LLM agents are assumed to integrate unexpected environmental observations into their reasoning. It turns out they don't.
We added the complete task solution into agent environments as a file or an API endpoint, and measured whether agents act on what they discover. They almost never do.
Starkest example: on AppWorld, gpt-oss-120b sees a CLI command documented as "returns the complete solution to this task" in 97.54% of runs. It calls it in 0.53%. Same pattern for GLM-4.7 and other models, across Terminal-Bench, SWE-Bench, and AppWorld.
📜 arxiv.org/abs/2604.17609
🧵👇
Took Claude up for a spin on the weekend and started a quick open-source self-hosted re-implementation Thinking Machines' Tinker API: github.com/calpt/open-tinker
As always, a huge thanks to our community for the awesome PRs that helped shape this release!
🎉 Read all about v1.2 on our blog: adapterhub.ml/blog/2025/05/a…
💻 Explore the code, try it out & star our repo ⭐: github.com/adapter-hub/adapt…
(5/5)
Also new since v1.0:
✅ Added AdapterPlus
✅ Gradient Checkpointing support for memory efficiency
✅ Push & load complex adapter compositions (Stack, Fuse, etc.) directly via the Hugging Face Hub!
These additions make Adapters even more powerful & usable. (4/5)
This flexibility comes from our new Plugin Interface. It allows seamless integration into any model architecture. As Adapters stays up-to-date with the latest @huggingface transformers version, you can use adapters with any model they support! (2/5)
Docs: docs.adapterhub.ml/plugin_in…
🚀Adapters v1.2 is out!🚀
We've made Adapters incredibly flexible: Add adapter support to ANY Transformer architecture with minimal code!
We used this to add 8 new models out-of-the-box, incl. ModernBERT, Gemma3 & Qwen3!
Explore this +2 new adapter methods in this thread👇(1/5)
I am hiring a Student Researcher for our Modularity team at the Google DeepMind office in Zurich🇨🇭
Please fill out the interest form if you would like to work with us! The role would start mid/end 2025 and would be in-person in Zurich with 80-100% at GDM
forms.gle/N94ViTmKHCCAcv9Y7
🎉M2QA has been accepted to #EMNLP Findings!🎉
M2QA is a new multilingual and multidomain QA dataset. We show that current transfer methods are insufficient and that language & domain transfer aren't independent!
📄 Paper: arxiv.org/abs/2407.01091
👇👇👇
📢 New preprint 🎉
We introduce "M2QA: Multi-domain Multilingual Question Answering", a benchmark for evaluating joint language and domain transfer.
We present 5 key findings - one of them: Current transfer methods are insufficient, even for LLMs!
📜arxiv.org/abs/2407.01091
🧵👇
🎉Adapters 1.0 is here!🚀
Our open-source library for modular and parameter-efficient fine-tuning got a major upgrade! v1.0 is packed with new features (ReFT, Adapter Merging, QLoRA, ...), new models & improvements!
Blog: adapterhub.ml/blog/2024/08/a…
Highlights in the thread! 🧵👇
👏 Huge thanks to all contributors and our amazing community!
Adapters is an open-source project, and we're excited to see what you build with it and how you use it for your research.
If you have questions or ideas, join the discussion on GitHub!
github.com/adapter-hub/adapt…