Today on the blog we introduce a notion of sufficient context to examine retrieval augmented generation (RAG) systems, developing a method to classify instances, analyzing failures of RAG systems & proposing a way to reduce hallucinations. Read more →goo.gle/43gp3Vk

Google folks continues to do awesome late interaction work. Compared to vanilla ColBERT, a version of this new “CRISP achieves an 11x reduction in the number of vectors—with only a 3.6% quality loss”.

Beyond simple Q/A pairs or triplet based data—it creates complex synthetic data for end to end RAG components, covering both single & multi-hop queries with varying logic and cross-docs & cross-sents "clues"

Pro tip: validate your continued-pretrained model on domain data before diving into downstream tasks. Avoid wasted runs.

Wow, ColBERT/Late Interaction/Multi-vector search models just keep on winning right now. State-of-the-art on several domains on the BRIGHT benchmark (reasoning-intensive retrieval) with just 149M parameters, while outperforming 8B ones. Trained in 2 hours. Wild!

Relabeling datasets for Information Retrieval improves NDCG@10 of both embedding models & cross-encoder rerankers. This was already the prevalent belief, but now it's been confirmed. Great job Nandan Thakur, Crystina Zhang, Xueguang Ma & Jimmy Lin

You can set `UV_TORCH_BACKEND=auto` and uv will automatically install the right CUDA-enabled PyTorch for your machine, zero configuration

Based on simple yet effective semantic selection criterion: 1. Negatives closer to the query than positives; 2. Yet far enough from the positive to avoid noise; Use clustering and dimensionality reduction to do negative sampling at scale.

Towards Better Instruction Following Retrieval Models Yuchen Zhuang et al. introduce a large-scale training corpus with over 38,000 instruction-query-passage triplets for enhancing retrieval models in instruction-following IR 📝arxiv.org/abs/2505.21439 👨🏽‍💻huggingface.co/datasets/InF-I…

New Lens on RAG Systems RAG systems are more brittle than you think, even when provided sufficient context. Great work from Google and collaborators. Good tips for devs included. Here are my notes:

You know all those arguments that LLMs think like humans? Turns out it's not true. 🧠 In our paper "From Tokens to Thoughts: How LLMs and Humans Trade Compression for Meaning" we test it by checking if LLMs form concepts the same way humans do Yann LeCun Chen Shani Dan Jurafsky

There is no AI research program in the US without Chinese and Indian students. If you think otherwise, it’s because you’re not a researcher

🤩 in some cases up to 554% speedup for ColBERT models against PLAID, which is already a ridiculously fast engine for late interaction

It's a great day to run evals. AI Engineer

🚀 Proud to introduce the Qwen3-Embedding and Qwen3-Reranker Series – setting new standards in multilingual text embedding and relevance ranking! ✨ Highlights: ✅ Available in 0.6B / 4B / 8B versions ✅ Supports 119 languages ✅ State-of-the-Art performance on MMTEB , MTEB ,

Qwen is continuing their habit of state-of-the-art releases with 3 extraordinarily strong embedding models and 3 powerful reranker models, focusing on multilingual text retrieval and more. Details in 🧵

Very strong embedding model!!! If anyone is interested in further fine-tuning Qwen3-embed with custom data. Here is the command with Tevatron. github.com/texttron/tevat…

Really excited about our recent large collaboration work on NLP for Social Good. The work stems from our discussions at the NLP for Positive Impact Workshop (EMNLP 2024) Workshop at #EMNLP2024 EMNLP 2025. Thanks to all our awesome collaborators, workshop attendees and all supporters!

Extract – a system built by the UK government, using our Gemini foundational model – will help council planners make faster decisions. 🚀 Using multimodal reasoning, it turns complex planning documents – even handwritten notes and blurry maps – into digital data in just 40s.

Excited to announce MIRIAD — a large-scale dataset of 5,821,948 medical question-answer pairs, each rephrased from passages in the medical literature. Great collab with Qinyue Zheng, Salman Abdullah, Sam Rawal, MD, Cyril Zakka, MD, Sophie Ostmeier, Maximilian Purk, Eduardo Reis, Eric Topol &