<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>AI Benchmark</title>
    <link>https://aibenchmark.gr/en/</link>
    <description>Technical news about language models with real numbers: benchmarks, research, evaluations and technical reports.</description>
    <language>en</language>
    <lastBuildDate>Fri, 20 Nov 2026 23:19:18 GMT</lastBuildDate>
    <atom:link href="https://aibenchmark.gr/en/feed.xml" rel="self" type="application/rss+xml"/>
    <item>
      <title>LoRA Adapter for Qwen3.8-27B: Terminal-Bench 2.1 Rises From 39.5% to 55.8% With In-Harness RL</title>
      <link>https://aibenchmark.gr/en/news/lora-adapter-qwen3-8-27b-terminal-bench-2-1-rl-inside-harness/</link>
      <guid isPermaLink="true">https://aibenchmark.gr/en/news/lora-adapter-qwen3-8-27b-terminal-bench-2-1-rl-inside-harness/</guid>
      <description>Hugging Face trained adapters for coding agents with reinforcement learning. In pi, the Terminal-Bench 2.1 success rate rises from 39.5% to 55.8%, with roughly half the tokens per task.</description>
      <pubDate>Fri, 20 Nov 2026 22:21:43 GMT</pubDate>
      <category>Research</category>
    </item>
    <item>
      <title>DaisyRose-27B: pass@1 on Terminal Bench 4.1 rises from 5.6% to 12.0% with 1,200 tasks</title>
      <link>https://aibenchmark.gr/en/news/daisyrose-27b-terminal-bench-4-1-pass-at-1-rises-from-5-6-to-12-0/</link>
      <guid isPermaLink="true">https://aibenchmark.gr/en/news/daisyrose-27b-terminal-bench-4-1-pass-at-1-rises-from-5-6-to-12-0/</guid>
      <description>RadixArk has released DaisyRose-27B, a model based on Qwen3.8-27B and trained with reinforcement learning. The company reports gains on all seven benchmarks it publishes.</description>
      <pubDate>Thu, 19 Nov 2026 22:21:36 GMT</pubDate>
      <category>Technical Reports</category>
    </item>
    <item>
      <title>GPT-6.2 Astra and Sol: Astra Scores 62 on the Artificial Analysis Intelligence Index, with API Pricing of $8/$40</title>
      <link>https://aibenchmark.gr/en/news/gpt-6-2-astra-and-sol-astra-scores-62-on-artificial-analysis-intelligence-index/</link>
      <guid isPermaLink="true">https://aibenchmark.gr/en/news/gpt-6-2-astra-and-sol-astra-scores-62-on-artificial-analysis-intelligence-index/</guid>
      <description>OpenAI released GPT-6.2 Astra and Sol on 17 November 2026. Artificial Analysis ranks Astra second with 62 points, behind Claude Fable 5.2 (65), while Astra&apos;s API price has been cut by 20%.</description>
      <pubDate>Wed, 18 Nov 2026 22:23:05 GMT</pubDate>
      <category>Benchmarks</category>
    </item>
    <item>
      <title>MLPerf Training v6.1: 250 submissions, systems of up to 10,240 GPUs and Vera Rubin up to 2x per GPU versus GB300</title>
      <link>https://aibenchmark.gr/en/news/mlperf-training-v6-1-250-submissions-vera-rubin-2x-per-gpu-vs-gb300/</link>
      <guid isPermaLink="true">https://aibenchmark.gr/en/news/mlperf-training-v6-1-250-submissions-vera-rubin-2x-per-gpu-vs-gb300/</guid>
      <description>MLCommons has published the v6.1 round with two new benchmarks and larger systems. NVIDIA says the Vera Rubin NVL72 reaches up to 2x the per-GPU performance of the GB300 NVL72.</description>
      <pubDate>Tue, 17 Nov 2026 22:21:48 GMT</pubDate>
      <category>Benchmarks</category>
    </item>
    <item>
      <title>Alibaba&apos;s Qwen4-Max: 2.4 Trillion Parameters, 71 Billion Active and a Cost per Task of $3.27</title>
      <link>https://aibenchmark.gr/en/news/alibaba-qwen4-max-2-4-trillion-parameters-71-billion-active-cost-per-task-3-27/</link>
      <guid isPermaLink="true">https://aibenchmark.gr/en/news/alibaba-qwen4-max-2-4-trillion-parameters-71-billion-active-cost-per-task-3-27/</guid>
      <description>Alibaba has unveiled Qwen4-Max, a closed mixture-of-experts model with a 1 million-token window. The company claims a 62% lower cost per Work360 task, while Artificial Analysis gives it 43 points on its index.</description>
      <pubDate>Mon, 16 Nov 2026 20:45:21 GMT</pubDate>
      <category>Technical Reports</category>
    </item>
    <item>
      <title>TokenEconomizer: 18.4–30.9% Cost Reduction for Coding Agents Across 277 SWE-Bench Tasks</title>
      <link>https://aibenchmark.gr/en/news/tokeneconomizer-coding-agents-cost-reduction-swe-bench/</link>
      <guid isPermaLink="true">https://aibenchmark.gr/en/news/tokeneconomizer-coding-agents-cost-reduction-swe-bench/</guid>
      <description>Microsoft Research presents TokenEconomizer, an open-source plugin that learns from earlier sessions in a repository. In the team&apos;s measurements, average cost per task fell by 18.4–30.9%.</description>
      <pubDate>Mon, 16 Nov 2026 20:45:21 GMT</pubDate>
      <category>Speed and Cost</category>
    </item>
    <item>
      <title>LLMON: a markup language for language models in Granite 4.1 3B and 8B, aimed at indirect prompt injection attacks</title>
      <link>https://aibenchmark.gr/en/news/llmon-markup-language-granite-4-1-3b-8b-indirect-prompt-injection/</link>
      <guid isPermaLink="true">https://aibenchmark.gr/en/news/llmon-markup-language-granite-4-1-3b-8b-indirect-prompt-injection/</guid>
      <description>Two experimental versions of Granite 4.1, with 3B and 8B parameters, learn a markup that separates instructions from data. The results provided show near-zero attack success in one test.</description>
      <pubDate>Mon, 16 Nov 2026 16:19:25 GMT</pubDate>
      <category>Technical Reports</category>
    </item>
    <item>
      <title>MLPerf Training v6.1: 153 Systems, 250 Results and Two New Benchmarks, Including Qwen3.5-397B</title>
      <link>https://aibenchmark.gr/en/news/mlperf-training-v6-1-153-systems-250-results-two-new-benchmarks-qwen3-5-397b/</link>
      <guid isPermaLink="true">https://aibenchmark.gr/en/news/mlperf-training-v6-1-153-systems-250-results-two-new-benchmarks-qwen3-5-397b/</guid>
      <description>On 16 November 2026, MLCommons announced the MLPerf Training v6.1 results, with 153 systems and 250 results, plus DLRMv4-HSTU and Qwen3.5-397B as new benchmarks.</description>
      <pubDate>Mon, 16 Nov 2026 16:19:25 GMT</pubDate>
      <category>Benchmarks</category>
    </item>
    <item>
      <title>Token pooling-aware training: retaining 97.8% of nDCG@10 on NanoBEIR with 4x fewer vectors</title>
      <link>https://aibenchmark.gr/en/news/token-pooling-aware-training-97-8-ndcg10-nanobeir-4x-fewer-vectors/</link>
      <guid isPermaLink="true">https://aibenchmark.gr/en/news/token-pooling-aware-training-97-8-ndcg10-nanobeir-4x-fewer-vectors/</guid>
      <description>In a Hugging Face post, the author reports that pooling-aware training raises nDCG@10 retention from 94.6% to 97.8% at pool_factor=4, at the same storage size.</description>
      <pubDate>Mon, 16 Nov 2026 16:19:25 GMT</pubDate>
      <category>Research</category>
    </item>
    <item>
      <title>Lazarus: Microsoft Research describes an LLM backdoor created by reprogramming fewer than 10 neurons</title>
      <link>https://aibenchmark.gr/en/news/lazarus-microsoft-research-llm-backdoor-fewer-than-10-neurons/</link>
      <guid isPermaLink="true">https://aibenchmark.gr/en/news/lazarus-microsoft-research-llm-backdoor-fewer-than-10-neurons/</guid>
      <description>Microsoft Research says changing the weights of fewer than 10 dormant neurons is enough to backdoor an agent, with no training. The model keeps passing standard evaluations.</description>
      <pubDate>Mon, 16 Nov 2026 16:19:25 GMT</pubDate>
      <category>Research</category>
    </item>
    <item>
      <title>XYBench: new 8,115-question benchmark measures whether models spot the user&apos;s real problem</title>
      <link>https://aibenchmark.gr/en/news/xybench-new-benchmark-8115-questions-models-real-user-problem/</link>
      <guid isPermaLink="true">https://aibenchmark.gr/en/news/xybench-new-benchmark-8115-questions-models-real-user-problem/</guid>
      <description>Hugging Face has unveiled XYBench, which tests whether five models identify the mistaken assumption behind a question. In the results, all of them fall short of humans at steering the user toward the real goal.</description>
      <pubDate>Mon, 16 Nov 2026 16:19:25 GMT</pubDate>
      <category>Benchmarks</category>
    </item>
  </channel>
</rss>
