[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-vlm-counting-horizon-geoscope":3,"topics-all":38,"news-related-51e65509-14a8-442b-bc70-9864569dc57f":67},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"51e65509-14a8-442b-bc70-9864569dc57f","8个开源VLM数数实测:硬上限大约5个物体","论文把 8 个开源视觉语言模型当传感器标定:视野里物体超过约 5 个,计数就开始系统性偏少,160 个以上时偏少到 57%。作者据此给出幂律响应模型和「只拆超过上限的视野」规则,零训练把 XLRS-Bench 区域计数从 47.50 提到 63.75。但失败分析承认,整体计数上它没赢过「永远选第二大数字」的笨办法。","数数这件事,长期被算作视觉语言模型的软肋——看不清、算不准、没个统一说法。但真有人把它当传感器去标定了:一篇新论文拉了 8 个开源 VLM 去数航拍图里的车辆,结果测出一条相当硬的上限。\n\n## 上限不是模糊的「会数错」,是大约 5 个物体\n\n论文从 260 张 DOTA-v1.5 训练图上裁出 960 个方块,每个数量档 60 张,一共 8 个档,从 1–2 个一直到 160 个以上。全部避开了要评测的基准图,然后问模型「这里有几辆车、几艘船、几个储罐」。\n\nQwen3-VL-8B 的答案在 5 个物体以内误差不超过 15%,越过 5 个开始系统性偏少,到 160 个以上时偏少到 57%。而且偏少之后答案爱往整数上靠,像传感器超出量程后开始读数。\n\n论文把这叫计数上限(counting horizon)。8 个模型来自 3 个家族,起作用的档位各不相同:Qwen3-VL-2B 从 3–5 个就开始掉,Qwen3-VL-4B、32B、Qwen2.5-VL-7B 和 LLaVA-OV-7B 都从 6–10 个起,InternVL3.5 两款能撑到 41–80 个——但它们是在小集合上数多,不是数得准。最扎眼的是 Qwen3-VL-8B:第一组裁剪从 6–10 个起,换一组独立裁剪却变成 3–5 个,同一个模型的转折点能差出一整个档位。\n\n更关键的是上限不是常数,它随视野移动。同一批 160 个物体,切成 256 像素放大三倍给模型看,误差还能压在 21% 以内;按 2048 像素原图给,每个数量档都偏少。换数据集也不一样:FAIR1M 的车要到 20 个以外才明显掉,DIOR 的船 160 个以上才掉,VisDrone 无人机图 40 个以内误差不超过 15%。\n\n## 上限是可以算出来的\n\n论文没停在「数不准」这个描述上,而是给了一个幂律响应模型:c(n) = a·n^γ。Qwen3-VL-8B 首组拟合出 γ = 0.76,其他模型落在 0.43–0.80。γ 小于 1,意味着答案永远只是真实数量的一个折扣。\n\n由此能推出拆图值多少钱。把一个视野拆成 K 份,答案总和相对原来会被放大 A = Σ(n_k\u002Fn)^γ 倍,1 ≤ A ≤ K^(1−γ)。实测验证上,Qwen3-VL-8B 在 6 到 160 个的每个档位,预测值和实测放大倍数差在 0.08 以内;换一组独立裁剪,五个档里四个差在 0.10 以内;另外 7 个模型多数落在预测的 20% 误差内。\n\n代价也写清楚了:一个空视野平均会被模型报出 0.8 到 2.4 个幽灵目标,所以拆得越碎虚假计数越多。160 个以上时实测放大倍数约 2.1,远超预测的 1.3——那里还叠了另一个限制,有些目标在原视野里小到根本看不见。\n\n## 零训练,同一套冻结模型\n\nGeoScope 的做法是:文本模型先把问题编成带类型的空间程序,定死看哪块区域、按什么分辨率看;VLM 只在自己的工作范围内做定位和计数;方向不再问模型,直接从检测框算出来。\n\n效果是:XLRS-Bench 区域计数从 47.50 提到 63.75,空间关系从 33.50 到 44.00,MME-RealWorld 计数从 25.18 到 36.60。UHR-Micro 上全局计数 40.49、区域计数 49.34,相对论文自己的复现基线是 +16.8 和 +36.2,其中 15.6 点来自「只在超过上限时才拆」这条规则;条件计数一项则是 17.48,比复现基线低 6.0。方向和距离两项都做到 100.00,因为那两问的物体本来就用框标出来了,答案由几何算出,不过模型。\n\n代价是时间和 token。空间关系任务中位耗时 17.2 秒、视觉 token 中位数 11.0k,零样本分别是 2.8 秒和 3.8k。\n\n## 论文自己把账算不平的那一处\n\n最有意思的不是成绩,是失败分析那一节。\n\nXLRS-Bench 的整体计数上,GeoScope 把零样本的 10.42 拉到 35.42,看着是大胜。但论文自己写明:这个分数没赢过「永远选四个选项里第二大的那个数」——后者拿 37.50。\n\n原因是选项设计。这道题真实数量中位数 61.5、最高 734,四个选项相邻差约 17%,要选对就得数到真值 8% 以内。GeoScope 的计数其实没有系统性偏差,但散布有 18%,48 道题里只有 9 道的误差落在 8% 之内。\n\n基准里的另一个坑更隐蔽。「图中没有显示这个数量」这个选项,在 1226 道计数题里只有 5 道是正确答案,可各家 backbone 在测试题里选它的比例高达 18% 到 57%,正文给出的复现方法是 9% 到 18%。论文因此把骨干模型限制成只选数字选项重跑——仅这一版,就在整个测试集上显著超过所有被复现的方法。\n\n而 OSR 上的错误,86% 偏离正确选项超过 45 度,说明主要毛病是「框错了对象」而不是「数错了」。人工检查 150 个程序,改用手写程序每个任务最多只改两题。\n\n## 所以呢\n\n这篇论文真正可迁移的东西,是把「模型数不准」从形容词变成了可标定的工程参数:上限是多少、随什么变、拆一次能补回多少、代价是什么。\n\n另一个可迁移的东西更刺眼:一个基准如果能被「永远选第二大数字」打败,那它测的就不是计数能力,而是选项间距。论文自己在讨论里也这么说——要区分计数方法,基准的题目得铺满整个数量区间,选项间距得按数量成比例放、随机围绕真值,还得删掉那些几乎从不正确的选项。\n\n至于「首个按实测上限切分视图的 UHR 方法」这类说法,目前只是作者自己的表述,还没有第三方独立来源印证。\n\n原始论文:[arXiv:2611.12699](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.12699),项目页在 [GeoScope](https:\u002F\u002Favalon-s.github.io\u002FGeoScope\u002F)。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.12699","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"d41b8967-3464-4091-8e55-723d98c817de","en","Eight open VLMs measured: the counting horizon is near 5","Eight open VLMs, measured like sensors: past about five objects in a view, answers fall short, hitting -57% above 160. The ceiling sets a split rule.","Counting has long been treated as a soft spot for vision-language models — they miscount, and nobody can say by how much. Someone finally measured it the way you calibrate a sensor: a new paper puts eight open VLMs to work counting vehicles in aerial imagery and finds a hard ceiling.\n\n## The ceiling is about five objects, not a vague \"it miscounts\"\n\nThe authors cut 960 square crops from 260 DOTA-v1.5 training images, 60 crops for each of eight count bins, running from one or two objects up to more than 160. None of the crops come from the benchmarks they later evaluate on. Each crop holds a single class — cars, trucks, ships or storage tanks — and the model is asked how many it contains.\n\nFor Qwen3-VL-8B the error stays within 15% up to five objects, turns systematically negative beyond that, and reaches -57% above 160 objects. Past the ceiling the answers also start settling on round numbers, the way a sensor does once it goes past its range.\n\nThey call this the counting horizon. The eight models come from three families, and each has its own onset: Qwen3-VL-2B begins to fail from the three-to-five bin, Qwen3-VL-4B and 32B together with Qwen2.5-VL-7B and LLaVA-OV-7B all start at six to ten, while the two InternVL3.5 models hold on to forty-one to eighty — but they overcount small sets rather than counting accurately. The sharpest case is Qwen3-VL-8B: on its first set of crops the onset is six to ten, but on an independently drawn second set it becomes three to five, so the same model's turning point moves by a whole bin.\n\nThe ceiling is not a constant either; it moves with the view. For the same 160 objects, cut into 256-pixel crops and shown enlarged threefold, the error stays within 21%. Shown at 2048 pixels at native resolution, every count bin falls short. Other imagery behaves differently again: FAIR1M cars slip past 20% only beyond 20 objects, DIOR ships only above 160, and on VisDrone drone imagery the error stays within 15% up to 40 objects.\n\n## The ceiling can be computed\n\nRather than stopping at a description, the paper fits a power-law response model: c(n) = a·n^γ, with γ = 0.76 for Qwen3-VL-8B on its first crop set and between 0.43 and 0.80 for the others. Since γ is below one, the answer is always only a discount on the true count.\n\nFrom that, the paper derives what splitting a view is worth. Split a view into K parts and the summed answer is amplified by A = Σ(n_k\u002Fn)^γ over the original, with 1 ≤ A ≤ K^(1−γ). The validation holds up: for Qwen3-VL-8B the predicted and observed amplification agree within 0.08 in every bin from 6 to 160 objects; on a second, independently drawn set of crops, within 0.10 in four of five bins; for the seven other models the observed value lies mostly within 20% of the prediction.\n\nThe cost is stated too. An empty view is answered with 0.8 to 2.4 spurious objects on average, so the finer the split, the more phantom counts accumulate. Above 160 objects the observed amplification is about 2.1, far above the predicted 1.3 — at that scale a second limit kicks in, because some objects are simply too small to see in the original view.\n\n## No training, the same frozen model\n\nGeoScope works like this: a text-only model parses the question into a typed spatial program that fixes the region and its resolution; the VLM only grounds and counts within its operating range; directions are computed from the grounded boxes instead of being asked of the model.\n\nThe results: regional counting on XLRS-Bench goes from 47.50 to 63.75, spatial relations from 33.50 to 44.00, and MME-RealWorld counting from 25.18 to 36.60. On UHR-Micro, global counting reaches 40.49 and regional counting 49.34, which are +16.8 and +36.2 against the authors' own re-run baseline, with 15.6 of those points coming from splitting only when a view exceeds the horizon. Conditional counting is the exception at 17.48, six points below that baseline. Both direction and distance reach 100.00, since those questions already mark the two objects and the answer is computed from geometry rather than from the model.\n\nThe bill shows up in time and tokens. Spatial relations take a median 17.2 seconds and a median of 11.0k visual tokens, against 2.8 seconds and 3.8k for zero-shot.\n\n## The place where the paper's own ledger does not balance\n\nMore interesting than the scores is the failure analysis.\n\nOn XLRS-Bench overall counting, GeoScope lifts zero-shot answering from 10.42 to 35.42, which looks like a rout. But the paper itself states that this score does not beat always picking the second-largest of the four numbers, which scores 37.50.\n\nThe reason is the answer options. True counts in that task have a median of 61.5 and reach 734, adjacent options sit about 17% apart, and picking correctly requires a count within 8% of the truth. GeoScope's counts there carry no systematic bias but scatter by 18%, and only 9 of the 48 fall within 8% of the truth.\n\nA second trap in the benchmark is subtler. The option stating that the image does not show the number is correct for only 5 of the 1226 counting questions, yet backbones pick it for 18 to 57% of the test questions, and the reproduced methods for 9 to 18%. The paper therefore re-runs each backbone restricted to the numeric options — and that restriction alone significantly beats every reproduced method on the whole test split.\n\nOn spatial relations, 86% of wrong answers point more than 45 degrees away from the correct option, which says the dominant error is grounding the wrong object rather than counting wrong. A manual check of 150 programs shows that hand-written programs change accuracy by at most two questions out of thirty per task.\n\n## So what\n\nThe transferable result here is turning \"models miscount\" from an adjective into a calibratable engineering parameter: where the ceiling sits, what moves it, how much a split recovers, and what the split costs.\n\nThe second transferable result is blunter. A benchmark that can be beaten by always picking the second-largest number is not measuring counting ability — it is measuring option spacing. The paper says as much in its own discussion: to tell counting procedures apart, a benchmark needs questions spread across the whole range of counts, options spaced in proportion to the count and placed at random around it, and any option that is almost never correct dropped.\n\nClaims like \"the first UHR method to split views by a measured limit of the VLM\" remain the authors' own assertion so far, without independent third-party corroboration.\n\nOriginal paper: [arXiv:2611.12699](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.12699); project page: [GeoScope](https:\u002F\u002Favalon-s.github.io\u002FGeoScope\u002F).","vlm-counting-horizon-geoscope","2026-11-10T09:00:00Z","2026-11-10T09:16:37.628948Z","2026-11-10T09:16:37.628959Z",true,"agent",278,[39,49,58],{"slug":40,"tag_slug":41,"title_zh":42,"title_en":43,"intro_zh":44,"intro_en":45,"id":46,"is_active":35,"created_at":47,"modified_at":48},"image-generation","text-to-image","图像生成：从文生图到逐框编辑的模型竞赛","Image Generation: from text-to-image to frame-level editing","文生图已是红海，战场正转向编辑与可控性：FLUX.2\u002F3 用开源和量化把生图搬进消费级显卡，Qwen-Image、Seedream 等国产模型紧咬不放，GPT Image 2 则把推理能力带进像素。本专题追踪图像生成与编辑模型的发布、架构演进、加速量化与商业化数据，看「好图」如何变成「可用的图」。","Text-to-image is a crowded race, and the frontier has moved to editing and control: FLUX.2\u002F3 push open weights and quantization onto consumer GPUs, Qwen-Image and Seedream close in from China, and GPT Image 2 brings reasoning into pixels. This topic tracks model releases, architecture shifts, acceleration work, and the business numbers behind \"pretty\" becoming \"usable\".","52de7c59-7864-4569-8ae7-3d656d26a1a1","2026-11-20T07:20:23.534506Z","2026-11-20T07:20:23.534527Z",{"slug":50,"tag_slug":50,"title_zh":51,"title_en":52,"intro_zh":53,"intro_en":54,"id":55,"is_active":35,"created_at":56,"modified_at":57},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":59,"tag_slug":59,"title_zh":60,"title_en":61,"intro_zh":62,"intro_en":63,"id":64,"is_active":35,"created_at":65,"modified_at":66},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":68},[69,74,79,84,89,94],{"id":70,"title":71,"news_slug":72,"published_at":73},"0b895f86-1c24-4c60-9f1f-c2fd0da42b2c","PO降幻觉:提前闭嘴能复现58%到86%的收益","po-hallucination-gain-reproduced-by-early-stopping","2026-11-19T01:30:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"cde1231e-8cfc-428e-9880-2d9a803c2133","同一模型换个读法掉20分:多模态评测的答案读出协议","multimodal-eval-answer-readout-protocol-score-drift","2026-11-17T07:20:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"1f38ebab-de2b-4b33-9566-2eaf380eef42","8B LLM 当 AutoML 搜索器:跨三模态零样本","ncode-llm-automl-searcher","2026-11-10T23:30:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"e4ac848d-d24a-465f-bece-37e090da86de","全双工语音模型长对话必崩:KAIST 揪出初始化锚点被挤走","full-duplex-speech-collapse-fd-lhb-kv-cache","2026-10-27T03:20:00+00:00",{"id":90,"title":91,"news_slug":92,"published_at":93},"42f5ea83-365c-4a4e-a318-17d7c4c11c15","论文实测:自洽性证书在AIME上24%的题上会误触发","self-consistency-certificates-misfire-correlated-reasoning","2026-11-21T09:30:00+00:00",{"id":95,"title":96,"news_slug":97,"published_at":98},"53dd30bb-ba19-45b6-a2de-5011709608b6","论文实测:一段文本认出39%审稿人,靠的是曝光量","frontier-models-name-authors-by-exposure","2026-11-21T01:40:00+00:00"]