Vision-language models often read a diagram correctly and then reason about it badly, or skip the diagram and answer from the wording of the question. Recent work shows that fine-tuning on a small, well-chosen subset of problems fixes much of this. The subset should hold problems that genuinely require the image and sit at the edge of what the model can currently solve. Every such method, however, picks its subset once, before training, using expensive probes: tree search per problem, or paired runs with and without the image. The model changes as it trains, so a subset chosen at step zero goes stale. Re-running those probes during training costs about as much as it saves, which is the familiar trap of dynamic data selection.
We propose to keep the selection current and nearly free. In reinforcement learning (RL) fine-tuning, the model already samples several answers per problem, and each problem's success rate falls out of that step at no extra cost. It also already encodes each problem while doing so. We reuse both. We then add one deliberately cheap probe: a short, text-only attempt at the question with the image removed. This probe is cheap because it skips the visual tokens entirely. From these signals, a small bookkeeping model estimates, for every problem in the pool, two things: how likely the current model is to solve it, and how likely it is to solve it without looking. We sample problems that are learnable now and that actually require the image. The central claim concerns total compute, counting both selection and training. Under equal total compute, this online, image-aware curriculum should beat offline curation. It should also beat curricula that track only success rate, and the gains should appear specifically on the problems that force the model to use the diagram.
Core Claim
Three offline methods target the multimodal reasoning gap by curating a small subset of the training pool before training begins. ThinkLite-VL runs tree search with up to 50 iterations on each of about 70,000 candidates and keeps the hard ones. RAP (Truth in the Few) runs each candidate with and without its image and discards those solvable from text alone. TIGER mines cases where the model solves a problem as text but fails it when the same text is rendered as an image. All three are motivating evidence that the right subset matters.
All three are also static by design. ThinkLite-VL openly concedes its search is affordable only as a one-time step. Its own finding that difficulty is model-specific (data curated for a 7B model hurts a 72B model) implies that difficulty is also checkpoint-specific. Wiki-R1 supplies the closest methodological precedent. It keeps difficulty current by spreading observed successes across a graph, and it targets problems with roughly even odds of success. But its graph is built from Wikipedia text similarity, which does not exist for geometry diagrams. It also never asks whether the image was needed.
On the text-only side, EDCO shows that cheap short-prefix entropy probes can drive a live curriculum. Its reported RL fine-tuning accuracy on Datacom is 46.96%, against 40.43% for random sampling and 44.78% for a static curriculum. Even so, EDCO still pays about 10% wall-clock overhead and has never been tested on images.
What would be new here is an online estimate of two quantities, kept inside a fixed probe budget, using signals the RL loop already produces. The first is current solvability. The second is image necessity. We also evaluate honestly, charging selection compute against the method.
Research Program
The base setting is GRPO, a common RL fine-tuning recipe that samples a group of answers per problem and rewards each against a verifiable ground truth. It is run on an open vision-language model such as Qwen2.5-VL-7B, the model family used by RAP and ThinkLite-VL. The system maintains three pieces of state for each problem $i$ in the pool.
The first piece is a success estimate $\hat p_i$. It is a Beta posterior updated from the group's rewards whenever the problem is trained on. Old counts are shrunk by a forgetting factor so that evidence from early checkpoints fades. This costs nothing beyond arithmetic.
The second piece is an embedding of the problem, taken from the policy's own hidden states when the problem is encoded for rollouts. It is used to spread observed successes to neighboring, unvisited problems in the style of Wiki-R1. Because the embedding is the model's own multimodal representation, the spreading works for diagrams. It needs no external text corpus.
We handle embedding drift, the concern raised by Sanakoyeu and Kim et al. for retrieval, lazily. Embeddings are refreshed only when a problem is touched, since the refresh then comes free. A region of the pool is re-embedded only when observed success rates there diverge from predicted ones beyond a threshold. This replaces the fixed re-scoring intervals the dynamic-sampling literature relies on with a trigger tied to measured prediction error.
The third piece is a no-image success estimate $\hat q_i$. It comes from a short, text-only attempt with the image removed, using EDCO's quick-answer prompt and a capped output length. The probe is run on a small fraction of each batch and spread to neighbors the same way. This is RAP's with-versus-without-image test made online and subsampled. We will measure its cost directly rather than assume it; the cheapness argument rests on image tokens making up much of the input for high-resolution diagrams.
Sampling weights combine the two estimates:
$$w_i \propto \hat p_i (1-\hat p_i)\,(1-\hat q_i),$$
The first factor, $\hat p_i(1-\hat p_i)$, peaks at even odds, where group rewards vary most and GRPO gets a useful gradient. This is Wiki-R1's rationale. It also resolves the file's tension between EDCO's hardest-first rule and the sparse-reward failures of overly hard RL data. The second factor, $1-\hat q_i$, down-weights problems the model can answer without the image.
All probing is held under a budget,
$$C_{\text{probe}} \le \rho\, C_{\text{train}},$$
where $C_{\text{probe}}$ is the extra compute spent on text-only probes and re-embedding, $C_{\text{train}}$ is the rollout and update compute, and $\rho$ is set to a few percent. The value of $\rho$ is swept.
The analysis side asks three questions. First, how fast do step-zero difficulty labels go stale? We measure the correlation between the offline ThinkLite-style or RAP labels and live success rates across checkpoints. Second, how accurate are the spread estimates for unvisited problems, measured as calibration against held-out full rollouts? Third, does image-necessity weighting change what the model learns? We look at this through the benchmark splits described below.
Baselines
The naive reference is uniform random sampling from the full pool under the same GRPO recipe and step count. RAP reports that a random 20% subset loses only 0.8% on average. That makes random sampling a serious baseline, not a strawman.
The strongest existing competitors are the offline curators.
ThinkLite-VL reports 75.1 on MathVista for its 7B model. We reproduce its search-based filtering on our pool. We charge its search compute to its budget and give it the same number of training steps. No reported result was found on Lacuna for ThinkLite-VL's selection compute, so we will measure it ourselves.
RAP reports a 56.04% average over six benchmarks, including MathVista, MMStar and MathVerse, using 9.3% of its pool. We run its with/without-image and attention filters offline on the same pool and base model.
TIGER reports roughly 15% improvement on MathVerse for Qwen3-VL-8B. It is included on the MathVerse comparison with its rendered-text counterfactual set. It is also run as an additional training mixture, since it is complementary rather than competing.
Fairness is enforced on two axes. At equal training steps, every method gets the same GRPO hyperparameters and the same tuning budget, a small grid over learning rate and group size shared by all. At equal total compute, selection plus training, offline methods are allowed to spend their selection cost however they like, including by training longer on a random subset instead. Beating them means one of two outcomes. The first is higher accuracy on MathVista and on the vision-only versions of MathVerse at equal total compute. The second is matching their accuracy at clearly less total compute. In either case the result must be reported with seeds so that differences exceed run-to-run spread.
Four ablations isolate the ingredients. The first uses only the free success-rate posterior, with no spreading and no text-only probe. The second adds spreading. The third adds the text-only probe. The fourth periodically re-runs RAP's full offline filter every fixed number of steps, which is the brute-force way to stay current. A final ablation freezes the embeddings from step zero to test whether lazy refresh matters.
Why not just…?
Why not just filter once offline with RAP or ThinkLite-VL? Offline filtering has zero in-loop cost. If difficulty barely shifts during a short fine-tuning run, it may be enough. The evidence that it shifts is indirect but consistent. ThinkLite-VL itself shows difficulty is model-specific, and EDCO's dynamic curriculum beats its static one on Datacom, 46.96% versus 44.78%. The staleness measurement in our analysis settles this directly. If step-zero labels stay well correlated with live success rates throughout training, the offline methods win on cost, and we will say so.
Why not just use the success rates GRPO already computes and skip everything else? This is the most serious threat. It is free, and it already drops problems the model finds trivial, including many that are solvable from text alone. Our bet is narrower. Problems that sit at medium success partly because the wording gives the answer away will still be sampled heavily. Training on them rewards the text shortcuts documented by MathVerse, where some models score 5% or more higher with the diagram removed. It also rewards the text-favoring behavior measured by Deng et al..
The prediction is falsifiable. A success-rate-only curriculum should roughly match ours on aggregate MathVista. It should lose on MathVerse's vision-dominant and vision-only versions, and it should show a larger gap between its text-dominant and vision-only scores. If it matches us on those splits too, the image-necessity probe is not worth its budget, and the contribution shrinks to the compute accounting and spreading results.
Why not just train on a random subset or the full pool? RAP's own 20% random result shows random subsets are strong on averages. Meanwhile, Zhao et al. show that adding noisy data can hurt, costing 3.7 points on MATH from raw web-crawled problems, and ThinkLite-VL reports that easy problems degrade RL fine-tuning. The equal-compute comparison against random and full-pool training decides this. We will report per-split results so that a random-sampling win on shortcut-friendly splits is not hidden by averaging.
Why not just use a blindfolded judge model to separate seeing from thinking? MoCA does exactly that. It routes the model's written perception through a 14B text model on every rollout and reports 86.6 on V*. That is a credit-assignment method with large per-rollout cost, and it assumes the diagram can be fully transcribed into text. Our probe asks a cheaper, coarser question: did the problem need the image at all? We treat MoCA as orthogonal and do not claim to match its perception gains.
Proposed Experiments
Training uses a mixed pool of math, chart and natural-image problems with verifiable answers, matching the composition ThinkLite-VL describes. The auto-verified scientific figures of VeriSciQA are added for a harder, figure-heavy slice.
Evaluation covers several benchmarks. MathVista is the headline comparison with ThinkLite-VL. MathVerse is reported per information version, since the text-dominant versus vision-only gap is our main mechanistic readout. MATH-Vision supplies competition diagrams. Its human baseline is 68.82% and GPT-4o scores 30.39%, and its error taxonomy lets us check whether our model's remaining errors shift from reasoning toward perception. GAOKAO-MM serves as a small out-of-distribution check, reported cautiously given that its scores vary by up to 20% across exam years. MMStar is included as used in RAP.
Every result is reported against two cost measures: training steps, and total accelerator-hours including selection. We also record the probe budget $\rho$ actually spent.
The decisive checks are three. First, at equal total compute, the full method must beat offline RAP and ThinkLite-style filtering on MathVista and on MathVerse's vision-only versions. Second, the text-only probe must add accuracy on MathVerse's vision-only versions beyond the success-rate-only ablation, together with a narrower text-dominant versus vision-only gap. Third, measured probing overhead must stay below EDCO's roughly 10% wall-clock figure while estimates for unvisited problems remain calibrated. If the first check fails, dynamic selection does not pay for itself in this regime. If only the second fails, image-aware selection is unnecessary once success rates are tracked. Either negative result would be worth reporting, because the field currently assumes both without an equal-compute test.
References
- SoTA with Less: MCTS-Guided Sample Selection for Data-Efficient Visual Reasoning Self-Improvement
- Truth in the Few: High-Value Data Selection for Efficient Multi-Modal Reasoning
- TIGER: Bridging the Multimodal Reasoning-Access Gap via Modality Counterfactuals
- Wiki-R1: Incentivizing Multimodal Reasoning for Knowledge-based VQA via Data and Sampling Curriculum
- EDCO: Dynamic Curriculum Orchestration for Domain-specific Large Language Model Fine-tuning
- Improving Deep Metric Learning by Divide and Conquer
- Saving Dense Retriever from Shortcut Dependency in Conversational Search
- MATHVERSE: Does Your Multi-modal LLM Truly See the Diagrams in Visual Math Problems?
- Words or Vision: Do Vision-Language Models Have Blind Faith in Text?
- More Data or Better Data? A Critical Analysis of Data Selection and Synthesis for Mathematical Reasoning
- Bad Seeing or Bad Thinking? Rewarding Perception for Multimodal Reasoning
- MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts
- Measuring Multimodal Mathematical Reasoning with MATH-Vision Dataset
- GAOKAO-MM: A Chinese Human-Level Benchmark for Multimodal Models Evaluation