Top-nsigma sampling enables long-context coding agents to solve GitHub issues
November 18, 2026
Testing 29 models from 1B to 122B shows that standard samplers like top-p or top-k fail at high temperatures during long generation. While top-p and top-k agents resolved 0 of 100 GitHub issues at temperature 2, top-nsigma samplers resolved between 46 and 62 issues.
HOW THIS AFFECTS YOU
●
builderSwitching to top-nsigma sampling can prevent your coding agents from failing during long-form generation.
●
researcherThe study identifies that sampler stability at high temperatures is a key factor in long-context agentic performance.