维基百科资深编辑威胁罢工抗议裁员风波

维基百科作为互联网上备受信赖的知识平台，由全球志愿者无偿维护。上周，支持其运营的非营利组织维基媒体基金会突然裁撤了一个小型但关键的工程师团队，此举立即引发社区震动。志愿者编辑和贡献者认为，裁员不仅切断了基金会与编辑社群之间的紧密纽带，更涉嫌打压工会权益，引发广泛质疑。经过多日激烈讨论，数百名活跃的资深维基百科编辑已公开表态支持罢工行动，以表达不满。然而，在一个以自愿贡献为主的平台上，罢工如何具体实施仍是未知数，凸显了非营利组织与社区之间平衡治理的挑战。 #维基百科 #编辑罢工 #裁员争议 #工会权益 #互联网信任 #科技新闻

AI初创公司免费清洁服务以训练家用机器人

AI训练初创企业Shift近日推出一项特殊服务：为用户提供免费的家庭清洁。其核心模式是通过录制清洁人员工作过程的视频，来训练未来能够执行家务的机器人。具体而言，该公司会派遣人员上门进行全面清洁，包括擦拭、吸尘、除尘、整理和清洗等工作，同时对全程进行拍摄记录。这些获取的宝贵数据将用于人工智能模型的训练，以期开发出能完成复杂家务劳动的智能机器人。此举旨在利用真实场景数据加速家庭服务机器人的研发进程。 #AI #机器人 #初创公司 #免费服务 #科技新闻 #家庭清洁 #数据训练

How to Stress-Test LLM Judges Fairly

A Fixed-Budget, Cluster-Aware Standard for LLM-as-a-Judge Evaluation: A Multi-Hop RAG Stress Test | alphaXiv []( []( []( []( []( []( We're hiring Paper 4 / 11 Hide Tools ⌘ / Open Tools A Fixed-Budget, Cluster-Aware Standard for LLM-as-a-Judge Evaluation: A Multi-Hop RAG Stress Test A Fixed-Budget, Cluster-Aware Standard for LLM-as-a-Judge Evaluation: A Multi-Hop RAG Stress Test Camilo Chacón Sartori*José H. García Catalan Institute of Nanoscience and Nanotechnology (ICN2), CSIC and BIST, Campus UAB, Bellaterra, Barcelona, Spain @ @ Code and materials: *Corresponding author Abstract Retrieval-augmented generation (RAG) sys- tems are often compared by asking a large lan- guage model (LLM) judge which answer is better. For multi-hop RAG, this has become a measurement problem as much as a modeling problem: the same score can reflect retrieval

quality, answer length, lexical overlap, or a sta- tistical test that ignores clustered data. We ask what happens when these choices are made ex- plicit. We propose a minimum measurement stan- dard for LLM-as-a-judge comparisons in RAG. The standard fixes the top-100 candidate pool, evidence budget, answer cap, generator, and prompt; it also requires pre-registered hypothe- ses, cluster-aware inference, an exact clus- ter sign-flip check when feasible, and second- judge replication. Clustered benchmarks can overstate progress; the field should adopt this standard. We stress-test it with Genetic Algo- rithm Decoder for Multi-hop Evidence Compo- sition (GADMEC), an evolutionary evidence selector, on 400 multi-hop questions in com- puter science/machine learning (CS/ML) and Materials Science. The protocol changes the empirical story.A binomial test makes all four semantic-baseline comparisons look sig- nificant; cluster-aware inference leaves only one Bonferroni-significant result. BM25 beats pure semantic GADMEC under the same bud- get, while a lexical-semantic hybrid recovers in CS/ML and narrows the Materials Science gap. 1 Introduction RAG systems are increasingly compared with pair- wise LLM-as-a-judge protocols: two answers are shown to a judge model, and the preferred answer is counted as a protocol is attractive because it is simple and cheap to also hides important choices. A method can look bet- ter because it selected better evidence, because it induced longer answers, because it matched lexi- cal cues that dense retrieval missed, or because the statistical test treated clustered examples as inde- pendent. In multi-hop RAG, these mechanisms are easy to mix. This paper asks whether the headline of a RAG comparison survives when those choices are con- trolled. In our experiment, the answer is mixed. A binomial test would make all four pre-registered semantic-baseline comparisons look significant. Cluster-aware inference leaves only one compar- ison significant after Bonferroni correction, with two more significant only before correction. The empirical story therefore depends not only on the selector, but also on the measurement protocol used to evaluate it. We make two claims. First, pairwise LLM-as-a- judge evaluation for multi-hop RAG is more fragile than standard reporting suggests. Second, the field should adopt a minimum measurement standard for cluster-structured LLM-as-a-judge benchmarks. The standard fixes evidence and answer budgets. It separates confirmatory from exploratory analy- ses through pre-registration and a public deviation log, uses cluster-aware inference, adds an exact sign-flip check when the cluster count permits it, and replicates headline results with a second judge. These components are not individually new. The contribution is to make them operate together in one controlled RAG comparison and show how the conclusions change. We stress-test the standard with Genetic Al- gorithm Decoder for Multi-hop Evidence Com- position (GADMEC), a Biased Random-Key Ge- 1 arXiv:2605.27789v1 [ ] 27 May 2026 []( " " "mailto: @ ")[](mailto: @ "mailto: @ ")[]( " " netic Algorithm (BRKGA; Gonçalves and Resende, 2011) for evidence subset selection. GADMEC is the instrument, not the protagonist. Its evolu- tionary search is separated from the decoder that enforces budget and diversity constraints, so a random-fitness ablation can test whether the fitness function contributes beyond the decoder machin- ery. Contributions. 1.

A fixed-budget evaluation design for LLM- as-a-judge comparisons in RAG: same top- 100 candidate pool, 2000-token evidence bud- get, 300-token answer cap, and generator set- tings for all methods. 2. A pre-registered analysis protocol:four Bonferroni-corrected primary hypotheses, an addendum for length matching and ablations, and a deviation log for post-hoc analyses. 3. A cross-domain benchmark: 687 arXiv pa- pers (2024–2026), a 3-level taxonomy, 10 cross-subfield combinations per area, 20 ques- tions per combination, and 200 contrastive multi-hop questions per area. 4. A methodological demonstration: the head- line of a multi-hop RAG comparison can change when inference respects the cluster binomial test would have reported all four primary tests as signifi- cant; wild-cluster bootstrap leaves only one Bonferroni-significant result. The same proto- col exposes a lexical-vs-semantic axis: BM25 beats pure semantic GADMEC, while a hy- brid recovers in CS/ML. The paper unfolds as follows. Section 2 situates fixed-budget evaluation relative to LLM-as-a-judge evaluation, multi-hop question answering (QA), subset selection, and pre-registration. Sections 3 and 4 define GADMEC, the baselines, and the controlled evaluation protocol. Section 5 reports the pre-registered results, length controls, content- distance diagnostics, and ablations. Sections 6–8 interpret the mechanisms, conclude, and state the limitations. 2 Related Work LLM-as-a-Judge Evaluation for RAG. Reference-free LLM-as-a-judge evaluation is now common in RAG formalised the setup for retrieval-augmented generation (Es et al., 2024), and recent surveys catalogue its variants (Li et al., 2025). The closest concern for this paper is length bias (Zheng et al., 2023; Dubois et al., 2024). We control both input and output budgets, then ask what signal remains after length matching. Multi-hop QA , 2Wiki- MultiHop, MuSiQue, and GRADE target chained- evidence multi-hop reasoning (Lee et al., 2025). Our questions have a different shape:contrastive composition(“how do X and Y differ on as- pect Z”). This makes evidence selection budget- sensitive because a good answer must cover several sub-aspects at once. Combinatorial subset selection in retrieval. Maximal Marginal Relevance (MMR; Carbonell and Goldstein, 1998) and Determinantal Point Pro- cesses (DPPs; Kulesza and Taskar, 2012) are clas- sical instances. Genetic algorithms have been ap- plied to RAG primarily for adversarial attack and corpus poisoning (Cho et al., 2024).We use a BRKGA constructively: the candidate pool is the search space, and the selected subset is the evidence plan. Pre-registration in -registration is rare in natural language processing (NLP). We follow social-science practice with a primary registration, timestamped addenda, and a deviation log separat- ing confirmatory and exploratory analyses. Together, these strands motivate an evaluation instrument rather than only a new retriever. The selector is allowed to vary; the evidence budget, answer cap, judge protocol, and evidential status of each analysis are not. 3 Method We first define the selector, then the controls that make the comparisons interpretable. Figure 1 sum- marizes the full pipeline. The key rule is simple: every method sees the same candidates and uses the same generation settings. Only the evidence plan changes. 3.1 GADMEC fitness GADMEC’s fitness over a candidate evidence plan P combines five components: f(P) =α COV+β DIV+γ COST+δ COH+ε SUB 2

[]( "Gonçalves and Resende")[]( "2011")[]( "Es et al.")[]( "2024")[]( "Li et al.")[]( "2025")[]( "Zheng et al.")[]( "2023")[]( "Dubois et al.")[]( "2024")[]( "Lee et al.")[]( "2025")[]( "Carbonell")[]( "and Goldstein")[]( "1998")[]( "Kulesza and Taskar")[]( "2012")[]( "Cho et al.")[]( "2024")[]( "Figure 1") 1. Corpus construction Taxonomy 3 levels 2 areas×8 subareas×5 topics arXiv search per combo 700 candidate combos 10 approved combos / area 5 TOP + 5 NICHO 687 papers chunked + embedded 2. Question generation Contrastive prompt “How do X and Y differ in. . . ” Sample 3 papers / combo ×2 paper triplets GPT-5.4-mini T= 0.7 200 questions / area multi-hop, no method names 3. Evaluation Retrieval top-100 chunks by cosine 4 evidence plans GADMEC / Greedy / MMR / BM25 GPT-5.4-mini answers T= 0, seed=42 Opus 4.7 judge (API) randomized A/B positions Figure 1: GADMEC pipeline. The protocol first builds a taxonomy-filtered corpus, then generates contrastive multi-hop questions, and finally evaluates four selectors under the same candidate pool, evidence budget, generator, answer cap, and judge (Claude Opus 4.7, randomised answer-A/answer-B positions). The only intended source of variation is the evidence selector. We re-judge the six headline comparisons with a second strong proprietary judge from a different provider (DeepSeek V4 Pro, thinking mode) to assess inter-judge robustness (§5.5). Here COV rewards mean query similarity, and DIV rewards pairwise dissimilarity among selected is a normalised token penalty,COH is centroid–query similarity, and SUB measures cover- age of GPT-5.4-mini query sub-aspects embedded with all-MiniLM-L6-v2. We use α= 0.30, β= 0.15, γ= 0.00, δ= 0.15, ε= 0.40, with sub-coverage threshold 0.40. The BM25 lexical component ζ is 0 for pure se- mantic GADMEC and activated only in the hybrid analyses. 3.2 BRKGA decoder BRKGA (Gonçalves and Resende, 2011) repre- sents each plan as random keys in[0,1]n and leaves feasibility to a decoder. The decoder sorts chunks by key and accepts them greedily until the plan reaches the 2000-token budget. It also enforces minimum query similarity 0.15, redundancy thresh- old 0.80, and at least three k-means thematic clus- search uses population size 20, elite fraction 0.24, elite-inheritance probability 0.70, at most 50 generations, early stopping after 15 stag- nant generations, and seed 42. 3.3 Baselines We compare against greedy top-k budget-fill, Max- imal Marginal Relevance (MMR) with λ= 0.5 (Carbonell and Goldstein, 1998), and the BM25 lexical retriever (k 1= 1.5, b= 0.75). All meth- ods operate on the same top-100 cosine-pre-filtered candidate pool. Win-rate differences therefore re- flect how methods choose evidence from the same pool, not whether one method had access to better candidates. 3.4 Question generation and judging Question generation uses a contrastive multi-hop prompt with GPT-5.4-mini at T= area has 10 combinations,with 20 questions per use GPT-5.4-mini at T= 0, seed 42, evidence-only prompting, and max_completion_tokens=300. Claude Opus 4.7 judges randomised answer-A/answer-B pairs in a single pass using a fixed pairwise prompt asks for a global preference based on factual correctness, completeness, evidence sup- port, clarity, specificity of evidence-backed claims, multi-source synthesis, and coverage of the ques- tion are allowed but reserved for answers that are truly equivalent; ties are recorded and excluded from the win-rate denominator. 4 Experimental Setup

The setup applies the fixed-budget design to two scientific domains. Each pairwise judgment com- pares evidence selection under the same retrieval substrate, input budget, and answer-generation bud- get. 4.1 Corpus The corpus is organised by a 3-level taxonomy: area→subarea→topic. We use two areas, Com- puter Science/Machine Learning (CS/ML; 341 pa- 3 []( "§5.5")[]( "Gonçalves and Resende")[]( "2011")[]( "Carbonell and Goldstein")[]( "1998") Assistant My Notes Comments Similar New Chat History Highlight & Ask Select any part of the paper to ask specific questions Add Context Type @ to reference other papers and expand the discussion Additional • Explore the paper's Github implementation • See how others cite this work • Literature reviews • Community context Try asking "What's the intuition behind section 3.2?" Smart Search Style

Jony Ive’s funky Ferrari

Most people will never own, drive, or even sit inside a Ferrari Luce . (If you can, or do… hit us up.) There’s still no question that Ferrari’s first electric vehicle is one of the most interesting, surprising cars of the year. With a decidedly un-Ferrari look , and lots of new technology and designs courtesy of Sir Jony Ive, the Luce is a lot of big ideas in a single swoopy package. A lot of people really hate it .

智谱和Minimax，市值差了4000多亿港元

5月 29日收盘，当天最高上涨 23.18%的智谱，股价突然跳水，以-1.42%收盘，当日振幅超过20%。 截至2026年5月29日收盘，智谱市值突破7000亿港元，从年初至今累计涨幅超过1000%，也就是翻了10倍。同日，MiniMax微涨0.36%，收盘840港元，市值2634.54亿港元。 前者是后者的近2.7倍。但2025全年两家营收分别为7.24亿元和5.69亿元人民币，比值仅1.27倍。两家都在2026年1月登陆港交所，从财报数字来看都处于巨额亏损状态。 两家同期上市的公司，营收规模、主营业务高度相似，为何资本市场的态度会如此不同？ MiniMax在3月2日发布 2025年年报，智谱在3月31日发布，这是两家今年1月上市以来的首份年度财报。但是从2025年报的关键数据来看，智谱和MiniMax更像是一对处境相似的难兄难弟，并非实力悬殊的对手。 智谱全年总收入7.24亿元人民币，同比增长131.9%；MiniMax全年总收入7903.8万美元（约合5.69亿元人民币），同比增长158.9%。MiniMax的增速反而更快。亏损方面，智谱经调整净亏损31.82亿元，MiniMax经调整净亏损约2.51亿美元（约17.3亿元人民币）。 因为港股无强制季报要求，两家都没有发布正式的一季报，但管理层在业绩说明会和运营更新中披露了加速度惊人的最新数据。 智谱CEO张鹏在3月31日业绩会上透露：2026年一季度API调用定价累计上调83%，调用量仍增长400%，截至3月ARR突破2.5亿美元，“量价齐升”。 MiniMax5月28日披露最新运营数据：全球企业和开发者客户超100万（半年增长5倍），全球用户规模约3亿，过去两个月ARR实现超过100%的增长，若以其2月披露的ARR超过1.5亿美元为基数推算，MiniMax当前ARR已超过3亿美元。 用市销率（PS）衡量，智谱7000亿港元市值对应7.24亿元人民币收入，PS约890倍；MiniMax 2625亿港元市值对应7903.8万美元收入，PS约425倍。由

于股价和汇率每天波动，实际倍数会有差异，但两家公司都远高于成熟科技公司的常规估值区间。 如果改用最新ARR口径重算，智谱2.5亿美元ARR约合19.5亿港元，对应7000亿港元市值，约为360倍ARR；MiniMax若按当前ARR超过3亿美元、约合23.4亿港元计算，对应2625亿港元市值，约为110倍ARR。这个角度看，MiniMax相对智谱更便宜，但二者估值都处在极高区间。 作为参照，OpenAI最新一轮融资后的投后估值已达到8520亿美元。OpenAI的收入公开口径并不完全统一，若按2025年全年收入约130亿美元估算，对应市销率约66倍；若按年化收入超过200亿美元估算，对应倍数约43倍。即便采用更乐观的年化收入口径，智谱和MiniMax的估值溢价仍然十分突出。 从财报数字来看，只能得出两家公司都不便宜的结论，但是，完全无法说明，究竟为什么两家公司的估值差距能够如此大。 从K线图来看，两家公司的走势分歧发生在三月中旬。 3月18日，两家公司同步大涨。当天，阿里云宣布对AI算力、存储等产品最高涨价34%，百度智能云随后跟进，市场的判断是，AI推理需求、Token调用量和算力供需紧张正在被价格验证。港股AI产业链在当天午后集体拉升，智谱收涨19.47%，MiniMax收涨19.85%。 但这一天对两家公司的意义并不完全相同。MiniMax3月18日也恰好发布新一代模型M2.7，股价盘中一度涨超28%，最高触及1330港元，创下上市以来高点。 分歧随后开始出现。MiniMax在3月18日冲高后进入调整，此后再未有效回到3月18日高点。 智谱在短暂调整后，等来了新的基本面催化。3月31日晚，智谱发布上市后首份年报，管理层在业绩会上披露，截至2026年3月，MaaS API ARR已达到约17亿元人民币，约合2.5亿美元；一季度API调用定价提升83%，调用量仍增长400%。这组“量价齐升”的数据，强化了市场对智谱未来的想象。 “Token调用—API收入—毛利率改善”的商业闭环刚好踩中全球Coding Agent、Agentic AI和企业API调用放量的交易主线。研报也放大了这种预期。中金随后上调智谱目标价至900港元，理由是API ARR大超预期、Agentic AI需求放量，并上调2026年和2027年收入预测。 4月1日，智谱股价再次大涨31.94%，此后继续沿着新的上涨趋势线运行，并在4月中旬盘中突破1000港元，截止 5月 29日，股价已经突破 1500港元。 一位二级市场的资深分析师说：“现在市场最大的关注点是模型的coding能力”。 大洋彼岸OpenAI 的最强对手Anthropic已经成为智谱 AI的对标方向。5月28日，Anthropic宣布完成650亿美元H轮融资，投后估值9650亿美元，超过OpenAI此前8520亿美元的估值，成为全球估值最高的AI创业公司之一。 与Anthropic对标，“编程能力全球领先”是资本市场为智谱支付更高溢价的核心逻辑之一。过去两年AI编程领域引用最多的基准是SWE-bench系列评测：给AI一个真实GitHub仓库和一个bug报告，看它能否独立生成正确的修复补丁。 在SWE-bench Pro上，智谱GLM-5.1以58.4%的通过率拿下全球最高分，高于GPT-5.4的57.7%和Claude Opus 4.6的57.3%。MiniMax披露，M2.7在SWE-Pro上得分56.22%，并在VIBE-Pro、Terminal Bench 2等工程类评测中取得较高分数。两家公司的旗舰模型差距依然不大。 问题在于，这个benchmark 已经接近饱和了。2026年4月，OpenAI在官方博客中公开承认：SWE-bench Verified已无法有效区分前沿模型的能力差异。 2026年5月27日，旧金山数据公司Datacurve发布了DeepSWE，一个为揭示前沿编程Agent真实能力差异而设计的新基准。它与SWE-bench的关键区别在于：113道题全部由工程师从零原创编写，从未出现在任何公开代码库中（零数据污染）；任务横跨91个仓库、5种语言；提示词只有SWE-bench Pro的一半长度，但参考解决方案的代码量是后者的5.5倍，意味着模型没法靠记忆蒙对，必须真正理解跨文件依赖、自主规划修改路径。 结果如上图所示，GPT-5.5以70%±4%断层领先，Claude Opus 4.7得54%±5%，月之暗面Kimi K2.6 24%，小米MiMo V2.5 Pro 19%，智谱GLM-5.1 18%，DeepSeek V4 Pro 8%。 在DeepSWE 的官方博客中显示，MiniMax M2.7参与了评测，但没有显示在这个榜单上 。参与测评的共有 16个模型，榜单上只列出了前12个。 这个结果对中国大模型也不太友好，在长链路、零污染的真实工程任务上，中国大模型与GPT-5.5之间存在3-4倍的能力鸿沟。 当然，DeepSWE自身也有明确的局限性，发布之后也引起了不少争议。样本量113题偏小，±4-5%的误差带在低分区间影响很大。所有模型都用统一的mini-swe-agent框架测试而非各自的原生工具链，可能系统性压低每家的上限。审计SWE-bench Pro“作弊”行为的判定本身由另一个LLM完成，存在幻觉风险。 Datacurve作为给大模型提供训练数据的公司，与OpenAI的商业距离比与Anthropic更近，这让“GPT清白、Claude作弊，GPT更优”的结论多了一层需要特殊看待的利益背景。 但是从一个侧面也可以反映，在模型能力这件事上，仅靠 benchmark 很难给出评价。DeepSWE确实“能看出一定的问题”，但它同样“说明不了本质问题”，那 4000 亿港元的市值鸿沟，到底是为什么？ 如果一定要问这是不是“市场不理性”，答案可能更复杂。几个经典的金融理论，也许能解释这种背离。 第一个是耶鲁大学罗伯特·席勒提出的“叙事经济学”：资产价格常常不是被数字驱动，而是被能够传播的故事驱动。智谱的故事正好踩在全球 Coding Agent、Agentic AI 的交易主线上。MiniMax 的故事很难被压缩成一个单一的交易主题。在主题驱动的市场里，叙事越集中，越容易吸走资金。 再看一下同样在港股上市的小米和阿里。截至 2026 年 5 月 28 日，阿里巴巴股价约 121.8 港元，总市值约 2.3 万亿港元；2025 财年收入 9963.47 亿元，按收入算 PS 约 2 倍。小米集团截至 2026 年 5 月 27 日股价约 28.40 港元，总市值约 7400 亿港元，2025 全年收入4573 亿元，PS 还不到2 。 智谱在5月 29日超越小米市值，而小米的收入是智谱的约 630 倍。Qwen和Mimo 大模型也不差，但放在了阿里和小米这里，成了市场认为的“负资产”。 第二个是凯恩斯的“选美理论”。投资者买的是自己判断别人会去买的模型。这句话有点绕，当智谱被默认为中国大模型中“Agentic AI 王座”最可能的候选人，资金便会自我强化地向它聚集。 最后一个角度，是把这种估值看成“实物期权”，当 PS 高达 890 倍、ARR 倍数高达 360 倍时，市场可以说在买一张“赢家通吃”的看涨期权，王者拿走全部期权溢价；其它的，都按一家普通公司定价。 在市场选定的叙事之下，4000亿的差距才能自洽。 AI的发展，是一场“快速”的长跑。速度和耐力，都要求极高。每家大模型企业都在高强度的技术研发节奏中。资本市场，也在长跑竞速中剧烈波动。 本文来自微信公众号 “腾讯科技” ，作者：晓静，36氪经授权发布。 该文观点仅代表作者本人，36氪平台仅提供信息存储空间服务。

The Download: unlocking lithium and controlling Ebola

This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology. How a new extraction process could unlock the world’s lithium A new method for extracting lithium could cut costs and emissions from one of the world’s most important materials for EVs…

Trump Mobile’s website still says the T1 Phone is ‘American

Where’s the Trump phone? We’re going to keep talking about it every week . We’ve reached out, as usual, to ask about the Trump phone’s whereabouts. As the phone’s supposed launch continues, we noticed its website may be violating FTC rules.

Acer’s answer to the MacBook Neo is a $699 laptop with Intel chips and 8GB of RAM

Acer is aiming to take on the MacBook Neo more directly with a new Swift Air 14 , a new 14-inch laptop that starts at $699. The laptop will use Intel’s lower-cost Core Series 3 “Wildcat Lake” chips — either Core 5 or Core 7 models, both of which are six-core processors. It comes in green, pink, purple, and a gray-ish blue with an all-aluminum chassis While the Neo is often criticized for its limited 8GB of RAM, the Swift Air will start with the same amount and be configurable with up to 16GB.