Toward Guarantees for Clinical Reasoning: Solving the VLM "Illusion of Logic" with SMT Verifiers

TL;DR

Vision-Language Models (VLMs) are increasingly used to draft radiology reports, but they often struggle with logical consistency—saying there is a "blunted costophrenic angle" (a sign of fluid) but then failing to list "pleural effusion" in the diagnosis. This paper introduces a neurosymbolic verification framework that uses the Z3 SMT solver to mathematically prove whether a VLM's conclusions are actually supported by its observations.

The Problem: Stochastic Parrots in White Coats

The fundamental issue with current VLMs like Llama-Med or MedGemma is their autoregressive nature. They predict the next token based on statistical probability, not logical necessity. In clinical settings, this leads to two dangerous failure modes:

1. Hallucinated Impressions: Claiming a disease that the visual findings don't support.
2. Omitted Conclusions: Seeing all the symptoms but "forgetting" to name the diagnosis.

Existing metrics (BLEU, ROUGE) are useless here because they only check if the text looks like a human report, not if the logic is sound.

Methodology: Decoupling Perception from Logic

The authors argue that we should let VLMs handle the "messy" visual perception but let a symbolic solver handle the "rigid" clinical reasoning.

The Pipeline:

1. Autoformalization: A constrained LLM converts free-text "Findings" into a structured vector of atomic predicates (e.g., costophrenic_blunting = True).
2. Knowledge Base (K): Clinical guidelines are converted into propositional logic (e.g., Costophrenic_Blunting → Pleural_Effusion).
3. SMT Verification: The system uses the Z3 Solver to check: ${\Phi }_{Findings}\wedge \mathcal{K}\models Diagnosis$.

Figure 1: The framework decouples probabilistic perception from deterministic logic.

Experiments: Auditing the Brains of VLMs

The researchers audited 7 different models. By using the solver, they could categorize models in ways traditional metrics never could:

• The Conservative (Qwen3-VL-8B): Extremely high "Soundness" (doesn't lie), but low "Completeness" (misses obvious conclusions).
• The Stochastic (Llava-Vicuna-7B): Functions as a text generator, often hallucinating diagnoses that have zero support in its own findings.
• The Balanced (MedGemma-27B): Showed the best alignment between what it "saw" and what it "concluded."

Symbolic Filtering as a Safeguard

By using the solver as a post-hoc filter—simply deleting any diagnosis the solver couldn't prove—the authors achieved a massive jump in Precision across the board.

Table 1: The "Ours" column shows the consistent increase in Soundness and Precision when symbolic filtering is applied.

Critical Insights & Takeaways

This paper marks a shift from Empirical Evaluation (does it look right?) to Formal Verification (is it provably right?).

• Logic over Lexicon: Overlap metrics like BLEU are near-zero in clinical settings because doctors use diverse synonyms. Formal logic ignores synonyms and focuses on the underlying truth.
• The "Assume-Guarantee" Paradigm: The system is only as good as the "Autoformalizer." If the model misidentifies the visual finding, the logic will be "correctly wrong." This highlights the need for radiologists to verify the Findings section while the machine guarantees the Impression.
• Trade-offs: Improving precision via symbolic filtering leads to a slight drop in recall. In medicine, this "conservative" bias is usually preferred over reckless hallucination.

Conclusion

As we move toward deploying AI in hospitals, "black box" models aren't enough. By wrapping probabilistic VLMs in a "logical cage" of SMT solvers, this research provides a concrete path to clinical tools that clinicians can actually trust.