Smarter LLMs, Faster—thanks to metadata What if training a large language model didn't just rely on text, but on the context around it? This study shows that adding fine-grained metadata—not just URLs—can meaningfully speed up pretraining and improve quality. * Beyond URLs: detailed quality signals (e.
LLMs can learn to reason—without task verifiers Many real-world problems don’t have automatic checkers to grade answers, even though we have lots of expert solutions. RARO (Relativistic Adversarial Reasoning Optimization) shows how to train reasoning skills from those examples alone. How it works: * A policy (the model) tries
Robots guided by Vision-Language-Action (VLA) AI are getting better at everyday tasks—but most still use simple two-finger grippers. That limits them on smooth, flat, or handleless objects. VacuumVLA is a low-cost robot hand that merges a standard two-finger gripper with a vacuum suction cup. The robot can switch between
Meet ChatDRex Finding new uses for approved drugs can save years and millions. But making those predictions usually takes teams of specialists and a tangle of tools. ChatDRex changes that. It’s a conversation-based, no‑code, multi‑agent system that lets clinicians and researchers ask complex bioinformatics questions in plain
Meet Qwen3-VL, our most capable vision-language model yet. It understands text, images, and video together—and keeps track of up to 256K tokens, so it can follow book-length docs and long videos without losing the thread. What’s new * Stronger brains: Even better at pure text tasks, beating many text-only
Large models are powerful, but solving deep, multi-step problems is pricey. ToolOrchestra trains a small “orchestrator” that decides which tools to use, when, and how—like a smart project manager for AI. It uses reinforcement learning with rewards for outcomes, efficiency, and user preferences. The result: an 8B Orchestrator that
What’s new Adding the right metadata can make LLM pretraining faster and more effective—and it’s not just about URLs. * Fine‑grained signals work: Prepending detailed quality indicators helps models learn quicker. * Append-and-predict: Appending metadata and training the model to predict it as an auxiliary task boosts efficiency.
Meet Qwen3-VL, the newest model in the Qwen series that natively understands text, images, and video—at the same time. * Long memory: up to 256K tokens (book-length) with interleaved text+media. * Flexible sizes: 2B/4B/8B/32B dense and 30B-A3B/235B-A22B MoE to balance speed and quality. * Stronger language skills
Small model, big wins Large language models are great generalists, but really tough, multi-step problems still strain both brains and budgets. ToolOrchestra flips the script: instead of one giant model, a small “orchestrator” coordinates other models and specialized tools. Trained with reinforcement learning that rewards outcomes, efficiency, and user preferences,
Ever scrolled past pages of near-identical listings? Those duplicates hurt user experience and waste ad spend. FITRep is an attention-guided, white-box way to represent items across image + text, so platforms can tell what’s truly important (primary) versus extra (auxiliary) and cluster near-duplicates with confidence. * CHIE: uses Multimodal LLMs to
What’s new As AI "teams" of LLM agents grow, the cloud bill can explode. BAMAS is a framework that designs these teams with a dollar cap in mind. How it works * Picks the right mix of models: Uses integer linear programming to balance accuracy vs. price. * Plans
TL;DR: RARO teaches language models to reason from expert demonstrations - no task-specific verifier needed. Many real tasks don’t have an automatic "checker." RARO (Relativistic Adversarial Reasoning Optimization) uses inverse reinforcement learning to learn from examples instead. How it works: a policy tries to produce answers