Introducing Recursion: the RL platform for enterprise specialist agents
Insights on AI research, reinforcement learning, evaluations, and enterprise AI systems
Recursion is a unified reinforcement learning platform for developing, evaluating, and deploying specialist AI models that improve from real enterprise execution.
Labelbox•June 24, 2026
Models rarely flag their own misbehavior, and when they do, they pick the most lenient monitor available. We introduce monitorability disposition: a model's willingness to stay monitored, a property that is measurable, undertrained, and missing from alignment evaluation.
Shahriar Golchin•June 30, 2026
We explore whether NLAs can surface internal patterns behind shortcut behavior in LLMs using branchpoint analysis. Our findings: signals are weak and distributed, with feedback and surface features strongly shaping behavior, suggesting useful directions for future interpretability work.
Almas Abdibayev•June 15, 2026
Meta Superintelligence Labs introduces GIM (Grounded Integration Measure), a benchmark shifting from isolated recall to integrated reasoning. It evaluates how models coordinate constraints, ambiguity, spatial logic, and epistemic judgment within a single problem.
Labelbox•May 20, 2026
Security is not a checklist, it is constantly evolving. As supply chain attacks grow, one compromised dependency can ripple across ecosystems. We are sharing best practices to contain risk, respond quickly, and design systems that limit blast radius by default.
Labelbox•April 10, 2026
We introduce EchoChain to advance audio evaluation by testing Dual-Stream Reasoning in scenario-driven conversations with mid-speech interruptions, constraint updates, and shifting objectives. The benchmark measures whether models sustain coherent, adaptive intelligence in real time.
Smit Nautambhai Modi•March 4, 2026
AI safety is often judged by refusal rates, but our study of datasets like AdvBench and HarmBench shows these scores rely on obvious trigger words, not real adversarial intent. Remove the cues and the supposed safety collapses, revealing a stark gap between benchmarks and real world risk.
Shahriar Golchin•February 20, 2026
We've acquired Upcraft to bring AI agent technology into Alignerr, scaling how elite domain experts train, evaluate, and improve the world’s most advanced AI models.
Labelbox•February 10, 2026
When startups shut down, their production code doesn’t lose value, it loses a home. Labelbox helps founders safely convert that dormant, real-world engineering work into value through a one-time, secure, founder-friendly transaction with no ongoing risk or obligations.
Labelbox•January 20, 2026
Takeaways on the themes and research directions likely to shape the year ahead. We focus on two core areas how to rigorously measure AI capabilities and how to build interactive systems that learn through experience over time.
Labelbox•December 16, 2025
Most real-world tasks are underspecified. We introduce Implicit Intelligence to test whether agents can infer hidden constraints, and Agent-as-a-World, a simple YAML framework for simulating environments without brittle, hard-coded worlds.
Ved Sirdeshmukh•November 21, 2025
Get started for free or see how Labelbox can fit your specific needs by requesting a demo
