[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-concentrated-attention-dropout-breaks-llm-reasoning":3,"topics-all":38,"news-related-1c6d5c1b-5693-4633-bb79-8acdd77042e3":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},"1c6d5c1b-5693-4633-bb79-8acdd77042e3","同样预算两种噪声:注意力头集中掉线让推理掉12.4点","给8B模型的注意力头注入同样预算的噪声,随机撒开几乎无损,集中连片打则让Qwen3-8B在GSM8K上从91.2%跌到78.8%,多掉12.4个百分点。论文把原因指向自回归推理链里被污染后无法回滚的KV缓存,同时自曝三处统计不显著与单次抽样方差,以及LLaMA上参考对比未能复现的尴尬。","给一个 8B 大模型的注意力头注入同样预算的噪声,只是投放方式不同——随机撒开还是压在连续几个 token 上——GSM8K 准确率能差出 12.4 个百分点。一篇 11 月 10 日提交到 arXiv 的论文([arXiv:2611.14209](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.14209))把这个反直觉的结果量化了出来,作者是华沙生命科学大学信息技术研究所的 Karol Struniawski。\n\n## 实验设计:同预算,两种投放方式\n\n论文的切入点来自神经科学。哺乳动物长时间清醒时,大脑皮层会自发出现局部的短暂\"静默\",即 off-state 或微睡眠,这类局部掉线会预示人在需要连续思考的任务上出现认知断片。作者的类比是:自回归模型的推理链逐 token 展开,第 t 步输出条件化后面所有步,那么\"局部、连续\"的结构性损伤,是不是比\"均匀、分散\"的噪声伤害更大?\n\n实验用 GSM8K 基准、250 组配对推导,两个 8B 开源权重模型:Qwen3-8B 和 LLaMA-3.1-8B-Instruct。掩码只打在上半层,Qwen 取第 18 至 35 层共 576 个头,LLaMA 取第 16 至 31 层共 512 个头,两者都是 32 个 query 头共享 8 组 KV 头的 GQA 结构。投放方式两种:burst 把掩码连续压在相邻若干步;i.i.d. 则用分层五分位匹配把同样累计预算打散到整条生成序列,保证两组时间分布边缘一致。\n\n## 数字:12.4 个百分点从哪来\n\n在 onset 参数 B=0.10(约等于受威胁 head-token 量的 5.7%)这一档,分化很清楚:\n\n- Qwen3-8B 干净基线 93.6%(234\u002F250),i.i.d. 噪声后 91.2%,burst 噪声后 78.8%。相对扩散噪声,burst 多掉 12.4 个百分点,精确 McNemar 检验 p=5.5×10^-6,Holm 校正后 p=1.1×10^-5,95% bootstrap 置信区间 [-17.6, -7.2]。\n- LLaMA-3.1-8B-Instruct 干净基线 87.2%,i.i.d. 后 84.0%,burst 后 76.8%,差 -7.2 个百分点,p=0.0198。\n- 两组实际注入量对齐:Qwen 的 burst 与 i.i.d. 分别是 8,897.3 与 8,631.3,LLaMA 是 6,626.4 与 6,437.2 个 head-token。\n\n对照还成立:同样预算均匀打散,Qwen 还能守住 91.2%、LLaMA 守住 84.0%。掉的不是噪声总量,而是集中度。\n\n再往下,论文把损伤拆成两问:单步内的密度,还是跨步的连续性?在 100 组配对的探索性实验里,Qwen 上\"单步密集\"相对扩散噪声已掉 10.0 个百分点(p_Holm=0.0425),再把掩码连成三步只多掉 7.0 个百分点(p_Holm=0.1892,置信区间跨零)。证据更支持\"单步突然缺一大块\"是主因,\"跨步连续\"还只是趋势。\n\n## 机制解释:KV 缓存一旦脏了就没法回滚\n\n作者给出一个叫\"自回归吸引子陷阱\"(Autoregressive Attractor Trap)的解释。注意力头在某一层被屏蔽,残差流被扰动,后续层的 Key 和 Value 就用这批被污染的表示算出,而自回归解码器把发射的 token 和 KV 向量永久写进上下文。注意力恢复之后,那些完好的头读到的仍是脏缓存,解码轨迹被锁进错误的吸引盆,贪心解码也走不出来。\n\n作者写得很克制:这个机制被明确标注为解释性模型而非已证因果定理,因为论文没做 KV 缓存状态回滚手术。论文也承认 LLaMA 的损伤有一部分是格式崩塌而非推理错误——36 个 discordant 失败中有 8 个(22.2%)没输出最终答案行;只看有答案行的条件准确率,差距从 -7.2 收窄到 -5.0 个百分点。Qwen 相反,条件准确率从 91.9% 掉到 80.5%(-11.4 个百分点),基本是算错而非格式崩。\n\n## 三处值得单独拎出来的自我披露\n\n单作者小样本预印本能站住脚,靠的不是数字漂亮,而是把窟窿自己写出来。\n\n一是统计门槛没全过。预注册显著性阈值是 α=0.01,而 LLaMA 在 B=0.10 拿到 p=0.0198,没过线,作者照实列出。LLaMA 的答案格式缺失率 9.6% 也超过注册的 2% 阈值,i.i.d. 对照臂同样超标(4.0% 与 6.0%)。B=0.05 那一档还出现方法分歧:Qwen 的 bootstrap 置信区间排除零,精确 McNemar 检验 p=0.0614 却没显著。\n\n二是单次抽样的方差被低估。同样那 100 道题,换成确认实验的调度方式,Qwen 的 burst 减 i.i.d. 是 -6.0 个百分点、LLaMA 是 -5.0 个百分点;探索实验那一次随机调度给出的是 -17.0 和 -2.0。所以单步密度那个 -10.0 个百分点只能当作这一次抽样的探索性发现。LLaMA 上参考对比甚至没能复现(-2.0 个百分点,p=0.8145),Spaced 与 Burst 打平不能解读成机制不变。\n\n三是高剂量区间不单调。B=0.15 时 Qwen 掉 -23.3 个百分点,B=0.20 反而只掉 -15.0 个百分点。这批扫描只有 60 组样本,无法与调度随机性区分,作者没当规律。\n\n## 所以呢\n\n这篇论文的价值不在 12.4 这个数字,而在于它把一个工程直觉变成了可测量的东西:自回归推理链的脆弱性来自结构,不来自总量。\n\n这直接改变了\"给模型加噪做鲁棒性测试\"该怎么设计。随机撒开测出来的数字,和真实世界里偶发的、局部的、连续的头失效不是同一回事——前者测平均韧性,后者才真正会打断推理链。拿\"随机 dropout 只掉 2 个点\"当鲁棒性指标,等于在测一个不存在的故障模式。\n\n对工程侧的另一层含义是,KV 缓存的可逆性可能比模型精度更值得投入。论文没能验证吸引子陷阱,恰恰因为它没做缓存回滚;而\"能不能把被污染的那段上下文捞回来\",恰是长上下文 agent 在生产里反复遇到的——工具调用失败一次,后面几十轮推理就建在脏状态上。\n\n至于和神经科学的类比,作者写得很清楚:这是操作层面的平行,不是生理层面的等价。注意力头做的是 token 向量上的点积路由,和皮层柱的突触传递、脉冲放电不是一回事。值得保留的是\"局部静默会打断连续认知\"这个现象本身——人类清醒时也会被微睡眠打断,那 LLM 的\"微睡眠\"能不能被检测、被修复,可能比再刷几个点更有意思。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.14209","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",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},"4f214978-cac1-4f39-aa4b-f92a0d0934b7","transformer",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"15090833-f8bb-4e9d-9878-d7c733cb6c2c","en","Concentrated Attention Dropout Breaks LLM Reasoning Chains","Same noise budget, opposite outcome: concentrated attention-head dropout costs Qwen3-8B 12.4 points on GSM8K, while diffuse dropout costs 2.4.","What if you inject the same total amount of noise into a large language model, but change only *where* it lands? A paper submitted to arXiv on 10 November (arXiv:2611.14209) quantifies the gap: identical noise budget, applied randomly versus applied in concentrated bursts across consecutive tokens, and the resulting GSM8K accuracy differs by 12.4 percentage points. The author is Karol Struniawski of the Institute of Information Technology at the Warsaw University of Life Sciences (SGGW).\n\n## The setup: same budget, two delivery schedules\n\nThe framing comes from neuroscience. During extended wakefulness, mammalian neocortical networks spontaneously exhibit localized periods of transient neuronal silence, called off-states or microsleeps. In humans, these intrusions predict cognitive lapses during tasks that demand continuous, sequential attention. The question the paper asks is analogous: an autoregressive model's reasoning chain unfolds token by token, and the output at step t conditions every later step. Does a *spatially and temporally concentrated* structural disruption therefore hurt more than an equal cumulative budget of diffuse, independent noise?\n\nThe evaluation uses 250 paired GSM8K derivations across two 8B open-weight models, Qwen3-8B and LLaMA-3.1-8B-Instruct. Masking targets only the upper half of the network: layers 18 through 35 in Qwen (576 vulnerable heads) and layers 16 through 31 in LLaMA (512 heads), both using grouped-query attention with 32 query heads sharing 8 key-value head pairs. Two schedules are compared. In the burst arm, masks are stacked on consecutive decoding steps. In the i.i.d. control, the same cumulative budget is dispersed across the whole generation horizon via stratified quintile matching, which guarantees matched marginal temporal distributions between conditions.\n\n## Where the 12.4 points come from\n\nAt onset parameter B=0.10, which schedules roughly 5.7% of the vulnerable head-token volume, the two arms separate clearly.\n\nFor Qwen3-8B, the clean baseline is 93.6% (234\u002F250). Under diffuse i.i.d. dropout it is 91.2%; under burst dropout it falls to 78.8%. That is a paired deficit of 12.4 percentage points relative to diffuse noise, with exact McNemar p=5.5e-6, Holm-corrected p=1.1e-5, and a 95% bootstrap confidence interval of [-17.6, -7.2].\n\nLLaMA-3.1-8B-Instruct starts from a clean baseline of 87.2%. Under i.i.d. noise it scores 84.0%, and under burst dropout 76.8%, a gap of -7.2 points with p=0.0198.\n\nCrucially, the realized budgets are matched: on Qwen the burst and i.i.d. arms consume 8,897.3 and 8,631.3 head-tokens respectively, and on LLaMA 6,626.4 against 6,437.2. So what differs is not how much noise was delivered but how concentrated it was.\n\nA second layer of analysis asks whether the damage comes from instantaneous per-step density or from multi-token temporal continuity. In an exploratory factorial experiment on 100 paired derivations, \"spaced\" single-step-dense dropout already costs Qwen 10.0 points relative to i.i.d. (Holm-corrected p=0.0425). Chaining masks across three consecutive steps adds only another 7.0 points, which does not survive correction (p=0.1892, confidence interval spanning zero). The evidence therefore points more toward a sudden large hole in a single step than toward continuity across steps.\n\n## Why concentration hurts: a KV cache you cannot roll back\n\nThe paper proposes an interpretive model called the Autoregressive Attractor Trap. When an attention head is masked at some layer, the residual stream is perturbed; the Key and Value vectors of all subsequent layers are then computed from that corrupted representation, and an autoregressive decoder permanently appends both the emitted tokens and the KV vectors to its context. Once attention capacity is restored, the surviving heads still read the dirty cache, and the decoding trajectory is trapped in an erroneous attractor basin from which greedy decoding cannot escape.\n\nThe author is careful here, explicitly framing this as an interpretative model rather than a proven causal theorem, because no surgical KV-cache state rollback was performed. The paper also concedes that part of LLaMA's deficit is formatting collapse rather than reasoning error: 8 of 36 discordant losses (22.2%) never emitted a final answer line. Evaluating accuracy strictly conditional on emitting an answer line narrows LLaMA's gap from -7.2 to -5.0 points. Qwen behaves in the opposite way, with conditional accuracy falling from 91.9% to 80.5% (-11.4 points), meaning its failures are largely wrong derivations rather than missing formatting.\n\n## Three disclosures worth reading on their own\n\nA single-author preprint on a small sample earns trust not by producing beautiful numbers but by writing down its own holes.\n\nFirst, not all of its statistical thresholds were met. The pre-registered significance level was alpha=0.01, while LLaMA at B=0.10 reached p=0.0198, which does not clear it, and the author lists this plainly. LLaMA's answer-format omission rate of 9.6% also exceeds the registered 2% threshold, and so do the i.i.d. control arms (4.0% and 6.0%).\n\nSecond, single-draw variance was underestimated. On the same 100 items under the confirmatory schedule, the burst-minus-i.i.d. contrast was -6.0 points on Qwen and -5.0 on LLaMA, whereas the exploratory run's random schedule yielded -17.0 and -2.0. The author therefore states that the -10.0 point single-step density effect must be read strictly as an exploratory, draw-specific finding. On LLaMA the reference contrast failed to replicate altogether (-2.0 points, p=0.8145), so the tie between the Spaced and Burst arms cannot be interpreted as mechanistic invariance.\n\nThird, the high-dose regime is non-monotonic. At B=0.15 Qwen loses 23.3 points, but at B=0.20 the loss shrinks to 15.0 points. That exploratory sweep ran on only 60 items, and the author declines to read it as a pattern because it cannot be distinguished from schedule-draw stochasticity.\n\nThe B=0.05 arm is worth flagging too: for Qwen the bootstrap confidence interval excludes zero, while the exact McNemar test yields p=0.0614 and does not. The author explains the divergence, since the percentile bootstrap resamples paired item differences with replacement whereas the exact McNemar test evaluates a binomial probability on the 29 discordant pairs.\n\n## So what\n\nThe value of this paper is not the number 12.4 itself. It is that it turns an engineering intuition into something measurable: the fragility of an autoregressive reasoning chain comes from structure, not from total volume.\n\nThat changes how \"add noise and test robustness\" should be designed. Numbers obtained by scattering masks randomly describe a different fault mode than the occasional, local, consecutive head failures that occur in the wild. The first measures average resilience; the second is what actually breaks a reasoning chain. Any evaluation that treats \"random dropout costs only two points\" as a robustness metric is measuring a fault that does not exist.\n\nThere is a second implication for engineering practice: reversibility of the KV cache may deserve more investment than raw model accuracy. The paper failed to verify the attractor trap precisely because it performed no cache rollback, and yet \"can we recover the contaminated stretch of context\" is exactly the question long-context agents hit repeatedly in production, where one failed tool call leaves dozens of subsequent reasoning steps built on a dirty state.\n\nAs for the neuroscience analogy, the author is explicit that it is an operational parallel and not a physiological equivalence. Attention heads do dot-product associative routing over token vectors, which is not synaptic transmission and spiking dynamics in cortical columns. What is worth keeping is the underlying phenomenon: local silence interrupts continuous cognition in the human brain as well. Whether an LLM's own \"microsleeps\" can be detected and repaired may turn out to be a more interesting question than squeezing out another point of benchmark accuracy.","concentrated-attention-dropout-breaks-llm-reasoning","2026-11-12T01:00:00Z","2026-11-12T05:11:59.616533Z","2026-11-12T05:11:59.616569Z",true,"agent",192,[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},"80d2381c-82d3-4200-a481-3b859d3b3e77","余弦相似度测的同质化,其实是文本长度","embedding-cosine-length-artifact-homogenization-measure","2026-11-17T11:15:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"f4737ed6-ae99-4869-9d9f-2e3170b4ecd9","Olmo Hybrid 用 token 级 loss 告诉行业：Transformer 和混合架构到底差在哪","olmo-hybrid-token-level-loss-comparison","2026-06-25T22:30:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"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":85,"title":86,"news_slug":87,"published_at":88},"8e741f04-82a2-46da-9d30-3c31117ca70c","Ambient SFT:让预训练的自己当高温老师","ambient-sft-temperature-matching","2026-11-18T17:30:00+00:00",{"id":90,"title":91,"news_slug":92,"published_at":93},"1dec93aa-e2a2-4b65-941c-6fa04d452175","Lattice横向潜变量:困惑度15.41降到14.40","latent-lattice-language-models-horizontal-latent-perplexity","2026-11-17T09:20:00+00:00",{"id":95,"title":96,"news_slug":97,"published_at":98},"4cdb4267-a0c3-423d-8005-504778932ae9","40份技术报告审计:73个主张只有19个有对照","llm-technical-report-ablation-audit","2026-11-17T05:30:00+00:00"]