Multimodal Fact-Level Attribution for Verifiable Reasoning

AI summary

Overview

• Research area: Multimodal NLP, specifically the evaluation of grounding and attribution in multimodal large language models (MLLMs) that perform multi-step reasoning over video, audio, and figures.
• Technical level: Intermediate — the paper assumes familiarity with MLLMs, retrieval-augmented generation, and metrics like precision/recall, but its core ideas are explained with concrete examples.
• Scope: The paper introduces a benchmark (MuRGAt) and an automatic scorer (MuRGAt-Score) for testing whether MLLMs can cite the exact modality and timestamp supporting every factual claim they make while reasoning over multimodal inputs.

What This Paper Is About

Most existing benchmarks for multimodal grounding test simple observation: a model sees an image or video and points to the region or moment that answers a question. Real-world tasks are harder, because the model must reason over what it observes and cite evidence for each intermediate claim, not just the final answer. This paper builds a benchmark that requires multimodal models to produce explicit reasoning with fine-grained citations, and an automatic evaluation method that checks whether each cited segment actually supports the claim it is attached to.

Key Contributions

1. The MuRGAt benchmark. A task where a model receives heterogeneous inputs (video, audio streams, figures) plus a question, and must produce a response containing explicit reasoning plus citations that each name a specific modality and a timestamped segment. Unlike prior work, it requires self-selected evidence and joint grounding across modalities rather than source-level attribution.

2. A three-stage evaluation protocol. Responses are evaluated by (1) identifying which sentences are verifiable (directly observable in the inputs) versus which are reasoning steps, (2) decomposing verifiable sentences into atomic facts with pronoun decontextualization and citation propagation, and (3) measuring attribution quality via recall (do the cited segments jointly entail the fact?) and precision (is each cited segment strictly necessary?).

3. MuRGAt-Score, an automatic metric. A pipeline combining the best-performing model configurations for each subtask (Gemini-3-Pro for verifiability, Gemini-3-Flash for decomposition, Gemini-2.5-Flash for entailment). Coverage multiplied by attribution F1 yields a single number that correlates with human judgments at an average Pearson r = 0.84, well above the next-best LLM-as-judge baseline at r = 0.59, and reaching r = 0.97 on coverage.

4. A systematic diagnostic study of MLLM attribution. Experiments across eight models (Gemini, Qwen3-Omni, Qwen3-VL, Molmo2) reveal where grounding fails, including hallucinated citations, task-dependent costs of citing, and diverging effects of test-time compute.

Main Findings

• Correct answers do not imply correct grounding. Models frequently hallucinate citations even when their final answer is right. Peak MuRGAt-Score was 69.2 on WorldSense and 56.9 on Video-MMMU (Gemini-3-Flash), meaning roughly 30–35% of claims remain ungrounded.

• Accuracy and attribution decouple. On Video-MMMU, Gemini-3-Pro and Gemini-3-Flash reached identical accuracy (86.0), yet Gemini-3-Flash scored far higher on MuRGAt-Score (56.9 vs. 41.8). Stronger models appear to lean on parametric knowledge while fabricating supporting evidence.

• Citation generation is task-dependent. Requiring citations imposes a "reasoning tax" on recognition-heavy WorldSense (Gemini-3-Pro drops from 71.4% to 70.0% accuracy) but scaffolds reasoning on Video-MMMU (Gemini-3-Pro rises from 85.3% to 86.0%, Qwen3-VL-Thinking jumps from 51.0% to 60.0%).

• Post-hoc attribution splits by domain. Applying citations after generation boosts MuRGAt-Score on WorldSense (Gemini-3-Pro: 51.7 to 65.2) but hurts on Video-MMMU (Gemini-2.5-Flash: 41.5 to 38.0), because it force-aligns abstract reasoning steps to arbitrary segments.

• Vision-language models fake audio grounding. VL models with no audio encoder still produced audio citations in up to 31.6% of their references, indicating their apparent grounding often reflects visual proxies or hallucination rather than genuine auditory understanding.

• Thinking effort scales inconsistently. On WorldSense, Gemini-3-Flash declined as reasoning effort increased (69.7 to 64.4 MuRGAt-Score), while Gemini-3-Pro improved by 6.1 points. On Video-MMMU, Gemini-3-Flash peaked at medium effort.

• Programmatic grounding trades accuracy for attribution. Extending program-aided generation to the multimodal setting improved attribution by an average of +9.6 MuRGAt-Score on WorldSense (logic-imperative variant reached 76.4), but answer accuracy fell by 7.4 points on average.

• CoT models have a format failure mode. Qwen3-Omni-Thinking gained accuracy from citations (+9.0% on Video-MMMU) yet produced invalid timestamps, collapsing to a MuRGAt-Score of 4.8 without post-hoc recovery.

Methodology in Plain English

The researchers first built a benchmark from two existing multimodal datasets, WorldSense (recognition-focused) and Video-MMMU (reasoning-intensive), covering inputs such as video, audio, and on-screen plots. Models must answer questions while annotating their responses with citations like "(audio, 0:42–0:46)" or "(video, 1:15–1:22)".

To grade these responses, they designed a three-step decomposition. First, an LLM judge labels each sentence as verifiable or not, filtering out pure reasoning statements so the model is not penalized for ungrounded thinking. Second, verifiable sentences are split into atomic facts, with pronouns resolved and citations inherited from the parent sentence. Third, each fact is checked against its cited evidence twice: once to see if the union of citations entails the fact (recall), and once to see whether each individual citation is necessary (precision).

They collected human annotations across all three stages (580 sentence-level verifiability examples, 635 decomposition examples, 917 entailment examples) and used them to select which models and prompts to plug into each stage of the automatic pipeline. They then compared this pipeline's scores against human scores using Pearson correlation and benchmarked eight MLLMs under three settings: base generation, generation with citations, and post-hoc attribution. Finally, they varied reasoning effort and tested program-aided grounding frameworks.

Why This Matters

• Research impact. The paper separates two capabilities that prior benchmarks conflated: reasoning correctly and grounding claims in evidence. MuRGAt-Score gives the field a scalable way to measure the second independently of the first, which matters for any work on hallucination, retrieval-augmented generation, or trustworthy multimodal systems.

• Real-world applications:

  • Medical or scientific video analysis, where a model summarizing a procedure must point to the exact moment and modality supporting each statement.
  • Educational content and tutoring systems that walk through reasoning over lecture videos and diagrams while citing sources students can verify.
  • Legal and journalistic review of recorded evidence, where every claim drawn from footage or audio must be traceable to a timestamp.
  • Multimodal assistants that summarize meetings or surveillance footage and need to defend each factual assertion with a citation.

• Industry relevance. Companies deploying MLLMs in regulated or high-stakes settings need auditability. This work provides both a diagnostic that reveals when a model's confidence is unearned and a demonstrated trade-off between structured grounding and raw accuracy, which directly informs prompt design and system architecture choices.

Future Directions

• Reconciling reasoning and grounding. The observed trade-offs (citations help some tasks and hurt others; programmatic scaffolding improves attribution but lowers accuracy) suggest a need for architectures that preserve reasoning flexibility while enforcing verifiable attribution.

• Fixing the thinking-effort disconnect. Smaller models degrade in attribution as reasoning effort rises, implying latent reasoning drifts away from cited evidence. Understanding and correcting this divergence is an open problem.

• Making attribution cheaper at scale. The pipeline currently relies on multiple strong proprietary models for different subtasks. Distilling or training a dedicated open evaluator could make benchmarking practical for the broader community.

• Handling cross-modal evidence properly. The finding that VL models fabricate audio citations shows current systems conflate modalities. Better protocols and training signals are needed to force genuine per-modality grounding rather than visual proxies.

Target Audience

Researchers and practitioners working on multimodal LLMs, hallucination detection, retrieval-augmented generation, and evaluation methodology will benefit most. The paper is also relevant to engineers building verifiable or auditable multimodal systems in domains like healthcare, education, and media analysis, and to anyone designing benchmarks that need to distinguish correct answers from trustworthy ones.

Authors’ abstract

Multimodal large language models (MLLMs) are increasingly used for real-world tasks involving multi-step reasoning and long-form generation, where reliability requires grounding model outputs in heterogeneous input sources and verifying individual factual claims. However, existing multimodal grounding benchmarks and evaluation methods focus on simplified, observation-based scenarios or limited modalities and fail to assess attribution in complex multimodal reasoning. We introduce MuRGAt (Multimodal Reasoning with Grounded Attribution), a benchmark for evaluating fact-level multimodal attribution in settings that require reasoning beyond direct observation. Given inputs spanning video, audio, and other modalities, MuRGAt requires models to generate answers with explicit reasoning and precise citations, where each citation specifies both modality and temporal segments. To enable reliable assessment, we introduce an automatic evaluation framework that strongly correlates with human judgments. Benchmarking with human and automated scores reveals that even strong MLLMs frequently hallucinate citations despite correct reasoning. Moreover, we observe a key trade-off: increasing reasoning depth or enforcing structured grounding often degrades accuracy, highlighting a significant gap between internal reasoning and verifiable attribution.