AuthorsZhifei Xie, Zongzheng Hu, Fangda Ye et al.
TL;DR
Pask uses the DD–MM–PAS paradigm with the IntentFlow demand detector and a three-level memory system to reach 84.2% balanced accuracy on LatentNeeds-Bench, 3.4 points above Gemini-3-Flash.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
LatentNeeds-Bench shows a 41.9% demand rate, yet many LLMs score only 18.8–38.9 on demand turns, even when prompts encourage help.
This failure means real-time assistants either interrupt unnecessarily or stay silent when users need support, breaking timing, context sensitivity, and long-term understanding.
HOW IT WORKS
Pask centers Demand Detection (IntentFlow), Pask-MM hierarchical memory, and Pask-PAS execution, with IntentFlow built on Qwen3-30B-A3B-Instruct and MemLoader on Qwen3-4B-Instruct.
You can think of Pask like a computer stack: User Memory as cache, Workspace Memory as RAM, and Global Memory as disk, orchestrated by an always-on OS-like PAS.
This design lets Pask infer when to help, ground decisions in evolving memory, and coordinate tools, going far beyond a plain context window chatbot.
DIAGRAM
This diagram shows how Pask’s IntentFlow processes each information segment and chooses between silence, fast intervention, or full assistance with memory.
DIAGRAM
This diagram shows how Pask builds LatentNeeds-100k and LatentNeeds-2K, then trains IntentFlow with supervised fine-tuning and reinforcement learning.
PROCESS
Demand Detection
Pask’s IntentFlow Demand Detector ingests streaming inputs plus user profile and predicts <silent>, <fast_intervention>, or <full_assistance> at each turn.
Memory Access and Inference Mechanism
When full assistance is needed, Pask-MM combines User Memory, Workspace Memory, and Global Memory with localized RAG to assemble relevant evidence under 1 second latency.
MemLoader Refinement
MemLoader distills retrieved memory into compact core information, returning refined evidence to IntentFlow for final proactive response or abstention.
Memory Self-Evolution and Maintenance
After the session, Pask-MM updates User Memory with time-decayed Bayesian conflict resolution and lazily merges Global Memory tree nodes to keep long-term storage bounded.
KEY CONTRIBUTIONS
DD–MM–PAS Proactive AI Paradigm
Pask defines DD–MM–PAS, unifying Demand Detection, Memory Modeling, and Proactive Agent System into a coherent architecture for real-world proactive assistance.
IntentFlow Model and LatentNeeds-Bench
Pask introduces IntentFlow with a dual Qwen3-30B-A3B and Qwen3-4B architecture, trained on LatentNeeds-100k plus LatentNeeds-2K, and evaluated on LatentNeeds-Bench with 3,936 turns.
Self-Evolving Hierarchical Memory Modeling
Pask-MM designs User Memory, Workspace Memory, and Global Memory as a cache–main–storage hierarchy with asynchronous RAG, lazy merging, and time-decayed Bayesian updates.
RESULTS
Balanced Accuracy
84.2%
+3.4 over Gemini-3-Flash
Demand Accuracy
83.1%
+16.0 over GPT-5-Mini
No-Demand Accuracy
85.2%
+2.2 over Gemini-3-Flash
Per-turn Latency
1.3–1.5 s
fastest among all compared models
LatentNeeds-Bench is a 100-session, 3,936-turn benchmark with a 41.9% demand rate across Work, Learning, and Daily domains. On this benchmark, Pask’s IntentFlow shows that task-specific demand detection and memory integration can beat frontier assistants while staying within ~1.5 s latency.
BENCHMARK
LatentNeeds-Bench is a 100-session, 3,936-turn benchmark with a 41.9% demand rate across Work, Learning, and Daily domains. On this benchmark, Pask’s IntentFlow shows that task-specific demand detection and memory integration can beat frontier assistants while staying within ~1.5 s latency.
BENCHMARK
Balanced Accuracy across all domains on LatentNeeds-Bench.
KEY INSIGHT
Gemini-2.5-Flash-Lite achieves only 18.8 demand accuracy despite 2.0–2.4 s latency, while Pask’s IntentFlow reaches 83.1 demand accuracy with 1.3–1.5 s latency.
This breaks the assumption that faster lightweight models are good enough for proactive help; Pask shows that specialized demand training can be both faster and far more accurate.
WHY IT MATTERS
Pask unlocks proactive agents that can continuously sense context, infer latent needs, and act through tools while maintaining long-term, person-level memory.
Builders can now design assistants that sit in meetings, lectures, or daily life streams and intervene only when truly helpful, instead of waiting for explicit prompts.
Guilin Zhang, Wei Jiang et al.
· 2026
A-MAC scores candidate memories using Utility, Confidence, Novelty, Recency, and Type Prior combined by a learned linear admission policy with Algorithm 1 A-MAC Memory Admission. On the LoCoMo benchmark, A-MAC achieves F1 0.583 and 2644 ms latency, improving F1 by 0.042 and reducing latency by 1187 ms compared to A-mem.
Robbyant Team, Zelin Gao et al.
arXiv 2026 · 2026
LingBot-World combines a Data Engine, Fundamental World Model, Action-Conditioned World Model, and Post-Training causal adaptation to turn a 28B-parameter video generator into a real-time interactive world simulator. On the VBench benchmark, LingBot-World achieves a dynamic degree of 0.8857 versus 0.7612 for Yume-1.5, while also improving imaging quality to 0.6683.
Manoj Madushanka Perera, Adnan Mahmood et al.
· 2026
AgenticAI-DialogGen chains ChatPreprocessor, KnowledgeExtractor, TopicAnalyzer, KnowledgeGraphBuilder, PersonaGenerator, DuelingChat Agent, ConversationValidator, ConversationRefiner, QAGeneration, and PostProcessing to turn raw multi-session chats into topic-guided, persona-grounded conversations with explicit short- and long-term memories. On the TGC / KG memory QA benchmark, Mistral-7B fine-tuned within AgenticAI-DialogGen achieves 87.36 F1, compared to GPT-4’s 83.77 F1 in a zero-shot setting on the same task.
Answers use this explainer on Memory Papers.
Checking…