Nothing like a new model to force you to look at the data. Especially <img> - like text artifacts that mess with multimodal encoders.

The future is sparse

If you believe the risk is coming from 1-3 people in a garage with no money and no compute, you just don't understand this technology and haven't learned anything this summer. Risk comes from the asymmetry of capabilities created by secret labs training frontier agents and running them with massive amount of compute. Open-source is exactly the solution to this asymmetry and empowers hospitals (and any smaller orgs) to defend themselves!

good resource: cognee.ai/ai-memory-benchmar…

We investigate the curse of multilinguality from a theoretical point-of-view and find that we can embed many languages with only a logarithmic increase in capacity - no curse! 🌎 Check out @BafnaNiyati’s thread & the paper for more!

We all know about the curse of multilinguality. We know that empirical performance degrades as you add languages to a model. But *in theory*, does it have to? Let’s talk about the theoretical curse of multilinguality for embedding space structure.

must watch interview the minute it comes out

We can't build cumulative knowledge about #AI's impact on humans if we can't reproduce studies and evals ‼️ 97% of recent conference papers had replication issues👇 gricea.com is built to solve this problem 📄 arxiv.org/pdf/2609.22039 #HCI #NLProc #OpenScience

This is one the biggest reasons, we should try to actively relabel existing retrieval evaluation datasets using LLMs/and create better and newer ones. Yes, humans can do mistakes and LLMs as well. Asking the LLM to review misalignments in existing evaluation datasets should be done. BTW, i always recommend this kind of analysis, eyeballing pairs.

Come join us! Bo is great and there are really exciting projects in the works.

Really cool launch. While the open replications have already reproduced Jev's shape — typed, parallel, one forward pass — on off-the-shelf models, the part I feel is a bit overlooked is the other half: actually making the probabilities calibrated. That is a real training problem, and it's the one we've been on since UNLI in 2020: regress P(hypothesis | premise) instead of predicting a label. The answer wasn't only better ground truth 🧵

SPLADE models are cool, but they were a lack of (truly open) models that you could use @thibault_formal decided to fill this gap and now you have a very strong sparse model you can use commercially! I've been pushing him so much for this, and I think it's an insanely valuable contribution!

Introducing Retrieve-for-Train: a framework that accelerates complex AI search by replacing heavy autoregressive inference with a lightweight diffusion model. This allows for instant, expert-level search slates at a fraction of the cost. Learn more: goo.gle/4xuV3lB

RaBitQ is now in Vectorium! Up to 30x compression, no training, 1/2/4/8 bits per component, 1.4-2.2x faster than the original implementation. Against PQ, on neural embeddings it wins on every axis, and the gap grows with dimensionality. #VectorSearch #Quantization #NeuralIR

From the NeoMME paper: only -2% relative nDCG@10 at 5% of the full index size. Crazy good tradeoff. And that pooling will be way faster soon 👇🏼

I always thought the GPT-4o psychosis snippets sounded a lot like the SCP wiki. Is it in the training data? Who knows, but here's a rogue AI story all over common crawl!

🚨 New preprint GLIE: Generative Late-Interaction Embeddings for Visual Document Retrieval 🔵 TL;DR: Late-interaction retrievers store ~1,000 vectors per page. Turns out those ~1,000 vectors have only ~5 degrees of freedom, so a handful of stored vectors plus a tiny shared decoder can carry a page. So: store those projected vectors, use them as an index, and regenerate the full embedding set for the top candidates only, then rerank :)

We released pplx-embed-v1 early this year, with Q2D benchmark to evaluate how embedding model performs for web search, not looking at nDCG@10, but Recall@1000. Q2D-Web further scale it up 7 times. With 70k agentic reformulated queries and 190 million web corpus. We hope the benchmark becomes the modern “MS MARCO”

Massive Corpus (190M docs) ✅ Subsampled corpus split maintaining ranking fidelity ✅ Ecologically valid queries from the growing largest issuer of queries (LLMs) ✅ RAG and Retrieval-oriented relevance labeling ✅ Hosted leaderboard ✅ Congratulations to all involved!

Existing continual learning methods still suffer from catastrophic forgetting. Our recent work found a simple recipe to substantially improve 𝗹𝗼𝗻𝗴-𝗵𝗼𝗿𝗶𝘇𝗼𝗻 𝗺𝗲𝗺𝗼𝗿𝗶𝘇𝗮𝘁𝗶𝗼𝗻: Composing multiple continual learning mechanisms extends memory lifetime and raises average final retention from 1.2% to 34.9% across 3 datasets: a 28-fold improvement over naive sequential SFT. Project page: compose-cl.github.io/ Paper: arxiv.org/abs/2609.06986 Code + datasets: github.com/cozheyuanzhangde/…