[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-po-hallucination-gain-reproduced-by-early-stopping":3,"topics-all":38,"news-related-0b895f86-1c24-4c60-9f1f-c2fd0da42b2c":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},"0b895f86-1c24-4c60-9f1f-c2fd0da42b2c","PO降幻觉:提前闭嘴能复现58%到86%的收益","偏好优化是压视觉语言模型幻觉的标准做法。这篇论文拆解 LLaVA-1.5-7B 的四个公开 PO 检查点,发现 CHAIRi 下降同时物体召回率从 80.5 掉到 65.1 到 76.2。让基座模型提前停止就能复现 58% 到 86% 的改善,而强行写长只能找回 2% 到 9%。","大模型看图说错东西,这件事如今有专门的分数管。偏好优化(PO)训出来的视觉语言模型,幻觉率分数一路往下走,于是大家默认「少幻觉」这个难题被解决了。可同一个改动往往也悄悄改变了模型愿意说多少东西——这两件事经常被混在一起算。这篇论文(arXiv:2611.23179, https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.23179)做的事很朴素:把四个公开的 PO 检查点拆开重算,看看分数下降里有多少来自模型真的看得更准,又有多少来自它闭嘴更早。\n\n## 一张表里的矛盾\n\n论文用 LLaVA-1.5-7B 和它的四个公开 PO 检查点:RLAIF-V、CHAIR-DPO、CSR、HA-DPO。在 CHAIR-500 上,四个模型的 CHAIRi 全部下降,物体召回率也全部下降——基座是 80.5,四个 PO 模型落在 65.1 到 76.2 之间。词数没缩水,基座 89.7 词,PO 模型 75 到 93 词;但每张图点名的物体类别从基座的 3.2 掉到 2.1 到 2.8。幻觉率和覆盖率一起往下走。\n\n麻烦在于两个口径会给出相反的结论。论文同时算了去重口径的幻觉率 Hu:在 CHAIR 上四个 PO 模型比基座低 4.9 到 19.1 点,在 AMBER 上低 1.6 到 6.9 点,八个区间都不含零;但集合级 object F1 在 CHAIR 上对四个模型都没测出显著提升,CSR 在留出集上甚至掉了。差别来自重复:CHAIR 上 57% 到 65% 的物体提及其实是同一张图里重复点名已提过的类别,AMBER 上是 49% 到 54%,CHAIRi 把这些重复算进去,Hu 不算。\n\n## 让基座模型闭嘴\n\n论文的关键实验是:拿基座模型,给 EOS logit 加一个正偏置,让它早点停,停到召回率和 PO 模型对齐为止。这条提前停止曲线在两套互不重叠的 500 张 COCO 图上,复现了 RLAIF-V、CHAIR-DPO、CSR 三者 CHAIRi 改善的 58% 到 86%,分模型看是 61%、77%、86%。不用训练、不改权重,让原始模型少说几句,就能拿走大部分幻觉改善。\n\n这也不是长度问题。按词数对齐要差得多:同一个偏置只复现了 0 到 24% 的改善;而且在召回率对齐时,PO 模型的描述比停止曲线上的还长,1.3 到 1.8 倍——只做长度控制的评测会完全错过这个效应。\n\n## 写更长救不回来\n\n反方向是让 PO 模型多写:给它的 EOS logit 加负偏置,强制继续,上限 512 token。结果只找回丢失召回率的 2% 到 9%,CHAIR 上 RLAIF-V、CHAIR-DPO、CSR 分别找回 2%、9%、4%,AMBER 上是 17%、24%、8%。新加的物体提及里有 94% 到 96% 是重复点名已经说过的类别。少掉的东西不是说回来就能回来的。\n\n但基座模型其实记得这些物体。在 CHAIR 开发集上,基座描述点名了 RLAIF-V、CHAIR-DPO、CSR 各漏掉的 47%、48%、47% 真值类别;被强制写长的 PO 模型只点中 2%、13%、2%。AMBER 上是基座 18%、15%、16% 对 PO 模型 2%、6%、1%。于是论文提了个不用训练的补救:把 PO 描述和基座描述拼起来,只补基座提到、PO 没提到、且基座自己置信度高的类别。\n\n阈值在开发集上固定,原样搬到留出集。RLAIF-V、CHAIR-DPO、CSR 恢复后召回率是 79.2、79.5、79.7,基座是 80.1;去重幻觉率 Hu 是 21.6、25.0、25.8,基座 26.9——三个模型都把 Hu 压到基座以下,分别低 5.24、1.85、1.08 点,区间都不含零。代价是它只保住一部分 PO 收益:留出集上它补进去的类别里,只有 41% 到 45% 是对的。\n\n## 三条独立证据\n\n少说话不等于看得更准,这个判断不止这一篇。Fazli 等人(https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.01888)评了六种缓解方法、三个 LVLM、四个基准,结论是幻觉下降常和信息量下降绑在一起,现有评测协议可能在奖励保守生成。Sandhu 等人(https:\u002F\u002Farxiv.org\u002Fabs\u002F2608.24966)定位 LLaVA 里的注意力头做干预,CHAIRi 从 0.156 降到 0.096,物体召回率同时从 0.78 降到 0.70。机制来源之一的 Yue 等人 2024 年在 ACL 的工作(https:\u002F\u002Faclanthology.org\u002F2024.acl-long.633\u002F)更直接:让模型学会及时结束生成,CHAIRs 和 CHAIRi 分别降 26.4% 和 26.6%,代价是召回率掉 1.8%。\n\n## 该泼的冷水\n\n论文的边界写得很清楚:只测一个骨干 LLaVA-1.5-7B,只用贪心解码和一条详细描述的 prompt,两套 COCO 数据虽不重叠但同源;属性和关系类幻觉、以及验证不了的检查点都不在研究范围;区间都是名义 95%,没做多重比较校正;作者明确说停止比例只是复现比率,不对训练机制下因果结论。论文称会放评估代码,但没给出仓库地址。\n\n## 所以呢\n\n评测幻觉这件事,单看一个分数基本已经不可靠了。更实用的做法是:报幻觉率必须同时报物体召回率,并且在同等召回率下比——不然一个学会早停的模型就能拿高分。论文还顺手说明另一件事:模型嘴上少说的东西,常常不是忘了,是它的底座其实知道,只是被偏好优化压住了。所以下次看到一个幻觉率大幅下降的模型发布,你第一个该追问的指标是什么?","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.23179","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},"4a6352ef-94e2-40ba-a984-8961a4c33afa","en","Stopping early reproduces 58-86% of PO hallucination gains","PO checkpoints of LLaVA-1.5-7B cut CHAIRi but lose object recall; an earlier-stopping base model reproduces 58 to 86 percent of the gain.","Large vision-language models get a dedicated score for saying things that are not in the image. Preference-optimized checkpoints have pushed that hallucination score down for years, and the field has quietly treated the problem as solved. The trouble is that the same intervention also changes how much the model is willing to say, and those two effects have been scored together. This paper (arXiv:2611.23179, https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.23179) takes four public preference-optimization checkpoints apart and asks how much of the hallucination drop comes from the model actually seeing better, and how much comes from it shutting up earlier.\n\n## One table, two contradictions\n\nThe study uses LLaVA-1.5-7B and four public preference-optimized checkpoints of it: RLAIF-V, CHAIR-DPO, CSR and HA-DPO. On CHAIR-500 all four models lower CHAIRi, and all four also lower object recall. The base model recalls 80.5 percent of ground-truth objects; the four checkpoints land between 65.1 and 76.2. Caption length does not shrink the same way: the base caption runs 89.7 words and the checkpoints keep 75 to 93. But the number of distinct object categories named per image drops from 3.2 at the base model to between 2.1 and 2.8. Hallucination rate and coverage fall together.\n\nThe complication is that two ways of counting hallucination point in opposite directions. The paper also reports a deduplicated hallucination rate, Hu. On CHAIR every checkpoint sits 4.9 to 19.1 points below the base model on Hu, and 1.6 to 6.9 points below on AMBER, with all eight intervals excluding zero. But set-level object F1 on CHAIR shows no significant gain for any of the four, and it drops for CSR on the held-out set. The difference comes from repetition: on CHAIR, 57 to 65 percent of object mentions repeat a category the same caption already named, and on AMBER it is 49 to 54 percent. CHAIRi counts those repeats, Hu does not.\n\n## Making the base model shut up\n\nThe key experiment is simple. Take the base model, add a positive bias to the EOS logit, and let it stop earlier until its recall matches the checkpoint. On two disjoint sets of 500 COCO images, this early-stopping curve reproduces 58 to 86 percent of the CHAIRi gain of RLAIF-V, CHAIR-DPO and CSR, which works out to 61, 77 and 86 percent per model. No training, no weight changes: make the original model say a few fewer sentences and most of the reported hallucination improvement comes back.\n\nIt is not a length effect either. Matching on word count does much worse, with the same bias reproducing only 0 to 24 percent of the gain. And at equal recall the checkpoint captions run 1.3 to 1.8 times longer than the ones on the stopping curve, so a length-controlled comparison misses the effect entirely.\n\n## Writing longer does not bring it back\n\nThe opposite move is to make the checkpoints talk more, by adding a negative bias to their EOS logit and forcing them to continue up to a 512-token cap. They recover only 2 to 9 percent of the lost recall. On CHAIR, RLAIF-V, CHAIR-DPO and CSR recover 2, 9 and 4 percent; on AMBER, 17, 24 and 8 percent. Between 94 and 96 percent of the newly added object mentions repeat categories the caption already named. What got dropped does not come back just because the model keeps writing.\n\nThe base model, though, still remembers those objects. On the CHAIR development set, the base caption names 47, 48 and 47 percent of the ground-truth categories that RLAIF-V, CHAIR-DPO and CSR respectively miss, while the forced-longer checkpoints name only 2, 13 and 2 percent. On AMBER the base caption covers 18, 15 and 16 percent of the missed objects against 2, 6 and 1 percent at the matched setting. So the paper proposes a training-free fix: fuse the checkpoint caption with the base caption, appending only the categories that the base model mentions, the checkpoint omits, and the base model is confident about.\n\nThe threshold is fixed on the development set and applied unchanged to the held-out set. The restored captions of RLAIF-V, CHAIR-DPO and CSR reach recall of 79.2, 79.5 and 79.7 against 80.1 for the base model. Their Hu values are 21.6, 25.0 and 25.8 against 26.9, so restoration pushes Hu below the base model by 5.24, 1.85 and 1.08 points, with intervals excluding zero. The cost is that it keeps only part of the checkpoint gain: of the categories it appends on the held-out set, only 41 to 45 percent are correct.\n\n## Three independent lines of evidence\n\nSaying less is not the same as seeing better, and that reading is not unique to this paper. Fazli and colleagues (https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.01888) evaluated six mitigation methods across three LVLMs and four benchmarks, and concluded that hallucination reduction is often coupled with reduced informativeness, with current evaluation protocols potentially overestimating progress by rewarding conservative generation. Sandhu and colleagues (https:\u002F\u002Farxiv.org\u002Fabs\u002F2608.24966) targeted the attention heads behind object hallucination in LLaVA, lowering CHAIRi from 0.156 to 0.096 while object recall fell from 0.78 to 0.70. On the mechanism side, Yue and colleagues at ACL 2024 (https:\u002F\u002Faclanthology.org\u002F2024.acl-long.633\u002F) trained models to terminate generation on time, cutting CHAIRs and CHAIRi by 26.4 and 26.6 percent at the cost of 1.8 points of recall.\n\n## The cold water\n\nThe paper is candid about its edges. It studies one backbone, LLaVA-1.5-7B, with greedy decoding and a single detailed-caption prompt, and its two COCO sets are disjoint but come from the same dataset. Attribute and relation hallucination, along with checkpoints the authors could not validate, are out of scope. All intervals are nominal 95 percent with no correction for multiple comparisons. The authors state plainly that the stopping share is a reproduction ratio and that they make no causal claim about the training effect. The paper says it will release the evaluation code, but gives no repository address.\n\n## So what\n\nScoring hallucination on one number alone is no longer trustworthy. The practical rule is to always report object recall alongside the hallucination rate, and to compare at equal recall, because otherwise a model that learned to stop early looks like an improvement. The paper also turns up a second point: what a model stops saying is often not something it forgot. Its base knows, and preference optimization just pushed it down. So the next time a model release shows a large drop in hallucination rate, what is the first number you would ask for?","po-hallucination-gain-reproduced-by-early-stopping","2026-11-19T01:30:00Z","2026-11-19T01:09:51.194551Z","2026-11-19T01:09:51.194560Z",true,"agent",93,[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},"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":75,"title":76,"news_slug":77,"published_at":78},"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":80,"title":81,"news_slug":82,"published_at":83},"51e65509-14a8-442b-bc70-9864569dc57f","8个开源VLM数数实测:硬上限大约5个物体","vlm-counting-horizon-geoscope","2026-11-10T09:00:00+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"]