John C Malone Professor at @JohnsHopkins @JHUCompSci @jhuclsp @jhumceh; Part time @techatbloomberg (tweets my own) Director @hopkinsdsai @mdredze.bsky.social🦋
Baltimore, MD
Joined August 2012
- Tweets2.8K
- Following797
- Followers6.7K
- Likes3.5K
Mark Dredze retweeted
I am attending #EMNLP2026 and recruiting PhD students for Fall 2027 at JHU @jhuclsp. Research topics include speech and audio foundation models, expressive speech, healthcare, and mechanistic interpretability.
If you’re interested, I will be at our posters with my students.
Mark Dredze retweeted
Do you feel like many agents have different personality? Is there a pattern?
Now, with Act*ONOMY (arxiv.org/abs/2605.13625), behavioral traits can be identified in the literature but also be inferred from what models do and say.
We found that some models are more talkative, while others are more responsible, planful, and energetic. Check the paper for more details (arxiv.org/pdf/2609.32776).
“Does anyone else feel like each AI model has a distinct personality?”
We stopped asking models and watched them work. A-B-D infers agent traits bottom-up from 345K real-world trajectories (80 models, 50 harnesses). 🧵
📄 arxiv.org/abs/2609.32776
💻 github.com/JoeZhao527/ABD
Mark Dredze retweeted
“Does anyone else feel like each AI model has a distinct personality?”
We stopped asking models and watched them work. A-B-D infers agent traits bottom-up from 345K real-world trajectories (80 models, 50 harnesses). 🧵
📄 arxiv.org/abs/2609.32776
💻 github.com/JoeZhao527/ABD
Mark Dredze retweeted
The final startup pitches of the morning at the @HopkinsDSAI Innovation Summit focused on a big challenge: turning AI advances into better care and smarter medical devices. Meet Aprentis and Inner Logic.
Mark Dredze retweeted
Want to know about the latest #naturallanguageprocessing research? See what our computer scientists are presenting next week at @emnlpmeeting’s #EMNLP2026: (🧵 1/12)
ALT EMNLP2026 Budapest logo.
Mark Dredze retweeted
“The commercial promise of AI at Johns Hopkins isn’t in one place. It’s everywhere.”
@HopkinsDSAI Director @mdredze highlighted the breadth of AI innovation across Johns Hopkins and the importance of translating breakthrough research into real-world impact.
Mark Dredze retweeted
This is the last week to apply to our Bloomberg CTO AI Internship Program!
If you are a current graduate student interested in applying cutting edge AI, consider applying!
👉 Apply here: forms.cloud.microsoft/r/ZLC3…
Mark Dredze retweeted
Congratulations to @isabelcachola, @mdredze, and @magmar_11 on their @hcomp_conf Best Paper Award! 🏆 Learn more about their award-winning work: cs.jhu.edu/news/johns-hopkin…
Mark Dredze retweeted
Schmidt Sciences is hiring Fellows in Residence!
You help us with grantmaking decisions, while continuing your own research
- $50k compute, $50k for a workshop
- based in NYC
- past fellows kept affiliations with Stanford, MIT, JHU
Apply by Dec 15
Link 👇 Please share/rt :)
Mark Dredze retweeted
If an LLM agent notices conflicts, will it tell you?
Our #EMNLP2026 Findings paper finds that higher accuracy doesn't always mean greater epistemic humility: agents can spot conflicts early yet leave uncertainty out of wrong answers.
📄 arxiv.org/abs/2610.12360 🧵👇 #NLProc #LLM
Mark Dredze retweeted
How useful is GPT-6 Astra for robot control, really? 🤔
Lin and I burned plenty of tokens to explore questions:
- How does Astra compare with recent VLAs across different tasks?
- Which control interfaces work best with Astra (linear motion vs. action chunking)?
- Can it learn from its own failures or others' demonstrations?
- Does higher reasoning effort (more thinking) improve performance?
- What happens when we remove its memory? (Many VLAs have little or no memory)
Check Ditto-Bench and our findings in the blog post. Full code included! 👇
llong-cs.github.io/llms-for-…
This was a fun analysis led by Lin, one of my first PhD students at JHU. He joined without much robotics experience but quickly produced a thorough benchmark and thoughtful analysis! 💪
GPT-6 Astra has shown impressive results in simple robot control tasks.
But what if the tasks become more challenging and involve complex geometric constraints, precise contact, or improvised tool use? To find out, we:
- Built Ditto-Bench, where simple goals meet challenging physics, and tested both Astra and MolmoAct2.
- Systematically ablated key design choices to understand what makes these robot-control agents succeed, or get stuck.
- Explored what new possibilities these models could open up for robotics, such as in-context learning.
More details 🧵👇
Wow! My students are on a roll! 🧵
In the past month:
🏆 HCOMP Best Paper (@isabelcachola)
🏆 COLM Outstanding Paper (@KaiserWhoLearns @JentseHuang )
🏆 DEX Best Abstract, People’s Choice (Ahmed Hassoon)
🎤 NeurIPS GenAI4Health Oral, top 4% (@YuanCosette)
Details below 👇
Retrieve-then-verify medical fact-checking fails in systematic ways, and bigger models, more reasoning, better sources, and medical fine-tuning don’t fix them.
Mark Dredze retweeted
🥳🥳 Let's go! Yay!!!
Honoured to receive an outstanding paper award for our work! Couldn’t have done it without all my coauthors @KaiserWhoLearns @millicent_li @wordscompute @lltjuatja @JentseHuang and my advisor @gneubig 🫡