RUBRIC-MME: Real-User Behavior-grounded Rubric for Multimodal Interaction Capability Evaluation

RUBRIC-MME is a benchmark for evaluating whether multimodal large language models (MLLMs) can sustain useful, goal-directed interactions over multiple turns. Unlike benchmarks that focus mainly on isolated visual questions or single-turn correctness, RUBRIC-MME evaluates both individual responses and complete interactions, and provides structured evidence about recurring model weaknesses.

Paper: RUBRIC-MME: Real-User Behavior-grounded Rubric for Multimodal Interaction Capability Evaluation, NeurIPS 2026, Evaluations and Datasets Track.

Dataset: https://huggingface.co/datasets/RobinTeng243/RUBRIC-MME
Compact sample: https://huggingface.co/datasets/RobinTeng243/RUBRIC-MME_small
Code repository (as listed in the current OpenReview record): https://github.com/IntMeGroup/RUBRIC-MME

1. What the benchmark measures

RUBRIC-MME evaluates multimodal interaction competence across two interaction modes:

• Streaming-video dialogues: the model receives visual information incrementally, aligned with the time of each user query.
• Multi-image dialogues: images are revealed across turns, so the model must use the visual context available at each point in the conversation.

The evaluation uses a 12-dimensional rubric at two levels:

• Turn level: evaluates the quality of individual assistant responses in their local dialogue and visual context.
• Session level: evaluates the complete interaction, including consistency, adaptation, recovery from intent changes, and progress toward the user's goal.

Results can also be analyzed by scenario, capability, and modality. The benchmark covers seven real-user-oriented scenarios and a 61-category capability taxonomy.

2. Dataset composition

The released benchmark contains 734 dialogues and 5,859 turns:

Interaction mode	Dialogues	Turns
Streaming video	426	3,658
Multi-image sequences	308	2,201
Total	734	5,859

These counts describe the benchmark reported in the paper. The hosted files should be checked against these totals when preparing a new release or version.

3. Construction and provenance

The benchmark is behavior-grounded but does not reproduce or release private user conversations.

1. The authors analyzed approximately 35,000 records from the LingGuang multimodal assistant deployment (January–February 2026) to derive aggregate patterns in scenarios, user intents, and interaction structures.
2. These aggregate patterns were used as design priors for selecting and balancing public visual sources.
3. Multimodal dialogues were instantiated through a scenario–persona–intent–dialogue construction process.
4. Candidate interactions were audited by 15 full-time quality-assurance staff with experience in multimodal-content review, dialogue quality control, and policy compliance. The paper reports an overall retention rate of approximately 73%.

Raw deployment logs are not included in this dataset. No user-submitted prompts, responses, images, videos, audio, or voice recordings from the deployment logs are released as benchmark items. Only aggregate behavioral patterns inform benchmark construction.

4. Intended use

RUBRIC-MME is intended for:

• Benchmarking MLLMs on multi-turn multimodal interaction.
• Comparing turn-level and session-level interaction quality.
• Diagnosing model weaknesses across capabilities, scenarios, and modalities.
• Research on rubric-based evaluation, MLLM-as-a-judge protocols, and structured failure attribution.
• Studying behavior-grounded benchmark construction.

The paper reports that the full 12-dimensional rubric predicts independent human preferences more accurately than answer accuracy alone (82.6% versus 65.8% in the reported study).

Uses not validated or recommended

RUBRIC-MME is an evaluation benchmark, not a training corpus. The authors do not recommend using its samples for pretraining, supervised fine-tuning, instruction tuning, preference optimization, or reinforcement learning.

The benchmark has not been validated as:

• A safety or deployment-readiness certification tool for medical, legal, financial, industrial, autonomous-driving, or other high-stakes applications.
• A fairness, demographic-bias, or harm benchmark.
• An adversarial robustness or jailbreak benchmark.
• An evaluation of audio, real-time bidirectional audio-visual interaction, screen-based tool use, or tool-using agents.

A strong score should not be interpreted as a guarantee of safe or reliable performance outside the benchmark's documented scope.

5. Data sources, privacy, and sensitive information

Deployment data

The deployment records were used only within the applicable institutional governance process to derive aggregate scenario, intent, and interaction-mode distributions. Raw records and direct identifiers are not part of the public release.

Public visual sources

Benchmark visual inputs are drawn from public web sources and are subject to asset-level provenance, terms-of-use, license, and privacy screening. Public accessibility does not imply permission to redistribute a video or extracted frame.

Only media assets whose applicable rights and screening results permit redistribution may be included as files. For other sources, the release uses source references and permitted derived annotations. The asset-level manifest, where provided, is the authoritative record for the source, applicable terms, attribution requirements, permitted release form, and takedown status.

Public videos may contain incidental images of identifiable people. The authors describe human screening and removal of samples that fail privacy review; residual risk cannot be ruled out completely.

Synthetic information

Persona specifications and dialogue content are synthesized. Persona attributes such as age, gender, occupation, knowledge level, and emotional state do not describe identified deployment users. They should not be interpreted as demographic labels for real people.

6. Known limitations and biases

Users should account for the following limitations when interpreting results:

• Deployment-source skew: behavioral priors come from one deployed assistant and a geographically and linguistically concentrated cohort, primarily Mandarin-speaking users in mainland China.
• Public-source selection bias: public videos may over-represent well-lit, well-framed, demonstrative content and under-represent casual, low-quality, first-person footage.
• Limited interaction scope: the benchmark covers streaming video and multi-image sequences, but not audio, real-time bidirectional interaction, or tool use.
• Synthetic-content limitations: dialogues, personas, and reference responses are model-assisted and human-audited; residual generation artifacts and generator biases may remain.
• Judge dependency: the reported rubric scoring uses Gemini 2.5 Pro. Potential judge-family preference effects are not assumed to be fully eliminated; evaluations of Gemini-family models should include alternative judges where possible.
• Population coverage: the benchmark is not designed to estimate average global deployment-traffic performance. Demographic, cultural, accessibility, and other subgroup performance remains incompletely validated.

The dataset is therefore best used as one component of a broader evaluation process, alongside human review and domain-specific testing.

7. License and third-party materials

The current Croissant metadata declares CC BY-NC-SA 4.0 for the dataset release:

https://creativecommons.org/licenses/by-nc-sa/4.0/

This declaration must be interpreted together with the asset-level release manifest. It does not override the copyright, license, platform terms, or attribution obligations of third-party videos, frames, or other source materials. Do not download, copy, modify, or redistribute a third-party media asset unless its individual release record permits that action.

The benchmark is intended for research evaluation. The evaluation-only recommendation above is a statement of intended use and should not be confused with a separate legal restriction in the CC BY-NC-SA license.

8. Download

The full dataset is hosted on Hugging Face:

pip install -U huggingface_hub

hf download RobinTeng243/RUBRIC-MME \
  --repo-type dataset \
  --local-dir ./RUBRIC-MME

A compact sample is hosted separately:

hf download RobinTeng243/RUBRIC-MME_small \
  --repo-type dataset \
  --local-dir ./RUBRIC-MME_small

The compact sample is provided for inspection. It is not a replacement for the full benchmark. Please consult the sample's accompanying documentation for its exact selection procedure and contents.

9. Code and reproduction

The companion code repository is intended to provide the evaluation and analysis implementation, including the scripts and instructions needed to reproduce the paper's reported results:

https://github.com/IntMeGroup/RUBRIC-MME

Please consult that repository for the current dependency list, execution commands, model/API configuration, and expected outputs. Do not commit API keys, access tokens, private deployment logs, or other credentials to the repository.

10. Citation

Please cite the paper as:

Jiajie Teng, Jianping Jiang, Huiyu Duan, Jingdong Chen, Yi Yuan, Guangming Yao, Sijing Wu, Yuqin Cao, Yixuan Gao, Xiongkuo Min, and Guangtao Zhai. RUBRIC-MME: Real-User Behavior-grounded Rubric for Multimodal Interaction Capability Evaluation. NeurIPS 2026, Evaluations and Datasets Track.

---

For questions about the benchmark, please use the issue tracker of the canonical code repository after the repository URL has been confirmed.