rodriguescarson/polychart-shown-is-not-supported

Shown Is Not Supported

A chart is a completion claim. It renders cleanly, states a confident finding, and the underlying data may not support it. A model reading that chart inherits the gap: it answers from the visual impression because it never extracted the values.

This adapter reads the data first and corrects the chart when its encoding misleads. It ships with the first continuous score for how much a chart lies.

Trained with AutoScientist by Adaption for the AutoScientist Challenge 2026, Part 2, Data Visualization.

Results

Evaluation	Base	Adapted
On the dataset	15	85
Across the Data Visualization category (held-out)	22	78

Pairwise win rate, LLM judge, against the same untuned base (Llama-3.3-70B). 78% on the held-out category clears the gate; the Legal category was won at 73.

The number is the least interesting part. Here is what the work actually is.

The failure this probes: a shortcut, not a knowledge gap

Geirhos et al. (2020) call a shortcut a decision rule that works on the benchmark and fails when the test gets harder. Their network learned an object's location instead of its shape, and broke the moment location stopped predicting the label.

Chart models do the same thing. We ran a base VLM on 60 matched pairs, the same series drawn faithfully and drawn to deceive, and the result is not uniform:

Distortion	Wrong direction, faithful to distorted
Inverted axis	17% to 93%
Truncated axis	0% to 0% (magnitude error fell)
Aspect ratio	4% to 4% (no effect)

The model takes direction from the geometry and magnitude from the axis labels, and never reconciles them. Invert the axis and the geometry shortcut breaks: the trend reads backwards 93% of the time while the magnitude, read from the untouched labels, survives with identical error (52.1 both sides). Squash the aspect ratio, which changes steepness but not label values, and nothing happens, because steepness was never being read.

That is a clean, isolated shortcut. The dataset is the intervention that closes it.

The intervention: counterfactual augmentation, built exact

Kaushik, Hovy and Lipton (2020) fix shortcut learning by pairing each example with a counterfactual: the same content, minimally edited to flip the label, so a model cannot lean on the spurious pattern. Their edits were written by humans, which is slow, approximate, and does not scale.

We build the same thing from parameters. Every chart is rendered twice from one OWID series: once honestly, once with a distortion whose parameters we set. The distortion is the counterfactual edit, and the label is derived from it, not annotated. Same data, altered encoding, exact label, no human in the loop.

Composition, 2,129 rows, weighted to the failure cells (see next section):

• supported (54%) accurate extraction and interpretation
• severity (27%) the lie factor, see below
• contradicted (13%) the encoding overstates or reverses the data; the answer gives the true figure and names the mechanism
• unanswerable (5%) the chart cannot answer the question; the model says so

Targeted, not uniform: hard-example mining

Winning data-centric submissions (KAIST, Data-Centric AI Competition) do not add uniform coverage; they find where the model fails and concentrate data there. We ran the base model across every question-type by distortion cell and scored it against exact ground truth. The failures cluster in three places: severity estimation, precise value extraction under a distorted axis, and mechanism identification on inverted and cherry-picked charts. Comparison and abstention were already solved. We re-weighted the dataset to those failure cells and removed the solved ones.

The effect is measurable. The uniform version of this dataset trained on a 109B base reached 76 on the held-out category. This failure-weighted version reached 78 on a smaller 70B base. Targeted data compensated for less model capacity, and on-dataset win rate rose from 77 to 85. Curation and targeting beat both scale and capacity, the result the data-centric literature predicts.

The new part: deception has a magnitude, and it is computable

Every misleading-chart dataset, including Misviz (ACL 2026), answers a yes/no question: is this chart misleading, by which of N mechanisms. None answers how much.

Edward Tufte defined the Lie Factor in 1983: the size of the effect shown in the graphic over the size of the effect in the data. One is honest. Because we construct every distortion, we compute it exactly rather than estimate it. 471 rows carry a lie factor and a severity band, from honest (1.0) through a truncation that triples the apparent slope (3.0) to an inverted axis that keeps the magnitude and reverses the sign (-1.0).

This turns deception detection from a label into a graded, calibrated target. No released chart dataset provides it.

The Deception Atlas: the measure on charts that actually aired

The distortions above are constructed, which is what makes their Lie Factors exact. The atlas shows the task transfers to the real world: 21 published charts, verified against primary sources, across 11 distinct mechanisms, none reconstructed. This is the part of the dataset that is not a chart-QA benchmark. It is a quantified corpus of real chart lies, a thing no released dataset provides.

Chart	Mechanism	Severity	The real reading
Fox News, top tax rate (2012)	truncated bars	LF 4.95	35% to 39.6%, a 13% rise, drawn 5.6x taller
Fox News, welfare vs work (2013)	truncated bars	LF 4.69	108.6M vs 101.7M, a 7% gap, drawn 5x taller
Planned Parenthood, AUL chart (2015)	dual mismatched axes	LF 2.86	screenings outnumber abortions 2.9 to 1, drawn as equal
Reuters, Florida gun deaths (2014)	inverted axis	LF -1.0	deaths rose 521 to 825 after 2005, drawn as a fall
National Review, global temperature (2015)	expanded axis	LF 0.017	a real 1.1 C rise flattened to under 2% of the frame
Ted Cruz, satellite "pause" (2015)	cherry-picked window	severe	window anchored to the 1998 El Nino peak; +0.44 C over 1997-2015
Steve Jobs, smartphone share (2008)	3D pie perspective	severe	Apple 19.5% tilted forward to rival RIM's 39%
Georgia DPH, COVID counties (2020)	reordered x-axis	severe	dates out of order manufacture a decline that is not in the data
Lipitor, "36% risk reduction" (2008)	percent-absolute swap	severe	a 1.1 point absolute drop (3.0 to 1.9) sold as 36%
OpenAI, GPT-5 "coding deception" chart (2025)	mis-scaled bars	severe	a bar labeled 50.0 drawn shorter than one labeled 47.4; corrected to 16.5
Meta, Llama-4 on LMArena (2025)	claim vs data	severe	rank earned by an experimental variant; the released model scored ~1370, not 1417
Adaption, "+16 beyond 20k rows"	claim vs data	severe	this project's runs show 76 to 72, a minus 4

The severities span the full signed range the measure defines: -1.0 where an inverted axis reverses a real rise, 0.017 where an expanded axis hides one, 4.95 where a truncated baseline invents one. A yes/no "is it misleading" label collapses all of these into the same bit. Where a single Lie Factor is not defined (a reordered axis, a 3D pie) the row names the mechanism and bands the severity rather than invent a number. Where a baseline was unlabelled it is marked inferred with the arithmetic shown.

The atlas also holds the AI industry's own charts. The GPT-5 launch bar chart drew a bar labeled 50.0 shorter than one labeled 47.4, geometry against its own printed values, in Sam Altman's words a "mega chart screwup." The Llama-4 leaderboard rank was earned by a non-public variant the fine print called Experimental. And the sharpest case is the platform that trained this model: its augmentation panel advertises +16 win-rate points beyond 20,000 datapoints, and this project's controlled runs, at matched base and epochs, measured 76 falling to 72. The atlas audits the tools of its own field, on the dataset's own thesis: shown is not supported. Full corpus, sources, and computation are in data/atlas/ and data/real/.

A controlled study, not a single run

We ran the task through five base models AutoScientist selected and read the training curves out of every adapter.

Base	Params	On-data	Category	What the curve shows
Llama-3.3-70B (this model)	70B	15 to 85	22 to 78	3 epochs, hard-mined data, eval loss down
Llama-4-Scout (uniform data)	109B	24 to 77	24 to 76	same recipe, uniform coverage
Mistral	7B	34 to 66	47 to 54	held slightly
gemma	4B	52 to 48	n/a	eval loss 1.28 to 2.73; severe overfit
Llama-3.2	3B	49 to 51	54 to 47	overfit; category fell below base
Qwen3.5	0.8B	41 to 59	56 to 45	underfit and category below base

Two findings.

Capacity dominates for uniform data, but targeted data closes the gap. With uniform coverage, only the large bases held their held-out performance: every base under 10B overfit at the platform default of 10 epochs, eval loss rising from the first checkpoint while train loss fell toward memorization. Yet the failure-weighted data let a 70B base reach 78, above the 109B's 76 on uniform data. Capacity sets the floor; targeting the data raises it. The Hardware Lottery (Hooker 2020) is real, and better data is the lever a competitor actually controls.

The headline number hides who gets hurt. Every small model improved on its own data while degrading on the held-out category. This is the effect Hooker et al. (2019) documented for compression: aggregate accuracy barely moves while the underrepresented distribution is quietly forgotten. Our two-panel split, on-data against held-out category, is the instrument that exposes it, the same move as their Pruning Identified Exemplars. We report both panels so the concealment is visible.

More data made it worse. We tested the platform's own claim that crossing 20,000 datapoints yields +16 win-rate points. Expanding this dataset from 1,882 to 21,882 rows with the platform's synthetic augmentation dropped the held-out category from 76 to 72, at matched epochs and base model, with eval loss still descending. It is dilution by unverified synthetic data, not overfitting: 1,882 verified rows beat 21,882 augmented ones by four points. Curation beats scale.

The metric probes itself

The eligibility number is LLM-judge pairwise win rate, so it inherits the biases Zheng et al. (2023) documented for that judge, verbosity among them: a preference for longer answers regardless of quality. We probe it directly with a terse-answer ablation, and we write answers to be correct-and-complete rather than terse, so substance and length are separated rather than confounded.

Nine silent failures, caught by an independent verifier

The dataset is about models stating confident findings their input does not support. We built a verifier that shares no code with the generator and re-derives every claim from the source table. It caught nine such failures in our own pipeline before anything shipped, including:

• an answer calling a 183% rise "far more modest than it looks" on a 1.28x exaggeration
• an ingest that returned unusable data for 22 of 28 sources while every request returned 200 and the build reported success
• three failures inside the verifier itself, which failed the standard it applies

The last point is the thesis reproduced inside the tool built to study it: a grader that is confident, well-formed, and wrong until an independent check catches it. Every failure and its fix is in the repository.

Honest limitations

• Epoch count did not matter on this base. At 3 epochs this model's eval loss fell throughout (0.944 to 0.861, no overfit); an earlier 10-epoch run on related data scored within noise (78 vs 76 category). The large bases are insensitive to the epoch schedule, unlike the sub-10B bases in the study, which overfit badly at 10.
• Editor overrides did not persist. AutoScientist re-selected the base model regardless of the manual recipe, so the five-model comparison is confounded by base as well as composition. We report the base per run rather than hide it.
• Text specifications for the headline model. Charts are given as plotted values plus axis parameters. A multimodal image variant is released separately.
• Single evaluation source, platform-reported, with a documented plus or minus 8 to 10 point run-to-run noise we did not average out.

Related work

Misviz and Misviz-synth (ACL 2026) release misleading visualizations and classify which of 12 design rules a chart breaks. Chart-QA benchmarks (ChartQA, CharXiv, ChartMuseum) test reading. Our contribution is orthogonal: a written correction grounded in the source table, and a continuous severity target, neither of which a released dataset provides.

Artifacts

• Seed dataset (verified): polychart-shown-is-not-supported
• Adapted dataset (trained this): chart-qa-with-axis-tricks
• Real-world validation slice: data/real/, three published charts with cited data
• Weights (this repo)
• Generation and verification code: public, reproducible from one command
• Source data: Our World in Data, CC-BY, per-row citation

References

• Geirhos et al. Shortcut Learning in Deep Neural Networks. Nature Machine Intelligence, 2020.
• Kaushik, Hovy, Lipton. Learning the Difference that Makes a Difference with Counterfactually-Augmented Data. ICLR, 2020.
• Hooker et al. What Do Compressed Deep Neural Networks Forget? 2019.
• Hooker. The Hardware Lottery. 2020.
• Zheng et al. Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena. NeurIPS, 2023.
• Mazumder et al. DataPerf: Benchmarks for Data-Centric AI Development. NeurIPS, 2023.
• Tufte. The Visual Display of Quantitative Information. 1983.