IdePHICS lab @idephics

@LightOnIO @tomaarsen wrote a super insightful blog covering it all: MaxSim, search, indexing, visual retrieval, interpretability, token pooling, faster inference, etc Check it out -> huggingface.co/blog/multi-vec… Congrats to @raphaelsrty & @antoine_chaffin for this milestone!

RAG quality is decided before the prompt reaches the LLM. LightOn’s open-source stack strengthens every layer between raw documents and grounded answers: 📄 LightOnOCR-2 for parsing at scale 🔍 mLateOn, a state-of-the-art retriever 🎯 LightOn-rerank, a state-of-the-art reranker Better parsing. Sharper retrieval. Cleaner context. Use each component independently, or access the full production-ready pipeline through LightOn Console. The bricks are already in your dependency tree. Start building: console.lighton.ai/signup

Sentence Transformers v6.0 is here with a very special addition: a 4th model type, MultiVectorEncoder, for ColBERT-style late interaction! 🎉 IMO, Sentence Transformers played a huge role in making dense retrievers more popular: easier to use, fine-tune & eval.

x.com/i/article/2086…

Yesterday, an interesting new exact-recovery theorem, and a remarkable AI-assisted proof appeared here: x.com/DimitrisPapail… While the result was interesting, the AMP algorithm, rigorous state evolution/replica formula, and the 2 log N prediction have a substantial prior history. I tried to separate what was known, what is new, and what AI may really have changed :

Krzakala Florent @KrzakalaF

a week ago

x.com/i/article/2086…

3 12 97 31K 70

@NoahChrein They are:

Surya Ganguli @SuryaGanguli

a month ago

Check out all the amazing work from our @SimonsFdn Collaboration on the Physics of Learning and Neural Computation (physicsoflearning.org) presented at the main meeting of @icmlconf #ICML2026 Tuesday Efficient Learning of Compositional Targets with Hierarchical Spectral

0 32 138 27K 76

Because the relevant physics is that of disordered systems, which was and is woefully understudied in US/Can. Many Europeans coming out of Paris/Rome schools are expert and have indeed moved to this (see Krzakala/Zdeberova, Biroli, Mezard and descendants), but the few US folks ..

∞-modal Noah @NoahChrein

2 weeks ago

I haven't been able to crack this psychological nut, please help me to understand: Why aren't math / physics theorists, on the whole, DYING to know what is going on inside of these large scale cognitive systems, e.g. ai interpretability from a pure theory perspective? Why is it

78 13 249 27K 82

ICTP just announced the Dirac medal winners Derrida, Dhar, Mézard, Sompolinsky: all for contributions to the physics of disordered systems, now relevant to understand AI. Excellent writeup with the citations ictp.it/news/2026/8/ic…

Eric De Giuli @EricDeGiuli

2 weeks ago

Because the relevant physics is that of disordered systems, which was and is woefully understudied in US/Can. Many Europeans coming out of Paris/Rome schools are expert and have indeed moved to this (see Krzakala/Zdeberova, Biroli, Mezard and descendants), but the few US folks ..

1 7 59 105K 50

@LightOnIO also we took the BEIR leaderboard hostage rn

Right time to announce the release of the 1.6.0 release of FastPlaid @paulomouraj made FastPlaid more memory efficient which was very needed for advanced agentic search evaluation with multi-vector model @chaochunh accelerated dramatically the number of QPS for more recent torch version Beautiful version

Paulo Moura @paulomouraj

2 weeks ago

Hey, having trouble to evaluate late-interaction models on long context MTEB tasks, like MLDR? Check out the new eval scripts I added to github.com/lightonai/mden…. You can now easily reproduce all evals we ran for mLateOn and mDenseOn :)

1 2 29 3K 5

These look like *very* strong multilingual retrieval models. The best part: multilingual training with late interaction generalizes to languages unseen during contrastive training. At this point it's not even surprising anymore :-) Congrats on the release everyone @LightOnIO !

Antoine Chaffin @antoine_chaffin

3 weeks ago

Blog: huggingface.co/blog/lightonai… Collection: huggingface.co/collections/li… DenseOn & LateOn pushed the frontier of English retrieval in the open We extended this recipe to multilinguality using translate-train and pushed long-context and code retrieval capabilities

1 1 26 2K 2

The generalization capabilities of late interaction models (first multimodal with ColPali, and now multilingual) never cease to amaze me!

Antoine Chaffin @antoine_chaffin

3 weeks ago

The most important result is this one: Translate-train gives good performance on target languages Late interaction makes it viable for unseen languages and scripts, essentially lifting translate-train's biggest limitation!

1 2 13 4K 0

Maybe it looks like "yet another SOTA release from LightOn" (which is already wild), but this one is personal When I joined LightOn, I looked at the top models and thought: the amount of work to get there is unreachable But it's only truly impossible if you don't try, so we set up a plan and methodically followed it, piece by piece We built ModernBERT & Ettin, setting a new standard in the encoder world We built PyLate, so everyone could experiment with late interaction and push the frontier of what was doable We used it to create strong models, such as LateOn-Code (whose data proved very helpful here!) and ColBERT-Zero, the first fully pre-trained ColBERT model at large scale We built DenseOn and LateOn, showing that carefully gathered and curated data can create frontier models even outside of closed labs And now: mDenseOn and mLateOn, compiling everything we've built and extending it to multilinguality, so non-English speakers are not excluded anymore Everything is public. Every model, every dataset, every piece of code, for anyone to go from random weights to production-ready models When I look back at the person who thought this was too hard, I could not have dreamed of these resources. I hope they help someone out there who's thinking the same thing right now to give it a shot and come up with their own awesome models Nothing but pride and gratitude towards everyone who contributed to making this possible. Open source is incredible, and I feel so lucky to be part of it!

Antoine Chaffin @antoine_chaffin

3 weeks ago

"Multilingual when?" Glad to finally answer today! Introducing mDenseOn & mLateOn, the next generation of fully open models for multilingual, long-context, and code retrieval Same openness, same frontier performance, but now for everything

12 37 119 14K 41

another great model from LightOn team:

Antoine Chaffin @antoine_chaffin

3 weeks ago

"Multilingual when?" Glad to finally answer today! Introducing mDenseOn & mLateOn, the next generation of fully open models for multilingual, long-context, and code retrieval Same openness, same frontier performance, but now for everything

12 37 119 14K 41

Tenet (2020) Christopher Nolan

New State-of-the-art Retrieval Model / Новая передовая поисковая модель / 全新最先进检索模型 / 최첨단 검색 모델 / 最先端の新しい検索モデル / نموذج استرجاع جديد متطور One retriever family for the languages your data actually speaks. Multilingual retrieval, built for production. A few months ago, LightOn open-sourced DenseOn and LateOn, its 149M-parameter retrievers. A new BEIR state of the art for their size class, Apache 2.0. Today, the multilingual pair: mDenseOn and mLateOn. 307M parameters. SOTA on BEIR, MLDR, MIRACL and MTEB Code. Nine languages, one quality bar: English, French, German, Spanish, Italian, Portuguese, Swedish, Norwegian, Arabic. Late interaction compounds. mLateOn stays competitive retrieving in languages and scripts it never saw in training: Cyrillic, Japanese, Korean, Chinese. It breaks the ceiling of translate-train pipelines. Everything open: models, datasets, training code. 👨‍🍳Kudos to @raphaelsrty @antoine_chaffin @paulomouraj @AmelieTabatta 🔗Read the full story: lighton.ai/lighton-blogs/… ⚡️Try it now: console.lighton.ai/signup

Yesss! I've been waiting for this release for some time. Multilingual dense and multi-vector embedding models for retrieval. They look extremely strong, with all data public! - lightonai/mDenseOn - lightonai/mLateOn Great work by @antoine_chaffin & @LightOnIO, as usual.

Antoine Chaffin @antoine_chaffin

3 weeks ago

"Multilingual when?" Glad to finally answer today! Introducing mDenseOn & mLateOn, the next generation of fully open models for multilingual, long-context, and code retrieval Same openness, same frontier performance, but now for everything

12 37 119 14K 41

Today, we are releasing the largest piece of work we have done on retrieval at @LightOnIO A training corpus of 2.8B curated query and document pairs across 9 languages 🇫🇷🇮🇹🇪🇸🇳🇴🇸🇪🇵🇹🇩🇪🏴󠁧󠁢󠁥󠁮󠁧󠁿🇺🇸🌍 العربية We are also releasing two SOTA models trained on it: mLateOn and mDenseOn 🧵

DenseOn & LateOn validated our rebuilding of an open English retrieval recipe: 665M curated pairs from 34 public sources. But we were not done! We translated it into 8 languages & gathered extra fine-tuning data, resulting in 2.8B pre-training pairs + 16.3M fine-tuning samples.

Drumroll please! After DenseOn and LateOn, we release mDenseOn & mLateOn: two fully open 307M retrievers for multilingual, long-context AND code search! Same backbone, same data, using translate-train. But very different behaviour once we leave the training languages 👀