Cursor Composer 1.5 在AI轉移中被忽略

🔑 增強重點摘要

• •Composer 1.5 scales reinforcement learning (RL) 20x on the same pretrained model, with post-training compute exceeding base model pre-training, improving real-world coding benchmarks and balancing speed with deep thinking[1].
• •Introduces self-summarization capability trained via RL, allowing the model to produce summaries when context runs out, maintaining accuracy on long tasks through recursive triggering[1].
• •User feedback highlights faster performance, better context management, less invalid code, and effectiveness for multi-file refactoring in languages like Laravel + Vue.js, though scope control issues persist for complex tasks[2][3].
• •Increased usage limits: Composer 1.5 has 3x the limit of Composer 1, temporarily boosted to 6x through February 16, 2026, with separate Auto+Composer and API pools amid rising agentic coding demand[1].
• •Scores 47.9% on Terminal-Bench 2.0 as standalone model, outperforming Claude Sonnet 4.5 on agentic coding benchmarks, released February 9, 2026[6].

🛠️ 技術深入

• •Built by scaling RL 20x further on the same pretrained model; post-training compute surpasses base model pre-training[1].
• •Self-summarization trained as part of RL: model produces useful summary when context exhausts, triggers recursively on hard examples to maintain accuracy across varying context lengths[1].
• •Improves real-world coding benchmarks; balances speed and intelligence for interactive daily use, with upgrades in release modeling for faster, smarter coding on long tasks[1].
• •Better context window management: seamless compaction, less accuracy loss as context grows; reduced invalid code output[3].
• •Integrated with features like Plan Mode (strategic thinking before coding), Checkpoints (snapshots for reversion), and Instant Grep (millisecond codebase search)[4].

🔮 前景展望AI analysis grounded in cited sources

⏳ 時間線