AIAny
Icon for item

Omni Interaction Agent Technical Report

Unifies continuous streaming perception, full‑duplex real‑time interaction, and agentic tool use across video, audio, and text. Proposes a Cerebellum–Brain split—streaming Thinker‑Talker cerebellum for low‑latency interaction plus a Brain for complex reasoning—and is released with models, code, and data.

Introduction

Gander attacks a practical gap in multimodal agents: most systems either process offline video or operate turn‑by‑turn, which breaks natural interruptions, proactive feedback, and continuous monitoring. The core insight is to separate low‑latency conversational control from heavyweight reasoning so a single system can both respond immediately to streaming inputs and execute complex agentic workflows.

Key Findings
  • Cerebellum–Brain architecture: a streaming Thinker‑Talker "Cerebellum" handles chunk‑level low‑latency perception and dialog act decisions, while a separate "Brain" handles higher‑latency, complex reasoning and tool orchestration — this decoupling preserves responsiveness without discarding long‑horizon reasoning.
  • Native full‑duplex streaming interface: flattening inputs/outputs into ordered token chunks lets the model detect multimodal triggers, accept user interruptions, and proactively emit intermediate feedback; in practice this reduces response latency compared with turn‑based baselines.
  • Evaluation across four axes: conversational quality, omni‑modal understanding, interactive robustness (interruptions, backchannel, noise), and agentic intelligence (tool calls, multi‑step tasks). Human evaluations show competitive spoken dialogue fluency while exposing gaps in streaming mathematical reasoning and long‑step tool plans.
  • Release and reproducibility: models, datasets, and code are provided to encourage follow‑up work on streaming benchmarks and agent orchestration.
Who it's for and tradeoffs

Great fit if you need a research‑grade reference for continuous multimodal interaction—especially scenarios with live video/audio, user interruptions, or agents that must interleave conversation with tool use. Look elsewhere if your primary goal is lightweight deployment on constrained devices (the architecture and training are resource‑intensive) or purely offline batch video understanding where low‑latency turn taking is irrelevant. Expect remaining limitations in robust long‑horizon reasoning under full‑duplex constraints and in extreme multi‑party audio environments.

Information

  • Websitearxiv.org
  • OrganizationsHunyuan Speech Team, Tencent, Zhejiang University, Shanghai Jiao Tong University, The Chinese University of Hong Kong, Nanyang Technological University
  • AuthorsOrantqing, Shengpeng Ji, Junlong Tong, Jialong Zuo, Dongjie Fu, Di Cao, Yangzhuo Li, Shangda Wu, Franz, Evan
  • Published date2026/09/08

Categories

More Items

Explores a practical mechanism for recursive self-improvement by post-training LLMs: uses a routing harness to record agent executions and convert traces into curriculum-guided supervised fine-tuning and on-policy distillation data, closing an evaluation-selection-update loop and improving benchmark performance.

Wraps vision–language–action policies into executable skills that are runtime-validated, executed as bounded low-level action chunks, outcome-verified, and logged as structured trajectories. A fixed skill interface enables swapping or adapting low-level VLA policies and provides component-level supervision for training and optional online adaptation.

Studies when and how an assistant should ask clarification questions before converting natural-language problem descriptions into optimization models. Introduces OR-Clarify, a benchmark for pre-formulation clarification, and InterOPT, a two-stage framework that diagnoses formulation-critical gaps and guides targeted questioning and stopping decisions.