PhD Student @LightOnIO | Post-Training
Paris, France
Joined September 2024
- Tweets109
- Following268
- Followers80
- Likes638
Pinned Tweet
🚀 Introducing ColGrep-Agent-2B!
An open 2B model, now integrated into NextPlaid via ColGrep, that runs locally to make coding agents more efficient.
With Claude Code:
→ 34% fewer API tokens
→ 40% lower API costs
All by delegating code search to a tiny agent.
Here's how 🧵
Maxence Lasbordes retweeted
🚨 Tested @LightOnIO's ColGrep-Agent (2B local code-search model) on our own repo across 3 coding agents.
Same 22 questions, same model, OFF vs ON, 3 runs each:
• pi: 43% fewer tokens
• OpenCode: 37%
• Claude Code: 21%
ACCURACY UNCHANGED.
The model needs ~3 GB of VRAM. 🧵
Maxence Lasbordes retweeted
We surfed the dense and late-interaction waves… now it's time to SlideOn sparse retrieval 🐕
🚀 Meet SlideOn and mSlideOn, fully open SOTA sparse retrievers at 174M and 339M parameters, trained by @LightOnIO R&D
Models:
huggingface.co/collections/l…
Blog: huggingface.co/blog/lightona…
Maxence Lasbordes retweeted
🚨Our new paper, TrustMI, is out! Watch Théo's video 👇
We study how LLMs represent trust in their users—and how to steer it.
TL;DR: more trust often meant less safety; lower trust reduced harm in our tests.
🧵1/X
Did you ever wonder if your AI model trusted you ?
We just release a new paper TrustMI where we finds a linear trust direction in 6 LLMs.
📄 arxiv.org/abs/2610.06064
🤗 huggingface.co/TrustMI
Maxence Lasbordes retweeted
Did you ever wonder if your AI model trusted you ?
We just release a new paper TrustMI where we finds a linear trust direction in 6 LLMs.
📄 arxiv.org/abs/2610.06064
🤗 huggingface.co/TrustMI
Did you ever wonder if your AI model trusted you ?
Introducing TrustMI, our latest paper! 🚀
We learn activation-steering matrices that causally shift assistants’ trust in users and study the impact on agent safety.
Thread below 🧵
We also found that trust steering directions are shared across languages!
On Qwen3.5-9B, steering works in English, Chinese, Spanish & French.
Huge thanks to my co-authors @Blyzi1, @froger_romain, and @zehavoc for doing such an amazing job, especially @Blyzi1 and @froger_romain for leading this work!!
📄 Paper: arxiv.org/abs/2610.06064
💻 GitHub: github.com/Blyzi/trustmi
🤗 Hugging Face: huggingface.co/TrustMI/
Maxence Lasbordes retweeted
Should you waste tokens having your coding agent use grep, or hand the search to a tiny local model?
We release ColGrep-Agent-2B! A code-search agent that runs locally, searches with ColGrep, reads files, & returns locations + code to your main agent
-> 34% fewer API tokens
Maxence Lasbordes retweeted
Introducing colgrep-agent-2B ☺️
ColGREP is a multi-vector search tool built for coding agents. It’s a hybrid grep that combines grep features with multi-vector retrieval. Runs 100% locally
Open-sourcing a 2B-parameter ColGREP expert model to serve as a model-in-the-middle 🧵
📄 In parallel, we also released a research paper!
It explores how small, specialized agents can leverage semantic search for efficient, multilingual code localization.
Check out the paper for more details! ↓
arxiv.org/abs/2610.05099
Huge thanks to my co-authors @raphaelsrty, @Aarush1003, and @AmelieTabatta for their incredible work! 🚀
🤗 Model:
huggingface.co/lightonai/col…
📝 Blog post:
huggingface.co/blog/lightona…
💻 GitHub:
github.com/lightonai/next-pl…
📄 Paper:
arxiv.org/abs/2610.05099
