Multimodal models, which answer questions about images, often fail in ways a single accuracy number cannot locate. A model may misread a value on a chart axis, correctly read every value and then mis-relate them, or relate them correctly and then botch the final multi-step deduction. When a verifier tells such a model "this step scores 0.3," the model learns that something is wrong but not what. It often responds by shortening or hedging its answer rather than looking at the image again.
This note proposes a critic that does three things. It finds the first faulty step in a visual reasoning trace. It says at which level of a three-tier hierarchy the step failed: perception, relation, or composition. It then describes the problem without supplying the fix. The design borrows one testable distinction from self-determination theory, a theory of human motivation that separates informative feedback (it explains what went wrong and leaves the learner to act) from controlling feedback (a bare score, or a dictated correction). The prediction is that informative, tier-located critique beats both scalar rewards and prescriptive corrections in two ways. It should improve visual reasoning at inference time. More importantly, it should transfer better when the model is later trained on its own critiqued revisions and then evaluated with no critic at all.
Core Claim
Two kinds of prior work motivate the claim. The first is evidence that the failure location varies and is invisible to flat metrics. On MATH-Vision, GPT-4V's errors were 42.2% reasoning and 31.9% vision recognition (MATH-Vision). On VCBench, roughly 62% of Gemini 2.0 Flash's errors were perceptual (VCBench). On We-Math, GPT-4o drops from 72.8% on one-step problems to 43.6% on three-step ones (We-Math). A useful critic therefore has to say which kind of error occurred.
The second is evidence about feedback form, from text-only agents. SWE-PRM, a trajectory critic for coding agents organized around a taxonomy of agent inefficiencies, beat both unguided scalar scoring and action-prescriptive feedback. It raised SWE-bench Verified resolution from 40.0% to 50.6% (When Agents go Astray). In multimodal math, URSA trains a step-level scalar reward model and needed special machinery (relative "drop-moment" scoring and synthetic misreading injection) to avoid reward hacking and length bias (URSA).
Nobody has tested whether SWE-PRM's finding carries over to visual reasoning. The question matters more there because the characteristic multimodal failure is visual de-grounding. Long reasoning chains drift away from the image and invent visual features to justify textual logic (Decoupling Reasoning and Perception).
The specific hypothesis is that a critique saying "step 3 uses a value read from the y-axis that is inconsistent with the image" re-grounds the model. A score or a supplied correct value would not. The self-determination lens is used only as a source of falsifiable design contrasts:
- Informative versus controlling: a located diagnosis, compared with a scalar score or a dictated fix.
- Autonomy: the policy decides whether to re-inspect the image, rather than having re-inspection forced on it.
- Internalization: whether gains persist after the critic is removed.
No claim is made that models have motivation.
Research Program
The critic. We build a tiered critic on an open multimodal backbone of about 7–8B parameters. Given an image, a question, and a reasoning trace, it outputs three things: the index of the first faulty step, a tier label, and a one-to-two-sentence diagnosis. The diagnosis may name the step's dependency on the image. It may not state the correct value or answer. A string-level filter checks diagnoses against the gold answer and gold intermediate quantities, and any diagnosis containing them is rejected.
Tier labels. The three tiers follow the hierarchy common to MVU-Eval, PhD, and We-Math:
- Level 1, perception: misreading an object, a quantity, or text.
- Level 2, relation: spatial, attribute, or cross-image binding.
- Level 3, composition: logic, calculation, or combining steps.
Training data. Two existing techniques supply the training signal. We-Math's decomposition of 6.5K problems into 10.9K sub-problems gives verifiable ground truth for Level 1 and Level 2 steps. A trace that contradicts a solved sub-problem can therefore be labelled at the right tier. URSA's two generation methods supply the faulty steps. Binary error location by tree search finds the first incorrect step. Misinterpretation insertion injects controlled perceptual misreadings, which give clean Level 1 negatives. Level 3 negatives come from traces whose sub-steps are all correct but whose composite answer is wrong, which is exactly We-Math's "inadequate generalization" state. Diagnoses are written by a stronger model conditioned on the located step and tier, then filtered for leakage.
The policy. The policy is Qwen2.5-VL-7B. It receives the critique and may revise freely. It also has an optional tool that returns a crop or zoom of a region it names, and it alone decides whether to call it.
Two-stage test. First, at inference time, we run critique-and-revise loops under different feedback forms. Second, to test internalization, we fine-tune the policy on its own successful revisions under each feedback form, using identical example counts. We then evaluate the fine-tuned policy with the critic removed.
Baselines
Do-nothing reference. The base policy with standard chain-of-thought and with direct answering. MATH-Vision reports that chain-of-thought sometimes lowers multimodal accuracy, so both variants are needed.
Compute-matched sampling. Self-consistency over N samples, where N equals the total number of model calls our loop uses. This controls for the fact that our method simply spends more inference.
Strongest scalar-verifier competitor. URSA-style best-of-N with a step-level scalar reward model. The reported figure is 31.5% on MathVision for URSA-8B-PS-GRPO, against 30.39% for GPT-4o (URSA). We retrain an equivalent scalar critic on exactly our critic's data, with the same backbone and tuning budget. That way the only difference is what the critic outputs.
Strongest structural competitor. The training-free reasoner–observer loop, which reports 52.6 on MathVision and 54.4 on MathVerse with Qwen2.5-VL-72B and Qwen3-32B (Decoupling Reasoning and Perception). Those numbers come from far larger models. We therefore rerun the loop at matched scale, with a 7B observer and a reasoner of comparable size. We also report our critic attached to the larger pair to check that it adds value even there.
Perception-only competitor. RAPID raises Qwen2.5-VL-7B from 42.0% to 53.2% averaged over seven benchmarks by RL-tuning the perceiver's captions (RAPID). It is a competitor on the perception tier only.
Ablations isolating the key ingredient. The ablations hold the critic's step localization fixed and vary only the message:
- scalar score only;
- tier label only;
- free-text diagnosis with no tier;
- prescriptive fix that supplies the corrected value, which is the "controlling" condition;
- the full tier-located, non-prescriptive critique.
A separate ablation removes autonomy by forcing a zoom call after every Level 1 critique.
Success criterion. Beating the baselines means the full critique exceeds the matched scalar critic and the matched reasoner–observer loop on MathVision and on VCBench, at equal call budget, with confidence intervals that do not overlap.
Why not just…?
Why not just do self-consistency or best-of-N with a scalar reward model? URSA shows that a scalar reward model beats majority voting, so it is a real threat. Our claim is narrower: holding the critic's training data and localization fixed, words beat a number. The matched scalar ablation tests this directly. If the scalar critic ties the full critique at inference time, the remaining claim is internalization. If it ties there too, the proposal fails.
Why not just decouple perception from reasoning, as the reasoner–observer loop and RAPID do? These designs fix the perceptual tier well. However, GeoSQA found that models still fail when diagrams are perfectly transcribed into text (GeoSQA), and We-Math's composition cliff persists. Decoupling has no mechanism aimed at Level 3. We will report results per tier. If decoupling matches us on Level 1 and Level 2 errors but loses on Level 3, the critic's value lies in composition. If decoupling wins everywhere at matched scale, we will say so.
Why not just prompt the model to self-check against an error list? Self-rethinking with a six-class error taxonomy raised PaLM2's GSM8K accuracy from 56.29% to 65.13% in text (Can LLMs Learn from Previous Mistakes?). The failure mode we target, however, is a model that is confidently wrong about the image. Self-checking relies on the same faulty perception that made the error. We include self-rethinking with our three-tier list as a baseline. No reported result was found on Lacuna for self-rethinking on multimodal benchmarks.
Decisive Experiments
Benchmarks. Evaluation uses MathVision, MathVerse (especially its vision-only versions), VCBench, and We-Math. We-Math's four states let us track transitions directly. The prediction is that Level 3 critique converts "inadequate generalization" into "complete mastery." "Rote memorization" cases, where the final answer is right but sub-steps are wrong, should not increase.
Grounding check. On MathVerse, removing the diagram sometimes raises accuracy by more than 5%. A re-grounded policy should instead lose more accuracy when the diagram is removed than the baselines do. Our policy is evaluated with diagrams and with diagrams ablated.
Multi-image check. On VCBench, models score 42.3% higher when separate images are stitched into one. Our policy should narrow this gap, on the reasoning that Level 2 critiques target cross-image binding.
Internalization check. This is the decisive test. After fine-tuning on revisions, a critic-free policy trained under informative critique should beat policies trained under scalar or prescriptive feedback with the same number of examples. The scalar-trained policy should also show the shortening and evasion URSA reported. If prescriptive feedback internalizes equally well, the self-determination contrast is refuted for models, and we will report it that way.
Critic quality. We report tier-label accuracy against We-Math sub-problem ground truth, and the leakage rate after filtering. As a calibration point for whether trace-level labels carry signal, Bloom-level step labels predicted text reasoning correctness at AUC 0.676 (Cognitive Profiling). We expect located tier labels to exceed this.
References
- Measuring Multimodal Mathematical Reasoning with MATH-Vision Dataset
- Benchmarking Multimodal Mathematical Reasoning with Explicit Visual Dependency
- We-Math: Does Your Large Multimodal Model Achieve Human-like Mathematical Reasoning?
- When Agents go Astray: Course-Correcting SWE Agents with PRMs
- Unlocking Multimodal Mathematical Reasoning via Process Reward Model
- Decoupling Reasoning and Perception: An LLM-LMM Framework for Faithful Visual Reasoning
- MVU-Eval: Towards Multi-Video Understanding Evaluation for Multimodal LLMs
- PhD: A ChatGPT-Prompted Visual Hallucination Evaluation Dataset
- Reasoning-Aligned Perception Decoupling for Scalable Multi-modal Reasoning
- GeoSQA: A Benchmark for Scenario-based Question Answering in the Geography Domain at High School Level
- Can LLMs Learn from Previous Mistakes? Investigating LLMs' Errors to Boost for Reasoning
- Cognitive Profiling of LRMs' Reasoning Traces Using Bloom's Taxonomy