← News

Agents & MCPSources: arXiv · 16 Nov 2026

RL training on 1,700 coding tasks lifts Kimi K2.7 Code across six benchmarks

The authors report that reinforcement learning alone, applied to expert-built agentic coding tasks, improved pass@1 on every external benchmark they tested, including sets released after training data was collected.

In plain words

A research team post-trained a large open-weight model using reinforcement learning on coding tasks an agent must complete inside repositories and terminals. The trained model solved more tasks on six outside benchmarks than before, suggesting the learned skills carry over beyond the exact training material.

What was announced

The paper targets what the authors call the last-mile problem: coding agents that build most of a feature yet miss a requirement, test only what they already handle, break working behavior, or trust an unverified assumption. To address this, the team post-trained Kimi K2.7 Code, a one-trillion-parameter open-weight mixture-of-experts model with 32 billion active parameters, using reinforcement learning on 1,700 tasks. One thousand were repository tasks scored by hidden tests that must newly pass plus tests guarding existing behavior; 700 were terminal tasks scored by expert-written hidden verifiers. The reward was the share of target checks passed, falling to zero if any behavior-preservation test failed.

After one epoch of GSPO on a rank-32 LoRA adapter, the authors report pass@1 gains on all six external benchmarks across three agent harnesses, for example SWE-Bench Pro rising from 60.1 to 64.8 and DeepSWE from 31.0 to 43.4. Pooling five independent task sets, the gain was significant, and it held for the three sets released after training data was gathered. The model also improved under two harnesses never seen in training, and median trajectories on DeepSWE and Terminal-Bench 3 were 24-35% shorter in agent steps. Paired runs showed the trained model avoiding the four failure modes on tasks it newly solved. These are the authors' own reported results, not independent replications.

What it does not show

  • The source is an abstract, so method details and full results are not available for verification.
  • All figures come from the paper's authors; no independent confirmation is included.
  • Terminal-Bench 4 revises Terminal-Bench 3, so the authors treat them as overlapping rather than fully independent sets.

For security and IT teams

Scope
Findings cover one open-weight model post-trained with RL on 1,700 expert-built tasks; transfer was measured on six external benchmarks across three agent harnesses.

Sources

Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 17 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.