TL;DR

A technical report on Composer 2, a specialized coding model trained through continued pretraining followed by large-scale reinforcement learning on long-horizon software-engineering tasks.

Why it matters

Composer 2 is a concrete example of environment fidelity as a scaling variable: training uses the same tool harness and task structure as deployment, with realistic repositories and multi-step execution rather than isolated code generation.

Key findings

  1. 01

    The RL phase targets end-to-end coding performance, multi-step execution, long-horizon planning, and trajectory coherence.

  2. 02

    The training infrastructure is designed to match the deployed Cursor harness, including its tools and interaction structure.

  3. 03

    The authors report strong results on CursorBench, Terminal-Bench, and SWE-bench Multilingual; comparisons remain sensitive to harness and evaluation setup.

Scaling dimensions

Models, methods & benchmarks

Models
Composer 2
Algorithms
Large-scale reinforcement learning, Continued pretraining
Benchmarks
CursorBench, Terminal-Bench, SWE-bench Multilingual

Topics

Read the original sourcearXiv