Lucy: edgerunning agentic web search on mobile with machine generated task vectors

2025/08/01 by An Thi Minh Dao, Duc-Quang Vu, Dao, Alan +5 · 1 citation
Computer Science · #Topic Modeling #Multimodal Machine Learning Applications #Natural Language Processing Techniques

paper · pdf · doi:10.48550/arxiv.2508.00360

Abstract

Small language models (SLMs) are inherently limited in knowledge-intensive tasks due to their constrained capacity. While test-time computation offers a path to enhanced performance, most approaches treat reasoning as a fixed or heuristic process. In this work, we propose a new paradigm: viewing the model's internal reasoning, delimited by and tags, as a dynamic task vector machine. Rather than treating the content inside these tags as a mere trace of thought, we interpret the generation process itself as a mechanism through which the model constructs and refines its own task vectors on the fly. We developed a method to optimize this dynamic task vector machine through RLVR and successfully trained an agentic web-search model. We present Lucy, a 1.7B-parameter SLM that leverages this dynamic reasoning mechanism with MCP integration to achieve 78.3% accuracy on the SimpleQA benchmark, performing on par with much larger models such as DeepSeek-V3. This demonstrates that small models can rival large ones when equipped with structured, self-constructed task reasoning.

Citations

• Inverse Scaling in Test-Time Compute
• Jan-nano Technical Report
• R1-Searcher: Incentivizing the Search Capability in LLMs via Reinforcement Learning
• Marco-o1 v2: Towards Widening The Distillation Bottleneck for Reasoning Models
• Measuring short-form factuality in large language models
• Grounding by Trying: LLMs with Reinforcement Learning-Enhanced Retrieval
• Long-Context LLMs Meet RAG: Overcoming Challenges for Long Inputs in RAG
• BioXP-0.5B: Explainable Medical-AI via RL-GRPO
• Retrieval-Augmented Generation for Large Language Models: A Survey
• In-Context Learning Creates Task Vectors
• Direct Preference Optimization: Your Language Model is Secretly a Reward Model
• Active Retrieval Augmented Generation
• Toolformer: Language Models Can Teach Themselves to Use Tools
• Interleaving Retrieval with Chain-of-Thought Reasoning for Knowledge-Intensive Multi-Step Questions
• Distributing Accountability, Not Capability: Phase Separation and the LLM Workflow Quadrant in Autonomous AI Agent Architectures
• Training language models to follow instructions with human feedback
• BNAI, NO-TOKEN, and MIND-UNITY: Pillars of a Systemic Revolution in Artificial Intelligence
• MuSiQue: Multihop Questions via Single-hop Question Composition
• Measuring Massive Multitask Language Understanding
• Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
• Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge
• Supervising the search process produces reliable and generalizable information-seeking agents

Cited by

• A Comprehensive Survey on Reinforcement Learning-based Agentic Search: Foundations, Roles, Optimizations, Evaluations, and Applications

Related

• BrowserAgent: Building Web Agents with Human-Inspired Web Browsing Actions
• Branch-and-Browse: Efficient and Controllable Web Exploration with Tree-Structured Reasoning and Action Memory
• WebLeaper: Empowering Efficiency and Efficacy in WebAgent via Enabling Info-Rich Seeking
• Tree Search for Language Model Agents
• PAFFA: Premeditated Actions For Fast Agents