SWE-INTERACT: Reimagining SWE Benchmarks as User-Driven Long-Horizon Coding Sessions
2026-06-29 • Machine Learning
Machine Learning
AI summaryⓘ
The authors created SWE-Interact, a new way to test coding AI that works with users over multiple steps instead of just following fixed instructions. Unlike past tests where the AI knows everything upfront, their system lets a simulated user reveal requirements bit by bit and give feedback, mimicking real coding projects. They found that models good at one-step tasks often struggle with this back-and-forth approach, making mistakes or forgetting details. The best models still have trouble with complex interactions but do better at adapting and writing cleaner code. This test highlights how important it is for coding AIs to understand changing goals and work closely with users.
SWE-Interactcoding agentsmulti-turn interactionuser-driven software engineeringrequirement elicitationdeveloper workflowAI coding benchmarksinteractive refinementlarge language modelstask adaptation
Authors
Mohit Raghavendra, Anisha Gunjal, Aakash Sabharwal, Yunzhong He
Abstract
We introduce SWE-Interact, a new testbed for evaluating coding agents on multi-turn, interactive, user-driven software engineering tasks. Existing frontier SWE benchmarks typically provide complete requirements upfront and evaluate agents on autonomous implementation. In contrast, SWE-Interact places agents in a realistic developer workflow: a carefully designed user simulator starts with vague or incomplete instructions, progressively reveals requirements, inspects the agent's workspace, and provides targeted feedback, revisions, and new constraints until the full task goal has been handed off. Grounded in large-scale studies of real coding-agent interactions, this setup tests whether agents can discover user intent, adapt to evolving requirements, and build on their own prior work. Across a suite of frontier and open-weight models, we find that strong performance on single-turn SWE tasks does not reliably transfer to multi-turn, user-driven workflows: the best-performing models solve roughly 50% of single-turn baseline tasks but only 25% of the corresponding SWE-Interact tasks. The strongest models in our evaluation, including Opus 4.8 and GPT 5.5, start strong even in the face of vague initial instructions, persevere until all the requirements are surfaced by the user, integrate them better and write clean code. However, they still suffer from over-agentic coding, forgetting requirements and technical mistakes. Weaker models start poorly under ambiguity, give up early, forget or ignore instructions and rework their code more. Overall, SWE-Interact measures an orthogonal, real-world capability axis for frontier model development: interactive goal discovery and iterative refinement with a user in the loop.