ChartNet entrena a la IA para leer gráficos con más precisión ift.tt/vL0W6xp

Language models still can’t look at a chart and understand what “up” means because they process graphs as a grid of unrelated tokens rather than structured information. Researchers from MIT and IBM built ChartNet, a synthetic dataset that teaches models to interpret charts by exposing them to the underlying code and data tables, and the result pushed a modestly sized model past GPT‑4o on key chart tasks. See how specialized training might outperform brute‑force scaling for enterprise document intelligence: zpr.io/KQArsyBGQFEJ #AI #LLM #ChartNet

It sounds small, but charts are where a lot of knowledge hides. MIT’s ChartNet work teaches AI to read and rebuild charts more reliably. That matters for accessibility, research, and anyone who has ever stared at a terrible PDF wondering what the graph actually says.

The new ChartNet training dataset from MIT and @MITIBMLab could improve the accuracy of vision-language models that help analyze business trends or interpret scientific figures. news.mit.edu/2026/mit-resear…

MIT just trained a small open source model to read charts, and it beat the big commercial models. A team at MIT and the MIT IBM Watson AI Lab built ChartNet, a million chart synthetic dataset, and used it to fine tune small vision language models. On chart reading, the small models won. The bottleneck was not compute. The bottleneck was data. And data is something the open source community can produce at scale. Red Hat called it earlier this year. In 2026 the question is not "what model do I use." The question is, "How many agents do I have running?" Every business will run a swarm of them. A 3B or 7B model, fine tuned on one task with synthetic data, will outperform a 400B model asked to do that same task on the side. You do not need a 400B model to file an invoice. You need a 3B model that has seen ten million invoices and never gets bored. The trillion dollar capex is building the cathedral. The swarm is doing the parish work. So what is the swarm in your business going to do first?

MIT @techreview: MIT researchers teach AI to read charts with ChartNet, a million synthetic chart data points enabling open-source models to outperform larger rivals. Swedish/European AI aura: boosts EU access to robust, affordable chart understanding t… news.mit.edu/2026/mit-resear…

ChartNet: a 1.5M-sample open-source dataset for chart understanding With an aligned chart image, plotting code, CSV/table data, summary, and QA with reasoning, Improves chart reconstruction, data extraction, summarization, and chart QA across model sizes. Huge thanks to respected collaborators @kondic_jovana, @RogerioFeris, @AudeOliva, @ZihanWang123, and everyone involved! The dataset is out on Hugging Face. Test it, break it, and we'll see you at #CVPR2026 to discuss!

ChartNet, our paper developed with @IBM and @MIT has been accepted at #CVPR2026. It introduces 1.5M multimodal chart samples across 24 chart types and 6 plotting libraries. Core question: how do we train models to connect a chart image with the numbers, code, text, and reasoning behind it?🧵

🚨CVPR 2026 Accepted ChartNet: A Million-Scale, High-Quality Multimodal Dataset for Robust Chart Understanding in collaboration with @MIT & @IBM 1.5M-sample open-source dataset for robust chart understanding Each sample aligns a chart image, plotting code, CSV/table data, natural-language summary, and QA with reasoning Full breakdown👇: #CVPR26 #CVPR2026

My curated papers from this week Mathematical methods and human thought in the age of AI paper: arxiv.org/abs/2603.26524 Proofdoors and Efficiency of CDCL Solvers paper: arxiv.org/abs/2603.26286 On merge-models paper: arxiv.org/abs/2603.26570 Short proofs in combinatorics and number theory paper: arxiv.org/abs/2603.29961 From logπ to π: Taming Divergence in Soft Clipping via Bilateral Decoupled Decay of Probability Gradient Weight paper: arxiv.org/abs/2603.14389 SKILL0: In-Context Agentic Reinforcement Learning for Skill Internalization paper: arxiv.org/abs/2604.02268 Loop Control Management in Tightly Coupled Processor Arrays (TCPAs) paper: arxiv.org/abs/2603.28645 LongCat-AudioDiT tech report: github.com/meituan-longcat/L… LongCat-Next: Lexicalizing Modalities as Discrete Tokens paper: arxiv.org/abs/2603.27538 Kernel-Smith: A Unified Recipe for Evolutionary Kernel Optimization paper: arxiv.org/abs/2603.28342 Announcing 1-bit Bonsai: The First Commercially Viable 1-bit LLMs whitepaper: github.com/PrismML-Eng/Bonsa… blog: prismml.com/news/bonsai-8b Preference-Aligned LoRA Merging: Preserving Subspace Coverage and Addressing Directional Anisotropy paper: arxiv.org/abs/2603.26299 Realtime-VLA V2: Learning to Run VLAs Fast, Smooth, and Accurate paper: arxiv.org/abs/2603.26360 DIAL: Decoupling Intent and Action via Latent World Modeling for End-to-End VLA paper: arxiv.org/abs/2603.29844 Avoid Routing Polarization for OCS-based GPU Clusters paper: arxiv.org/abs/2603.28168 Weight Tying Biases Token Embeddings Towards the Output Space paper: arxiv.org/abs/2603.26663 AgentFixer: From Failure Detection to Fix Recommendations in LLM Agentic Systems paper: arxiv.org/abs/2603.29848 ChartNet: A Million-Scale, High-Quality Multimodal Dataset for Robust Chart Understanding paper: arxiv.org/abs/2603.27064 VAREX: A Benchmark for Multi-Modal Structured Extraction from Documents paper: arxiv.org/abs/2603.15118 PruneFuse: Efficient Data Selection via Weight Pruning and Network Fusion paper: arxiv.org/abs/2603.26138 DataFlex: A Unified Framework for Data-Centric Dynamic Training of Large Language Models paper: arxiv.org/abs/2603.26164 Rethinking Language Model Scaling under Transferable Hypersphere Optimization paper: arxiv.org/abs/2603.28743 A Family of LLMs Liberated from Static Vocabularies paper: arxiv.org/abs/2603.15953 VideoZeroBench: Probing the Limits of Video MLLMs with Spatio-Temporal Evidence Verification paper: arxiv.org/abs/2604.01569 Marco DeepResearch: Unlocking Efficient Deep Research Agents via Verification-Centric Design paper: arxiv.org/abs/2603.28376 daVinci-LLM:Towards the Science of Pretraining paper: arxiv.org/abs/2603.27164 Sharp Capacity Scaling of Spectral Optimizers in Learning Associative Memory paper: arxiv.org/abs/2603.26554 Working Notes on Late Interaction Dynamics: Analyzing Targeted Behaviors of Late Interaction Models paper: arxiv.org/abs/2603.26259 On Strengths and Limitations of Single-Vector Embeddings paper: arxiv.org/abs/2603.29519

IBM has released Granite 4.0 3B Vision, a multimodal model specifically optimized for enterprise document extraction and structured data parsing The technical release highlights include: -- Architecture: The model is delivered as a LoRA adapter (~0.5B parameters) designed to run on top of the Granite 4.0 Micro (3.5B) dense backbone. -- Vision Encoder: It utilizes the google/siglip2-so400m-patch16-384 encoder. -- DeepStack Injection: Rather than a single projection point, the model employs a variant of the DeepStack architecture with 8 injection points. This routes abstract semantic features into earlier layers and high-resolution spatial details into later layers for precise layout awareness. -- Specialized Training: The model was refined using ChartNet, a million-scale dataset developed via a code-guided data augmentation pipeline (aligning plotting code, rendered images, and source tables). -- Benchmarks: VAREX: 85.5% zero-shot Exact Match (EM) accuracy for KVP extraction. Chart2Summary: 86.4% accuracy on the human-verified ChartNet test set. Table Extraction: Leads on PubTablesV2 (92.1 TEDS cropped) and OmniDocBench (64.0 TEDS). Full analysis: marktechpost.com/2026/04/01/… Model weight: huggingface.co/ibm-granite/g… Technical details: huggingface.co/blog/ibm-gran… @IBM @IBMwatsonx @IBMData @IBMResearch @IBMcloud @IBMDeveloper @IBMNews

【IBM】ビジネス文書に特化した最強の軽量VLM「Granite 4.0 3B Vision」登場！ 👉 x.com/mervenoyann/status/203… IBMから驚きの視覚言語モデルが登場しました！ 30億パラメータという軽量さながら、表やグラフの読解力で世界最高水準を記録しています。 ■注目のポイント ・ChartNet採用で複雑な折れ線・棒グラフも正確に理解 ・DeepStackにより画像内の細かい文字やレイアウトを逃さずキャッチ ・商用利用可能なフリーライセンスで自社ツールへの組み込みも自由 決算資料の自動データ化や、高精度なOCR、グラフからの分析コメント生成など、ビジネスの実務を劇的に効率化できる可能性を秘めています！✨ #IBM #AI

今日はChartNet論文（arxiv.org/abs/2603.27064）の手法でグラフ解析機能を実装。 使用技術: ・DINOv3 Transformers ・Hugging Face Dataset API ・Claude Code for data pipeline 100万枚のチャートデータを前処理中。 明日はマルチモーダル推論部分に着手。 着実に進んでる。

Granite-4.0-3B-Vision IBM is extending Granite-4.0 into a document-first VLM. LoRA on top of Granite 4.0 Micro, tuned for chart extraction, table extraction, and semantic KVP extraction rather than generic multimodal chat. Main signal is specialized enterprise extraction: Chart2CSV / Chart2Code / Chart2Summary, table extraction to JSON / HTML / OTSL, and 85.5 exact match on VAREX, a multimodal structured extraction benchmark built from 1,777 government-form documents with 1,771 schemas. Under the hood it uses SigLIP2, window Q-Former projectors, and Deepstack-style multi-layer visual feature injection across 8 vision-to-LLM injection points, trained on extraction-heavy mixtures plus ChartNet, IBM’s 1.5M-sample chart dataset spanning 24 chart types and 6 plotting libraries with aligned image, code, table, summary, and reasoning supervision.

ift.tt/335kaNL ChartNet: Visual Reasoning over Statistical Charts using MAC-Networks. (arXiv:1911.09375v1 [cs.CV]) #NLProc

Börsenfreaks bevorzugen das Chartnet.