CURV: Enhancing Chart Understanding Through Curriculum Visual Grounded Reasoning
Xuehang Guo, Pingyue Zhang, Ruiyi Zhang, Zhenhailong Wang, Hanrui Lyu, Heng Ji, Tong Sun, Qingyun Wang, Manling Li
cs.CV, cs.AI, cs.CL
2026-08-04
CURV trains models to ground each reasoning step on the relevant chart region. A 7B model trained this way beats GPT-4o by 12 points on the synthetic chart benchmark and adds 4 on real-world CharXiv.
Chart question answering (CQA) asks a multimodal model to read a chart and answer a question about its data. It sounds easy and isn't. The authors run a telling probe: on 60 CharXiv questions, GPT-4o scores 43.3% answering directly. Hand it the relevant visual values and it rises to 61.7%. Ask it to reason step by step first and it manages 53.3%. Give it both the visual cues and the reasoning prompt and it hits 86.7%.
Two failures surface: the model misreads values in the chart, and it loses the logical thread of what to compute in what order. Supplying both visual grounding and a reasoning chain nearly doubles accuracy. CURV's goal is to bake those two crutches (hand-fed visual info, chain-of-thought prompting) into the model itself, so shifting focus and reasoning become one learned behavior rather than an inference-time patch.
CURV replaces the direct map of image plus question to answer with a chain where every reasoning step Rt is paired with a visual region Vt marking which part of the chart that step is looking at: (I, Q) -> {(R1, V1), ..., (RT, VT)} -> A.
Training has two stages:
Three grounding variants exist: applied (mask-highlight the predicted region directly on the image), boxed (draw a bounding box), and cropped (zoom in). The key design choice is treating grounding as a bridge between reasoning and vision rather than as the final objective; an ablation shows this explicit grounding beats the implicit version (grounding learned internally, never output) by 8.78 points.
Curriculum learning scales difficulty across three levels measured by reasoning depth D, the number of nested functions f1(f2(...fD(x))) a question requires. Level 1 is a single operation, level 2 is nested, level 3 is multi-subplot composition. Training uses only single-plot charts (1 <= D < 3) but generalizes to multi-plot.
The companion dataset CCQA is synthetic: seven chart types (bar, histogram, scatter, line, heatmap, pie, radar) across 30 domains, with GPT-4o generating plotting data and templates generating the QA. The notable trick is using only 30 base images per chart type and deriving many query-reason-region-answer quadruplets from each, forcing the model to learn transferable reading skills rather than memorize specific visuals.
On CCQA (the authors' synthetic benchmark) the two-stage training gains up to 20.92 points. CURV@Applied on Qwen2.5-VL-7B is the strongest setting:
| Model | Level 1 | Level 2 | Level 3 (multi-chart) |
| GPT-4o | 57.64 | 34.04 | 22.14 |
| Qwen2.5-VL-7B (base) | 54.21 | 28.68 | 19.01 |
| CURV@Applied (Qwen2.5-VL-7B) | 69.86 | 40.21 | 26.11 |
(acc@MLLM, judged by GPT-4.1-mini.)
The 7B model beats GPT-4o across all three levels, by up to 12.22 points. Multi-chart scenes it never saw in training still improve 7.1 points.
On real chart benchmarks the synthetic-trained gain shrinks but holds:
| Model | CharXiv (reasoning) | ChartMuseum | MMMU-Pro (out-of-domain) |
| Qwen2.5-VL-7B (base) | 32.50 | 21.62 | 28.21 |
| CURV@Applied | 36.70 | 25.12 | 32.08 |
CharXiv rises 4 points; the out-of-domain MMMU-Pro also rises about 4, which suggests the model learns a general skill of grounding reasoning in the image rather than overfitting synthetic questions. Switching to RL training adds more (up to +17%) but at far higher compute cost, so the authors push SFT for its price-performance ratio.
For anyone building document or chart understanding, this is a reusable recipe: with only 7x30 = 210 synthetic base images and templates, a small model can match or beat GPT-4o on chart reasoning. The core insight, grounding each reasoning step explicitly to an image region, ports onto any multimodal backbone, and the authors verify it across the InternVL3 family (1B/2B/8B).
It also offers a diagnostic angle for multimodal failures: do not grade only the final answer. Aligning the reasoning chain with visual focus separates a misread chart from faulty reasoning, which is more useful for knowing where the model breaks.