<?xml version="1.0" encoding="utf-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
    <channel>
        <title>Tuned Qwen 8B vs Gemini and GPT: Six AI-Judging Checks (Workshop Preprint)</title>
        <link>https://peertube.watch/videos/watch/9c69c189-6a3c-4f1b-beef-e8220b32089b</link>
        <description>Peace Call. Mr. Peace is an AI-made character; voice is AI-generated; script written with AI assistance. No affiliate links in this video. How this site is run: https://getpeacecall.com/verify/?utm_source=peertube&amp;utm_medium=video&amp;utm_campaign=underdog-judge In a preprint accepted at a NeurIPS 2026 workshop (not the main conference), a tuned Qwen3-8B judge (trained with CPA) scored higher than Gemini 3.6 Flash and GPT 5.6 across six narrow AI-answer checks: 90.4, versus 78.6 and 76.1 macro-F1 (Table 2). Across these six checks, the tuned judge led on five and trailed on tool-parameter checking. On the hallucination check it scored 99.6 macro-F1 (a score, not "% caught"). It lost the tool-parameter check: 89.9 vs 92.0 (Gemini) and 91.2 (GPT 5.6). This was not a matched-run comparison: the small judge's numbers are 10-run means; the frontier models got a single zero-shot run. Qwen3-8B also appeared in an earlier video - different study. Source: Zangooei, Teymoori, Robati Shirzad, Shahtalebi (Huawei), 'Counterfactual Pairwise Augmentation Improves Small Pointwise LLM Judges', arXiv:2610.41149v1 - PREPRINT (Oct 30, 2026), accepted at a NeurIPS 2026 workshop. https://arxiv.org/abs/2610.41149 . Per-criterion macro-F1 from Table 2. Names are trademarks of their respective owners; Peace Call is not affiliated with or endorsed by them. Charts drawn by code from the source's numbers. Paper: https://arxiv.org/abs/2610.41149 (version v1, posted Oct 30, 2026). AI Disclosure: Mr. Peace is a fictional AI character using a synthetic voice. This video is AI-assisted and does not depict a real person speaking. Research Disclosure: Based on a research preprint that has not been peer-reviewed. Results apply to the reported benchmarks and experimental conditions, not overall model superiority. Trademark Disclosure: Names are trademarks of their respective owners. Peace Call is not affiliated with or endorsed by them.</description>
        <lastBuildDate>Wed, 04 Nov 2026 13:42:33 GMT</lastBuildDate>
        <docs>https://validator.w3.org/feed/docs/rss2.html</docs>
        <generator>PeerTube - https://peertube.watch</generator>
        <image>
            <title>Tuned Qwen 8B vs Gemini and GPT: Six AI-Judging Checks (Workshop Preprint)</title>
            <url>https://peertube.watch/lazy-static/avatars/3b284627-b75f-409e-ac2e-10f4c2946f43.png</url>
            <link>https://peertube.watch/videos/watch/9c69c189-6a3c-4f1b-beef-e8220b32089b</link>
        </image>
        <copyright>All rights reserved, unless otherwise specified in the terms specified at https://peertube.watch/about and potential licenses granted by each content's rightholder.</copyright>
        <atom:link href="https://peertube.watch/feeds/video-comments.xml?videoId=9c69c189-6a3c-4f1b-beef-e8220b32089b" rel="self" type="application/rss+xml"/>
    </channel>
</rss>