Research
Data quality makes all the difference.
We're driven by the conviction that model performance is fundamentally bounded by training data quality. Through expert collaboration, rigorous curation methodologies, and deep domain expertise, we research datasets that power tomorrow's models.
Research and Blog
How We Improved Terminal-Bench 2.0 Scores by Over 5x Using Tinker and Harbor
Technical Blog
IDE-Bench: Evaluating Large Language Models as IDE Agents on Real-World Software Engineering Tasks
Research Paper and Benchmark
Joining AfterQuery
Blog
Market-Bench: Evaluating LLMs on Introductory Quantitative Trading
Research Paper and Benchmark
App-Bench: Evaluating Coding Agents on Generating Economically Useful Web-Apps
Benchmark
The AfterQuery Thesis
Blog
UI-Bench: A Benchmark for Evaluating User Interface Understanding
Research Paper and Benchmark
LeetBench: A Benchmark for Competitive Programming & Algorithmic Reasoning
Benchmark
VADER: Vulnerability Assessment, Detection, Explanation, and Remediation
Research Paper and Benchmark
FinanceQA: A Question Answering Benchmark for Financial Data
Research Paper and Benchmark
Core Research Areas
†AI Safety and Security
Our research focuses on developing novel approaches to AI training that help models understand and respect human values without sacrificing capability.
÷Multimodal Learning
We advance AI's ability to understand and reason across visual, audio, and textual modalities simultaneously.
>Computer Use & Automation
We've created training data that teaches AI agents to understand context, anticipate user needs, and execute complex multi-step workflows across diverse software environments.
∈Data Quality & Curation
We've developed rigorous methodologies for identifying, filtering, and enhancing training data quality to drive superior model performance.
∑Model Evaluation
Our evaluation frameworks go beyond traditional benchmarks to rigorously assess real-world AI performance across diverse real-world scenarios.