Multimodal Fact-Level Attribution for Verifiable Reasoning

Abstract: Multimodal LLMs (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.

• The paper introduces MuRGAt, a benchmark evaluating fact-level attribution in multimodal reasoning by assessing precise citation and evidence traceability.
• It presents a protocol that decomposes model responses into verifiable claims and atomic facts, using the MuRGAt-Score to measure citation coverage and attribution quality.
• Empirical results reveal a trade-off between narrative synthesis and grounding precision, underscoring the need for programmatic frameworks to improve citation reliability.

Multimodal Fact-Level Attribution for Verifiable Reasoning: Expert Summary

Motivation and Context

The paper "Multimodal Fact-Level Attribution for Verifiable Reasoning" (2602.11509) introduces MuRGAt, a benchmark and evaluation protocol for fact-level multimodal attribution, targeting the verifiability of claims in the outputs of Multimodal LLMs (MLLMs). Previous benchmarks have largely focused on observational or retrieval-based tasks, emphasizing direct grounding in restricted modalities—most notably visual inputs—without rigorous assessment of complex reasoning chains or explicit evidence tracking across heterogeneous multimodal sources (audio, video, graphs). The authors identify that the risk of factual hallucinations is amplified in multi-step, long-form generation, necessitating fine-grained attribution evaluation that distinguishes directly observable claims from reasoning steps and ensures precise citation of supportive evidence.

MuRGAt moves beyond direct observation scenarios, demanding not only answer correctness but explicit, atomic-level reasoning with citations specifying both modality and temporal segments. The central observation is that even high-performing MLLMs often synthesize correct answers but fail to provide sufficient and accurate supporting evidence, exhibiting a substantial gap between internal reasoning and verifiable attribution.

Figure 1: Overview of MuRGAt and its evaluation protocol, featuring input of multimodal sources, generation of explicit reasoning with precise citations, and three-stage fact-level attribution evaluation.

The MuRGAt Benchmark and Protocol

Task Definition

MuRGAt presents inputs spanning video, audio, and figures, together with questions requiring reasoning that cannot be satisfied by direct retrieval. Models generate responses containing explicit reasoning and precise citations delineating the supporting modality and temporal/segment information for each claim.

The evaluation pipeline decomposes responses into:

1. Verifiable Claim Identification: Differentiating sentences containing directly observable content (subject to grounding) from reasoning steps or meta-statements.
2. Atomic Fact Decomposition: Further breaking verifiable sentences into atomic, minimal claims, supporting fine-grained evaluation and propagation of citation information.
3. Attribution Quality Assessment: Determining whether the cited evidence strictly and sufficiently entails each atomic fact, scoring both recall (completeness) and precision (necessity).

Figure 2: Annotation UI for atomic fact decomposition, facilitating the breakdown of complex sentences into minimal, verifiable claims.

Figure 3: Annotation UI for attribution quality, enabling annotators to assess support and necessity of citations for each atomic fact.

Evaluation Metric: MuRGAt-Score

The core evaluation metric, MuRGAt-Score, is defined as the product of citation coverage (proportion of verifiable sentences with at least one citation) and attribution F1 (harmonic mean of recall and precision of evidence support for atomic facts). This metric penalizes models that hallucinate citations for a subset while neglecting most claims, offering a stringent holistic assessment.

The protocol leverages both human and automatic evaluation. Automatic methods utilize strong LLMs to simulate the three annotation stages, achieving high correlation with human judgments (Pearson $r = 0.84$, well above standard LLM-as-judge baselines).

Empirical Analysis: Model Behavior and Failure Modes

Human Evaluation Results

The initial human-annotated experiments show that state-of-the-art MLLMs (Gemini and Qwen variants) lag far behind expert-level performance on MuRGAt-S. For instance, Gemini-3-Pro achieves only 49.7 MuRGAt-S on WorldSense (video/audio reasoning) and 16.3 on Video-MMMU (visual/graph reasoning), despite high QA accuracy. Attribution remains the bottleneck, encompassing both missing citations and spurious/irrelevant ones.

Qualitative analysis reveals a fundamental trade-off between narrative synthesis and grounding precision: large models frequently hallucinate spatial/temporal details to maintain coherent narratives at the cost of grounding faithfulness. Minimalist, shot-by-shot models (e.g., Gemini-2.5-Flash) perform better in attribution, avoiding contextualization traps.

Figure 4: Gemini models' performance across thinking levels, illustrating divergent trends in attribution quality as reasoning effort increases.

Automatic Evaluation Pipeline

Variants of the prompting protocol (Simple, Chain-of-Thought, JSON-structured) are benchmarked for verifiable claim identification and atomic fact decomposition. Sentence-level decomposition and explicit decontextualization steps consistently outperform response-level approaches and single-pass prompts, underscoring the necessity of pipeline separation for high-quality, interpretable attribution.

Generation Experiments and Analysis

On both WorldSense and Video-MMMU, requiring citation generation delineates a task-dependent impact:

• On recognition-focused datasets, explicit citation requirements act as a "reasoning tax," slightly decreasing accuracy.
• On reasoning-intensive tasks, citation generation scaffolds reasoning chains and boosts accuracy.

High answer accuracy does not correlate with reliable grounding. Models frequently produce correct answers without verifiable supporting evidence—the attribution scores remain low due to hallucinated or missing citations.

Post-hoc attribution (forcing citation generation after answer synthesis) improves coverage and recall in perception-driven tasks but severely degrades precision in reasoning-intensive domains, leading to erroneous or performative citation (mapping deductions to random segments).

Omni models capable of audio processing outperform vision-LLMs on multimodal tasks, but modality-agnostic grounding remains problematic, with hallucinated audio citations prevalent in vision-only models.

Figure 5: Gemini-3-Flash performance with program-aided generation on Worldsense, highlighting attribution gains via programmatic structures.

Scaling, Reasoning, and Programmatic Grounding

Increasing model reasoning depth exhibits non-monotonic effects: smaller models see attribution quality drop as thinking effort increases (latent reasoning disconnects from verifiable evidence), while larger models align chains of reasoning with external evidence more effectively.

Programmatic approaches (plan-then-execute paradigms) imposing explicit separation between reasoning and citation generation yield substantial attribution improvements (+9.6 points MuRGAt-S), but consistently reduce answer accuracy. Specifically, imperative paradigms (executor-discovered grounding) outperform declarative (planner-defined grounding) in attribution but suffer from noise propagation and restricted reasoning flexibility.

Implications and Future Directions

Theoretical and Practical Implications

The work demonstrates a fundamental disconnect between answer synthesis and evidentiary grounding in MLLMs. MuRGAt-Score exposes critical model failure modes: hallucination in citations, performative attribution, and trade-offs between narrative fluency and grounding faithfulness.

Existing benchmarks and metrics insufficiently capture these issues; MuRGAt and its protocol enable robust evaluation for real-world multimodal reasoning tasks, advancing towards models capable of truly faithful and interpretable outputs. The findings imply that simply scaling models or increasing reasoning depth is insufficient—novel architectures and training strategies are needed to reconcile internal reasoning processes with verifiable, grounded evidence.

Prospects for Future AI Developments

Future research must address the reconciliation of accuracy and faithfulness by:

• Developing modular and program-aided reasoning frameworks that structurally enforce citation discipline.
• Investigating agentic and iterative multimodal inference strategies to align latent reasoning with evidence retrieval.
• Incorporating dynamic refinement mechanisms that optimize attribution throughout execution and synthesis.
• Designing evaluation protocols capable of decomposing complex outputs into interpretable, verifiable substrates.

Advances along these axes are likely necessary for deploying reliable MLLMs in applications demanding transparency, auditability, and trustworthy decision support.

Conclusion

MuRGAt constitutes a rigorous benchmark and protocol for evaluating multimodal fact-level attribution in MLLMs, focusing on the verifiability and faithfulness of generated claims. The framework's automatic evaluation achieves high correlation with human assessment and exposes persistent gaps in attribution quality among leading models. The results underscore the distinction between correct answer synthesis and verifiable grounding, motivating programmatic and pipeline-based advances in model architectures and training. MuRGAt and MuRGAt-Score provide essential infrastructure for evaluating and improving multimodal reasoning, with significant implications for the future of interpretable and faithful AI systems.