LongCapHarness boosts frozen models' long-context reasoning without training
Microsoft Research Asia proposes LongCapHarness, which builds a harness around a frozen model rather than post-training it.
Tao Zhang and Yuhang He of Microsoft Research Asia report that a frozen Llama3-8B gains 64.18 points on average across long-context benchmarks. It does so without any training, by running inside a harness that LongCapHarness builds automatically. The arXiv paper, submitted on Nov 7, 2026, describes a method that leaves the model's weights untouched. It changes only the code and control logic wrapped around the model.
The process starts with a harnessing agent that studies a development set and breaks each capability the task needs into primitives. The paper's example is reranking, which splits into relevance judgment and ordering, with sequential control logic tying the steps together. The agent then rewrites the primitive implementations, the state they track, and the control logic over many rounds of recursive self-improvement on the development set. Once a harness is finished, it is frozen and saved to a library that the model draws on at test time. The gains therefore come from the scaffolding, not from any change to the model.
The authors position this against two common routes to better long-context reasoning, data synthesis and follow-up post-training. They describe both as empirically driven and computationally intensive. In their experiments, the harness built for Llama3-8B also transfers to Qwen2.5-3B, a smaller model, without being rebuilt for it.
On the HELMET and RULER benchmarks, the paper reports that LongCapHarness beats NExtLong and PolicyLong, which rely on data synthesis and training. The authors put its development cost at less than 50% of the estimated training costs of those two methods. The paper also compares against an RRSI self-improvement baseline.
The method has a dependency that the authors state directly. The harnesses were written by GPT-5.6-Sol, and the authors report that performance drops sharply when Llama3-8B or Qwen2.5-3B build their own harnesses. The abstract does not say whether the development cost estimate counts the calls to that stronger harnessing agent.
What changed
LongCapHarness improves a frozen model's long-context reasoning by building and saving a harness around it rather than post-training the model's weights.
Type: Research · Applies to: Llama3-8B, Qwen2.5-3B, HELMET, RULER
Why it matters
Before post-training a small model for long context, teams could try improving the harness around it, which the authors say is cheaper, though a strong harnessing agent is still needed.
Original source
Where we saw it
DAIR.AI (Elvis Saravia)
AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works