TL;DR
Large vision-language models (LVLMs) are notorious for hallucinating facts, and while standard conformal prediction can guarantee safety by filtering out ungrounded claims, it does so at a brutal cost—abstaining more than 80% of the time. This paper introduces Budgeted Conformal Evidence Acquisition (BCEA), a framework that allows models to "look again" via targeted visual interventions (like zooming or flipping) to gather more evidence when uncertain. Crucially, BCEA folds this active perception policy directly into the calibration pipeline, significantly boosting model utility while mathematically preserving strict, distribution-free error guarantees.
What Problem Does This Paper Address?
To deploy LVLMs in high-stakes environments like medicine or science, we need rigorous guarantees that their outputs are factual. Conformal prediction offers a solution by treating model claims as hypotheses and filtering out those that fall below a calibrated uncertainty threshold. This provides a finite-sample, distribution-free bound on the hallucination rate.
However, selective prediction is incredibly conservative. To guarantee a hallucination rate below 5% on standard benchmarks, a state-of-the-art conformal filter has to abstain on over 80% of all claims. A model that refuses to answer four out of five questions is safe, but practically useless.
The authors observe that human beings don't just give up when they are unsure about a small detail in an image; they look closer. While LVLMs can easily re-examine zoomed crops or alternative views, doing so adaptively at test time alters the score distribution. This breaks the exchangeability assumption that conformal prediction relies on, silently destroying the statistical safety guarantees.
Methods
BCEA transforms the standard binary "answer/abstain" decision into a three-way choice: answer, abstain, or acquire more evidence under a strict compute budget.
The pipeline works as follows:
- Base Scoring: A claim's grounding is initially scored using "global ungrounding"—measuring how much the model's confidence drops when the image is replaced with a blank black box.
- Three-Way Gating: If the base score is high, the model answers; if it's very low, it abstains. If it falls into a predefined "borderline band," it triggers the Acquire action.
- Evidence Acquisition: The model generates new views of the image. For object existence, it uses CLIP to select the top 5 most relevant zoomed crops. For spatial relations (e.g., "A is left of B"), it applies a structured horizontal flip.
- Conformal Re-calibration: To fix the broken guarantee caused by adaptive acquisition, BCEA folds the entire active perception policy (the crops, interventions, and the gate) directly into the scoring function. By calibrating the conformal threshold $\tau$ on these post-acquisition scores, calibration and testing pass through the exact same sample-independent map, fully restoring the mathematical guarantee.
Datasets
The authors evaluated BCEA using:
- COCO val2017: Used to construct controlled, ground-truth-verified claims covering object existence, spatial relations, and counting.
- POPE Benchmark: A standard object hallucination benchmark consisting of random, popular, and adversarial splits across four open-source LVLMs (LLaVA-1.5, Qwen3-VL, LLaVA-NeXT, and InternVL2).
Key Insights and Ideas
- Naive Acquisition is Dangerous: Simply taking a closer look when unsure and applying the original conformal threshold leads to an anti-conservative bias. It inflates the scores of both true and false claims, violating the risk guarantee across almost every data split.
- Global Ungrounding Fails on Relations: Prior conformal filters check if a claim's likelihood drops when the image is removed. This works for object existence but fails for spatial relations; removing the image drops the likelihood of "A is left of B" and "A is right of B" equally.
- Structured Interventions Restore Signal: Applying a claim-specific visual transformation like a horizontal flip explicitly reverses the truth value of a spatial relation. If a "left of" claim's confidence rises after a horizontal mirror flip, the model knows the original claim was likely a hallucination.
Key Results
- Higher Certified Coverage: On COCO existence claims with a guaranteed maximum error rate of $\alpha = 0.10$, BCEA boosted certified coverage from 23% (no acquisition) to 32% while maintaining zero empirical risk violations.
- Spatial Relation Breakthrough: Introducing the horizontal flip intervention lifted spatial-relation verification from near-chance (0.57 AUROC) to a highly effective 0.77 AUROC.
- Success on Free-Form Generation: When evaluated on open-ended descriptions generated by LLaVA-1.5, BCEA successfully transferred to self-generated claims. At a strict $\le 10\%$ guaranteed hallucination rate, BCEA certified 34% of claims compared to just 24% for standard abstention.
Limitations and Future Work
- Perceptual Boundaries: BCEA can only certify what the underlying backbone is capable of perceiving. For properties where the VLM is completely blind (e.g., complex counting or fine-grained color-object binding), the coverage collapses to zero.
- Text Parsing Constraints: The free-form caption pipeline currently relies on a parser that only extracts object mentions. Attributes, complex relations, and counts within captions remain unverified.
- Future Directions: The authors point to the creation of a comprehensive, automated atomic-fact decomposer and a broader library of per-claim-type visual interventions as the next logical steps for the framework.
Summary updated