Microsoft's LongCapHarness boosts Llama3-8B long-context by 64 points with zero training
omarsar0 · x · 2026-11-10
Microsoft Research proposes LongCapHarness (arXiv:2611.10463), improving long-context reasoning without synthetic data or post-training:
- Decomposes long-context tasks into capability primitives; a strong agent iteratively rewrites the primitives, tracked state, and control logic via self-improvement on a dev set
- Llama3-8B gains 64.18 points on average across long-context benchmarks with no training at all
- The harness transfers to Qwen2.5-3B; on HELMET and RULER it beats NExtLong and PolicyLong at less than half their estimated cost ($0.54K vs thousands)
- The harness was written by GPT-5.6-Sol; results drop sharply when small models build their own
Takeaway: before post-training a small model for long context, try improving the harness instead.
More from coding & agent
- Free RAG Zero to Hero Guide: 1.5k-star GitHub roadmap from basics to evaluation — nevrekaraishwa2 · 2026-11-10
- WorkflowArena: First Benchmark for Multi-Agent Workflow Orchestration With 145 Real-World Tasks — _akhaliq · 2026-11-10
- Klint 1.0 ships: AI-ported TS to Rust is 25x faster with 28x less memory — rchardkovacs · 2026-11-10
- Pi's Virtual Models auto-route queries between local Qwen and remote Kimi or GLM — ggerganov · 2026-11-10
- Unlimited-OCR-2 tested: parsing 30-page PDFs in one request without breaking tables — KhuyenTran16 · 2026-11-10
- Screaming Frog tip: your AI assistant can fully configure your SEO crawler — VeryWellVersed · 2026-11-10