Correct answers from multimodal large language models (MLLMs) do not guarantee that their stated reasoning is grounded in what they actually saw or heard. That is the central finding of ANCHOR, a new benchmark introduced by Julin and colleagues and accepted at NeurIPS 2026. Existing audio-visual QA benchmarks typically score only final answers, leaving it unclear whether a model's rationale correctly identifies the visual regions, time intervals, and audio events that support its prediction.
The benchmark pairs 2,319 videos with 8,272 question-answer items, each accompanied by structured reasoning traces, temporal evidence, spatial bounding boxes, and salient audio cues. A smaller human-refined test set of 224 videos and 975 questions allows the authors to separately score answer correctness, hallucination, logical divergence, and conciseness, with direct spatial and temporal grounding measurements on a focused subset.
Evaluating twelve contemporary MLLMs on the full benchmark, the authors found a substantial alignment gap: models with comparable answer performance could differ markedly in whether their explanations were supported by annotated evidence. On a matched subset of 100 videos and 446 questions, model rankings were highly stable across human, Gemini-generated, and GPT-generated references (mean Pearson r = 0.99), indicating the results are not an artifact of reference circularity.
Ablation experiments on three MLLMs revealed a consistent modality hierarchy. Removing video caused the largest accuracy drops (38.4–41.8 points), while removing bounding boxes (3.2–7.7 points) and audio (4.2–5.6 points) produced smaller, complementary, and model-dependent declines. The authors argue that answer-only evaluation overstates model competence and positions ANCHOR as a testbed for assessing evidence-aligned audio-visual reasoning rather than superficial correctness.