AIAny
Icon for item

AVA-Encoder: Towards Agent-Native Video Representation Learning

Encodes videos into a Film Knowledge Graph and reconstructs them to learn agent-native, editable video representations for agentic reasoning and manipulation. Uses agentic auto-encoding with dual-loop textual-gradient optimization, reports large reconstruction gains, and releases a benchmark and dataset.

Introduction

Why this matters

High-quality filmmaking contains dense, cross-shot structure (characters, events, camera language, audio design) that current video agents cannot reliably capture or edit. The core insight of AVA-Encoder is to map films into a text-centered Film Knowledge Graph (Story–Event–Shot hierarchy with typed state nodes and linked assets) so agents can read, query, and perform topology-aware edits without losing multimodal fidelity. Treating reconstruction fidelity as a direct measure of representational faithfulness converts generation residuals into optimization signals that both refine an input-specific Film KG and pseudo-train a shared Agentic Video Encoder policy.

Key Findings
  • Film KG + agentic auto-encoding improves reconstruction fidelity: AVA-Encoder achieves an Overall reconstruction score of 49.0%, a 20.7-percentage-point absolute improvement over the strongest external baseline (28.3%).

    • So what: structured, editable representations materially close the gap between agentic reasoning and cinematic fidelity.
  • Efficient policy pseudo-training: the pseudo-trained Agentic Video Encoder policy reaches 45.8% (vs 44.4% for a human-tuned policy) while using 74.3% fewer shot-level and 70.1% fewer keyframe-level system-prompt tokens.

    • So what: the learned encoding rules transfer and reduce expensive prompt/system-token usage for agent policies.
  • Robust benchmark and human agreement: the paper releases a reconstruction-based benchmark covering narrative, visual, temporal, and multimodal consistency; its automatic metrics agree with human judgments on 97.3% (710/730) of blinded triples.

    • So what: enables reproducible, quantitative evaluation of agent-native video representations.
  • Graph-based editing and asset consistency: the Film KG separates structured textual nodes from a linked asset layer and preserves typed edges, enabling topology-aware edits (character replacement, visual-treatment propagation) that remain consistent across affected shots.

    • So what: supports controllable film remixing and subgraph-level changes without manual cross-shot patching.
Who it's for and trade-offs

Great fit if you work on multimodal agent systems, controllable video generation/editing, or representation learning for long-horizon creative media. AVA-Encoder is specifically aimed at cinematic-quality films where cross-shot narrative and production dependencies matter.

Look elsewhere if you need low-latency, real-time video capture on-device, or solely pixel-level enhancement: the approach emphasizes structured, text-centered representations and reconstruction-faithfulness over raw per-frame GAN-style photorealism. Also note that reconstruction scores depend on the fixed decoder used in experiments; representation quality and downstream generation quality will vary with decoder and computational budget.

Method snapshot
  • Representation: Film Knowledge Graph (nodes: story, event, shot, state, keyframe textual descriptions; assets: linked generated images/audio/video; typed edges preserve hierarchical, temporal, cross-shot relations).
  • Training loop: Agentic Video Encoder maps input film -> Film KG; a fixed decoder reconstructs video; reconstruction residuals drive gated dual-loop textual-gradient refinement (data-dependent KG refinement + data-agnostic encoding-policy pseudo-training).
  • Artifacts: authors release the AVA-Encoder framework, a reconstruction benchmark with fine-grained evaluation dimensions, and a dataset of high-quality Film KG representations and an editing framework.

Bottom line: AVA-Encoder reframes cinematic video understanding as a structured, agent-operable representation problem and demonstrates measurable gains in reconstruction faithfulness and editability, at the cost of dependency on decoder quality and focus on film-like inputs.

Information

  • Websitearxiv.org
  • OrganizationsAffiliation: Qwen Business Unit of Alibaba, Affiliation: ShanghaiTech University, Affiliation: The Hong Kong University of Science and Technology, Affiliation: Institute of Computing Technology, Affiliation: Southeast University
  • AuthorsChuyue Li, Jinpeng Yu, Haozhe Wang, Tian Xueyun, Zhijing Zhang, Bingnan Li, Shuqi Gu, Kan Ren, Jiaming Liu, Ruihua Huang
  • Published date2026/08/12

More Items

Fine-tunes long-horizon LLM agents with evolution strategies so full-model updates run at inference-level GPU memory. Emphasizes trajectory-level credit via black-box rewards, online prompt–parameter co-evolution, and a cosine decay for perturbation scale to balance exploration and adaptation; suited for limited-GPU settings.

Turns embodied navigation into 2D visual prompting where a vision-language model selects image pixels that are projected to 3D actions; adds selective chain-of-thought, compressed anchor-trajectory memory, and a two-level alignment objective to improve sample and runtime efficiency.

Evaluates whether AI systems can independently carry out project-level scientific research by progressively removing human methodological guidance across 60 tasks in 11 domains. Built with expert review, sandbox execution, and multi-agent–model scoring to measure innovation and autonomous experimental execution.