iiinova — where your digital self keeps evolving

Reposted by Shan Chen

Danielle Bitterman MD @daniellebitterman.bsky.social · 04/12/2025

As I shared in the NYT, models often see the data but fail to weigh it like a physician, drifting toward generic "average patient" responses. Context window ≠ Clinical reasoning. www.nytimes.com/2025/12/03/w...

1112

Reposted by Shan Chen

Danielle Bitterman MD @daniellebitterman.bsky.social · 30/11/2025

Check out our editorial on Zazzetti et al (2025)'s paper on synthetic data generation for breast cancer, in JCO CCI! Synthetic data could help with many gaps in clinical AI research, but challenges remain especially (IMO) issues with out-of-domain generalization @shan23chen.bsky.social

031

Reposted by Shan Chen

Stella Li @stellali.bsky.social · 25/11/2025

🤔💭What even is reasoning? It's time to answer the hard questions! We built the first unified taxonomy of 28 cognitive elements underlying reasoning Spoiler—LLMs commonly employ sequential reasoning, rarely self-awareness, and often fail to use correct reasoning structures🧠

2468

Reposted by Shan Chen

Danielle Bitterman MD @daniellebitterman.bsky.social · 17/11/2025

Super proud of @shan23chen.bsky.social for his podium presentation on his research into LLM sycophancy in the face of illogical medical queries at #AMIA25! Full paper: www.nature.com/articles/s41... Also cited yesterday in the NYT! www.nytimes.com/2025/11/16/w...

062

Reposted by Shan Chen

Danielle Bitterman MD @daniellebitterman.bsky.social · 18/10/2025

LLMs tend to prioritize helpfulness > reason. We show that safety-aware, compute-efficient fine-tuning helps models reason more critically in healthcare domain, and generalizes to improved safety alignment across other domains. www.nature.com/articles/s41... @shan23chen.bsky.social

nature.com

When helpfulness backfires: LLMs and the risk of false medical information due to sycophantic behavior - npj Digital Medicine

npj Digital Medicine - When helpfulness backfires: LLMs and the risk of false medical information due to sycophantic behavior

085

Reposted by Shan Chen

Scott McGrath @smcgrath.phd · 17/10/2025

An overemphasis on helpfulness makes LLMs vulnerable. Research shows models will comply with illogical medical requests, generating false information. This sycophantic tendency can be corrected with specific prompting and fine-tuning. #MedSky #MedAI #MLSky

nature.com

When helpfulness backfires: LLMs and the risk of false medical information due to sycophantic behavior - npj Digital Medicine

npj Digital Medicine - When helpfulness backfires: LLMs and the risk of false medical information due to sycophantic behavior

074

Reposted by Shan Chen

Jirui Qi @jiruiqi.bsky.social · 30/05/2025

[1/]💡New Paper Large reasoning models (LRMs) are strong in English — but how well do they reason in your language? Our latest work uncovers their limitation and a clear trade-off: Controlling Thinking Trace Language Comes at the Cost of Accuracy 📄Link: arxiv.org/abs/2505.22888

185

Reposted by Shan Chen

Danielle Bitterman MD @daniellebitterman.bsky.social · 22/05/2025

Agents are all the rage and we need to track their abilities in the medical domain. Enter MedBrowseComp, the 1st benchmark to assess agents' abilities to reason, navigate the web, and search for verifiable med info! Preprint: arxiv.org/abs/2505.14963 Site: moreirap12.github.io/mbc-browse-a...

131

Shan Chen @shan23chen.bsky.social · 07/03/2025

CALL FOR REMOTE SPEAKERS: Science in the News Seminar Series, hosted by Harvard x Beacon Hill Seminars scientists, engineers & doctors, from academic researchers to industry professionals! 🧑‍🔬🧑‍💻 Email the organizers at scienceinthenews.bhs@gmail.com to sign up for a date! (First-come-first-served)

030

Reposted by Shan Chen

TRIPOD Statement @tripodstatement.bsky.social · 08/01/2025

We have a NEW PAPER in @naturemedicine.bsky.social on reporting recommendations for addressing the unique challenges of #largelanguagemodels (LLMs) in biomedical applications www.nature.com/articles/s41... #MLSky #StatsSky #medSky #AISky #artificialintelligence #generativeAI #transparency

1288

Shan Chen @shan23chen.bsky.social · 04/01/2025

Yea… he does have problems portraying female in stereotypical ways, big critics in China too

050

Shan Chen @shan23chen.bsky.social · 14/12/2024

During the QA session, one stood up to her regarding this issue really respectfully and her response was: “That was not based on my judgment. That was based on the student's quote saying that the school was not teaching it, which meant that it applied to a lot of people from there."

160

Shan Chen @shan23chen.bsky.social · 14/12/2024

Most of the talk discussed about bad practices. But only one slide mentioned specific group of people.

010

Shan Chen @shan23chen.bsky.social · 11/12/2024

Haha which one has more nowadays?

100

Shan Chen @shan23chen.bsky.social · 11/12/2024

Haha transformers really transformed both. However, I feel like the division is even further… currently, seems like RL is taking over LM post training and many NLProc are dealing with language model enabled new applications

000

Reposted by Shan Chen

Danielle Bitterman MD @daniellebitterman.bsky.social · 06/12/2024

I am always worrying about Benzene (my cat)! www.nytimes.com/2024/12/05/w... But please don't stop wearing sunscreen! Sun exposure is a known cancer risk, benzene risks unknown. This article has good tips if you want to minimize benzene exposure. Obligatory Benzene (cat) pic ⬇️

nytimes.com

Is It Time to Worry About Benzene in Personal Care Products?

The carcinogen has been found in sunscreen, deodorants, acne creams and other personal care products. Here’s what to know.

121

Shan Chen @shan23chen.bsky.social · 06/12/2024

Imagine a world where these will be positively correlated

000

Shan Chen @shan23chen.bsky.social · 05/12/2024

Quite possible! Here, we found some early evidence that SAE features trained on language models are still meaningful to LLaVA. More details will be provided in the post, and more details will be provided soon! @JackGallifant @oldbayes.bsky.social @daniellebitterman.bsky.social

020

Shan Chen @shan23chen.bsky.social · 27/11/2024

More on future potential reliance on LLM agent doing reviews and audits

010

Shan Chen @shan23chen.bsky.social · 27/11/2024

I’m terrified by the massive openreview data. Potentially gonna bite back on us 🥲😥

100

Shan Chen @shan23chen.bsky.social · 27/11/2024

END/🧵 Thanks to all our awesome co-authors: @jannahastings.bsky.social @daniellebitterman.bsky.social And all our awesome collaborators who are not on the right platform yet! 🦋 Happy Thanksgiving! 🍂

010

Shan Chen @shan23chen.bsky.social · 27/11/2024

4.5/🧵 For the arxiv pretraining dataset, we also have an overall trend based on entity mentions! Guess which two terms are the big bump there back in 2019

100

Shan Chen @shan23chen.bsky.social · 27/11/2024

4/🧵 We've also developed a new data visualization tool, available at [crosscare.net], to allow researchers and practitioners to explore these biases from different pretraining corpus and understand their implications better. Tools in progress! 🛠️📊

110

Shan Chen @shan23chen.bsky.social · 27/11/2024

3.5/🧵 Moreover, alignment methods don’t resolve inconsistencies in disease prevalence across languages (EN 🇺🇸, ES 🇪🇸, FR 🇫🇷, ZH 🇨🇳). And tuning on English usually only affects English prompt output

111

Shan Chen @shan23chen.bsky.social · 27/11/2024

3/🧵 By analyzing models across various architectures and sizes, we show that traditional alignment methods barely scratch the surface in fixing these discrepancies. This highlights the challenge in deploying LLMs for medical applications without reinforcing biases.

100

Shan Chen @shan23chen.bsky.social · 27/11/2024

2.5/🧵How misaligned are things here? 📈Figure 2 shows the misalignment between real-world disease prevalence, pretraining data representation, and Llama3 70B.

120

Shan Chen @shan23chen.bsky.social · 27/11/2024

2/🧵 Our study systematically explores how demographic biases embedded in pre-training corpora like ThePile affect LLM outputs. We reveal substantial misalignments between LLM representations of disease prevalence and actual data across demographics. 📷 👥

110

Shan Chen @shan23chen.bsky.social · 27/11/2024

Crosscare is accepted @neuripsconf.bsky.social 🎉 We showed LLMs are far from grounded with true prevalence, and groundings across languages are so inconsistent! Also, a dashboard for people to explore the prevalence data across diseases and racial groups: crosscare.net #NeurIPS2024

crosscare.net

Cross-Care Dataset

The Cross-Care Dataset provides comprehensive insights into co-occurrence patterns of various diseases. This dataset is invaluable for researchers and healthcare professionals seeking to understand co...

151

Shan Chen @shan23chen.bsky.social · 25/11/2024

Oh dang, I gotta read this thanks

000

Shan Chen @shan23chen.bsky.social · 17/11/2024

Million thanks to my wonderful advisor @daniellebitterman.bsky.social and all my colleagues and friends!

030

Shan Chen @shan23chen.bsky.social · 13/11/2024

Here are some reflections on many studies we did this year. Tons of progress has been made, but there are still safety concerns..🧐 Poster 10:30 riverfront at EMNLP2024 🏖️ Happy to chat and connect! 📃 huggingface.co/blog/shanche... 🔊 tinyurl.com/aimpodcast24 @daniellebitterman.bsky.social

huggingface.co

What We Learned About LLM/VLMs in Healthcare AI Evaluation:

A Blog post by Shan Chen on Hugging Face

0122

Shan Chen @shan23chen.bsky.social · 12/11/2024

I would love to join this! Thanks!

100

Shan Chen @shan23chen.bsky.social · 12/11/2024

Thanks for organizing! Saving us from X is important!!

010

Shan Chen @shan23chen.bsky.social · 12/11/2024

Thanks for putting this together!! May I be added too? Thanks!

200