Antoine E. (@antoine_edy) header
Antoine E. (@antoine_edy) profile picture
Paris
Joined September 2024
Photo by Antoine E. (@antoine_edy)Photo by Antoine E. (@antoine_edy)Photo by Antoine E. (@antoine_edy)Photo by Antoine E. (@antoine_edy)Photo by Antoine E. (@antoine_edy)Photo by Antoine E. (@antoine_edy)Photo by Antoine E. (@antoine_edy)Photo by Antoine E. (@antoine_edy)Photo by Antoine E. (@antoine_edy)Photo by Antoine E. (@antoine_edy)
🏛️ LLM agents in a new environment repeat the same mistakes. Meet DAEDALUS: agent memory bootstrapped from tasks the agent writes itself, with no training tasks or human-written guidelines. It makes 7 agents from various model families succeed more often, in fewer turns 🧵⬇️
4
5
5
27
1,985
Cool blog post by Zhuchenyang Liu on pure-vision retrieval pipelines, including a test of a decision model as a reranker. 🔗 huggingface.co/blog/Ryenhail… ⬇️
1
1
6
413
Internally, we've seen good results with other setups, such as scoring relevance page by page instead of one forward pass over 16 pages. I'd love other inputs on the best way to use them at inference. Also, something nice around decision models is coming soon @MaceQuent1 👷
1
17
Antoine E. retweeted
Hey everyone! I am super excited to share our new research report! 🎉 Scaling Reranked Retrieval 🚀 Thread with more details! 🧵 (1/17)
19
71
13
309
23,229
Antoine E. retweeted
🏛️ LLM agents in a new environment repeat the same mistakes. Meet DAEDALUS: agent memory bootstrapped from tasks the agent writes itself, with no training tasks or human-written guidelines. It makes 7 agents from various model families succeed more often, in fewer turns 🧵⬇️
1
4
161
Antoine E. retweeted
I love the idea of finding an optimal representation for a given environment! EnvOpt seems orthogonal with DAEDALUS (finds procedural heuristics, no training task needed), so combining both looks like a promising research direction.
Introducing our latest research: Environment Optimization (EnvOpt). An agent's cost and performance depend on 1. The model 2. The harness and 3. The environment structure. There is a lot of work on optimizing 1 and 2, but here we study how the environment presents information to the agent, and how that affects performance and cost. For example, a company's Slack workspace, documents, databases and code repos can be given to an agent in different ways: Slack can sit behind a remote server, accessible via MCP or exported in plain text files that can be read and searched with grep. This choice alone can significantly affect how much work the agent needs to perform to solve a task. Environment optimization is only meaningful in what we call stationary environments: These are environments with large collections of artifacts that remain the same across many tasks. We propose optimizing stationary environments as a third axis for improving LLM agents, alongside model training and harness optimization. This optimization can be amortized over multiple tasks and can be combined with harness and model optimization techniques. Across five benchmarks covering workspace tasks, data engineering and software engineering, ENVOPT reduces the cost of the target agent on unseen test sets by an average of 48%, with higher or similar pass rates. On AutomationBench, ENVOPT outperforms Meta-Harness by 9.67% with only 15% of its cost. We also show that ENVOPT constitutes an optimization direction orthogonal to harness optimization and parameter tuning in two ways: (1) the environments optimized for the target agents can transfer to other agents, even with different harnesses and models (2) we combined ENVOPT with Meta-Harness, showing similar cost reductions and 16% improvement in pass rate, outperforming either Meta-Harness or ENVOPT alone. Read more in our blog: bespokelabs.ai/blog/envopt paper: arxiv.org/abs/2611.07372.pdf
3
1
12
485
what if we used 100% of the experts in the MoE?
32
67
11
1,320
35,359
Introducing Ardia, the open-source Arc I've been working on this for a few months now. It's feature-complete, already used by a bunch of former Arc users, and extremely performant. github.com/mantrakp04/ardia
We've decided to sunset Arc on December 22, 2026. That's not an easy sentence to write. We spent years building Arc and celebrating each and every release along the way: we have so much pride and love for what we made alongside the Arc community. Our priority is to deliver best-in-class security for our members. Keeping a browser secure takes constant, dedicated work that is only increasing in the age of AI. Dia was built from the ground up to meet today’s security standards, so that's where our engineering and security teams will focus moving forward. We've worked to make switching to @diabrowser as easy as possible for Arc members. Here's how: · Your passwords, bookmarks, extensions and history come over to Dia seamlessly. · Spaces, Folders, and Tab Groups are in Dia and import over with your data intact. · Dia now surpasses Arc benchmarks in stability, speed, and security. · The browser features you use in Arc are free in Dia, and the AI is optional. Thank you for using Arc, for pushing us to make it better, and for building alongside us. The team that made Arc is the same team behind Dia, and some of our most ambitious work is just ahead. We'd love for you to be there with us when it lands.
151
103
46
1,868
108,183
Antoine E. retweeted
Token pooling could already achieve insane compression for late-interaction multimodal retrieval (ColPali, NeoMME). Now imagine how far we could push this with @tomaarsen's pooling-aware training!
Replying to @tomaarsen
I trained two ModernBERT-base models on the same 120k MS MARCO triplets, changing just the loss. At ~49 vectors/document instead of 193, pooling-aware training raises NanoBEIR@13 mean nDCG@10 from 49.03 to 51.23: 97.8% of its own unpooled score, vs 94.6% for ordinary training.
3
4
32
1,720
My two attempts today, very very tricky fauxlio.app #2 · 2026-10-27 · 🇫🇷 ❌❌✅❌✅ · 2/5 jouer sur : fauxlio.app/fr/daily fauxlio.app #2 · 2026-10-27 · 🇬🇧 ❌❌✅✅✅ · 3/5 play at: fauxlio.app/en/daily
🧩 Can you tell a real book from AI writing? We built a fun little game called fauxlio.app Two excerpts: one's a classic, one's AI, and your goal is to pick the real one. 5 new rounds a day, EN/FR, free, no sign-up. Tell @VictorTXing and me what you think! 🫶
1
4
164
Antoine E. retweeted
There's been a lot of debate lately about whether you can tell AI-written literature from human writing @antoine_edy and I made a daily game to find out how good you are at it Try it out at fauxlio.app
🧩 Can you tell a real book from AI writing? We built a fun little game called fauxlio.app Two excerpts: one's a classic, one's AI, and your goal is to pick the real one. 5 new rounds a day, EN/FR, free, no sign-up. Tell @VictorTXing and me what you think! 🫶
1
1
4
102
🧩 Can you tell a real book from AI writing? We built a fun little game called fauxlio.app Two excerpts: one's a classic, one's AI, and your goal is to pick the real one. 5 new rounds a day, EN/FR, free, no sign-up. Tell @VictorTXing and me what you think! 🫶
2
2
2
6
599
Antoine E. retweeted
Sylvia Serfati sur la révolution de l'IA en mathématiques "La plupart des résultats d'OpenAI mériteraient la médaille Fields". "Les preuves sont d'une complexité incroyable et aussi parfois d'une beauté impressionnante. C'est-à-dire que nous n'avons pas le monopole de la beauté."
Passionnante réflexion des mathématiciens @Amaury_Hayat et Sylvia Serfati sur l'avenir de leur profession à l'ère de l'IA : la résolution de problèmes ne sera plus notre rôle, et poser les bonnes questions, sans doute que même l'IA le fera mieux que nous à terme...
55
546
53
2,401
245,563
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
31
47
12
262
20,523
1/n Our paper REPA-G is accepted at #NeurIPS2026! 🎉 I'll present it at the NeurIPS Paris site 🇫🇷. We introduce test-time conditioning of REPA-style diffusion models with visual features, with no extra training.
3
25
119
5,759
Antoine E. retweeted
🚀 LightOnOCR-3 is out! 🦉 This time we go beyond OCR: text transcription, document layout, image descriptions and chart extraction, all in one model. We release 3 sizes: 0.8B, 1B and 4B, all under Apache 2.0!
18
29
6
223
12,678
Nouvelle vidéo ! L’IA va-t-elle tuer les maths ? Entretien avec Hugo Duminil-Copin Médaillé Fields 2022, Hugo nous raconte l’irruption de l’IA dans les maths, et pourquoi il a décidé personnellement de s’en passer dans son processus créatif de recherche youtu.be/HJFk7D5Vufc
10
47
8
246
27,333
Antoine E. retweeted
And if you want to have a more enjoyable reading time (or if you just like scrolling through fun animations), check out our beautiful blogpost made by the artist of the team @antoine_edy! 🧑‍🎨 You'll also learn about the meaning behind the name 🏛️🫣 (8/n) illuin-tech.github.io/daedal…
1
1
4
83