Skip to content
Trending storyDeveloping

Microsoft presents LongCapHarness, a self-improving harness for long-context reasoning

1 article1 sourcesince Nov 10Last article 2h ago ·

Overview

AISummary of 1 article

Microsoft researchers present LongCapHarness, a harness that improves long-context reasoning without synthetic data or post-training.

The authors say it raises Llama3-8B by 64.18 points on average across long-context benchmarks and transfers to Qwen2.5-3B.

The authors also say LongCapHarness beats the NExtLong and PolicyLong baselines on the HELMET and RULER benchmarks at less than half their estimated cost. These are the authors' claims, reported by Elvis Saravia on X; the report does not include independent verification.

Written by AI from the articles below · updated Nov 10, 9:36 AM ET

Check the sources:

Latest developmentNov 10, 9:23 AM ET
Microsoft presents LongCapHarness, a self-improving harness for long-context reasoning

Article timeline

The articles in this story. Times are ET.

Nov 10
  1. elvisX
    Microsoft presents LongCapHarness, a self-improving harness for long-context reasoning

    AIMicrosoft researchers present LongCapHarness, a harness that improves long-context reasoning without synthetic data or post-training. Llama3-8B gains 64.18 points on average across long-context benchmarks, and the harness transfers to Qwen2.5-3B. The authors say it beats NExtLong and PolicyLong on HELMET and RULER at less than half their estimated cost.

    Image from @omarsar0's post

Heat Trend

Current comparable heat
9
Peak comparable heat
10Nov 10, 10:00 AM ET
Change in past 24 hours
–

The trend compares only continuously observed participants, so its scope may be smaller than the current heat total. Hover or tap for hourly heat; use arrow keys with a keyboard.