Skip to main content
Back

Xiaomi's PRPF Doubles Mobile AI Success Rate While Cutting Compute by 69%

Researchers from Xiaomi's HyperAI team and partner universities propose a two-stage framework that cuts mobile agent false triggers nearly in half while slashing inference compute by 69%

Paper, Full Access
  • Two-stage framework: PRPF separates "when to intervene" (lightweight MPP gate) from "how to help" (Qwen3.5-9B reasoner), cutting wasted inference.
  • Big efficiency gains: 69.3% compute reduction and 60.1% latency reduction vs. the ProactiveMobile (7B) single-stage baseline.
  • False trigger rate halved: FTR drops from 13.76% to 7.21% overall; to just 1.75% in text-only mode — critical for real-world usability.
  • Beats frontier models: PRPF outperforms GPT-5.5, o3, Gemini-3.1-Pro, and Claude-Opus-4.7 on the ProactiveMobile benchmark.
  • Plug-and-play perceptor: MPP (~0.1B params) can be added to existing VLMs without retraining them, boosting their SR and reducing false triggers.
  • Multimodal gap persists: Absolute SR on GUI screenshot tasks stays at 17.19%; even the best models barely exceed 18%, pointing to a fundamental vision-grounding bottleneck.

A new research paper from authors affiliated with Xiaomi's HyperAI team, Zhongnan University of Economics and Law, Jilin University, and the Chinese University of Hong Kong proposes a rethink of how mobile AI agents decide when to interrupt you. The work, titled "Perceive Before Reasoning," introduces the Pre-Reasoning Perception Framework (PRPF), a two-stage architecture that separates the question of when to intervene from how to help , and the numbers are striking.

The problem with always-on reasoning

Today's mobile agents are mostly reactive: you tell them what to do, they do it. The next frontier is proactive assistance, where an agent watches what you're doing and jumps in before you ask. MLLMs have made significant progress in mobile agent development, yet their capabilities are predominantly confined to a reactive paradigm. The emerging paradigm of proactive intelligence, where agents autonomously anticipate needs and initiate actions, represents the next frontier for mobile agents.

The challenge is that proactive agents face a two-part problem every time they observe your screen: first, should they intervene at all? Second, if yes, what should they recommend? Existing systems often implement these two decisions within a unified MLLM-based pipeline, leading to goal misalignment between conservative intervention filtering and comprehensive assistance generation, as well as redundant inference when the agent should remain silent.

This creates two concrete failure modes:

  • Goal misalignment: The "when" decision needs to be conservative and highly discriminative to avoid annoying the user. The "how" decision needs to be broad and generative. Squeezing both into one model means neither is done optimally.
  • Wasted compute: Even when the right answer is silence, the full reasoning pipeline runs anyway. On a benchmark spanning 14 mobile scenarios and 63 API functions, this becomes expensive fast.

A gatekeeper before the thinker

PRPF's core idea is architectural separation. Instead of one large vision-language model (VLM) doing everything, PRPF inserts a small, fast "perceptor" module in front of the big reasoner. A lightweight perceptual front-end first performs intervention gating and candidate function compression, assigning the conservative "when" judgment to an efficient discriminative module. The VLM-based reasoner is then activated only when intervention is warranted, allowing it to focus on the generative "how" stage.