goals. Not optimized for in RL finetuning. AI's don't invent new areas of math or new programming languages to solve problems

[Pascal Iversen @Iv_Pascal]

@prz_chojecki That is an awesome idea! I love the lore/drama behind this "proof," and the philosophical implications are interesting (math seems to become postmodern...). But what are the "funny things" that happen here?

[Pascal Iversen @Iv_Pascal]

@Plinz @maier_ak Because it is a much more difficult study to do

[Pascal Iversen @Iv_Pascal]

@Quasilocal Makes sense. In BioML it is vastly different. There are papers in reputable journals with many citations that are completely sloppy. You have to be an expert in the field to spot this

[Pascal Iversen @Iv_Pascal]

@AlexShtf Even worse: when they use max to mean argmax.

[Pascal Iversen @Iv_Pascal]

Now published in Nature Communications:
https://t.co/23eFOVOHmo

DrEval could also function as an unbiased reward signal for AI agent–based model development in biological ML. We are happy to chat about this if anyone is interested in details :)

[Pascal Iversen @Iv_Pascal]

ML-based cancer cell line drug response models are well-motivated, and significant research effort has gone into developing complex modeling approaches (over 100 papers in 2025).

The problem: under rigorous evaluation, we found none that actually works.

[Pascal Iversen @Iv_Pascal]

Most are published based on inflated metrics, break down when probed in realistic application scenarios, and are outperformed by simple baselines.

We built DrEval, a pipeline for unbiased evaluation of these models, to encourage more meaningful progress in the field.

[Pascal Iversen @Iv_Pascal]

@AlexanderKalian Fully agree. I see that in my field: Drug response prediction in cancer. The bottleneck is not papers, not intelligence, but just large datasets. People train large neural networks on 1k data points, and surprise Pikachu face that we haven't cured cancer

[Pascal Iversen @Iv_Pascal]

@Plinz The English movie title being so bad and boring ("The Misfortunates") shows your point.

[Pascal Iversen @Iv_Pascal]

@HistedLab Biggest problem for me is the papers that are "not even wrong". Too fluffy and unfalsifiable. They have names like "Navigating the landscape of xy" and noone will ever engineer something based on them.

[Pascal Iversen @Iv_Pascal]

@docmilanfar Interestingly, this paper itself also overstates some claims: https://t.co/V5pqV4C665