Multimodal models that caption photos fluently still fail on competition math with diagrams, charts, and geometry figures. On MATH-Vision, GPT-4o reaches 30.39% against a human 68.82%, and asking it to "think step by step" sometimes lowers accuracy (MATH-Vision). One documented cause is visual de-grounding: as the written reasoning grows longer, the model pays more attention to its own text than to the image and invents visual facts to support its argument (Decoupling Reasoning and Perception). Long-form story generation had a parallel problem, where stories drifted from their premises over thousands of words. That field addressed it by writing an outline first and checking each chapter against it (Re3), and by expanding each branch of a pre-built reasoning tree into its own short chapter (MuSR).
This note proposes carrying that plan-then-write idea into visual reasoning in one specific form. Before solving anything, the model writes a shallow plan tree. Its internal nodes are deduction sub-goals. Its leaves are visual obligations: atomic questions about the image ("what is the label on segment BC?", "are lines AD and BE parallel?") that the solution needs answered. Each obligation goes to a vision model in a fresh, single-question context. Each sub-goal is then solved in its own short context, using only its children's answers. When a check fails or a deduction produces an impossible state, only the affected subtree is repaired. In two sentences: de-grounding happens because the model decides what it needs to see too late, inside a long free-running trace. Declaring all visual dependencies up front, and solving in short chapters tied to verified percepts, should remove the long trace in which the drift occurs.
Core Claim
The claim is that a global plan whose leaves are explicitly verified perceptual queries beats both flat chain-of-thought and reactive perceiver dialogues on diagram-dependent problems. The advantage should be largest on problems where the text carries little information. It should also show up as a measurable gain in grounding, meaning that answers change correctly when the image is changed, and not only as a gain in accuracy.
The closest prior system is the LLM-LMM framework of Jia et al.. There, a text-only reasoner questions a vision "observer" turn by turn until a sufficiency check passes. Their result motivates this proposal, since decoupling reaches 52.6 on MathVision and 54.4 on MathVerse. The difference is structural. Their queries are reactive, issued from inside one growing reasoning transcript, and that transcript is exactly where drift can occur. Here, queries are fixed by a plan before reasoning starts, each answer is consumed by a bounded local context, and repair is localized to a subtree. If reactive questioning already captures all of the benefit, the experiments below will show it, and the proposal will have failed.
Three other lines of work motivate the design or constrain it. The first line converts the image into text for a reasoner. RAPID trains a perceiver with reinforcement learning to write captions that help a frozen reasoner. VCAR trains one module to describe the image and another to reason from that description. Both pass a single dense, unstructured description forward with no way to look again. The second line is a constraint. Primitive Vision and SVE-Math find that exhaustive textual descriptions of diagram elements hurt accuracy by up to 4.2% through information overload. This is the main reason the proposal asks only for what the plan requires instead of describing everything. The third line is also a constraint. MuSR reports that frontier models cannot build sound reasoning trees deeper than about 2. So the plans here are deliberately shallow, depth 2 to 3, and the system relies on Re3-style repair rather than deep upfront trees.
Research Program
The system has four stages and runs on open models so that comparisons can be controlled. The main pairing matches Jia et al.: a Qwen3-32B reasoner and a Qwen2.5-VL-72B perceiver. A smaller pairing matches RAPID: Qwen3-8B with Qwen2.5-VL-7B.
Stage 1: planning. The reasoner sees the question text but not the image. It writes a plan tree in a fixed format. Internal nodes are sub-goals, each listing which children it depends on. Leaves are visual obligations, each with a typed expected answer: a number, a relation, yes/no, or "not determinable." On MathVerse's vision-only split the question itself is inside the image, so the perceiver first transcribes the question verbatim. Keeping the image away from the planner is a deliberate choice. It forces every visual dependency to be named as a leaf, which closes the textual shortcut that MathVerse and Plot Twist show inflates benchmark scores.
Stage 2: perception. Each leaf goes to the perceiver alone, together with the image, in a new context. The perceiver may answer "not determinable."
Stage 3: chaptered realization. Each internal node is solved bottom-up in a short context containing only the node's sub-goal and its children's verified results. Arithmetic is delegated to executed Python, following the sub-question-plus-code rationales of Synthesize Step-by-Step.
Stage 4: repair. A subtree is regenerated under a fixed repair budget in three cases: a leaf returns "not determinable," two leaves contradict each other, or a node produces an impossible state such as a negative length or a probability above one. MathVista observed that GPT-4V sometimes backtracks on its own after hitting such states (MathVista); here that check is explicit and systematic.
The analysis has two parts. First, the full system is compared with the baselines below under matched model calls and token budgets. Second, the source of each improvement is located using a visual-claim provenance audit and a counterfactual image test, both described in the final section.
Baselines
The naive reference point is the Qwen2.5-VL-72B perceiver alone, prompted with standard chain-of-thought and with self-consistency sampled up to the same token budget the full system uses. I found no reported result on Lacuna for Qwen2.5-VL-72B with plain chain-of-thought on these benchmarks, so it will be measured here. For calibration against a frontier model, GPT-4o's 30.39% on MATH-Vision (MATH-Vision) is quoted, not rerun.
The strongest existing method on this question is the reactive decoupled framework of Jia et al.. It uses the same division of labor, a text-only reasoner and an on-demand vision observer, and it reports 52.6 on MathVision and 54.4 on MathVerse with Qwen2.5-VL-72B plus Qwen3-32B. The reimplementation will use identical models, the same total budget of perceiver calls and reasoner tokens, and a matched prompt-tuning budget of a fixed number of development-set iterations per method. The reported numbers serve only as a sanity check that the reimplementation is faithful. Beating it means a statistically significant accuracy gain on MathVerse vision-only and on MATH-Vision under equal budgets, and fewer unsupported visual claims. Matching its accuracy while using fewer perceiver calls would be a weaker but still useful outcome.
The second strong competitor is RAPID, which reports 53.2% average accuracy across its seven-benchmark suite with Qwen2.5-VL-7B and Qwen3-8B. It is the best caption-based approach. The comparison uses the same small model pair on the same suite, with RAPID's trained perceiver swapped in as the leaf answerer. That configuration also tests whether the two ideas combine.
Three ablations isolate the key ingredient. Plan without perception keeps the plan tree but lets the reasoner solve from a single dense caption. Flat plan replaces the tree with a linear list of visual questions answered in one pass, similar to the question-answer plans of Visual Storytelling with Question-Answer Plans. No repair removes Stage 4. If plan-without-perception matches the full system, the gain comes from outlining and not from grounding, and the core claim is wrong.
Why not just…?
Why not just prompt the model to outline before solving? This is a real threat, because plan-then-solve prompting is cheap. The MATH-Vision finding that chain-of-thought can hurt multimodal models suggests that more text alone does not help. A plain outline still leaves every visual fact to be recalled inside one long trace. The plan-without-perception ablation is exactly this alternative, and it settles the question directly.
Why not just let the reasoner question the perceiver reactively, as Jia et al. do? This is the strongest alternative, and it may capture most of the benefit. The hypothesis is that a reactive transcript still grows long and leaves room for drift, while chaptered contexts do not. The equal-budget comparison, the provenance audit, and the counterfactual test will show whether the upfront plan adds grounding beyond reactive questioning. The counterfactual test matters most, because accuracy alone could be similar.
Why not just write a better caption, or train the perceiver as RAPID does? Dense description runs into the 4.2% overload penalty reported by Primitive Vision. GeoSQA shows that even human-written transcriptions of diagrams leave multi-step reasoning failures intact. Caption quality and plan structure are separate factors, so the RAPID-perceiver-inside-the-plan configuration tests whether they add up rather than compete.
Why not just improve the vision encoder? Better encoders help: perception fixes yield about 12% in Primitive Vision. But 42.2% of GPT-4V's MATH-Vision errors are reasoning errors made on correctly perceived content. The proposal does not depend on any particular encoder, and a GeoGLIP-style perceiver could answer the leaves.
Proposed Experiments
Accuracy is measured on MATH-Vision, all six MathVerse splits, MathVista, and ChartQA's human-written split. The main question is whether the gap over the reactive baseline grows as text becomes less informative, from MathVerse's text-dominant split to its vision-only split. That trend is the signature of genuine grounding.
The decisive check is a counterfactual image test. Synthetic charts are generated in the style of Plot Twist, and geometry figures are re-rendered with changed numeric labels. Because the edits are programmatic, the correct new answers are known. For each image pair, the test records whether the system's answer moves to the new correct value. A grounded system should track the edit. A system relying on priors or text will not. If visual-obligation planning does not track edits better than the reactive baseline, the central claim fails.
The second check is the provenance audit. A judge model, validated against human annotation on 200 traces, labels every visual assertion in a final solution as either supported by a perceiver answer or unsupported. If the de-grounding explanation is right, unsupported claims should be rare in the full system and should increase with trace length in the flat chain-of-thought and reactive baselines.
Finally, error profiles will be broken down by plan depth and repair count. This tests the MuSR depth limit in the visual setting and shows whether localized repair recovers problems where the initial plan was flawed.
References
- Measuring Multimodal Mathematical Reasoning with MATH-Vision Dataset
- Decoupling Reasoning and Perception: An LLM-LMM Framework for Faithful Visual Reasoning
- Re3: Generating Longer Stories With Recursive Reprompting and Revision
- MuSR: Testing the Limits of Chain-of-thought with Multistep Soft Reasoning
- Reasoning-Aligned Perception Decoupling for Scalable Multi-modal Reasoning
- Describe-then-Reason: Improving Multimodal Mathematical Reasoning through Visual Comprehension Training
- Primitive Vision: Improving Diagram Understanding in MLLMs
- Open Eyes, Then Reason: Fine-grained Visual Mathematical Understanding in MLLMs
- MATHVERSE: Does Your Multi-modal LLM Truly See the Diagrams in Visual Math Problems?
- Plot Twist: Multimodal Models Don't Comprehend Simple Chart Details
- Synthesize Step-by-Step: Tools, Templates and LLMs as Data Generators for Reasoning-Based Chart VQA
- MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts
- Visual Storytelling with Question-Answer Plans
- GeoSQA: A Benchmark for Scenario-based Question Answering in the Geography Domain at High School Level