📢 Our model pre-release is now open!
🩺 Try it: https://t.co/ReoYVaEDVb
It supports EMR-based diagnosis & prescription suggestions with search over guidelines, reimbursement criteria, and drug references.
Use your own input or simply “Try Samples.”
Feedback welcome!
I’ll present our team’s recent work at #ICML2026#SD4H!
📄 Distilling Expert-level Planning for Real-world Diabetes Prescribing
We train medical LLMs to follow expert prescribing workflows using EMR context, insurance-aware search, and RL.
🗓️ Jul 11, 9:50–10:50
📍Hall D2, COEX
I’ll present our team’s recent work at #ICML2026#SD4H!
📄 Distilling Expert-level Planning for Real-world Diabetes Prescribing
We train medical LLMs to follow expert prescribing workflows using EMR context, insurance-aware search, and RL.
🗓️ Jul 11, 9:50–10:50
📍Hall D2, COEX
How can we train small agentic models that are highly capable of terminal use and coding?
Announcing OpenThoughts-Agent + OpenThinkerAgent-32B, the strongest Qwen-3 based open-data agentic model: 44.8% avg across 7 agentic benchmarks! (1/n)
Honored to join the @Gates_Cambridge community as a Gates Cambridge Scholar.
I’m deeply grateful for this opportunity, and I look forward to learning from and growing with such an inspiring cohort at @Cambridge_Uni .
Thank you to everyone who supported me along the way.
The new 2026 class of Gates Cambridge Scholars has been announced, comprising 68 Scholars from 25 different countries. Read the full story here - https://t.co/wZr7Fi8K6W
@Cambridge_Uni@GatesAlumni@gatesfoundation
🚨 #ICLR2026 🚨
Introducing ASGuard — a mechanistic safety alignment
LLMs can be jailbroken by simple tweaks → showing a gap in alignment.
So, we
• Identify vulnerable attention heads
• Apply activation scaling and preventative FT
• Decrease ASR without hurting utility
I’m delighted to share that I’ve accepted an offer of admission to the Ph.D. program in Computer Science at the @Cambridge_Uni, starting this fall.
I hope to continue this journey with humility and to work toward research that can contribute meaningfully to society!
How can we make a better TerminalBench agent?
Today, we are announcing the OpenThoughts-Agent project.
OpenThoughts-Agent v1 is the first TerminalBench agent trained on fully open curated SFT and RL environments.
OpenThinker-Agent-v1 is the strongest model of its size on TerminalBench, and sets a new bar on our newly released OpenThoughts-TB-Dev benchmark. (1/n)
📢Attention📢
Excited to share our new paper: Thinking Sparks!: Emergent Attention Heads in Reasoning Models During Post Training
With circuit, we uncover how post-training sparks the emergence of functionally specialized attention heads, supporting reasoning ability! 💡
#AI#RL
@nikhil07prakash@jeongminby98858 EAP-IG based circuit construction might make a difference, too. Meanwhile, the recent study (https://t.co/PeLyCtc71t) suggests a similar trend, showing SFT/Distillation introduce novel reasoning patterns in mathematical and coding benchmarks.
@nikhil07prakash@jeongminby98858 Thank you so much for your interest! We think both task domain and experimental setup could explain difference. We fine-tuned the model with solely mathematical reasoning datasets, and tracked internal computation triggered by <think> token, in contrast to entity-tracking task.
⏳Just 1 week to go until #ACL2025 in #Vienna!
Our paper, "Temporal Head" will be featured in the poster session:
🗓️ Tuesday, July 29 | 10:30 AM – 12:00 PM
📍 Hall 4/5
Feel free to reach out us! We'd love to hear your thoughts and share ours!✨
🔥We are on fire!🔥
Hot on the heels of many interests at #ICLR2025, our work, "Does Time Have Its Place? Temporal Heads: Where Language Models Recall Time-specific Information," has been accepted to ACL 2025 Main Conference!
Looking forward to see in Austria! 🇦🇹
#ACL2025
🎉 Countdown for #ICLR2025 in Singapore!
Our paper, "ChroKnowledge", has been showcased in the poster session!
🗓️ Friday, April 25 | 11:00 AM – 1:30 PM
📍 Hall 3 + Hall 2B #305
Come find us to talk about our new evaluation framework for LLMs handling time-variant knowledge!
🚀 Ready to Take the Next Step? 🚀
Here's the question: Does large language models (LLMs) specifically handle temporally facts within their internal structures? 🤔
We introduce Temporal Heads—attention heads within LLMs managing and processing time-sensitive information.