444 points • 4 days ago • Article Link

The Cognition Team 推出了 SWE-2，这是他们迄今为止最先进的编码模型，旨在显著改善能力与成本之间的平衡。通过将强化学习扩展到万亿参数级别，SWE-2 在 FrontierCode 1.1 Main 基准测试中达到了 50.0% 的成功率。该性能使其接近当前的顶尖模型，同时成本比 Fable 5.1 等竞争选项低 64% 。该模型以 Kimi K3 为基础，经过大规模的后期训练以强化其代理式编码能力。

SWE-2 在解决问题的效率和工程判断力上都有提升。与前代 SWE-1.7 相比，新模型能更快聚焦代码库中的关键部分，减少过度探索和重复读取代码，从而用更少的步骤就能开始实现，大幅降低每个任务的平均成本。此外，SWE-2 在编写端到端测试方面更可靠，也更善于变通：它能利用现有数据绕过受阻路径，而不只是照搬静态指令。

SWE-2 的一项核心技术创新是在强化学习过程中采用 Pareto-informed cost penalties 。团队在一次训练中同时覆盖所有推理努力级别，并根据基础模型的成本 - 性能曲线斜率调整惩罚，从而成功优化出整个权衡前沿。这一方法保证了模型在处理简单任务时具有成本效益，同时在面对更复杂的挑战时保持高强度的复杂规划能力。为支持该策略，团队还引入了基于长度加权的奖励基线，以减小梯度方差、提升训练稳定性。

训练基础设施也进行了优化，优先提升吞吐量并尽量减少推理与训练之间的不匹配。团队整合了在线草稿模型训练，并采用诸如 FP8 kernels 的低精度、量化感知方法以高效管理内存。这些改进带来了更稳定的训练过程，使模型在不显著增加延迟的前提下保持高性能。数据质量同样被高度重视：通过扩展 RL environments 并实施递归飞轮机制，不断增强验证器以防止奖励操纵。

在可信性与对齐方面，团队对 SWE-2 在宣传、审查和对上下文依赖表述的脆弱性等方面进行了评估。模型表现良好，在政治敏感提示上的通过率达到 98%，且无论用户如何表述或使用何种语言，都未表现出统计学上显著的偏见或脆弱性。 SWE-2 已通过 Devin Desktop 应用、命令行工具和 Web 平台向用户提供，这是将顶尖编码代理变得更强大且更易获取的重要一步。 The Cognition Team has introduced SWE-2, their most advanced coding model to date, designed to significantly improve the balance between capability and cost. By scaling reinforcement learning to the multi-trillion-parameter regime, SWE-2 achieves a 50.0% success rate on the FrontierCode 1.1 Main benchmark. This performance places it near the current frontier models while maintaining a cost profile 64% cheaper than competing options like Fable 5.1. The model is built on the foundation of the Kimi K3 base model, which underwent extensive post-training to refine its agentic coding abilities.

The intelligence of SWE-2 is characterized by more efficient problem solving and improved engineering judgment. Compared to its predecessor, SWE-1.7, the new model reduces over-exploration and redundant code reads by focusing more quickly on the critical parts of a codebase. This allows SWE-2 to begin implementation in fewer steps, significantly lowering the average cost per task. Furthermore, the model demonstrates enhanced reliability in writing end-to-end tests and exhibits greater resourcefulness by navigating around blocked paths using available data, rather than simply relying on static instructions.

A central technical innovation in SWE-2 is the use of Pareto-informed cost penalties during the reinforcement learning process. By training all reasoning-effort levels in a single run with penalties tuned to the slope of the base model's cost-performance curve, the team successfully optimized the entire trade-off frontier. This principled approach ensures that the model provides cost-effective performance for simple tasks while maintaining high-effort, complex planning capabilities for more difficult challenges. To support this, they also implemented a length-weighted reward baseline that reduces gradient variance and promotes stability throughout the training phase.

The training infrastructure for SWE-2 was refined to prioritize throughput and minimize inference-training mismatches. The team integrated online draft-model training and utilized low-precision, quantization-aware methods, such as FP8 kernels, to manage memory usage efficiently. These improvements allowed for a more stable training process, resulting in a model that maintains high performance without significantly increasing latency. Data quality was similarly prioritized through the expansion of RL environments and the implementation of a recursive flywheel that continuously hardens verifiers against reward hacking.

Finally, the team evaluated SWE-2 on its trustworthiness and alignment, specifically regarding propaganda, censorship, and vulnerability to context-dependent framing. The model performed strongly, achieving a 98% pass rate on politically sensitive prompts while showing no statistically significant bias or vulnerability regardless of the user's framing or the language used. SWE-2 is currently available for users through the Devin Desktop application, command-line interface, and web-based platforms, marking a significant step forward in making frontier-level coding agents both more capable and accessible.