We made Pask easy to try — just download the .exe, .dmg, or .apk for Windows, macOS, or Android. You’ll get free live subtitles in 41 languages, real-time summaries, and memory dashboard, and we cover all the API costs(Google ASR & Gemini-3-flash).
If you like it, please give us a Star~~~ memeda ⭐
Q: Proactivity is a core expectation for AGI. What should a truly proactive agent look like?
Academic work often studies agents in clean task settings: goals are explicit, inputs are controlled, and evaluation is straightforward. Real-world proactivity is much harder. People speak over each other. Context shifts fast. Scenarios range from learning, to sales, to legal consultation. More importantly, intent itself is messy, hidden, and constantly changing. If we want proactive AI to be more than a toy, it must have robust, real-time, deep intent understanding, together with a complex multi-process agentic execution loop. And over time, these interactions must be distilled into personal memory, so the system can become a true soul knower.
We model this capability through three core challenges:
- Precision → we train IntentFlow, a dedicated model for real-time latent intent classification
- Super Low Latency → we design IntentFlow with a streaming architecture for per-turn online inference
- Long-term Modeling → we propose a hierarchical memory module for self-evolving user modeling across sessions
This leads to a key insight: proactive AI cannot be achieved by stronger agents alone; it requires innovation at both the model layer and the agent layer.
In this paper, we instantiate this view in Pask.
We introduce DD-MM-PAS, a general paradigm of Demand Detection, Memory Modeling, and Proactive Agent System, for building low-latency, online proactive AI. It integrates model-level innovations in intent understanding and user modeling with agent-level innovations in proactive decision-making and action, forming a closed loop for real-world proactive systems.
Pask in Learning.
git clone https://github.com/xzf-thu/Pask.git
cd Pask
conda create -n pask python=3.10
conda activate pask
pip install -r requirements.txt# Run a model
python -m eval.run --models gpt-5-mini --level all
# Run on local vLLM
VLLM_BASE_URL=http://localhost:9000/v1 python -m eval.run --models qwen3-30b-a3b --level all
# Score with LLM-as-judge
python -m eval.score --models gpt-5-mini
# Summarize results
python -m eval.report
# Generate LaTeX tables & figures
python -m latex.latex_fill
python -m latex.plot
Table 1: Main benchmark results. IntentFlow achieves the best balanced accuracy of 84.2.
Table 2: Breakdown by demand type (Work, Learning, Daily).
Table 3: Multi-turn accuracy degradation.
Table 4: Latency. IntentFlow achieves 1.3–1.5 s per turn.
Proactivity is a core expectation for AGI. Prior work remains largely confined to laboratory settings, leaving a clear gap with real-world proactive agents: depth, complexity, ambiguity, precision, and real-time constraints. We study this setting, where useful intervention requires inferring latent needs from ongoing context and grounding actions in evolving user memory under latency and long-horizon constraints.
We propose DD-MM-PAS (Demand Detection, Memory Modeling, Proactive Agent System) as a general paradigm for streaming proactive AI. We instantiate this in Pask, with the streaming IntentFlow model for demand detection, a hierarchical memory system (workspace, user, global) for long-term modeling, and the PAS infrastructure that closes the loop from detection to action.
DD-MM-PAS decomposes proactive intelligence into three coupled functions: detecting what a user needs (DD), remembering who the user is over time (MM), and executing useful assistance (PAS). Each component can be studied and improved independently, while the full system operates as a closed loop.
IntentFlow reads a live conversation continuously and outputs one of three decisions at each turn: stay silent, respond immediately, or query memory before responding. The model is trained to make this judgment under streaming constraints, with outputs calibrated to minimize both false positives (unnecessary interruptions) and false negatives (missed needs).
The memory system operates at three levels:
- User Profile — compact summary injected into every inference call; always-on, near-zero latency
- Working Memory — session-level state tracking current task context
- Long-term Store — retrieved via search when session context is insufficient
The system updates continuously across sessions, allowing Pask to personalize responses based on past interactions without manual configuration.
PAS connects frontend devices (glasses, phone, desktop) through a server layer to the full model and tool suite — web search, code execution, vision, and speech recognition. Detected intents are routed to the appropriate tool and returned to the user within the latency budget.
@article{xie2025pask,
title={Pask: Toward Intent-Aware Proactive Agents with Long-Term Memory},
author={Xie, Zhifei and Hu, Zongzheng and Ye, Fangda and Zhang, Xin and Chai, Haobo and Liu, Zihang and Wu, Pengcheng and Zhang, Guibin and Liao, Yue and Hu, Xiaobin and Ye, Deheng and Miao, Chunyan and Yan, Shuicheng},
journal={arXiv preprint arXiv:2604.08000},
year={2025}
}