Can AI agents organize scientific investigations that lead to well-supported conclusions? To evaluate this capability, we introduce PhysClosureBench, spanning strong-field, attosecond, and neighboring physics areas.
The answer is Not yet, with only 32%.
https://t.co/7cFlMRzGvy
Reward Hacking also happens in OPD training?
Our new work analyzes the OPD process from the perspective of reinforcement learning to dive deeper!
https://t.co/lKJwKHCLmM
We were very excited by @inco_ai's Dflash2 release and started building with it immediately.
A few tweaks to the training recipe further improved acceptance length by another 33%. Full write-up below. https://t.co/i4PmWkNsWp
SimpleOPD is now available on arXiv! We’ve also updated the paper with extensive new experimental results demonstrating the effectiveness of our method.
📄 https://t.co/F7r8R8nSMR
@krasul Thank you for your interest and recognition! While the work you mentioned shares some similarities with ours, we expect to make long-context OPD more stable and effective, particularly when transferring capabilities across different model architectures.
We’re excited to share our latest findings on Long-Context Reasoning OPD!
We introduce SimpleOPD, a tokenizer-agnostic approach for transferring reasoning capabilities from SU-01-160K to short-context student models.
https://t.co/5LRIpVhHBq
We’re releasing a 30B-A3B reasoning model that reaches gold-medal level across both physics and math Olympiad evaluations: IPhO directly, and IMO/USAMO with test-time self-verification and refinement.
A simple, unified scaling recipe for proof search.
https://t.co/yc2ZlLVbD2
Thrilled to announce our paper has been accepted for an Oral Presentation at ACL 2025 (8% of accepted papers)! 🎉 https://t.co/MhdOtbEfy3
We dive deep into why current #LLM editing methods often fail robustness tests and propose a solution. #ACL2025#LLMs#ModelEditing
1/X
This work is accepted in NAACL 2024. We utilize the reasoning capabilities of LLM to enhance the efficiency and interpretability of smaller models in active learning. 😀
https://t.co/a8Egfxd2SS
📝 New preprint at @NlpWestlake
How stubborn are LLMs with user feedback during the interaction?
We design refuting instructions to evaluate 7 open-sourced and commercial LLMs, covering 3 different tasks.
https://t.co/rpUjPIkQOz
Our findings: