The authors post-trained a large (Kimi K2.7 Code) using reinforcement learning on 1,700 expert-built tasks. These tasks included repository changes graded by hidden tests and terminal tasks with complex verifiers, with a reward that gave partial credit for solved checks but zero for any regression. After one epoch of training, the model improved pass@1 on all six external benchmarks tested, across three different , with gains from 4.7 to 20 percentage points. The model also used 24-35% fewer agent steps on average. The authors show that the base model’s failures were often near-misses, and the trained model specifically avoided four common failure modes: dropping requirements, testing only its own implementation, causing silent regressions, and relying on weak ground truth. This indicates the training taught generalizable skills for completing coding work carefully.
Reinforcement learning helps coding agents finish tasks across benchmarks
Training on 1,700 expert-built tasks improved one open-weight model across six external coding benchmarks and three agent harnesses.
Industry
Sushant Mehta · Logan Ritchie · Edwin Chen
Surge AI
Research Digest··2 min read
Thread:RL for Tool Agents
In one line
RL on 1,700 expert coding tasks improves a 1T-parameter model on six benchmarks and shortens agent steps.
Before you read this
Three ideas this paper builds onreinforcement learning with verifiable rewardsReinforcement learning that uses rewards from objectively checkable signals like correct answers or verification feedback.163 papers →coding agentAn AI-powered software system that autonomously writes, modifies, or extends source code based on high-level instructions.167 papers →verificationChecking outputs for correctness before accepting them, often to catch errors or reward hacking.341 papers →
Read it as
Mehta, Ritchie and Chen post-trained Kimi K2.7 Code using reinforcement learning with verifiable rewards, without a supervised warm-up.
The authors trained a rank-32 LoRA adapter for the 1-trillion-parameter mixture-of-experts model Kimi K2.7 Code, which activates 32 billion parameters per inference.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.