LLM safety filters like Purple Llama promise responsible and safe deployment of AI, but how effective are they? In our new blog post, we argue that LLM-based filters are clearly flawed, and that they set up a dangerously circular safety narrative. Blog: lve-project.org/blog/how-effec…

A really nice attempt to build a community vulnerability database on LLM vulnerabilities. It could be useful to standardize format (probably interoperate with #OSV) and add versioning information (for regression/fixed if available). I really like how they thought about

When I talk with people about constrained decoding, I'm always asked: "Can it be applied to blackbox LLMs like GPT-4?" My response has been a bit pessimistic due to the limited logit access.🤔 🚀However, we're excited to announce Sketch-Guide Constrained Decoding (SGCD), a

I'm excited to represent LVE Project at the webinar. Hope to see you there!

We have started a new initiative called the LMQL developer survey. With this, we want learn more from everyone around here and the bigger LLM community. We are looking for a broader feedback signal on how and what people are using LMQL for or may be planning to. The goal of

Constraining LLM outputs (e.g. regex, JSON schema...) is key to develop robust LLM use cases. In this context, I'm happy to introduce a novel, fast, high-fidelity constrained decoding method. It applies to regex constraints and builds on the approach implemented in Outlines. 🧶

Just arrived at #iclr in Vienna. Feel free to reach out to talk about LLM constrained generation, LLM programming and agentic stuff. Looking forward to see all the cool papers and workshops. I’ll be co-presenting the paper below, on controlled text generation this afternoon.

Constraining LLMs ✅ Constraining Agents 🔄

We are delighted to announce Invariant Labs (invariantlabs.ai)! Our mission is to make AI agents secure and reliable! It is founded by Martin Vechev, Mislav Balunović, Luca Beurer-Kellner, Marc Fischer, Florian Tramèr and builds on years of experience in industry and academia.

On my way to #ICML2024 now. Two papers on LMQL (Language Model Query Language) in the main conference one from Invariant Labs in the workshops 🎉 🧵👇

We are presenting two papers at ICML and also a poster at the NextGenAISafety workshop on Fri. The papers focus on LMQL (Language Model Query Language) constrained decoding and Invariant Labs agents. Join us for one of the poster sessions (see below), or DM me if you want to chat more directly.

🎄 Ho ho ho 🎅 An early holiday celebration from Invariant Labs! We are launching (and open-sourcing!) Explorer, a visual and intuitive way of inspecting AI agent traces, and Testing, a new library for debuggable unit testing for AI agents.

We are releasing two open source tools that we've already been using for quite a while internally Invariant Labs. I hope they are useful and speed up all the folks building agents these days. Explorer is for trace viewing and Testing for agent unit testing. Happy Holidays!

Releasing Computer Use + Playwright: Let a Claude agent control a browser on your machine to achieve tasks. We've been working with computer use models for a while now and just released a helper library for it. (yes it can dismiss cookie banners) Repo: github.com/invariantlabs-…

🏷️ Are you working on error labeling and/or post training with LLM agents? Our OSS tool 'Explorer' allows you to easily collaborate on and annotate your captured agent traces within your team or community. Check it out 👇