Research

Today, we’re sharing details about the Snorkel Agentic Coding benchmark—a comprehensive evaluation suite designed to test whether agents can handle the full complexity of software engineering work.

We just returned from NeurIPS 2025, and we’re still processing everything we saw. The energy around data-centric AI has never been stronger—and we couldn’t be more grateful to the research community for pushing these ideas forward.

Explores how rubrics support agentic, multi-turn, tool-using, multimodal, and code-generating AI systems, and how they evolve with AI feedback and ensemble evaluation.

TL;DR: We stress-tested the “generate → criticize → improve” loop on 50 visual reasoning tasks. The results were counterintuitive: self-critique acts as a corrosive agent on high-performance tasks, turning 98% accuracy into 57%. Yet, for tasks where models fail completely, it works like magic. This difficulty-dependent behavior poses a critical, hidden risk for RLFT pipelines. The promise vs. the reality

Snorkel Chief Scientist Fred Sala and Kobie Crawford chat with the Terminal-Bench team to unpack the design behind Terminal-Bench 2.0 and the new Harbor framework.

Snorkel AI contributes specialized datasets to Hazy Research’s “Intelligence-per-Watt” study, advancing how efficiently AI turns energy into intelligence.

Terminal-Bench 2.0 launches today, marking a major leap in AI agent evaluation. Snorkel AI contributed key research and task design to this release.

We unpack what makes a high-quality RL environment for LLMs and show how we build realistic, enterprise-grade environments at Snorkel AI.

A procedurally generated and programmatically verified benchmark for evaluating spatial reasoning capabilities in LLMs Large language models (LLMs) are showing remarkable results on solving complex reasoning problems across domains—from mathematical proofs and logical puzzles to graduate-level science and engineering questions. On the other hand, their spatial reasoning capabilities are less understood, even though such reasoning underlies many everyday tasks. We

Snorkel’s “Trusted Scale” philosophy Welcome to Part 4 of Snorkel AI’s rubric series. In previous posts, we explored how rubrics enable structured evaluation (Part 1), the spectrum of rubric types and use cases (Part 2), and the science behind designing and validating them (Part 3). In this latest installment, we pull back the curtain on how Snorkel puts these principles

In recent months, there has been increasing interest in the area of multi-agent systems and how they can be used to solve more complex tasks than a single agent could accomplish on its own. The topic is particularly interesting and raises several questions and ideas to consider: Anthropic’s blog post about how they architected a multi-agent deep research system is

Terminal-Bench, developed through a collaboration between Stanford University and Laude Institute, has quickly become the gold standard benchmark for evaluating AI agent capabilities in a command line environment. This comprehensive evaluation framework measures how effectively AI agents can perform complex, real-world tasks within terminal environments. At Snorkel AI, we’re excited to share that we’re one of the top collaborators contributing

Behind every AI benchmark is a hidden choice: how to read the model’s answers. That choice—parsing—can quietly tilt results more than the model itself. Parsing is where we take an AI system’s raw response and extract the “answer” we use for scoring. It sounds mechanical, but as our research shows, the choice of parser can dramatically change measured accuracy. In

Part 3 of our rubric series explains the science of rubric design. We show why rubrics should be treated like models—structured, measured, and iterated—to maximize objective alignment and inter-rater agreement. Learn how to choose hierarchy and scale points, track agreement (IAA) and LLMAJ alignment, and refine with domain experts, with examples like PaperBench and HealthBench.

Rubrics turn fuzzy “good vs. bad” into measurable criteria for GenAI. In Part 2, we map what to measure (granularity and dataset-level vs instance-specific), where to measure (process vs outcome), and how to measure (humans, LLM-as-judge, code, reward models)—with examples like HHH, FLASK, HealthBench, and PaperBench.