# Mmkc Bench Eval

> Evaluates how large multimodal models (LMMs) handle factual knowledge conflicts between their internal parametric knowledge and external multimodal evidence. It probes both behavioral alignment (whether models follow internal knowledge or external context) and conflict detection capabilities across coarse- and fine-grained settings. Use when the user wants to benchmark on MMKC-Bench, or asks about evaluating this task. Reports Detection Accuracy.

- Skill: `qhjqhj00/mmkc-bench-eval` (Agent Skill)
- Install (CLI): `npx skillmds add qhjqhj00/mmkc-bench-eval`
- Raw SKILL.md: https://api.skillmd.com/api/skills/qhjqhj00/mmkc-bench-eval/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Productivity
- Author: qhjqhj00 (https://skillmd.com/u/qhjqhj00)
- Updated: 2026-09-08
- Page: https://skillmd.com/skills/qhjqhj00/mmkc-bench-eval

---


# mmkc-bench-eval

> Benchmarking Multimodal Knowledge Conflict for Large Multimodal Models — Jia et al. (2025) (arXiv:2505.19509, 2025)

## What this evaluates

Evaluates how large multimodal models (LMMs) handle factual knowledge conflicts between their internal parametric knowledge and external multimodal evidence. It probes both behavioral alignment (whether models follow internal knowledge or external context) and conflict detection capabilities across coarse- and fine-grained settings.

## Datasets

- **MMKC-Bench** — total 1573; splits: test (-1); repo https://github.com/MLLMKCBENCH/MLLMKCBench

## Metrics

- `Original Answer Ratio (OAR)` — range: percent
  - Proportion of model predictions that match the model's original answer in a non-conflict setting.
- `Counter Answer Ratio (CAR)` — range: percent
  - Proportion of model predictions that align with the provided external conflicting evidence.
- `Irrelevant Answer Ratio (IAR)` — range: percent
  - Proportion of model predictions that are inconsistent with both the internal knowledge and the external evidence. OAR + CAR + IAR = 1.
- `Detection Accuracy` **(primary)** — range: percent
  - Binary classification accuracy for determining whether a knowledge conflict exists in the provided context. Evaluated in both coarse-grained (full evidence) and fine-grained (single sentence subset) settings.

## Input / output format

**Input**: Interleaved multimodal context (images and text) representing either one conflicting piece of evidence (context-memory conflict) or two conflicting pieces (inter-context conflict), followed by a question. For detection tasks, the input is either full evidence or a single sentence subset, with a binary judgment prompt.

**Output**: For behavior analysis: a QA answer (multiple-choice selection or open-ended text). For conflict detection: a binary 'yes' or 'no' indicating the presence of a conflict.

## Scoring recipe

```python
baseline_answers = get_non_conflict_answers(model, questions)
conflict_answers = get_conflict_answers(model, context, questions)
OAR = sum(1 for p, g in zip(conflict_answers, baseline_answers) if p == g) / len(questions)
CAR = sum(1 for p, g in zip(conflict_answers, external_evidence_answers) if p == g) / len(questions)
IAR = 1 - OAR - CAR

predictions = model.predict('yes/no', context)
gold_labels = [1 if conflict else 0 for context in dataset]
detection_accuracy = sum(1 for p, g in zip(predictions, gold_labels) if p == g) / len(dataset)
```

## Common pitfalls

- OAR, CAR, and IAR require establishing a baseline non-conflict answer first; skipping this step invalidates the ratios.
- Open-ended QA formats yield significantly higher IAR than multiple-choice formats due to output variability, which can skew behavioral analysis if not controlled.
- Fine-grained detection uses only a single sentence subset of the evidence, making it harder than coarse-grained detection and potentially underestimating model capabilities if evaluated only at the coarse level.

## Evidence (verbatim from paper)

> For conflict behavior analysis, we assess how conflicting contexts influence the model's answers to QA pairs. Each model prediction under a conflict scenario is categorized into one of three types: (1) consistent with the model's answer in the non-conflict setting, (2) consistent with the external conflicting evidence, and (3) inconsistent with both, referred to as an irrelevant answer. To enable this, we first perform QA under a non-conflict setting to establish the model's internal knowledge. We then compute three ratios: Original Answer Ratio (OAR), Counter Answer Ratio (CAR), and Irrelevant Answer Ratio (IAO), with OAR + CAR + IAO = 1 across the dataset. For conflict detection, we treat this as a binary classification task. If a knowledge conflict exists, the model should output "yes"; otherwise, it should output "no", in both coarse and fine-grained settings. Accordingly, we report the detection accuracy as the evaluation metric.

## Citation

```bibtex
@misc{jia2025benchmarkingmultimodal,
  title={Benchmarking Multimodal Knowledge Conflict for Large Multimodal Models},
  author={Jia et al. (2025)},
  year={2025},
  note={arXiv:2505.19509}
}
```

- arXiv: 2505.19509