💬 NLP

PASK: Toward Intent-Aware Proactive Agents with Long-Term Memory

This paper introduces PASK, a proactive agent system featuring a DD-MM-PAS paradigm and a new LatentNeeds-Bench benchmark, designed to infer latent user needs and execute timely interventions in complex, real-world environments by leveraging streaming intent detection and hybrid long-term memory.

Original authors: Zhifei Xie, Zongzheng Hu, Fangda Ye, Xin Zhang, Haobo Chai, Zihang Liu, Pengcheng Wu, Guibin Zhang, Yue Liao, Xiaobin Hu, Deheng Ye, Chunyan Miao, Shuicheng Yan

Published 2026-04-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zhifei Xie, Zongzheng Hu, Fangda Ye, Xin Zhang, Haobo Chai, Zihang Liu, Pengcheng Wu, Guibin Zhang, Yue Liao, Xiaobin Hu, Deheng Ye, Chunyan Miao, Shuicheng Yan

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a personal assistant who is currently a bit like a very polite but slightly deaf waiter.

The Old Way (Passive AI):
Right now, most AI works like that waiter. You have to raise your hand, snap your fingers, and say, "Hey, I need a menu!" or "Can you explain this wine?" The waiter only moves when you ask. If you're staring at a menu, confused, but don't ask for help, the waiter just stands there waiting. They don't know you're stuck until you speak up.

The New Way (Pask - The Proactive Agent):
The paper introduces Pask, a new kind of AI that acts more like a mind-reading sous-chef in a busy kitchen.

You don't have to ask for help.
The chef sees you staring at a complicated recipe, notices you're sweating, and before you even panic, they slide a simplified version of the recipe onto your counter.
They see you're running out of time and pre-chop the vegetables for you.
They remember that you hate cilantro from three years ago, so they never suggest it, even if you don't remind them.

Here is how Pask works, broken down into three simple parts using everyday analogies:

1. The "Sixth Sense" (Demand Detection / IntentFlow)

This is the brain of the operation. In the past, AI had to wait for a question. Pask has a sixth sense.

How it works: It watches what you are doing in real-time (reading emails, watching a video, having a meeting).
The Analogy: Imagine a security guard who doesn't just watch the cameras; they watch people. If they see someone looking lost in a mall, the guard doesn't wait for the person to ask, "Where is the bathroom?" The guard approaches and says, "Hey, you look like you need directions."
The Magic: Pask uses a special model called IntentFlow. It's trained to spot "latent needs"—the things you need but haven't said yet. It decides: Should I stay quiet? Should I give a quick tip? Or should I pull out the big tools and do a deep dive?

2. The "Super Memory" (Memory Modeling / Pask-MM)

Most AI has a short attention span. If you talk to it for 20 minutes, it starts forgetting what you said at the beginning. Pask is different. It has a three-layer memory system, like a human brain:

The "Cache" (User Memory): This is like a sticky note on your fridge. It holds your permanent traits: "I am a software engineer," "I hate spicy food," "I am allergic to peanuts." This is always ready, instantly.
The "Workbench" (Workspace Memory): This is like your current desk. It holds everything happening right now in this specific conversation. It's messy and temporary, but it keeps the current task organized.
The "Library" (Global Memory): This is a giant, organized archive. If you mention a project you worked on last year, Pask can instantly pull up the files, notes, and context from that project to help you today.
The Evolution: The best part? This memory learns and updates itself. If you change your job or your hobbies, Pask updates its "sticky notes" and "library" automatically, so it never gets outdated.

3. The "Always-On Engine" (Proactive Agent System / PAS)

This is the muscle that makes the brain and memory actually do things.

The Analogy: Think of this as the conveyor belt in a factory. The "Sixth Sense" spots a problem, the "Super Memory" recalls the solution, and the "Engine" immediately starts building the fix.
It runs in the background 24/7. It doesn't crash. It can talk to other tools (like searching the web, writing code, or analyzing data) without you having to open a new tab or give a command.

Why is this a big deal? (The Results)

The researchers tested Pask against the smartest AI models available today (like the ones from Google and OpenAI).

The Problem: Even the smartest AIs are bad at "reading the room." They often interrupt when they shouldn't, or stay silent when they should help. They are like that waiter who either trips over your feet or ignores your hunger.
The Solution: Pask was trained on a massive dataset of real human interactions (called LatentNeeds-Bench). It learned the subtle art of timing.
The Outcome: Pask is faster (it reacts in about 1.3 seconds) and smarter at guessing your needs than the current giants. It stays silent when you are focused, but jumps in with the perfect help the moment you get stuck.

The Bottom Line

We are moving from an era where you have to talk to the computer, to an era where the computer understands you.

Pask isn't just a tool you use; it's a partner that grows with you, remembers your history, and anticipates your needs before you even realize you have them. It's the difference between a calculator (which waits for you to press buttons) and a personal trainer (who watches your form, adjusts your weights, and encourages you before you even ask).

1. Problem Statement

Current AI systems largely operate in a passive "query-response" paradigm, where the AI waits for explicit user input before acting. This creates significant limitations for Artificial General Intelligence (AGI) in real-world scenarios:

Contextual Mismatch: Users often cannot stop to query an AI during time-sensitive tasks (e.g., live meetings, video watching) or social situations where interruption is awkward.
Latent Need Gap: Users struggle to articulate implicit needs, and current models fail to infer "latent demands" from ongoing context without explicit prompting.
Lack of Long-Term Adaptation: Existing agents lack evolving memory mechanisms to accumulate person-level understanding over time, preventing true human-AI co-evolution.
Real-World Constraints: Proactive agents must operate under strict latency constraints (real-time responsiveness) while managing long-horizon interactions and ambiguity.

The paper defines the goal as building an agent that can continuously sense context, infer latent user needs, and proactively intervene with appropriate assistance without disrupting the user flow.

2. Methodology: The DD-MM-PAS Paradigm

The authors propose Pask, a complete proactive AI system built on a unified framework called DD-MM-PAS (Demand Detection, Memory Modeling, Proactive Agent System).

A. Demand Detection (DD): IntentFlow

The core proactive capability is IntentFlow, a streaming foundation model designed for real-time demand detection.

Architecture: A dual-model system consisting of:
- Demand Detector: A primary model (based on Qwen3-30B-A3B) that processes streaming inputs and predicts a control token: <silent>, <fast_intervention>, or <full_assistance>.
- MemLoader: An auxiliary model (Qwen3-4B) that distills retrieved memory evidence for the detector.
Decision Logic:
- Silent: No intervention; preserves natural flow.
- Fast Intervention: Low-latency response (<1s) based on immediate context (e.g., explaining a term).
- Full Assistance: Triggers a memory-grounded reasoning pipeline for complex, personalized help.
Training Strategy:
- Data: Constructed LatentNeeds, a dataset of 102k samples (100k synthetic, 2k real-world).
- Pipeline: Multi-stage generation involving public data grounding, multi-agent simulation, and iterative human-intent inference.
- Optimization: Two-stage training: (1) Separate Supervised Fine-Tuning (SFT) for the detector and loader; (2) Reinforcement Learning (RL) using a hybrid reward (rule-based + LLM-as-a-judge) to align with deep human intent and balance helpfulness vs. intrusiveness.

B. Memory Modeling (MM): Self-Evolving Hierarchical Memory

To support long-term adaptation, Pask introduces a hybrid memory architecture inspired by computer memory hierarchies (Cache-Main-Storage):

User Memory ( $M_{user}$ - Cache): Stores stable user traits (profile, preferences) injected directly into the system prompt for zero-latency access.
Workspace Memory ( $M_{wsp}$ - Main Memory): Maintains the current session's context, conversation history, and intermediate states. Updated continuously during interaction.
Global Memory ( $M_{global}$ - External Storage): A tree-structured, RAG-based system storing long-term episodic knowledge.
- Self-Evolution: Uses Lazy Merging and Bounded-Depth strategies to prevent compute explosion. Updates occur asynchronously post-session to ensure inference latency remains under 1 second.

C. Proactive Agent System (PAS)

The operational backbone integrating frontend (AI glasses, phones, PC), server backend, and AI infrastructure. It orchestrates over 20 models and agents, managing multi-process execution, tool use (search, code), and feedback loops to ensure stable, always-on operation.

3. Key Contributions

DD-MM-PAS Paradigm: A general architecture unifying demand detection, memory-based user modeling, and always-on execution, shifting AI from reactive to proactive.
IntentFlow Model & Benchmark:
- Introduced IntentFlow, an ultra-fast streaming model for proactive assistance.
- Released LatentNeeds-Bench, a real-world benchmark derived from user-consented data and refined by human editors, covering Work, Learning, and Daily life domains.
Co-Evolving Memory System: A novel hierarchical memory design that balances real-time latency with deep, long-term personalization through structured self-evolution.
End-to-End System: A fully deployable system demonstrating that proactive intelligence can function reliably in real-world settings with low latency.

4. Experimental Results

The system was evaluated on LatentNeeds-Bench against strong baselines (GPT-5, Gemini-3, Claude, DeepSeek, Qwen).

Performance:
- IntentFlow achieved a Balanced Accuracy of 84.2%, outperforming the best baseline (Gemini-3-Flash at 80.8%) and significantly beating other open-source models.
- It successfully balanced Demand Detection (83.1%) and Non-Demand Silence (85.2%), whereas many baselines were either too intrusive or too passive.
Latency:
- IntentFlow maintained an inference latency of ~1.3–1.5 seconds, significantly faster than larger models (e.g., GPT-5-Mini at ~7–8s).
Long-Term Stability:
- In multi-turn interactions (up to 60 turns), IntentFlow showed minimal performance degradation (-5.0%), while baselines like GPT-5-Mini degraded by -19.0% and Gemini-3-Flash by -17.3%.
User Study:
- Pask improved knowledge acquisition scores in learning scenarios by +1.0 to +1.5 points compared to a control group.
- User ratings indicated the system is most effective in structured learning tasks, with daily life scenarios remaining the most challenging.

5. Significance and Conclusion

Paradigm Shift: The paper argues that scaling model size alone is insufficient; proactive AI requires a unified approach combining intent detection, memory, and system engineering.
Real-World Viability: Pask demonstrates that open-source models, when trained with specific data pipelines and RL for intent alignment, can match or exceed proprietary frontier models in proactive tasks while maintaining lower latency.
Future Direction: The work moves the field beyond "shallow" proactive behaviors (e.g., simple reminders) toward deep intent awareness, where agents can detect unexpressed needs and intervene at the right moment with high-value assistance.

The authors conclude that Pask provides a concrete path toward AGI agents that are not just reactive tools but active, evolving partners in human workflows.