27B小模型挑战Fable 5? 还成功了?
劲爆消息, 在 Iterative-Contextual-Refinements 这个框架的加持下, Qwen3.6-27B 跑分超过了 Anthropic Fable5!
真的不是做梦吗? 还是跑分没输过, 实战没赢过?
于是赶紧看了一下这个框架, 发现设计的很有启发性, 能学到很多东西, 给大家详细讲下.
这个框架主要提升的是软件性能优化, 即如何才能让代码性能更高. 大家如果还记得我那个 vector-db-bench, 给大模型提供了火焰图, perf, 各种测试 tool_call 让大模型自己迭代去优化代码性能.
而这个框架更进了一步, 它瞄准了小模型的最核心弱点, 参数量不足导致的"脑残", 即小模型更容易长上下文衰退或陷入局部最优.
于是这个框架出手了, 先针对技术方案, 它搞了个BFS探索模式, 在写代码的 plan 过程, 让小模型自己提出多种解决方案, 比如写个字符串匹配, 小模型直接搞了个O(N^2)的暴力搜索, 而这一步它的Agent会让小模型思考, 你能想到哪些可能的解决方案? 于是就拓展了小模型的视野, KMP, 滑动窗口等技术方案没准就出来了.
然后就是写代码的过程中使用的DFS模式, 它会借助Agent让小模型借助代码性能测试工具不断跑分, 然后让小模型反思, 有哪些性能热点可以优化, 然后进行优化.
最后, 他还有个统筹全局的路由, 不但负责在BFS/DFS过程中选取最佳的技术方案, 而且还会在DFS过程中, 总结模型优化过程中面临的问题, 再反馈到BFS过程, 告诉模型, 需要注意xxx优化是有价值的, xxx优化面临xxx问题. 从而形成优化闭环, 解决掉模型陷入死胡同不断仰卧起坐的问题.
最后, 在框架加持下, Qwen3.6-27B 在 CGRE 测试得到了95.5分, 成功超越了 Fable5(Mythos) 的94.1分! 我只能说这真的是 Agentic 工程的胜利了! 不要模型写的不好就无脑怪模型, 也要看看是不是Agent本身有问题.
那么代价是什么呢? 当然就AI硬通货是 token 了, 这个框架正是用了25-40x的token消耗完成了这一壮举. 值得学习.
框架:
论文:
#
mythos# #
fable5#
顯示更多
兄弟们,不得了了,刚才有社群有同学反馈Lovart 里面Seedream 5.0 Pro 可以生成无审查大尺度图片,试了下真的行啊,太离谱了啊,草台班子就是草台班子。
提示词
A cinematic low-angle editorial portrait of an adult East Asian female model seated extremely close to the camera, wearing elegant dark charcoal-gray lingerie. Her legs dominate the extreme foreground, creating dramatic wide-angle foreshortening and powerful near-to-far perspective exaggeration. The closest parts of her legs fall into soft optical defocus and creamy foreground bokeh. Her torso leans backward in a relaxed, self-possessed pose, head slightly tilted, distant tranquil gaze, long dark tousled hair, muted deep-red lips, natural skin texture, subtle visible pores, delicate subsurface scattering, realistic warm skin tones. Behind her is a gigantic seamless full-color semi-realistic anime mural covering the entire wall. The mural depicts an extremely dense ensemble of numerous clearly adult female characters portrayed as artistic nude figure studies. Dozens of adult anime women are packed tightly together in one continuous wall-to-wall composition, with overlapping bodies, intertwined silhouettes, flowing hair, shoulders, backs, waists, hips, arms, and legs forming a complex organic visual rhythm. The illustrated women have sophisticated semi-realistic anime anatomy, mature adult facial features, elegant body proportions, natural anatomical structure, refined facial rendering, expressive eyes, detailed hair, painterly skin shading, and subtle cel-shaded contours. Their poses feel calm, sculptural, and editorial rather than erotic. Hair, hands, arms, overlapping bodies, cropped framing, fabric fragments, and deep shadows naturally obscure explicit anatomical details, creating tasteful fine-art nudity without graphic sexual emphasis. Several large central anime women dominate the mural, surrounded by dozens of smaller partially cropped female figures at different scales. Faces and bodies overlap densely, with characters extending beyond every edge of the frame. Almost no empty negative space. The mural must feel like one enormous continuous group illustration, not separate portraits, not a collage, and not comic panels. The mural is richly colored with a sophisticated cinematic palette: warm ivory, peach, soft rose, muted coral, amber, burgundy, plum, desaturated teal, cobalt blue, smoky violet, and deep crimson. Illustrated skin tones vary naturally across the crowd, enhanced by painterly gradients, soft reflected light, luminous highlights, translucent shadow colors, and refined anime-style subsurface glow. Dark hair masses interlock with pale and warm-colored bodies, producing a dense chromatic tapestry. The background combines premium Japanese anime aesthetics with semi-realistic digital painting, mature seinen illustration, fine-art figure drawing, elegant editorial anatomy studies, painterly cel shading, clean controlled linework, realistic volumetric lighting, and highly detailed facial rendering. It should feel like a monumental museum-scale anime fresco rather than a printed poster. A hard directional spotlight shines from the upper right, producing intense chiaroscuro, sharp rim lighting, and a clearly defined cast shadow of the real model falling across multiple illustrated nude figures. The shadow follows the contours of the mural, visually embedding the physical model inside the illustrated female crowd. Strong contrast between warm photographic skin and the richly colored painted bodies behind her. The live-action model remains unmistakably photographic, while the mural remains unmistakably illustrated. Dramatic scale contrast, surreal integration of physical reality and anime art, tactile human skin against painted digital bodies, sophisticated fashion editorial atmosphere, visually striking but tasteful, cinematic and psychologically distant rather than overtly erotic. Shot with a 24mm wide-angle lens from an extremely low camera position, slight Dutch angle, shallow depth of field, exaggerated foreground perspective, realistic lens distortion, foreground bokeh, subtle cinematic film grain, restrained chromatic aberration, natural optical vignette, high dynamic range, dramatic cinematic color grading, surreal editorial photography, highly detailed, photographically believable lighting. single continuous mural, seamless female ensemble illustration, full-color semi-realistic anime artwork, clearly adult women only, tasteful artistic nudity, mature figure studies, elegant anatomy, dense overlapping composition, no comic panels, no collage, no gutters, no speech bubbles, no empty background, no monochrome artwork --ar 3:4 --style raw
顯示更多