[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-llm-search-narrower-local-news-latam":3,"topics-all":35,"news-related-76c3f55f-f5b7-491e-bbdb-cc004dbc02b9":64},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":21,"news_slug":28,"published_at":29,"created_at":30,"modified_at":31,"is_published":32,"publish_type":33,"image_url":14,"view_count":34},"76c3f55f-f5b7-491e-bbdb-cc004dbc02b9","牛津实测:AI 搜索比搜索引擎更窄,本地新闻少 11.6 点","牛津大学互联网研究所用 GDELT 事件库构造 1,901 条拉美国家时事查询,分发给 3 家搜索引擎与 4 个大模型,累计 49,207 次请求、8,834 个域名。论文实测:大模型搜索返回的新闻源比搜索引擎更集中,对本地用户还把本国新闻占比再压低 3.58 到 11.6 个百分点。","AI 聊天机器人正在变成越来越多人获取新闻的第一入口。论文引用路透研究所 2026 年的调查说,48 个国家里有 10% 的人已经在用 AI 聊天机器人看新闻,一年前这个比例是 7%,35 岁以下人群里达到 16%。而在全球南方,拉丁美洲占 AI 平台全球流量的 14%,按消费者使用规模排在第三位。\n\n问题在于,当入口从「搜索引擎加一堆链接」换成「一段话」,信源到底变多了还是变少了。牛津大学互联网研究所四个人(Francisco W. Kerche、Franziska Sofia Hafner、Felix M. Simon、Luc Rocher)11 月 19 日挂上 arXiv 的论文给了一个偏反直觉的答案:变窄了。\n\n实验规模不小。从 GDELT 事件库取 1,901 条涉及九个拉丁美洲国家的时事,让大模型按八种开头各写一条口语化候选提问再随机取一条,翻译成当地语言后分发给三个搜索引擎(Google、Bing、Yahoo)和四个自带联网搜索的大模型,累计 49,207 次请求,涉及 8,834 个不同域名。\n\n## 搜索引擎本来就窄,大模型又叠了一层\n\n论文引用的数据显示,Google News 各国首页的本地信源占比,巴西是 88.9%,墨西哥 78.3%,哥伦比亚 63%。全球超过 78% 的域名注册在美国和欧洲。\n\n论文自己的统计更直接。在国别内部比较里,本地用户拿到的有效信源数平均比外国用户少 42 个,Google 上这个差距是 55.6 个。跨国家口径下,只有 Google 还保持这个模式:本地用户 76.5 个,外国用户 124 个,Bing 和 Yahoo 在不同用户条件下变化小得多。\n\n大模型搜索这一层,论文测出的是继续变窄。它的判断是,GPT-4o 和 Perplexity Sonar 在任何用户条件下返回的有效信源数都低于搜索引擎。本地用户这一档,三个搜索引擎分别是 Google 24.3、Bing 53.6、Yahoo 44.9,四个大模型是 Claude 29.2、Grok 42.3、Perplexity 18.4、GPT-4o 16.1。跨国家口径差距更大:GPT-4o 33.9、Perplexity 46.9,搜索引擎一侧是 Google 76.5、Yahoo 142.5、Bing 156.2。用本地语言查询时,论文的说法是 GPT-4o 和 Perplexity 低于搜索引擎,Grok 高于搜索引擎,Claude 结果不一。\n\n## 本地用户拿到的本国新闻,又少了一层\n\n多样性之外,论文还量了本国信源占比。本地用户用本国语言提问时,搜索引擎返回的本地新闻本来就只有 41.3%,大模型搜索还要再低 3.6 到 11.6 个百分点。分模型看,GPT-4o 少 3.58 个点,Perplexity 少 11.6 个点,Claude 少 9.2 个点,Grok 少 9.5 个点。\n\n但这里有个容易被标题掩盖的反转:受损的不是所有人。旅游者场景下,GPT-4o 的本地新闻占比反而比搜索引擎平均高出 5.9 个点,外国用户场景下,GPT-4o 高 3.4 个点、Perplexity 高 1.8 个点。也就是说,论文的结论是效果取决于模型和用户类型,不是 AI 搜索一律更差。另外,如果只跟 Google 这一个引擎比,只有 GPT-4o 没出现显著下降。\n\n国家之间也不齐。论文说九个国家里只有巴西和阿根廷,所有大模型和搜索引擎的平均本地信源占比都在 50% 以上,而且这个占比大体跟着该国注册的 ccTLD 域名数量走。\n\n## 机制层:这不是第一次有人发现\n\n大模型带地理偏置这件事,此前已经有独立团队在不同场合测到过。ICML 2024 上 Manvi 等人的论文(https:\u002F\u002Fproceedings.mlr.press\u002Fv235\u002Fmanvi24a.html)给出过类似判断:大模型做零样本地理预测时,和真值的 Spearman 相关最高能到 0.89,但在吸引力、道德感、智力这类主观题目上,对经济条件较差地区的偏见相关最高到 0.70。另一篇 Global-Liar(https:\u002F\u002Farxiv.org\u002Fabs\u002F2401.17839)用地理和时间都均衡的数据集测 GPT 系列,发现模型偏向采信来自全球北方的表述,非洲和中东的事实准确率明显更低。\n\n这篇论文的新意不在「模型有偏见」,而在于把偏置定位到了检索层:不是模型怎么描述一个国家,而是你根本看不见哪些国家的报道。\n\n供给侧也在动。Zhao 和 Berman 的研究(https:\u002F\u002Farxiv.org\u002Fabs\u002F2512.24968)用双重差分方法发现,选择用 robots.txt 屏蔽生成式 AI 爬虫的大型出版商,网站流量相比不屏蔽的同行是下降的,同时会转向更难被大模型复刻的富内容形式。这条线索接上前文:当 AI 成为主要入口,出版商的第一反应是关门,那么关门之后谁被关上,就成了结构性问题。\n\n## 该泼的冷水\n\n论文自己列的局限有几条值得照抄。它明说结论未必能推广到这九个拉丁美洲国家之外,因为各地媒体生态和政府管控差异很大。更关键的是,所有交互都是单轮、无用户历史、无个性化的,而搜索引擎会改排序、模型在加记忆机制,这些跨天跨周的用户级效应一个都没测到。\n\n方法链路本身也带一层模型依赖:事件来自 GDELT,提问由 gpt-4o-mini 生成并翻译,域名分类用 Cloudflare 内容分类法、同样由模型判读。四个被测系统的知识截止时间也不一样,Grok 3 是 2024 年 11 月到 2025 年 2 月,GPT-4o mini 是 2023 年 10 月,Claude 3.5 Haiku 是 2024 年 7 月,Perplexity 未公布。论文说事件都挑在这些截止时间之外,但四个系统面对的其实是不同的检索基线。\n\n## 所以呢\n\n对中文读者来说,拉丁美洲是别人的考场,但机制是通用的:你看到的「本地新闻」正在由一个你看不见的选择过程决定,而这个过程既不比搜索引擎更宽,也不比搜索引擎更懂你。对做检索和聚合的团队,论文还附了一份开放数据集,GDELT 加 DataForSEO 的组合把复现成本压到了边际成本级别。真正该问的不是 AI 会不会取代搜索,而是谁决定了你看不见的那部分。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.25955","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18],{"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",[22],{"id":23,"lang":24,"title":25,"summary":26,"content":27},"1e956681-cc0a-430e-83bf-649f5839ce08","en","LLM search returns fewer, less diverse news outlets","An Oxford audit of 49,207 queries on 1,901 Latin American events found LLM search returns fewer and less diverse outlets than Google, Bing and Yahoo.","AI chatbots are becoming the first doorway to news for a growing share of people. The paper cites a 2026 Reuters Institute survey finding that 10% of people across 48 countries now use AI chatbots for news, up from 7% the year before, with 16% of under-35s reporting the same. In the Global South, Latin America accounts for 14% of global traffic to AI platforms, making it the third-largest market by consumer presence.\n\nThe question this raises is simple: when the entry point changes from a list of links to a single paragraph, does the reader end up seeing more sources or fewer? A paper from the Oxford Internet Institute, posted on arXiv on 19 November by Francisco W. Kerche, Franziska Sofia Hafner, Felix M. Simon and Luc Rocher, gives a counterintuitive answer: fewer.\n\nThe scale of the experiment is not small. The team pulled 1,901 current events involving nine Latin American countries from the GDELT database, had a model write eight candidate conversational queries with different openings before picking one at random, translated it into the local language, and submitted it to three search engines (Google, Bing, Yahoo) and four web-search-enabled language models. That produced 49,207 requests across 8,834 unique domains.\n\n## Search engines were already narrow; the models stack another layer on top\n\nData cited by the paper shows the local-source share on each Google News country homepage: 88.9% for Brazil, 78.3% for Mexico and 63% for Colombia. More than 78% of global domains are registered in the United States and Europe.\n\nThe paper's own measurements are more direct. Within a single country, local users receive on average 42 fewer effective news outlets than foreign users, and the gap is wider for Google at 55.6 fewer. Across countries, only Google still shows that pattern: 76.5 effective outlets for locals against 124 for foreigners, while Bing and Yahoo vary much less across conditions.\n\nOn top of that, the language-model layer narrows things further. The paper finds that GPT-4o and Perplexity Sonar return fewer effective news outlets than search engines regardless of user conditions. For local users, the three search engines come in at Google 24.3, Bing 53.6 and Yahoo 44.9, while the four models come in at Claude 29.2, Grok 42.3, Perplexity 18.4 and GPT-4o 16.1. The gap widens in the across-countries comparison: GPT-4o at 33.9 and Perplexity at 46.9, against Google 76.5, Yahoo 142.5 and Bing 156.2. For local-language queries, the paper says GPT-4o and Perplexity fall below the search engines, Grok comes out above them, and Claude shows mixed results.\n\n## Local users get even less domestic news on top of that\n\nBeyond diversity, the paper also measures the share of domestic sources. When local users search in their own language, search engines already return only 41.3% local news, and language models return a further 3.6 to 11.6 percentage points less. Broken down by model, GPT-4o gives up 3.58 points, Perplexity 11.6 points, Claude 9.2 points and Grok 9.5 points.\n\nBut there is a reversal here that headline framing tends to bury: not everyone is worse off. In the tourist scenario, GPT-4o's domestic share comes in 5.9 points above the search-engine average, and in the foreign-user scenario GPT-4o is 3.4 points above and Perplexity 1.8 points above. The paper's conclusion is that the effect depends on both the model and the user type, not that AI search is uniformly worse. Separately, when the comparison is made against Google alone, only GPT-4o fails to show a significant reduction.\n\nCountries also differ. The paper reports that among the nine countries, only Brazil and Argentina have an average domestic news share above 50% for all language models and search engines, and that this share generally tracks the number of ccTLD domains registered in the country.\n\n## The mechanism has been measured before by other teams\n\nThat language models carry geographic bias is not a new observation. A paper by Manvi and colleagues at ICML 2024 (https:\u002F\u002Fproceedings.mlr.press\u002Fv235\u002Fmanvi24a.html) reached a similar conclusion: zero-shot geographic predictions correlate with ground truth at Spearman's rho up to 0.89, but on subjective topics such as attractiveness, morality and intelligence, the bias against locations with weaker socioeconomic conditions reaches rho up to 0.70. Another paper, Global-Liar (https:\u002F\u002Farxiv.org\u002Fabs\u002F2401.17839), evaluated GPT models on a dataset balanced across geography and time and found the models privilege statements from the Global North over the Global South, with Africa and the Middle East at a clear disadvantage and much lower factual accuracy.\n\nWhat this paper adds is not the observation that models are biased, but the relocation of that bias to the retrieval layer. The issue is not how a model describes a country, but that certain countries' reporting never becomes visible to you at all.\n\nThe supply side is moving too. Research by Zhao and Berman (https:\u002F\u002Farxiv.org\u002Fabs\u002F2512.24968) uses a difference-in-differences approach and finds that large publishers who block generative AI crawlers via robots.txt see reduced website traffic compared to those who do not, while shifting toward richer content that is harder for LLMs to replicate. Connected to the above, when AI becomes the main doorway, the first response from publishers is to close the door, and once it is closed, who gets shut out becomes a structural question.\n\n## The cold water\n\nThe paper's own limitations deserve to be quoted directly. It states plainly that its findings may not generalise beyond the nine Latin American countries presented, because media landscapes and government control differ. More importantly, all interactions were single-shot with no user history and no personalisation, even though search engines adapt rankings and models increasingly add memory mechanisms that learn preferences across days and weeks. None of those long-horizon, user-specific effects were measured.\n\nThe method chain also carries a layer of model dependency: events come from GDELT, queries are generated and translated by gpt-4o-mini, and domain classification uses Cloudflare's content taxonomy, applied by a model as well. The four systems under test do not share a knowledge cutoff. Grok 3 runs from November 2024 to February 2025, GPT-4o mini is at October 2023, Claude 3.5 Haiku is at July 2024 and Perplexity does not publish one. The paper says events were chosen to fall outside those cutoffs, but the four systems still face different retrieval baselines.\n\n## So what\n\nLatin America is someone else's test ground for Chinese readers, but the mechanism is general: what you see as local news is now decided by a selection process you cannot observe, and that process is neither broader than the search engine's nor better at knowing you. For teams building retrieval and aggregation systems, the paper also ships an open dataset, and the GDELT plus DataForSEO combination pushes the cost of reproducing the audit down to a marginal level. The question worth asking is not whether AI replaces search, but who decides the part you cannot see.","llm-search-narrower-local-news-latam","2026-11-19T11:43:22Z","2026-11-22T01:14:36.215131Z","2026-11-22T01:14:36.215140Z",true,"agent",22,[36,46,55],{"slug":37,"tag_slug":38,"title_zh":39,"title_en":40,"intro_zh":41,"intro_en":42,"id":43,"is_active":32,"created_at":44,"modified_at":45},"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":47,"tag_slug":47,"title_zh":48,"title_en":49,"intro_zh":50,"intro_en":51,"id":52,"is_active":32,"created_at":53,"modified_at":54},"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":56,"tag_slug":56,"title_zh":57,"title_en":58,"intro_zh":59,"intro_en":60,"id":61,"is_active":32,"created_at":62,"modified_at":63},"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":65},[66,71,76,81,86,91],{"id":67,"title":68,"news_slug":69,"published_at":70},"4f5d6e7d-b06c-4331-8e13-b916fa82bd2e","论文实测:70B模型说谎,探针认的是措辞","persona-direction-lie-probe-measures-wording","2026-11-22T09:00:00+00:00",{"id":72,"title":73,"news_slug":74,"published_at":75},"872f163b-5ed4-4405-91aa-48c0fe34c25f","论文实测:模型在哪一层想通这个词,能预测人读它多久","resolution-depth-word-reading-time-llm","2026-11-22T05:00:00+00:00",{"id":77,"title":78,"news_slug":79,"published_at":80},"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":82,"title":83,"news_slug":84,"published_at":85},"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",{"id":87,"title":88,"news_slug":89,"published_at":90},"c4623b85-37f5-435e-ae98-e67eb72eb1dc","别再逐篇问大模型:AI写程序当评委,每主题0.17美元起","execjudges-llm-written-program-judges-retrieval-eval","2026-11-20T01:40:00+00:00",{"id":92,"title":93,"news_slug":94,"published_at":95},"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"]