Code-Generation

DuoDiscover: Scaling Autonomous Discovery Beyond the Evaluation Bottleneck. Arxiv 2026.

DuoDiscover lets coding agents co-discover an optimization target and a fast, fidelity-checked proxy evaluator, so search is no longer bottlenecked by hours-long true evaluations. It ranks first on 121 of 235 kernels on NVIDIA’s SOL-Exec B200 leaderboard, doubles human-expert speedups on SWE-fficiency, and reaches 1.92× on vLLM serving. Work done at Google.

October 2026 · Shubham Agarwal, Akshat Gupta, Mike Heddes, Dilara Soylu, Rui Meng, Patrick Tser Jern Kon, Jiefeng Chen, Mert Cemri, Shu Liu, Abe Gonzalez, Akanksha Jain, Bhavana Dalvi Mishra, Linfeng Song, Vilobh Meshram, Xiangru Tang, Jinsung Yoon, Aditya Parameswaran, Ion Stoica, Vahab Mirrokni, Burak Gokturk, Parthasarathy Ranganathan, Tomas Pfister, Nanyun Peng and Chun-Liang Li

SWE-Protégé: Small Language Models for Software Engineering Through Learned Expert Collaboration. NeurIPS 2026.

SWE-Protégé is a post-training framework (SFT + agentic RL) that teaches a small language model to stay the decision-maker on SWE tasks while selectively asking a strong expert model for guidance. It brings Qwen2.5-Coder-7B-Instruct to 42.4% Pass@1 on SWE-bench Verified, +25.4% over the prior SLM state of the art, with Claude Opus 4.1 as the expert. Work done at Meta.

September 2026 · Patrick Tser Jern Kon, Archana Pradeep, Ang Chen, Alexander P. Ellis, Warren Hunt, Zijian Wang, John Yang and Samuel Thompson