As of March 2026, AI systems are able to post-train models to get about half as much of the uplift as ones trained by humans. The specific eval scores are derived by a 'weighted average is taken across all post-trained LLMs... The top-scoring systems as of April get 25%-28% (Opus 4.6, and GPT 5.4), compared to a human score of 51%.'
在模型微调任务上，AI系统已能达到人类研究员51%性能的一半，显示出AI在科研任务上的显著进步。
DeepSeek-V4-Pro-Max beats GPT-5.2 and Gemini 3.0-Pro on standard reasoning benchmarks and lands just behind GPT-5.4 and Gemini 3.1-Pro
DeepSeek V4-Pro-Max在标准推理基准测试中超越了GPT-5.2和Gemini 3.0-Pro，这表明了开源模型在性能上的巨大提升。
Two variants are available: **Sakana Fugu Mini 🐟**, optimized with latency in mind, and **Sakana Fugu Ultra 🐡**, the full orchestration system, optimized for performance for demanding tasks.
文章提到有两种变体：Mini（延迟优化）和Ultra（性能优化），但未提供具体的性能指标差异，如延迟降低百分比或吞吐量提升数据。这种缺乏具体量化参数的描述难以评估两种变体在实际应用中的性能差异。
GPQAD | 94.4 | 90.9 | 92.7 | 92.4 | **95.1** | LCBv6 | 90.3 | 92.1 | 92.4 | 90.4 | **93.2** | SWEPro | 48.4 | 51.2 | _53.4_ | 51.3 | **54.2**
性能对比表格显示，Sakana Fugu Ultra在三个基准测试中均优于竞争对手：GPQAD上达95.1%（超越Gemini 3.1的94.4%），LCBv6上达93.2%（超越GPT 5.4的92.1%），SWEPro上达54.2%（超越Opus 4.6的53.4%）。这些数据表明其多模型协调策略确实带来了性能提升，特别是在科学推理任务上优势明显。
The best-performing model across these three metrics was a pair of independent linear trends: one for reasoning models and one for non-reasoning models.
这个发现表明推理模型和非推理模型的发展轨迹确实存在显著差异。这种分离的线性趋势模型在三个指标上表现最佳，100%的情况下优于其他模型，提供了强有力的统计证据支持AI能力加速的论点。
GPT-4o operates at roughly 200 billion parameters and outperforms the original 1.8 trillion-parameter GPT-4
这一发现与行业普遍认为'更大模型必然更好'的共识相悖，暗示模型质量和架构可能比规模更重要。这可能是AI发展史上最令人惊讶的效率提升案例之一，挑战了我们对AI进步的理解。
It also surpasses all peer-scale dense models by a wide margin.
在多数情况下，人们可能认为更大规模的模型将具有更好的性能，但作者提出Qwen3.6-27B在同等规模密集模型中表现卓越，这一观点与主流认知相悖。
It also surpasses all peer-scale dense models by a wide margin.
大多数人可能认为模型性能与其规模成正比，但作者指出Qwen3.6-27B在同等规模模型中表现突出，超越了所有同规模密集模型，这挑战了规模与性能之间的传统认知。
Among the latest frontier models, GPT-5.4 over-edits the most.
大多数人认为GPT-5.4是最先进的模型，但作者指出它在最小化编辑任务上表现最差，这挑战了对其能力的普遍看法。
Opus 4.7 introduces a new `xhigh` ('extra high') effort level between `high` and `max`, giving users finer control over the tradeoff between reasoning and latency on hard problems.
引入'xhigh'努力等级显示了AI模型在推理深度与响应速度之间提供更精细控制的能力，这反映了用户对AI性能调优需求的增长，也表明AI系统正变得更加可定制和专业化。
I-DLM-8B is the first DLM to match the quality of its same-scale AR counterpart, outperforming LLaDA-2.1-mini (16B) by +26 on AIME-24 and +15 on LiveCodeBench-v6 with half the parameters
令人惊讶的是：I-DLM-8B模型仅用80亿参数就超过了160亿参数的LLaDA-2.1-mini模型，在AIME-24和LiveCodeBench-v6测试中分别高出26和15分。这表明扩散模型首次达到了与自回归模型相当的质量水平，同时参数减半，打破了人们对扩散模型质量不如自回归模型的普遍认知。
Cost (USD) to run the evaluation: GPT-5.4 (xhigh): $1,110, Claude Opus 4.6 (max): $1,055
运行一次 452 个任务的评测，GPT-5.4 花费 1110 美元，Claude Opus 4.6 花费 1055 美元——每个任务平均约 2.3 美元。而 Gemini 3 Flash 只需要 596 美元，实现了 27.7% 的成绩（vs 顶级模型的 33.3%）。这个性价比数据对 AI 选型决策极为关键：如果业务场景可以接受 27% 而非 33% 的成功率，Gemini 3 Flash 能节省近一半成本。在金融服务的大规模部署中，这个差异将被放大数千倍。
our approach improves Qwen3.5-4B from 63.8 percent to 66.7 percent (+2.9pp) and Qwen3-30B-A3B from 58.0 percent to 69.5 percent (+11.5pp)
大多数人认为在复杂的多轮任务中，只有大型语言模型才能通过强化学习取得显著进步，但作者展示了即使是较小的4B模型也能通过他们的方法获得实质性提升，而30B模型的提升更是惊人地达到了11.5个百分点，挑战了'规模越大越好'的普遍认知。
the trained 4B model exceeding GPT-4.1 (49.4 percent) and GPT-4o (42.8 percent) despite being 50 times smaller
大多数人认为AI模型的大小与性能直接正相关，更大的模型必然表现更好。但作者展示了一个仅40亿参数的模型通过强化学习训练后，性能超越了比它大50倍的GPT-4.1和GPT-4o，挑战了当前AI领域'参数规模决定一切'的主流观点。
The most intriguing result in the present study is thepositive effect of white noise on performance for theADHD children. This noise effect was present in boththe non-medicated and medicated children. Thissupports the MBA (Moderate Brain Arousal) model(Sikstro ̈m & So ̈derlund, 2007), suggesting that theendogenous (neural) noise level in children withADHD is sub-optimal. MBA accounts for the noise-enhancing phenomenon by stochastic resonance(SR). The model suggests that noise in the environ-ment introduces internal noise into the neural sys-tem through the perceptual system. Of particularimportance, the MBA model suggests that the peakof the SR curve depends on the dopamine level, sothat participants with low dopamine levels (ADHD)require more noise for optimal cognitive performancecompared to controls.
Author's self-described "most intriguing result"
The MBAmodel predicts that noise enhances memory perfor-mance for ADHD and attenuates performance forcontrols. We will also argue for a link between theeffects of noise, dopamine regulation, and cognitiveperformance.
Prediction of Moderate Brain Arousal model and author's additional argument.
(7) ReconfigBehSci on Twitter: “@ToddHorowitz3 probably- and I think there are many interesting questions around why he is there and whether he should be there. But to answer those properly, looking at the performance of the model seems important and interesting to me- that is all I am saying” / Twitter. (n.d.). Retrieved March 6, 2021, from https://twitter.com/SciBeh/status/1324389147050569734
Behavior Engineering Model This page has a design that is not especially attractive or user friendly but it does provide an overview of Gilbert's Behavior Engineering Model. This is a model that can be used to analyze the issues that underlie performance. A six-cell model is presented. Rating 5/5
Human Performance Technology Model This page is an eight page PDF that gives an overview of the human performance technology model. This is a black and white PDF that is simply written and is accessible to the layperson. Authors are prominent writers in the field of performance technology. Rating 5/5