Retriever wins, then reasoner adds Summaries beat full pages for retrieval (p@5: 0.76 vs 0.68). With k=5 retrieved docs, condition accuracy caps at 0.76 upstream; within that cap, Qwen2.5‑32B jumps from 0.38 to 0.54 with RAG, and to 0.56 after reasoning distillation. Frontier baselines with RAG land around 0.56–0.57.

Small models, big gains. Distilled “t0” models from 1.5B–32B retain strong condition accuracy with k=5 (e.g., 1.5B at 0.53; 32B at 0.56), narrowing the gap to frontier models while fitting in 3–64 GB GPU memory. The study highlights that reasoning distillation especially lifts the smallest models.

Retrieval-Augmented Reasoning with Lean Language Models Great paper showing how to fuse RAG and reasoning into a single small-footprint language model. Distillation works if done correctly. Very exciting results! Here are my notes: https://t.co/awRAdylVc2

Follow along this example of how we created an example vibe-coded @streamlit application for an invoice parser, built with LlamaExtract. In this quick overview, @tuanacelik shows you how we: - Create extraction agents based on a pre-defined schema - Use a carefully constructed prompt equipped with UI requirements that @cursor_ai can use as the main instruction to build an app with - Use sample code as additional context for Cursor to have in terms of what our application does! Check out the full repository here: https://t.co/wwtajEEn3E And watch the walk-through: https://t.co/ZXrIvD7fk0

+1 more

🚀 Excited to co-sponsor Agentic AI in Action in SF with @AWS, @elastic, and @twelve_labs on Aug 26 in SF! Catch our own @seldo (VP Dev Rel) with a live tech talk: “Building Document Agents with LlamaIndex: Effective Design Patterns”. Expect food, demos, partner insights, and AI networking magic. See you at the AWS Loft! 🔗 Register today: https://t.co/9nJHqZHcm1 #AgenticAI #LlamaIndex #AWS #Elastic #AI

@grrberr Probably unpopular opinion but Lenovo is probably the best laptop maker at the moment. Durable by also trying to innovate in a sense that adds utility. Eg the thinkbook with the rollable display. https://t.co/IKYHywol4o

@moskstraum21745 @grrberr They also have something for that demographic, ie the yoga series lol https://t.co/9oZCCgHWJt

Everything you need to know in one shortcut. https://t.co/zDIOKDjKOb

Everything you need to know in one shortcut. https://t.co/zDIOKDjKOb

AI in HR: in an experiment with 70,000 applicants in the Philippines, an LLM voice recruiter beat humans in hiring customer service reps, with 12% more offers & 18% more starts. Also better matches (17% higher 1-month retention), less gender discrimination & equal satisfaction. https://t.co/KGxstQSzBj

Playing with the new mystery "nano-banana" image generation model: "a photo where a woman with a pink mask covering just the left side of her face, and the right side is painted green, she is wearing a duck costume, but the feet are muddy. she stands next to a golden retriever with a remote control in its mouth, they stand in front of an abandoned space shuttle, a small fire burns in a blue vase in the lower left. the moon is reflected in a pond in the lower right, yet the sky is cloudy"

+1 more

I am sure Google people will correct me if I am wrong, because the actual interface is a bit confusing, but Gemini agrees with me. It also autodeletes chats after 18 months (which is not in itself a bad thing) but you can change the timing or remove the autodelete all together. https://t.co/wKRz4MK9v3

🚀 Excited to introduce Qwen-Image-Edit! Built on 20B Qwen-Image, it brings precise bilingual text editing (Chinese & English) while preserving style, and supports both semantic and appearance-level editing. ✨ Key Features ✅ Accurate text editing with bilingual support ✅ High-level semantic editing (e.g. object rotation, IP creation) ✅ Low-level appearance editing (e.g. addition/delete/insert) Try it now: https://t.co/r2Zcg4OjGc Hugging Face: https://t.co/jfCS6b0W5O ModelScope: https://t.co/NwbfHlXonE Blog: https://t.co/VGDdwKuwHy Github: https://t.co/A9yvJZ6TJc API: https://t.co/uaumRmgcGG

+2 more

New Workshop: Dynamic Sketching 🚀 https://t.co/LmtF2yheLA

Thyme Think Beyond Images https://t.co/JrwQcpuZG2

discuss with author: https://t.co/GNJvucjT0T

SSRL Self-Search Reinforcement Learning https://t.co/BMwRjfBLy3

discuss with author: https://t.co/d0GYf5lypi

XQuant Breaking the Memory Wall for LLM Inference with KV Cache Rematerialization https://t.co/UWCjF7Rtcv

discuss with author: https://t.co/kctQ7vXptD

Unstoppable Anycoder 🫡 https://t.co/SUMM00akuB

Unstoppable Anycoder 🫡 https://t.co/SUMM00akuB

Unlimited FREE Generations for Hailuo MiniMax are BACK in Higgsfield! Hailuo MiniMax 02 for Draw-to-Video. A HIGHEST Quality of 1080p. Enjoy it a FULL week! Previously for Ultimate and Creator, it is NOW unlocked for ALL Pro users! Retweet to get 1 of 30 Pro Plan promocodes in DMs!

🚀 Today at #GoogleNext19, we are launching Cloud Run. Allowing you to run any stateless http container in a fully managed environment, paying only for the exact resources you use. 📦☁️ 👉 https://t.co/sT9AVIScNM https://t.co/KRoVgTwizf

Qwen-Image-Edit is out in anycoder for image editing in your vibe coded apps Built on 20B Qwen-Image, it brings precise bilingual text editing (Chinese & English) while preserving style, and supports both semantic and appearance-level editing. https://t.co/tn5uzKmBNr

app: https://t.co/esPDyHE1YC https://t.co/dJX90GRJsk

@Alibaba_Qwen now available in anycoder: https://t.co/esPDyHDu94 https://t.co/NQNiSUBogJ

Noise Hypernetworks Amortizing Test-Time Compute in Diffusion Models https://t.co/RxfrgEpOQX

discuss with author: https://t.co/574ERkoK3v

1/Pretraining is hitting a data wall; scaling raw web data alone leads to diminishing returns. Today @datologyai shares BeyondWeb, our synthetic data approach & all the learnings from scaling it to trillions of tokens🧑🏼‍🍳 - 3B LLMs beat 8B models🚀 - Pareto frontier for performance https://t.co/MUittjMqOO

One of the most underrated players in AI models, @IBM, released 2 new extremely efficient embedding models: granite-embedding-english-r2 & granite-embedding-small-english-r2, commercially viable. Details in 🧵: https://t.co/MXCRUgav3m

The URL context tool is now generally available in the Gemini API and comes packed with new features! The tool lets you provide additional context to the models in the form of URLs and now also has image and PDF endpoint support! Makes adding context for Gemini much easier :) https://t.co/pIJUoj4TcC