DigiNews

Tech Watch by Johan Denoyer

← Back to articles

Introducing SWE-2: Pushing the Pareto Frontier

Quality: 8/10 Relevance: 9/10

Summary

Cognition announces SWE-2, their most advanced coding model, claiming it achieves near-parity with Fable 5.1 and GPT-Astra at a fraction of the cost. The post describes RL-driven training at multi-trillion-parameter scale, a single-run optimization across reasoning effort levels, and substantial improvements in efficiency, test coverage, and verification. It also covers the Pareto-frontier RL methodology, data enhancements, and trustworthiness evaluations across multiple models.

🚀 Service construit par Johan Denoyer