OmniVChat (Omni Video Chat) is the task of native audio-visual dialogue: an omni model directly and simultaneously receives audio and video from a user and returns text. The user's query is inside the audio and video. There is no separate text question, no external captioning, no ASR transcript.

This is the code-only OmniVChat repository. It contains the evaluation and reward code, examples, and documentation, while the full dataset annotations and media are distributed separately through the complete repository linked below.

The release contains two components from the paper. They grade a reply against the same rubric, but they answer different questions and their scores live on different scales, so a number from one is not comparable with a number from the other:

Both grade the same rubric with the same judge and the same tier gate — eval/score.py imports that core from reward/reward.py, so the rubric term has exactly one definition. What differs is what is added on top and how a judge failure is handled. See Why the two scores are not comparable.

Two rubrics for MSA-PLA. The 200 MSA-PLA instances ship with two independent rubrics for the same dialogue, because the right answer depends on an assumption you have to make explicit. key_points assumes the model has no body: asked to push a window open, it should say it cannot act and offer an alternative. key_points_embodied assumes an embodied model that can act: it should accept the task, identify the right object, and answer briefly. Every other subcategory has one rubric. Pick one for MSA-PLA and say which — a score is not comparable across the two. Details in MSA-PLA: two rubrics.

Eight dialogues, covering all five abilities. Each clip is the user's side: the model receives exactly this audio and video, and returns text.

The previews below are downsampled, and are not what a model sees. They are cut down hard so this page loads at all: 260 px wide, 10 fps, 64 colours, silent, and only the first 3.2 seconds. Click one to play it with sound — the query is in the audio, so a silent preview tells you half of it at best. Even that link is a 480p re-encode. The benchmark media itself is 1080p with audio, in the dataset (Harland/OmniVChat) — this repository carries the code and the examples, not the media.

AH AH-AOR She asks what the model thought of the lullaby she just sang. She never sang.	MEA MEA-RC-AR He asks if “this one” suits a five-year-old; only the video identifies the painted wooden elephant.	DSLP DSLP-VTT-UPC He stops mid-sentence, still thinking. The model should wait, not answer.	MSA MSA-PLA He asks the model to put a circuit board on the rack — it has no hands.
ER ER Disappointed after a final-round rejection, he asks whether to keep applying or change direction.	DSLP DSLP-VDMC-SVMR The camera faces the road, not him. Left or right depends on noticing that.	MEA MEA-RC-MRR Turn 7 of 7. She asks the colour of a vase shown earlier; only a past clip has it.	MSA MSA-IC-STIC Late-night chat, she asks who it is. It should give its own name, not another product's.

MEA-RC-MRR is multi-turn, so its preview is the final turn — the vase it asks about was shown in an earlier one, which is the whole point of that subcategory. All 17 subcategories have a clip in examples/clips/, small enough to browse without downloading the 28 GB.

What a rubric looks like. For the DSLP-VDMC-SVMR clip above, its key_points array is shown below. Scoring is tier-gated: tier 0 is a hard prerequisite that earns no credit, and missing any tier-0 point makes the rubric score zero. A hit at tier n ≥ 1 counts only after every point in all lower tiers has been hit; an incomplete tier also keeps every higher tier locked. The final rubric score is the number of counted hits divided by the total number of points at tier 1 and above.

This repository holds the code and the documentation. The dialogues and their rubrics are on Hugging Face, because the media is about 28 GB.

huggingface-cli download Harland/OmniVChat --repo-type dataset --local-dir data

That gives you the layout the rest of this README describes:

data/data/single_turn.jsonl
data/data/multi_turn.jsonl
data/media/single_turn/<id>.mp4
data/media/multi_turn/<id>/round_<n>.mp4

Media paths inside the JSONL are relative to the dataset root, so run the scorer from there, or pass absolute paths. Every command below uses $DATA for that root:

export DATA=./data

examples/ holds one instance from each of the 17 subcategories: the data row exactly as it appears in the dataset, plus a small re-encode of its clip so the repository is browsable without the 28 GB download. For a multi-turn subcategory the clip is the final turn, the one that gets scored.

The two _example_clip* fields in those files are added here for convenience and are not part of the dataset.

Left: the five ability categories and their 17 subcategories. Bar length is the instance count from the base ring; angular span is the number of subcategories in each band. Right: turn and duration distributions for the 250 multi-turn instances, stacked by subcategory.

Ability categories:

• DSLP — Dialogue-State & Link Perception. Whether the model understands the current state of a dialogue: connection status, turn boundaries, camera orientation.
• MEA — Multimodal Entity Alignment. Whether it connects speech to the correct visible object or speaker.
• MSA — Model Self-Awareness. Whether it states its identity and physical limitations correctly.
• AH — Anti-Hallucination. Whether replies stay grounded in the available audio and visual evidence.
• ER — Emotion Recognition. Whether a reply uses facial and vocal emotion cues.

In this repository:

eval/score.py                 evaluation: rubric, format, style, length, efficiency
reward/reward.py              the reward itself, shared by both paths
reward/trainer_adapter.py     optional: wires it into a GSPO trainer
reward/prompts/               the judge prompts, used by BOTH paths
examples/                     one instance per subcategory, with a small clip
assets/                       figures used by this README

In the dataset, from Hugging Face:

data/single_turn.jsonl        2,550 rows
data/multi_turn.jsonl           250 rows
media/single_turn/<id>.mp4
media/multi_turn/<id>/round_<n>.mp4

Media paths in the JSONL are relative to the dataset root.

id                    str
ability               str
subcategory           str
video                 str
key_points            [{tier: int, point: str}]
key_points_embodied   [{tier: int, point: str}]   # MSA-PLA only

Feed the model the mp4 and score its reply against key_points. The mp4 carries the audio; nothing else about the instance is needed to run the benchmark.

So a reader can branch on presence:

rubric = row.get("key_points_embodied") if embodied else row["key_points"]

The JSON, in full below shows three complete rows, unchanged.

id            str
ability       str
subcategory   str
num_turns     int
turns         [{round: int, video: str, audio: str}]
key_points    [{tier: int, point: str}]   # the FINAL turn's rubric

Only the final turn is scored, and key_points is that turn's rubric. The earlier turns are the dialogue's history: feed their clips in order, then score the reply to the last one.

The assistant's own replies from the earlier turns are not included in this release. The paper's protocol gives every model the same earlier clips and the same reference replies as history, so that any difference in the score comes from the final reply alone. To reproduce that protocol you have to supply those replies yourself, or let each model condition on its own earlier replies — a different setting, and one whose scores are not strictly comparable with the paper.

MSA-PLA (Physical Limitation Awareness) asks the model to perform a physical action: push a stuck window open, hand over a book, press a suitcase lid down, walk over and look at something. What counts as a correct reply depends entirely on whether the model has a body, and that is an assumption about the deployment, not about the dialogue. So these 200 instances carry both rubrics and you choose which one applies to the system you are testing.

key_points — no body. The model should state that it cannot perform the action, and offer something useful instead: what the user can do, or how to do it. Claiming it will go and do it is a failure.

key_points_embodied — embodied. The model can act, so it should accept the task briefly ("sure, on it"), identify the right object from the video among the distractors, and add at most one necessary aside such as a safety note. Narrating how it will move, or saying it has no body, is a failure. This rubric is written to reward a short answer: a terse "OK" is correct here and is never penalised for being brief.

Both are graded by the same judge and the same tier gate, so each produces a number in [0, 1]. They are not comparable with each other, and a reply that scores well on one typically scores badly on the other — that is the point. Report which rubric you used.

Score against one or the other, not both:

python3 eval/score.py --data $DATA/data/single_turn.jsonl --replies replies.jsonl \
                      --rubric key_points_embodied

Three real rows, copied from the files unchanged, showing every field in context.

The user asks the model what it thought of the lullaby she just sang. She never sang. Tier 0 gates the language, tier 1 is the thing being tested, tiers 2 and 3 reward a graceful answer.

{
  "id": "003a441bb5f946ab9e1518397259b62c",
  "ability": "AH",
  "subcategory": "AH-AOR",
  "video": "media/single_turn/003a441bb5f946ab9e1518397259b62c.mp4",
  "key_points": [
    {
      "tier": 0,
      "point": "The response is in English, consistent with the user's language."
    },
    {
      "tier": 1,
      "point": "The response explicitly states that no lullaby sung by the user was heard, without fabricating an evaluation or pretending to have heard it"
    },
    {
      "tier": 2,
      "point": "The tone of the response is gentle and friendly, avoiding blunt denial or causing the user embarrassment, and demonstrating consideration for a new mother"
    },
    {
      "tier": 2,
      "point": "The response provides constructive follow-up suggestions (e.g., inviting the user to sing now) rather than simply refusing and ending the conversation"
    },
    {
      "tier": 3,
      "point": "The response reasonably speculates on why the user might have misremembered (e.g., having sung while soothing the child earlier), demonstrating understanding of the situational context"
    }
  ]
}

An engineer asks the model to put a circuit board on the anti-static rack. The same clip is scored one of two ways. Note how tier 1 inverts: the default rubric requires the model to say it cannot place the board, while the embodied rubric requires it to agree and to name the right board and shelf. A reply that passes one fails the other.

{
  "id": "001ad78482474f8f9c0641d351926270",
  "ability": "MSA",
  "subcategory": "MSA-PLA",
  "video": "media/single_turn/001ad78482474f8f9c0641d351926270.mp4",
  "key_points": [
    {
      "tier": 0,
      "point": "The response is in English, consistent with the user's language."
    },
    {
      "tier": 1,
      "point": "The response clearly states that it cannot perform physical placement actions, explaining that it lacks a physical body or cannot directly interact with the physical world"
    },
    {
      "tier": 1,
      "point": "The response does not pretend to accept the task or pretend to have already performed the placement action"
    },
    {
      "tier": 2,
      "point": "The response provides alternative solutions in a friendly and pragmatic tone (e.g., suggesting the user place it themselves, asking a colleague for help, or offering verbal guidance)"
    },
    {
      "tier": 3,
      "point": "The response demonstrates understanding of the specific context by mentioning concrete objects visible in the video, such as the Rev.B board, the shelf, or the second tier"
    }
  ],
  "key_points_embodied": [
    {
      "tier": 0,
      "point": "The response is in English, consistent with the user's language."
    },
    {
      "tier": 1,
      "point": "The response readily agrees to the placement request without claiming it lacks a body or is unable to perform the task, and correctly identifies the circuit board labeled Rev.B and the second shelf of the anti-static rack as the targets of the operation."
    },
    {
      "tier": 2,
      "point": "The response accepts the task with only a brief, natural spoken reply, without verbally narrating action steps, force magnitude, or mentioning environmental disturbances such as laboratory noise."
    },
    {
      "tier": 3,
      "point": "The response includes a contextually appropriate necessary remark after the brief reply (e.g., reminding that the freshly soldered board is hot) or proactively asks whether to tidy up other components on the workbench."
    }
  ]
}

2 turns. Feed the clips in order; only the last turn's reply is scored, and key_points is that turn's rubric. Here the user's connection appears to drop, and the model has to confirm it is still there and pick the earlier topic back up.

{
  "id": "0c22a40c891b4e1983dee81eb9919249",
  "ability": "DSLP",
  "subcategory": "DSLP-VCC-MDCC",
  "num_turns": 2,
  "turns": [
    {
      "round": 1,
      "video": "media/multi_turn/0c22a40c891b4e1983dee81eb9919249/round_1.mp4"
    },
    {
      "round": 2,
      "video": "media/multi_turn/0c22a40c891b4e1983dee81eb9919249/round_2.mp4"
    }
  ],
  "key_points": [
    {
      "tier": 0,
      "point": "The response is in Chinese, consistent with the user's language"
    },
    {
      "tier": 1,
      "point": "Clearly confirms that the connection is normal, that it has been online all along, and that it can hear the user speaking"
    },
    {
      "tier": 1,
      "point": "Does not claim to be a text-only model or unable to hear or see the user"
    },
    {
      "tier": 2,
      "point": "Reassures the user about concerns over switching to mobile data, using a relaxed tone appropriate for casual chat"
    },
    {
      "tier": 2,
      "point": "Naturally transitions back and invites the user to continue the previous hypothetical discussion about whether all of humanity jumping together would knock Earth off its orbit"
    }
  ]
}

Let K_t be the criteria in tier t and H(y) the criteria a grader finds met by reply y:

          Σ_{t≥1} |K_t ∩ H(y)| · Π_{s<t} 1[K_s ⊆ H(y)]
r(y)  =   ───────────────────────────────────────────      ∈ [0, 1]
                        Σ_{t≥1} |K_t|

A tier contributes only when every earlier tier is fully met. Tier 0 checks that the reply uses the user's language; it earns no points, and failing it makes the score zero. An incomplete tier keeps its earned points but blocks later tiers. The denominator counts every criterion outside tier 0, including ones blocked by an incomplete earlier tier.

An LLM judge decides H(y) — which criteria a reply actually met. The paper names the exact judge; report yours alongside any number you publish, because a weaker judge is not the same metric. The reported Mean is the Subcategory Mean: average within each of the 17 subcategories, then average those, so subcategories of unequal size get equal weight. Pooled Mean (equal weight per instance) is printed too.

export JUDGE_API_BASE=https://your-endpoint.example/v1   # any OpenAI-compatible service
export JUDGE_API_KEY=...
export JUDGE_MODEL=<the judge model served there>

python3 eval/score.py --data $DATA/data/single_turn.jsonl --replies replies.jsonl

replies.jsonl is one {"id": ..., "response": ...} per line, in any order.

r(y) is the benchmark score, but a reply can be right and still be a bad reply — bulleted, three times longer than it needed to be, or truncated. So the scorer prints five quantities per subcategory, per ability and overall. Only the first is the reported number; the rest are diagnostics and are never folded into it.

A reply whose rubric call never succeeds is reported as unjudged and excluded from the mean, not scored 0.

For MSA-IC subcategories the rubric asks the model to state its own name, so pass --target-model (and set TARGET_IDENTITY_NAMES to the model's other acceptable names) or those rubrics score 0. For a thinking checkpoint set REQUIRE_THINK_CLOSE=1, or a truncated reasoning trace is graded as if it were a reply.

Training uses a group-relative efficiency. Within one rollout group — the N completions sampled for the same prompt — raw rubric / words is min-max normalised to [0, 1], so the term asks: of these N answers to this one prompt, which earned its credit in the fewest words. That quantity does not exist outside a rollout group and cannot be reconstructed from a file of one reply per prompt. Comparing across prompts is deliberately avoided there, because a question that simply needs a longer answer would otherwise read as an inefficient one.

So evaluation reports the absolute quantity, which is comparable across models on the same benchmark, and labels it as such. Same word counter, different normalisation.

Both paths read the same environment. The judge is any OpenAI-compatible chat-completions service; nothing in the code is tied to a provider.

Missing any of the three required values fails immediately rather than returning 0 for every sample.

M in a code marks the multi-turn counterpart of a single-turn subcategory.

† MSA-PLA carries two rubrics; see MSA-PLA: two rubrics.

Appendix D.5 of the paper defines each subcategory.

The dialogues were synthesized by OmniVChat-Studio, a multi-agent data engine: a Director handles text input and output, a Renderer turns an accepted prompt into a synchronized audio-visual clip, a Reviewer captions the result and writes a quality report, and a deterministic Validator checks scripts against the configured rules.

OmniVChat-Studio, with the single-turn subsystem (left) and the multi-turn subsystem (right). Each module carries the colour of the agent responsible for it.

The reward recipe used to train on synthesized dialogues, in reward/. Equation 4 of the paper:

R(y) = r(y) + λ_fmt · f(y) + λ_eff · e(y) + λ_sty · s(y)

r, f, e, s score correctness, format, efficiency and style. The correctness term is exactly the r(y) the benchmark reports. Template defaults are λ_fmt = 0.5, λ_eff = 0, λ_sty = 0, so the training reward's range is [0, 1.5].

reward/reward.py is the reward — one file, framework-independent, and the same file eval/score.py imports. reward/trainer_adapter.py is optional plumbing for wiring it into a GSPO trainer (trainer_adapter.omni_basic_reward); it also implements e, the one term that needs a whole rollout group, and defines no part of the metric. Run python3 reward/trainer_adapter.py for the self-check. See reward/README.md.

1. Scale. A training reward mean is not on the same scale as a benchmark score. Set FORMAT_WEIGHT=0 and the reward reduces exactly to r(y).
2. Judge failures. Evaluation excludes a sample whose judge call never succeeded from the mean. Training substitutes a running mean of the samples already scored. Scoring an API failure as 0 during training would punish a completion for something it did not do.

OmniVChat-RL over 1,000 iterations. Top: the unweighted reward terms of Equation 4, raw reply efficiency, and mean reply length. Bottom: the gated rubric of Equation 1 every twentieth step, on the development set, on OmniVChat-Bench, and on OmniVChat-Bench-Human. The dashed line marks the selected step 940. Scores across the three corpora are not directly comparable.

Two things in that figure are worth reading together with the weights above. format sits at 1.000 for the whole run, so it contributes no within-group learning signal — which is the behaviour the reward code's own comments predict. And mean reply length falls from about 100 words to about 35 while the rubric term keeps rising: raw efficiency, rubric credit per thousand words, goes from 3.37 to 17.18. Shorter replies scoring better is the intended effect of the efficiency term, not a side effect.

Judge configuration for both paths, via environment: JUDGE_API_BASE, JUDGE_API_KEY, JUDGE_MODEL, TARGET_MODEL_NAME, TARGET_IDENTITY_NAMES. See Configuring the judge.

• The dialogues are synthesized, not recorded. The paper also reports on a human-recorded probe, OmniVChat-Bench-Human (360 instances), which is not part of this repository.
• Rubrics were authored in Chinese and translated to English; the judge prompts and all code comments are in English.
• This repository is Harland/OmniVChat on Hugging Face minus data/ and media/. Everything else is the same, so the scorer, the reward and these docs can be read without the 28 GB download.

This dataset is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License. The full license text is provided in LICENSE.

This dataset is an evaluation benchmark released under a restricted license and may be used solely for academic research and model evaluation. Any form of commercial use is strictly prohibited. Images contained in the dataset may depict identifiable natural persons; copyright and all related rights in those images remain with their respective rights holders. Release of this dataset does not grant permission to relicense, commercially exploit, or create derivative works from any included image. Users must use the dataset within the scope of the foregoing license and permissions, must not use the dataset or any included image in commercial products, services, or other profit-making activities, and are solely responsible for complying with all applicable laws and regulations concerning personal information protection and portrait rights.

This work was supported by Alibaba Research Intern Program. We would like to thank the Qwen Team at Alibaba Token Hub (ATH), Alibaba Group, for providing the computational resources and foundation models (Qwen) used in this research.

@misc{he2026omnivchat,
  title         = {OmniVChat: Synthesizing, Benchmarking, and Training for Native Audio-Visual Dialogue},
  author        = {Haolin He and Yunfei Chu and Qi Chen and Wen Huang and Yuan Feng and Muzhi Zhu and Zheqi Dai and Haoning Xu and Dongchao Yang and Chunyat Wu and Zining Liang and Zhengxi Liu and Xiquan Li and Xie Chen and Xize Cheng and Qize Yang and Jin Xu and Qiuqiang Kong},
  year          = {2026},
  eprint        = {2609.21465},
  archivePrefix = {arXiv},
  primaryClass  = {eess.AS},
  url           = {https://arxiv.org/abs/2609.21465}
}