Research map / Vision-language and multimodal / Multimodal LLMs

Meow-Omni 1: A Multimodal Large Language Model for Feline Ethology

NeurIPS 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Benchmarks and analysis of MLLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces Meow-Omni 1, a quad-modal (text, video, audio, biological time-series) MLLM built by grafting a time-series encoder from Intern-S1-Pro onto MiniCPM-o, evaluated on a new MeowBench cat-intent MCQ benchmark.

Problem

Decoding feline intent is hard because identical signals (e.g., purring) map to different internal states depending on physiological context, and existing MLLMs cannot natively process high-frequency biological time-series alongside audio and video.

Approach

Integrate a time-series encoder and projector into MiniCPM-o so physiological accelerometer streams are interleaved with video and audio tokens in one LLM context, trained with projector alignment then LLM-only specialization on a curated feline dataset.

Key results

• Meow-Omni 1 reaches 71.16% top-1 accuracy on MeowBench (527 MCQs), vs 66.89% for Qwen3.5-Omni-Plus with all three modalities and 61.95% for Qwen3.5-122B video-only.
• Predictive entropy averages 1.28 bits on congruent samples vs 3.15 bits on 50 synthetic conflict samples (preliminary).
• Ablation: unimodal vision 69.97%, V+TS 70.82%, full quad 71.16%; TS-only 55.63%.

Surprising findings

• Audio-only and TS+audio configurations perform well below vision-only, and the full model adds only 1.2 points over vision-only on the authors' own ablation.
• Qwen3.5-Omni-Plus with biosignals as text summary underperforms the native model by 4.3 points, though the baseline receives only a textual summary of the TS data.

Limitations

• MeowBench conflict and quad-modal samples are synthesized by intent-matching unimodal data from different sessions, not naturally synchronized.
• Next-behaviour prediction assumes intent manifests as an immediate physical action; protracted intents are not captured.
• Evaluation is MCQ-based with only 527 samples and preliminary conflict-entropy experiments of 50+50 instances.

Experimental setup

Datasets: Meow-10K MeowBench AudioSet Freesound cat_class (taozi555) two accelerometer cat datasets Bain video dataset

Benchmarks: MeowBench

Metrics: top-1 accuracy predictive entropy

Baselines: Qwen3.5-122B-A10B Qwen3.5-Omni-Plus AVES Ntalampiras audio framework IMU-based 1D-CNN/LSTM baseline

Builds on: MiniCPM-o Intern-S1-Pro Qwen3-Omni

Closest papers by meaning

• Animal-Bench: Benchmarking Multimodal Video Models for Animal-centric Video Understanding (NeurIPS 2024 · similarity 0.87)
• UniAP: Towards Universal Animal Perception in Vision via Few-Shot Learning (AAAI 2024 · similarity 0.84)
• NatureLM-audio: an Audio-Language Foundation Model for Bioacoustics (ICLR 2025 · similarity 0.84)
• OmniBench: Towards The Future of Universal Omni-Language Models (NeurIPS 2025 · similarity 0.84)
• AniMo: Species-Aware Model for Text-Driven Animal Motion Generation (CVPR 2025 · similarity 0.83)
• Neural Encoding and Decoding at Scale (ICML 2025 · similarity 0.83)