💡 一句话总结：蚂蚁广见团队的 PolyOCR 想让 OCR 模型不只是「把字抄下来」，而是识别、定位、解析、抽取、理解、推理一把抓：6000 万实例按六大能力维度组织训练，后训练用 CGPO——先量出「教师在这个样本上到底比学生强多少、强得是否稳定」，再决定蒸馏开多重。OCRBench v2.1 英文榜 80.42 拿下第一，内部 KIE 甚至压过 GPT-5.6-Sol，还顺手把 OCRBench v2 的标注修了一遍。
导语：OCR 的尽头是「以文本为中心的视觉智能」
做 OCR 的朋友应该有感觉：这两年榜单刷得飞起，但刷榜的模型大多是「偏科生」——解析模型（MinerU、PaddleOCR-VL 这类）把 PDF 还原得炉火纯青，可你让它数一数图里有几个文字实例、或者对着图表做步推理，它就露怯；通用 VLM（Qwen3-VL、Gemini 这些）什么都懂一点，但细粒度认字、空间定位、密集文档抽取又不如专科生。
蚂蚁这篇报告给这个局面一个相当有野心的回应：PolyOCR——不是一个模型，而是一个家族（2B/9B 通用版 + 2.7B 多语言专精版），目标把 OCR 从「纯文本转写」扩成「文本中心的视觉智能」：识别、定位、文档解析、信息抽取、视觉文本理解、OCR 推理六大能力，全部塞进一个统一的指令跟随生成框架。想知道细节？资源在这：
想自己动手试？
- 📄 论文：arXiv 2609.37712
这篇论文到底想解决什么问题？
统一 OCR 基础模型有两个拦路虎，一个在数据，一个在训练。
数据侧：现有 OCR 数据集几乎都是按单一任务、单一领域、单一语言建的，拼起来覆盖碎片化、配比失衡——空间锚定的双语翻译、模板化 KIE、中文文档表格解析、场景文本定位、OCR 推理这些能力，要么没数据，要么标注贵到离谱。
训练侧更有意思：不同任务的学习难度、输出结构、优化目标天差地别——识别要逐字精确，定位要空间落点，抽取要格式严格，推理要语义理解。多任务混训时强势任务会淹没弱势任务，正迁移和任务干扰还会同时存在。更微妙的是 RL 阶段的一个死角，论文举了个特别生动的例子：
一张 1000 字符的金融单据，模型把金额 18,650 抄成了 18,850——就错这一个字符。8 个 rollout 的编辑相似度奖励全都是 0.999，一模一样。GRPO 的组内相对优势全是零：这个组对「纠正这个数字」贡献的梯度是零。
数字错了，奖励却看不出错——这叫等价奖励退化。转写任务里错一个字可能就是业务上的致命错，但 page 级奖励对它钝感。解药是有的：在线蒸馏（OPD）——让老师站在学生自己生成的路径上逐 token 给监督，「到这一位该是 6 不是 8」，信息密度完全不同。但直接加蒸馏又引入新问题：老师也可能 favor 1850 那个错误数字；更麻烦的是，等学生后来自己改对了、老师还在犯糊涂时，固定权重的蒸馏会拖着学生不让进步。
怎么让老师的监督「在该在的时候、以该有的力度」出现？这就是 CGPO 要回答的。
它的思路是什么？
📚 第一步：6000 万实例的「能力账本」
数据引擎两手抓：一手全谱收集（公开数据集 + 内部资源，按六大能力维度归账），一手定向补缺——对稀缺能力做可控合成和模型辅助标注：规则式计数合成直接继承已验证标注的精确答案（不引入模型标签噪声）；复杂推理数据用 DeepSeek-V4-Pro 基于已验证的 OCR 标注生成问答与思维链，再过任务一致性校验。所有管线共享三重质控：多模型一致性验证（分歧送人工）、结构感知验证（坐标/HTML/表格与渲染图对得上）、源头验证（翻译和推理链不得脱离源证据）。
6000 万 SFT 实例按四组配比：文本感知 32.5%、文档结构 30%、OCR 理解推理 22.5%、通用多模态 15%，再按三相课程推进——推理类占比从早期 10% 一路加到后期 35%，先练认字，再练动脑。
🎯 第二步：CGPO——给蒸馏装一个「能力路由器」
训练分两阶段：SFT 打底之后，进入 CGPO（Competence-Guided Policy Optimization）。它的骨架是 GRPO（一直在线），叠加一个样本级路由的在线蒸馏，路由逻辑分三步：
- 量出双方能力：教师在每个样本上离线采样 8 次、按任务验证器打分取均值，得到教师能力分 c_T(x)，训练前算好缓存（验证器按任务映射：文本用编辑相似度、定位用 IoU、公式用 CDM、表格用 TEDS、KIE 用字段 F1、开放问答用 LLM 裁判）；学生能力分 c_S(x) 用训练中的 8 个 rollout 在线估计。两者相减就是能力差 Δ(x)。
- 校准，别裸用估计值：8 次采样的均值毕竟有噪声，直接比大小不可靠。作者在校准集上用 A/B 独立划分拟合两个单调映射：一个从 c_T 预测「教师的真实质量」，一个从 Δ 预测「教师稳定赢过学生的概率」（平局算半分）。这一步纯离线推理，不训练任何东西。
- 按路由系数开蒸馏：α(x) = 教师质量 × max(2×胜率 − 1, 0)。读一下这个式子的脾气：胜率不到一半，蒸馏直接归零；教师只是「矮子里拔将军」赢过更差的学生（质量分低），权重也会被压住。最终损失 = GRPO + λ·α(x)·OPD。
训练动态验证了设计意图：路由系数的平滑值随训练一路下降，后期归零的决策越来越频繁——教师自己逐渐「下岗」，把优化权交还给验证器。这就是标题里「competence-guided」的含义：蒸馏权重跟着师生能力边界走，而不是跟着训练步数走。
🧭 第三步：顺手把考场修了——OCRBench v2.1
作者还审计了 OCRBench v2：部分参考标注有噪声、部分任务的计分方式与任务语义错位（公式、文档解析、翻译、推理这几类尤甚）。v2.1 做了人工核验的标注修正 + 按任务对齐的计分调整，任务覆盖不变——但要记住：旧 v2 的分数和新 v2.1 不可直接对比。
效果到底怎么样？
主战报（PolyOCR-9B）：
- OCRBench v2.1：英文 80.42 / 中文 78.37，英文全场第一，压过 GPT-5.6-Sol（72.92）和 Qwen3.8-Max（77.75）；HunyuanOCR-1.5 这种专用模型在 v2.1 上只有 32.63——旧 OCR 专精模型在「理解+推理」维度被拉开得很难看。2B 版拿到 72.33，已经摸到 GPT-5.6-Sol 的门槛。
- CC-OCR：82.43 全场第一，其中文档解析 71.71、KIE 94.02 两个单项都是第一。
- 内部 KIE 基准（学信网证明/身份证/录取通知书/护照）：Overall 93.60%，超过 GPT-5.6-Sol 的 90.98%——2B 版 93.49% 也几乎打平。
- MDPBench 多语言：2.7B 专精版 Overall 87.0%，超 GPT-5.6-Sol 0.4 个点、Gemini-3-Pro 0.6 个点，17 个语言组里 6 个第一、12 个前二。
- OmniDocBench v1.6（文档解析）：整页直推 91.57；外挂 PP-DocLayoutV3 版面检测后 95.52（公式 CDM 91.71→96.30）——但这个数字仍低于 OvisOCR2（96.58）、TeleOCR（96.87）这些解析专精模型。
CGPO 本身的消融很能说明问题（同 SFT 起点，CC-OCR 分数）：SFT 79.46 → 加 GRPO 81.72 → 加固定权重 OPD 82.18 → 换余弦衰减 82.32（但 OmniDocBench 反而掉到 90.12）→ CGPO 82.43，且三榜全面最高。按样本路由，赢过按步数调度。
能拿来做什么？
- 证件/票据结构化场景：KIE 93.60% 的成绩单（学信网、身份证、录取通知、护照）对国内证件类业务几乎是量身定做的参考。
- 多语言文档管线：2.7B 专精版 21 个语言/文字组的成绩，对出海业务的文档解析是个现成的选型参考。
- RL+蒸馏联训的调度范式：CGPO 的「验证器打分估能力 → 离线校准 → 样本级路由」三段式，和最近同期的 GAD-RL（按组奖励门控+衰减）思路同源但路由粒度更细——正在做后训练的团队可以直接把这两篇对着抄作业。
泼冷水时间
- 自修订的榜自己第一：OCRBench v2.1 的标注修正与计分调整都是本文做的，动机（噪声、计分错位）说得通，但「在自己修订的考场拿第一」这个循环需要第三方复核来闭环，目前没有。
- 文档解析深水区仍是专精模型的天下：OmniDocBench 整页直推 91.57、pipeline 95.52，距解析专精的第一梯队还有约 1 个点的差距——「六边形」意味着没有一条边是最长的。
- 内部基准占比高：KIE 是 in-house 四类证件，MDPBench 部分对照模型未评测（表中「–」），横向比较的完备度有限。
- CGPO 的隐性成本：教师要在全部训练样本上离线采样 8 次 + 构建校准集，教师还是 122B 级模型——省下的蒸馏浪费，换成的是前期的推理账单。
- 报告未附代码/权重发布链接，想复现只能等。
理性看待
PolyOCR-Venus 值得记住的有两件事。一是它把「OCR 基础模型」的能力边界从「转写 + 解析」正式扩到了「理解 + 推理」，并用六大维度的数据账本证明了这条路在 2B-9B 规模上走得通——OCRBench v2.1 上通用大模型和 OCR 专精模型的分数断层，被这个统一配方填平了大半。二是 CGPO 的路由思想：教师监督的价值不是常数，它取决于「教师此时此地是否真的比学生强、赢得是否稳定」——把这个判断做成校准过的概率再开蒸馏，比拍一个衰减曲线优雅得多。
统一的代价依然存在：最深的解析榜上它还不是第一。但方向已经清楚了——OCR 的竞争正在从「谁抄得准」转向「谁看得懂」，这一篇是很好的注脚。
（数据与结论均来自论文原文，解读部分为作者观点，欢迎交流。）
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
In summary, the PolyOCR of the ants' wide view team wants the OCR model to be more than "copying down" but rather to identify, locate, decipher, extract, understand, and reason: 60 million examples are organized in six dimensions, followed by CGPO — to measure "how strong and stable teachers are on this sample" and then decide how much distillation is left. OCRBench v2.1 80.42 takes first, and within KIE even presses GPT-5.6-Sol, and reworks the OCRBench v2.
Introduction: The OCR ends with "text-centric visual intelligence."
Friends who are OCRs should have the feeling that these two years of lists have come up, but most of the models on the lists are "scientific" — the decomposition model (MinerU, PaddleOCR-VL, etc.) has returned PDF to pure green, but you have allowed it to have a few examples of words in a number of graphs, or to make it look at graphs, and it's timid; the general VLM (Qwen3-VL, Gemini) has a little bit of an idea, but it is less nuanced, spatial positioning, and intensive document extraction than specialists.
Ant's report is an ambitious response to this situation: PollyOCR - not a model, but a family (2B/9B common version + 2.7B multilingual precision) with the objective of expanding OCR from "text transliteration" to "visual intelligence of the text centre": identification, positioning, document resolution, information extraction, visual text understanding, OCR reasoning, all of which is embedded in a unified command and production framework. You want details? The resources are here:
You want to try it yourself?
Thesis: arXiv 2609.37712
What is this paper trying to solve?
The unified OCR base model has two stopovers, one in data and one in training.
Data side: The existing OCR data sets are almost all created by a single task, single domain, single language, and are combined to cover fragmentation, ratio imbalance - spatial anchored bilingual translation, templateization of KIE, Chinese document table analysis, scene text positioning, OCR reasoning, either without data or with an indication of their value to the extreme.
The training side is more interesting: learning difficulty in different missions, output structure, optimization of target differentials - identification is word for word, location is spatially location, extraction is in a strict format and reasoning is semantic. While multi-mission mix-ups, strong missions can overwhelm vulnerable missions, and ongoing migration and disruption of missions can occur simultaneously. Even more subtle is a dead end of the RL phase, where the paper offers a particularly vivid example:
A 1000-character financial document, the model copied the amount 18,650 into 18,850 - the wrong character. Eight rollout editorial similarity awards are all 0.999, identical. The GRPO's group has zero comparative advantages: the group's contribution to “correcting this number” is zero.
The numbers are wrong, the rewards are not right — it's called the equivalent rewards for deterioration. A typographical error may be a fatal business mistake, but a page-level reward is blunt. The antidote is available: OPD-- to get the teacher to follow the path that the students themselves have created, "It should be six, not eight" and the information density is completely different. But direct distillation leads to a new problem: teachers can also fall 1850, the wrong number; even more troublesome, when students change themselves and teachers are still confused, the distillation of fixed weights drags students away from progress.
How can you let your teacher's supervision appear "at the time and with all due force"? That's what CGPO has to answer.
What's the idea?
First step: 60 million examples of "capable books"
Two hands on the data engine: first-hand spectral collection (open dataset + internal resources, accounted for by six capacity dimensions), first-hand orientation fills the gap - a manageable synthesis of scarce capacity and model-aided labelling: a rule-based synthesis of numbers directly inherits the exact answer to the validated label (no model label noise is included); complex reasoning data uses DeepSeek-V4-Pro based on validated OCR labels to generate questions and answers and think chains, and then verify mission consistency. All tubes share a triple mass control: multi-model consistency verification (disagreement delivery manual), structured perception validation (coordinates/HTML/tables against rendering), source validation (translation and reasoning chain cannot be separated from source evidence).
The 6 million SFT examples are based on four sets: text perception of 32.5 per cent, document structure of 30 per cent, OCR understanding reasoning of 22.5 per cent, generic multimodules of 15 per cent, then three-phase courses - reasoning of 10 per cent from the early stages to post-production of 35 per cent, first reading, then brain learning.
Step 2: CGPO -- a "capable router" for distillation
The training was conducted in two stages: SFT, after which it entered the CGPO. Its skeleton is GRPO, superseding an online distillation of a sample class route in three steps of logic:
- Measuring the skills of both parties: Teachers sampled offline on each Z111Z 8 times, graded the average values by the mission c T(x) and received a score of c T(x) before training. cache (Certifiers map by mission: text with editorial similarity, IoU with positioning, formula with CDM, table with TEDS, KIE with field F1, open with LLM judges); c S(x) with 8 rollout estimates in training. The difference between the two is the power gap (x).
- Calibration, do not use naked estimates: after all, the average of eight samples has noise and is not reliable in relation to size. The author uses the A/B separate division for the school quasiset to combine two single-map maps: one from c T predicts the "real quality of teachers" and one from Δ predicts the "probability of teachers to stabilize and win students" (half a score). This step is pure offline reasoning and does not train anything.
- Distillation by route factor: α(x) = teacher quality x max (2 x win - 1, 0). Read the temperament: less than half the odds, distillation is zero; the teacher is just a student who has won worse (low-quality) than the "General in the short" and will be pressurized. Final loss = GRPO + λ (x) OPD.
Training dynamics confirm design intentions: the smooth value of route coefficients declines along the course, and post-production zero decisions are becoming more frequent — teachers themselves gradually “decommissioned” and give optimisation back to the certifier. This is what the title of "competence-guided" means: distillation weights follow teacher-student boundaries rather than training steps.
Step 3: Fixing the exam - OCRBench v2.1
The author also audited OCRBench v2: Some references refer to noise, scores for some tasks and linguistic errors (formula, document resolution, translation, reasoning, etc.). v2.1 Manually verified label correction + the task is adjusted to the score for task alignment, and the task covers the same - remember: the score for old v2 cannot be directly compared with the new v2.1.
How did it work?
Main newspaper (PolyOCR-9B):
- OCRBench v2.1: 80.42 / Chinese 78.37, first in English, pressing GPT-5.6-Sol (72.92) and Qwen 3.8-Max (77.75); this specialized model of HunyuanOCR-1.5 has only 32.63 on v2.1 - the old OCR precision model is being pulled down in the `understanding + reasoning' dimension. The 2B edition is in 72.33 and the GPT-5.6-Sol threshold has been touched.
- CC-OCR: 82.43 First in the room, where documents are analyzed 71.71, KIE 94.02, both individual items are first.
Internal KIE benchmark (Customs/identity card/entry note/passport): Overall 93.60%, 9098% - 2B 93.49%, more than GPT-5.6 - Sol.
MDPBench Multilingual: 2.7B Edition Overall 87.0%, Super GPT-5.6-Sol 0.4 Point, Gemini-3-Pro 0.6 Point, 6 first and 12 first two in 17 language groups.
- OmniDocBench v1.6 (document resolution): page round: 91.57; pep-DocLayoutV3 external: 95.52 (formula CDM 91.71:96.30) - but this figure is still lower than OvisOCR2(96.58) and TeleOCR (96.87), which are deconstructed precision models.
The collapse of the CGPO itself (with the SFT starting point, CC-OCR fractions): SFT 79.46 → with GRPO 81.72 → with a fixed weight OPD 82.18 → with a change of cosine decline 82.32 (but OmniDocBench fell to 90.12) →CGP 82.43, with the highest overall ranking. By sample route, through step-by-step movement.
What can I do with it?
- Structured scene of documents/papers: KIE 93.60% of the report cards (clinics, identity cards, admission notices, passports) are almost customized references for domestic document-type operations.
- Multilinguistic document line: 2.7B performance in 21 language/text groups, an off-the-shelf reference for document resolution for offshore operations.
- Movement control paradigm for RL+ distillation combination training: CGPO's "Certifier scalability offline calibration sample class route" three-part pattern, and the recent GAD-RL (group-incentive door-control + decay) line at the same time, although the route size is much smaller — teams that are doing post-training can directly turn the two into copying.
Cold Water Time
- Self-revised list number one: amendments to the labels and points for OCRBench v2.1 are done here. Motives (noise, miscalculation) make sense, but the cycle of "first in a self-revised examination" requires third-party review to close the loop, which is not currently available.
- Document analysis of the deep water area remains a specialized model: OmniDocBench pushes the whole page 91.57, Pipeline 95.52, with a gap of about one point from the first stage of decomposition - the "hexagon" means that none of the sides is the longest.
- High percentage of internal benchmarks: KIE is in-house four types of documents, MDPBench part of which is not assessed against the model (table –), and the horizontal comparison is limited in completeness.
- The hidden cost of the CGPO: the teacher has to take offline samples eight times on all sample and build the calibration collection. The teacher is also a 122B grade model, which saves the waste of distillation and replaces it with a prior reasoning bill.
- The report does not contain a code/weight publish link, which can only wait to reproduce.
A rational view.
PolyOCR-Venus is worth remembering two things. The first is that it formally extends the capacity boundary of the OCR Foundation from "translation + resolution" to "understanding + reasoning" and that it is supported by six dimensions of data books that show that this path works on the scale of 2B-9B - the fraction faults of the generic LLM and OCR precision models on ORR. Secondly, the CGPO's philosophy: the value of teacher supervision is not constant, but depends on whether "teachers are really stronger and more stable than students at this point in time" — a much more elegant way to distill this judgment as a calibrated probability.
The price of unity remains: it is not the first on the list. But the direction is clear — the competition for OCR is moving from “who can copy” to “who can read”, a good note.
(The data and conclusions are taken from the original paper, read in part as the author ' s views and welcome the exchange.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.