MeowBench: Feline Intent Benchmark

• MeowBench is a quad-modal benchmark that defines feline intent as a latent, causally grounded state rather than purely observable behavior.
• It synthesizes video, audio, physiological time-series, and text data through expert-verified pairing, emphasizing the resolution of semantic aliasing.
• Evaluation is based on a 30-class taxonomy using multiple-choice questions measured by Top-1 Accuracy, showcasing the advantage of native time-series grounding.

Searching arXiv for the specified paper and any benchmark-specific context. arXiv Search Query: (Hu et al., 9 May 2026) MeowBench is a held-out, expert-verified quad-modal benchmark for decoding feline intention from combined video, audio, physiological time-series, and text. It was introduced in connection with "Meow-Omni 1: A Multimodal LLM for Feline Ethology" to evaluate inter-species intent reasoning rather than surface-level action recognition, with particular emphasis on resolving semantic aliasing, in which identical external signals can correspond to different internal states depending on physiological context. The benchmark presents intent inference as a multiple-choice classification problem over a unified 30-class taxonomy and is used strictly for evaluation; in its reported form, it contains 527 expert-verified samples and uses Top-1 Accuracy as the primary metric (Hu et al., 9 May 2026).

1. Conceptualization of feline intent

MeowBench is built around a specific problem formulation in which intent is treated as a causal latent state rather than a directly observable behavior label. In the underlying formulation, intent is defined using Pearl’s do-operator to decouple environment confounders:

$$\mathcal{I} = \arg \max_{i \in \mathbb{I}} P(a^*_{t+1} \mid \mathcal{H}_t, do(\mathcal{E}_{free}), i)$$

with $\mathcal{H}_t = \{V_{1:t}, A_{1:t}, B_{1:t}\}$ and $\mathbb{I}$ the discrete set of intent categories. The associated multimodal likelihood objective is

$$\mathcal{L}(\theta) = \sum_{(\mathcal{H}_t, \mathcal{I}) \in \mathcal{D}} \log P(\mathcal{I} \mid Encoder(\mathcal{H}_t); \theta).$$

Within this framework, MeowBench operationalizes intent inference by forcing models to reconcile audio-visual cues with biometrics rather than relying on superficial correlates. The benchmark was introduced in response to what the paper describes as a benchmarks and evaluation vacuum for native cross-modal reasoning among biometrics, vision, and audio in animal-intent inference. The paper also situates semantic aliasing in relation to prior ethological and neuroscience literature, including work by Green and Angelaki and by Merola and Mills (Hu et al., 9 May 2026).

This framing distinguishes MeowBench from ordinary behavior-recognition datasets. The task is not merely to identify an externally visible act; it is to infer a latent intention under conditions where externally similar signals may encode different internal states. A plausible implication is that performance on MeowBench is intended to reflect cross-modal grounding capacity more directly than standard action-classification accuracy.

2. Modal structure and synthesis procedure

MeowBench is quad-modal. Its inputs comprise video, audio, physiological time-series, and text. Video is derived from action-centric clips pre-processed through a VLM pipeline and appears in synthesized quad-modal samples. Audio consists of cat vocalization segments obtained from AudioSet-derived synchronized audio-video clips and standalone audio sources. Physiological time-series are accelerometer-based biosignals derived from triaxial accelerometry. Text is used to wrap all inputs in natural-language prompts, and MeowBench instances are presented as Multiple Choice Questions with four options (Hu et al., 9 May 2026).

The physiological stream undergoes explicit preprocessing. Original source sampling rates, such as 30 Hz or 60 Hz, are aggregated to second-level signals at 1 Hz by in-window averaging in order to reduce noise and harmonize sampling rates. The resulting signals are segmented into fixed-length windows of 5, 7, 10, or 15 seconds. Windows are constrained not to cross individuals or discontinuities, and “other” or ambiguous labels and transient intermediate movements are discarded. Units are not specified in the paper, although accelerometer magnitudes are implied.

Video preprocessing is likewise specified in temporal rather than spatial detail. Action onset detection and temporal localization use coarse sampling at 1.5 s intervals with a gap rescan at approximately 0.5 s, while dense localization uses 0.20 s frame intervals around a candidate anchor. Each verified clip uses an observation window $T_{obs} = 6.0$ s, asymmetrically centered around onset as $$[t_{anchor} - 0.85 T_{obs},\ t_{anchor} + 0.15 T_{obs}]$$. Resolution and frame rate are not specified. Audio sample rate and bit depth are also not specified, but audio verification is performed with an Audio Spectrogram Transformer classifier using cat-meow-purr-related labels and thresholds of 0.10 to retain, 0.03 to discard, and 0.03–0.10 to denoise and recheck.

A defining feature of MeowBench is that its quad-modal examples are synthesized by intent-matching rather than timestamp alignment. The paper states that no naturally synchronized quad-modal recordings exist. For a given intent, a synchronized audio-video clip from the same cat and session is paired with a time-series sample from a different session but annotated with the identical intent. Synchronization is therefore semantic rather than temporal. Eight professional feline ethologists review all combinations to ensure physical plausibility of the joint state (Hu et al., 9 May 2026).

3. Taxonomy, annotation, and benchmark composition

MeowBench uses a unified 30-class intention set. The paper consistently refers to this as a 30-class taxonomy and provides representative classes including feed, groom, rest, run, shake, trot, walk, maintenance_littering.* categories such as defecating, digging, urinating, and none, inactive_* categories such as lying, sitting, and standing substates, active_* categories such as walking, trotting, and jumping., maintenance_ categories such as scratching and shake.*, and other_social.allogrooming (Hu et al., 9 May 2026).

Each benchmark item is an MCQ with one correct intent and three distractors randomly sampled from the 30-class set. The benchmark’s primary supported task is intent recognition, and the format is designed to require cross-modal reasoning over video, audio, time-series, and the textual prompt. The paper also studies uncertainty under modality conflict using subsets derived from MeowBench, but MeowBench itself is defined as a classification suite.

The benchmark composition is tightly curated. A total of 645 initial intent-matched quad-modal candidates were reviewed by eight Professional Feline Ethologists organized into three groups of 3, 3, and 2 reviewers, with within-group consensus. This process yielded 527 high-fidelity samples. The guidelines emphasize physical plausibility of the cross-modal pairing for each intent. Inter-rater agreement metrics such as Cohen’s $\kappa$ are not reported.

MeowBench is held out and does not define train, validation, or test splits for itself. It is used strictly for evaluation. This is an important design distinction relative to Meow-10K, the 10,831-sample training set for Meow-Omni 1, which contains varying modality combinations and natural-language query-response pairs. The benchmark is therefore not a multitask corpus in the usual sense; it is an evaluation suite built to probe whether a model can perform intent-matched cross-modal synthesis under controlled conditions.

4. Evaluation protocol and formal measurement

The primary metric reported for MeowBench is Top-1 Accuracy on the 527 MCQs. Macro-F1 is not reported. The benchmark’s scoring protocol is deterministic: models are prompted with fixed MCQ formatting, must output a single-letter choice, and a regex extracts the first capital letter after the answer prefix to compute Top-1 Accuracy. Evaluation uses greedy decoding with temperature $T = 0$ and top-$p$ disabled (Hu et al., 9 May 2026).

The benchmark also supports an uncertainty analysis under modality conflict. In that setting, predictive Shannon entropy is estimated under temperature sampling with $N = 10$ samples at $\mathcal{H}_t = \{V_{1:t}, A_{1:t}, B_{1:t}\}$0:

$\mathcal{H}_t = \{V_{1:t}, A_{1:t}, B_{1:t}\}$1

where $\mathcal{H}_t = \{V_{1:t}, A_{1:t}, B_{1:t}\}$2 is the empirical frequency over sampled predictions. Although this analysis is reported on subsets derived from MeowBench rather than on the benchmark as a separate task definition, it is central to the paper’s claim that quad-modal reasoning should register contradiction rather than collapse to a single dominant modality.

The data-handling protocol preserves the benchmark’s semantic alignment principle. Each MeowBench sample bundles a synchronized audio-video clip from one session and a time-series window with the same intent from a distinct session, and this combination is vetted by experts. In the broader model framework, missing modalities are not padded with placeholders; the input sequence simply omits absent streams. Preprocessing references for modality construction are distributed across the appendices: 5, 7, 10, and 15 second time-series windows at 1 Hz after aggregation; video localization at coarse 1.5 s, gap rescan at approximately 0.5 s, dense localization at 0.20 s, with a 6.0 s window around onset; and AST-based audio verification with thresholds 0.10 keep and 0.03 drop, with marginal clips denoised and rechecked.

5. Relationship to Meow-Omni 1 and empirical results

MeowBench is tightly coupled to the architectural claims of Meow-Omni 1. The model integrates specialized scientific encoders adapted from Intern-S1-Pro for time-series, introduces time-series control tokens <|ts_start|>, <|ts_unit|>, and <|ts_end|>, and uses a linear projector to map time-series embeddings into the LLM hidden space. Training is organized in two stages: Stage 1 aligns the time-series projector using 383,853 time-series samples while freezing the LLM and time-series encoder and updating only the projector; Stage 2 fine-tunes on Meow-10K while freezing encoders and projector and updating only the LLM backbone. This design is presented as enabling native co-embedding of time-series with audio-visual tokens, which the benchmark requires for quad-modal reasoning (Hu et al., 9 May 2026).

The main reported MeowBench results are as follows:

Model or setting	Top-1 Accuracy
Audio SOTA (Ntalampiras et al., 2019)	36.86%
Vision SOTA (Qwen3.5‑122B‑A10B)	61.95%
TS SOTA (Chen et al., 2025, IMU-based)	48.98%
Qwen3.5‑Omni‑Plus (V+A)	65.36%
Qwen3.5‑Omni‑Plus (V+TS)	66.21%
Qwen3.5‑Omni‑Plus (TS+A)	42.15%
Qwen3.5‑Omni‑Plus (V+A+TS; TS as text summary)	66.89%
Meow‑Omni 1 (V+A+TS)	71.16%

The paper describes Meow-Omni 1’s 71.16% Top-1 score as state of the art. The ablation study further reports Meow-Omni 1 unimodal performance of 51.88% for audio, 69.97% for vision, and 55.63% for time-series; bimodal performance of 68.43% for V+A, 70.82% for V+TS, and 54.95% for TS+A; and 71.16% for the full V+A+TS configuration. The stated takeaway is that time-series consistently improves over vision-only by resolving aliasing, and that full integration yields the highest accuracy.

These numbers place MeowBench in a distinctive evaluative role. The benchmark is not merely used to compare generic multimodal systems; it is used to argue that native time-series grounding is decisive when identical or near-identical observable signals map to different latent intents.

6. Uncertainty behavior, limitations, and broader significance

The uncertainty analysis reported on MeowBench-derived subsets compares 50 congruent instances, in which video, audio, and time-series agree, with 50 adversarial instances, in which video and audio indicate one intent and time-series indicates another. The average predictive entropy is 1.28 bits for the congruent group and 3.15 bits for the conflict group, with $\mathcal{H}_t = \{V_{1:t}, A_{1:t}, B_{1:t}\}$3 stochastic passes at $\mathcal{H}_t = \{V_{1:t}, A_{1:t}, B_{1:t}\}$4. The paper interprets this as evidence that the model expresses higher uncertainty under cross-modal contradiction rather than defaulting to vision dominance (Hu et al., 9 May 2026).

Several limitations are explicit. Because native quad-modal data are unavailable, MeowBench relies on intent-matched synthetic composition. Although expert verification is used, the paper states that this may not capture all cross-modal nuances and that further validation on natively co-recorded data is needed. The next-behaviour prediction strategy may miss long-horizon intents without imminent action transitions. Potential domain, breed, age, and environment skew are not quantified, and these may affect generalization. Label noise may also enter through original source labels and automated mappings, although the authors mitigate this by discarding ambiguous labels and using expert consensus verification.

The benchmark is also situated within an ethics and deployment framework. The paper emphasizes uncertainty-aware decision making and human-in-the-loop usage for veterinary contexts. It states that the release includes a complete open-source pipeline with model weights, training framework, and the Meow-10K dataset, and that planned licensing requires attribution and restricts unlicensed commercial use, although specific repository URLs are not given.

In the benchmark landscape described by the paper, MeowBench’s significance lies in making physiologically grounded intent inference a concrete evaluation problem. It directly targets semantic aliasing by requiring reconciliation of video and audio with accelerometer time-series, and it provides a reproducible template for assessing whether a multimodal LLM can reason over semantically aligned but non-temporally synchronized biological and behavioral signals. A plausible implication is that MeowBench functions not only as a benchmark for feline ethology, but also as a prototype for cross-modal latent-state evaluation in computational ethology more broadly.