🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Care about LLM evaluation? 🤖 🤔 We bring you🕊️ DOVE a massive (250M!) collection of LLMs outputs On different prompts, domains, tokens, models... Join our community effort to expand it with YOUR model predictions & become a co-author!

In this paper, the authors show that an LLM can learn to use a search engine using reinforcement learning, which is especially cool when, to give the right answer, the model needs to run multiple searches, one based on the result of another: arxiv.org/pdf/2503.09516 Their code

A great AI application starts with choosing the right embedding type. Here are 6 embedding types and when you should use them: • Sparse embeddings: weaviate.io/developers/wea… • Dense embeddings: weaviate.io/developers/wea… • Quantized embeddings: weaviate.io/developers/wea… • Binary

A solution for better search results: Late interaction preserves contextual nuances that pooling (as in most dense vector retrieval models) destroys. How? There are three different ways retrieval models handle the "interaction" between your query and potential documents: 1️⃣

// Tracing LLM Outputs Back to Trillions of Training Tokens // Presents OLMOTRACE, the first system that can trace LLM outputs verbatim back to their entire multi-trillion-token training sets in real time!

Someone just reminded me of this lecture I gave in 2009 that described the evolution of Google Search from 1999 to 2009. People who are interested in how our search systems work might find this interesting. It touches on disk-based serving systems, in-memory indices,

A gentle reminder: two weeks left to submit to SCAI‘25. Don’t miss the opportunity to present your work in Montreal this summer! scai.info #IJCAI2025

i created a gist with some non-default LLM courses: gist.github.com/yoavg/95bbc576…

Training on wrong answers outpaces training on correct ones. 10 times more learning emerges from plausible errors than from truths. Large language models refine their accuracy slowly when they learn only from correct examples. This paper introduces Likra, which trains one

LMs often output answers that sound right but aren’t supported by input context. This is intrinsic hallucination: the generation of plausible, but unsupported content. We propose Precise Information Control (PIC): a task requiring LMs to ground only on given verifiable claims.

New Anthropic research: We elicit capabilities from pretrained models using no external supervision, often competitive or better than using human supervision. Using this approach, we are able to train a Claude 3.5-based assistant that beats its human-supervised counterpart.

1/24 I’m thrilled to share what my co-founder Gilad Mishne and I’ve been cooking up over the past year. Check out Yupp – a fun and easy way for anyone to discover, compare and get the best answers across the latest AIs, all for free! Yes, even the most powerful pro models.

Feels good to be back coding! Just picked a fun one from my “someday” side project list and finally added a KV cache to the LLMs From Scratch repo: github.com/rasbt/LLMs-fro…

CLIP embeddings for image search damian github.com/vdlm/meetups

AI just learned to fine-tune itself between questions. MIT introduces SEAL, a framework enabling LLMs to self-edit and update their weights via reinforcement learning, all by itself. LLMs consume whatever data they are given, so they stay frozen after pretraining. SEAL teaches

Web data, the “fossil fuel of AI”, is being exhausted. What’s next?🤔 We propose Recycling the Web to break the data wall of pretraining via grounded synthetic data. It is more effective than standard data filtering methods, even with multi-epoch repeats! arxiv.org/abs/2506.04689

Really thought-provoking new paper on representation learning and the notion of 'semantic compression' by Chen Shani Dan Jurafsky Yann LeCun Ravid Shwartz Ziv

AI tool that splits text, extracts entities, builds knowledge graphs from documents

The detailed program for #sigir2025 is available at: sigir2025.dei.unipd.it/detailed-progr…