Labelbox

Today, we’re introducing Recursion: the RL platform for enterprise specialist agents. AI agents are moving beyond experiments and into real enterprise workflows. As general intelligence becomes more accessible, the advantage will shift to companies that can turn their expertise, workflows, and decisions into AI systems that continuously learn and improve. With Recursion, enterprises get specialist agents tailored to their proprietary data, workflows, and expert feedback. The result is frontier-level performance from smaller, more efficient open-source LLMs at a fraction of the cost. The results: → A Recursion-tuned 35B model outperformed a 397B model on held-out agentic finance tasks, improving accuracy from 56% to 86% through reinforcement learning. → Specialist customer support models outperformed leading frontier models with higher resolution rates, 72% fewer hallucinations, 2.6x faster responses, and 5x lower cost per ticket. The future of AI won’t be driven by a handful of general-purpose models alone. It will be shaped by systems that combine foundation models, specialist agents, and proprietary knowledge to continuously improve. Read the full post to learn how Recursion helps enterprises build their own learning loop. https://lnkd.in/gvzgJZ-z

Where do models change their minds? Natural Language Autoencoders (NLAs) offer a promising way to translate a model’s internal representations into natural language. But the harder question is: where do meaningful decisions actually happen? We tested a workflow for finding these “branchpoints” by tracing activations across layers and comparing when models arrive at the right answer versus when they rely on shortcuts. The surprising finding: early signals can look convincing, but stronger controls showed that many apparent decision points may come from confounding factors rather than the process we’re trying to uncover. The takeaway: understanding model behavior takes more than spotting activation differences. It requires careful validation to separate real insights from artifacts. Check out the full post in the comments and special thanks to Almas A. for the collaboration and insights!

The next generation of AI models is built through continuous learning. Labelbox helps leading AI labs generate, evaluate, and refine high-quality reinforcement learning data that improves every checkpoint.

Frontier data isn't found—it's engineered. Labelbox combines expert-powered data generation, rigorous QA, and scalable operations to help the world's leading AI teams train better models.

Discover how we partner with researchers to fuel the next wave of AI advancements, powered by experts in post-training and model evaluation. Labelbox is where the world’s leading labs turn when they need frontier-grade data, fast. From multimodal data and agentic systems to advanced robotics, we start at the idea level and combine cutting-edge software, our expert network, and RL workflows to accelerate your most ambitious AI projects.

Generalization in AI doesn’t come from a single dataset - it comes from training across many realistic tasks and environments. Take a sales agent. The challenge isn’t teaching it to click the right buttons in Salesforce. It’s teaching judgment: Which leads are worth pursuing? How should the agent react when a prospect ghosts? How do you handle different personalities? To train these systems, Labelbox builds real-world reinforcement learning environments with domain experts. For example, dozens of Fortune 500 sales professionals helped create thousands of simulated sales scenarios where AI agents interact with customer personas powered by another AI. Experts monitor these interactions in real time and continuously refine the environment to keep models learning effectively. Labelbox applies this approach across industries from finance to healthcare to aviation.