Organizations: Columbia University, United States of America · UC Berkeley, United States of America · Texas A&M University, United States of America · Rice University, United States of America · Sun Yat-sen University, China · Stanford University, United States of America · Georgetown University, United States of America · Carnegie Mellon University, United States of America · University of Southern California, United States of America
Reliable vision-language reasoning requires a control decision that conventional grounding methods leave implicit: which input is entitled to determine the answer? Auxiliary text may be a defeasible claim about an image, a binding task instruction, or compatible context. Correcting every image-text conflict toward the image therefore repairs false claims but overwrites valid task rules. We formalize this setting as mixed-authority vision-language reasoning and introduce AITHER, an Authority-Informed Text-Image Hierarchy for Evidence Routing. Its final policy, the Semantic-Frame Router, is a training-free black-box control layer that (i) compiles explicit task directives, (ii) recovers standardized spatial frames from metadata-scrubbed pixels, (iii) resolves the semantic polarity of remaining notes, and (iv) routes among frozen observable candidates with fail-closed semantics. This ordered decomposition converts authority arbitration from an unconstrained prompt preference into a reproducible inference program. Across 1,620 MATRIX predictions, AITHER reaches 95.1% accuracy, exceeding the strongest task-and-frame baseline by 9.1 points with a 95% record-cluster interval of 6.8-11.5 points and no harmed records. It further reaches 100.0% on 1,800 controlled FRAME predictions and 93.3% on GQA-MATRIX. The results show that explicit authority routing can simultaneously improve false-claim correction, preserve task-defined answers, and eliminate clean-context overcorrection.
Figures & tables
Figure 1: Overview of AITHER. Binding task rules are resolved before spatial frame cues; unresolved notes are then assigned claim, context, or uncertain authority and routed to observable candidates. The hierarchy preserves valid instructions while selectively correcting false claims.
Vision-Language Models (VLMs) remain prone to hallucinations, producing fluent but visually unfaithful outputs. Existing chain-of-thought and retrieval-augmented methods only partially address this, as they neither enforce step-level citation grounding nor route verification failures back to retrieval for correction. We present CaVe-VLM-CoT, a modular reflection-based agentic-RAG framework that enforces evidence-grounded reasoning through a five-stage closed-loop pipeline: Extractor, Retriever, Solver, Citation Injector, and Verifier, in which detected ungrounded claims trigger structured feedback to the Extractor for targeted re-retrieval. Since no existing framework jointly measures retrieval quality, step-wise citation faithfulness, and cross-modal grounding, we propose a suite of 23 component-wise metrics across all stages, anchored by CaVeScore, a composite metric weighting accuracy, citation precision and recall, attribution, and evidence grounding. Without any architectural or prompt modifications, CaVe-VLM-CoT achieves 87.1% accuracy and 56.6% CaVeScore on ScienceQA , and 55.2% accuracy and 35.7% CaVeScore on MMMU (30 subjects).
Multimodal large language models increasingly reason over screenshots and documents where the task itself may be written in pixels. Yet benchmarks usually place questions in text, leaving it unclear whether models use the same instruction equally well across channels. We introduce Visualized Task Semantics (VTS), a controlled intervention that moves the question into the image while keeping the source problem and answer fixed. Across six MLLMs and four benchmarks, accuracy drops in all 24 model-task pairs, by 17.8 points on average. Models often transcribe the visual question correctly yet fail to use it, exposing a semantic channel gap beyond OCR. To reduce this gap, we present prompt-region grounding, whose core design aligns the question region with typed semantics and recovers its clean representation from a masked view. At matched training cost, our method raises four-benchmark VTS accuracy from 58.0 to 66.3 while preserving accuracy on the original interface, and requires no OCR or region metadata at inference. Reading task-bearing text and grounding it as an instruction for reasoning are distinct capabilities.
Yongxin Wang, Ruizhe Zhou, Yueling Tang +4
Mohamed bin Zayed University of Artificial Intelligence · Sun Yat-sen University · Transsion +1
Retrieval-augmented vision-language models (VLMs) answer from two sources of evidence that can each fail: the retrieved text can be wrong or noisy, and the image can be degraded or contradicted. We introduce GIB-200K, 200,000 image-question items built from COCO and Visual Genome, each expanded into controlled stress conditions: hard-negative retrieval noise (1-20 documents), visual ambiguity (Gaussian and motion blur, low-resource crops) and cross-modal conflict with counterfactual flip pairs. Instead of only measuring degradation, we ask how models combine the two modalities. From ideal Bayesian cue combination we derive a testable prediction that yields formal definitions of Knowledge Inertia (text over-weighted relative to the model's own image evidence) and Sensory Capture (the reverse), and we estimate the weights with a measurement-error-corrected regression whose recovery of known weights we verify. A counterfactual flip decomposition separates modality bias from concept priors, and a Grounding Illusion Score measures how much clean accuracy is not grounded in the image. We evaluate 4 open models on one frozen, seeded evaluation split of 2,000 GIB-200K items. All models are accurate on clean inputs (92.5%-94.2%) but adverse accuracy falls to 41.3%-59.9%. The two LLaVA models over-weight the text (Knowledge Inertia, w_T/w_I = 1.229 and 1.322), whereas InternVL3 favours the image (w_T/w_I = 0.680); between 34.8% and 71.4% of the clean answers that are correct are abandoned when the document lies. A training-free consistency check that compares the answer with and without retrieval trades coverage for accuracy under misleading documents, while falling back to the image fails when the image is heavily degraded. All per-instance outputs, code, and an audit that recomputes every reported open-model number are released.