Yunjue Agent Tech Report: A Fully Reproducible, Zero-Start In-Situ Self-Evolving Agent System for Open-Ended Tasks
Yunjue Agent achieves zero-start adaptability for open tasks via parallel batch evolution and tool optimization, outperforming baselines.
Key Findings
Methodology
Proposed the In-Situ Self-Evolving framework, treating task interactions as a continuous experience stream. Introduced Parallel Batch Evolution to optimize tool synthesis and integration.
Key Results
- On DeepSearchQA, Yunjue Agent achieved 73.5% accuracy, outperforming Gemini 3 Pro by 16.9 percentage points.
- In xSciQA warm-start settings, accuracy improved from 76.5% to 80.2%, with a 100% reduction in new tool synthesis.
- The proposed EGL metric effectively monitored evolution convergence, with the tool library stabilizing after 2,500 queries, generating only 97 new tools.
Significance
This research addresses the adaptability limitations of traditional agent systems in open tasks, providing a dynamic, unsupervised tool optimization approach foundational for general intelligence.
Technical Contribution
Introduced Parallel Batch Evolution, combining tool semantic clustering and merging to enhance efficiency. Proposed the EGL metric to monitor evolution convergence, filling a critical gap.
Novelty
First to achieve dynamic tool evolution in zero-start environments, overcoming the reliance on offline training and static toolsets.
Limitations
- Tool generation efficiency may drop under extreme task distribution shifts.
- Limited support for large-scale parallel tasks.
- Long-term storage and management costs of the tool library remain unoptimized.
Future Work
Future work could explore more efficient tool clustering algorithms and applicability to multimodal tasks.
AI Executive Summary
Yunjue Agent is a novel zero-start self-evolving agent system designed to address the adaptability challenges of traditional agents in open task environments. By leveraging the In-Situ Self-Evolving framework, it dynamically optimizes its tool library to tackle emerging challenges.
Key innovations include the Parallel Batch Evolution strategy and the EGL convergence monitoring metric. Parallel Batch Evolution enhances tool generation efficiency through semantic clustering and merging, while EGL provides a stability monitoring mechanism analogous to training loss.
Experimental results demonstrate Yunjue Agent's superior performance across five benchmarks, including DeepSearchQA and xSciQA, significantly outperforming existing baselines in zero-start settings. Warm-start experiments further validate the tool library's cross-domain transferability, showcasing efficient adaptation to novel domains. Despite its strengths, the system faces challenges in extreme task distribution shifts and large-scale parallel processing, offering avenues for future improvement.
Deep Analysis
Background
Traditional agent systems rely on static toolsets and offline training, limiting adaptability in open environments with task distribution shifts and scarce supervision. Recent advances focus on tool optimization and self-evolving methods.
Core Problem
The core challenge is enabling agent systems to dynamically expand capabilities through tool evolution in zero-start environments while maintaining efficiency and generality.
Innovation
- �� Proposed the In-Situ Self-Evolving framework to treat task interactions as an experience stream.
- �� Introduced Parallel Batch Evolution to optimize tool synthesis via semantic clustering and merging.
- �� Developed the EGL metric to monitor tool library convergence in real-time.
Methodology
- �� The system comprises Manager, Tool Developer, Executor, and other modules.
- �� Upon receiving a task, the Manager retrieves relevant tools or directs the Tool Developer to generate new ones.
- �� Parallel Batch Evolution reduces redundancy through clustering and merging.
- �� EGL monitors the convergence of the tool library.
Experiments
Evaluated on five benchmarks, including HLE, DeepSearchQA, and xSciQA, under zero-start and warm-start settings. Key metrics include accuracy, tool synthesis count, and EGL trends.
Results
- �� Achieved 73.5% accuracy on DeepSearchQA, surpassing baselines.
- �� Warm-start xSciQA accuracy improved to 80.2%, with no new tools synthesized.
- �� Tool library stabilized after 2,500 queries, generating only 97 new tools.
Applications
Applicable to open-task scenarios requiring dynamic adaptation, such as cross-domain information retrieval, complex problem-solving, and interdisciplinary research.
Limitations & Outlook
Performance under extreme task distribution shifts and large-scale parallel processing needs improvement. Long-term tool library management costs may pose challenges.
Plain Language Accessible to non-experts
Imagine a chef facing new recipes daily, needing to invent new tools on the spot. Yunjue Agent acts like a smart kitchen assistant, automatically designing, optimizing, and storing tools for future use. For instance, if a recipe requires a unique vegetable cutter, it creates one instantly and reuses it later.
ELI14 Explained like you're 14
Think of playing a game where you invent new gadgets to beat each level! Yunjue Agent is like a gamer who creates tools on the fly to solve problems and saves them for later. Cool, right?
Glossary
In-Situ Self-Evolving
A dynamic framework treating task interactions as a continuous experience stream to update the tool library.
Used for zero-start capability expansion.
Parallel Batch Evolution
Optimizes tool generation by clustering and merging tools across parallel tasks.
Reduces redundancy and accelerates convergence.
EGL (Evolutionary Generality Loss)
A metric to monitor tool library convergence, similar to training loss.
Tracks evolution stability.
Tool Developer
A module responsible for generating new tools based on task requirements.
Triggered when existing tools are insufficient.
HLE (Humanity’s Last Exam)
A multidisciplinary benchmark for evaluating advanced reasoning.
Tests system performance on complex tasks.
Open Questions Unanswered questions from this research
- 1 How can tool generation efficiency be further optimized for large-scale parallel tasks?
- 2 What strategies can reduce long-term storage and management costs of the tool library?
Applications
Immediate Applications
Cross-domain information retrieval
Dynamically generates tools for multi-domain query and integration tasks.
Complex problem-solving
Applicable to scientific research and engineering design requiring dynamic task adaptation.
Long-term Vision
General AI
Advances toward self-adaptive general intelligence systems.
Abstract
Conventional agent systems often struggle in open-ended environments where task distributions continuously drift and external supervision is scarce. Their reliance on static toolsets or offline training lags behind these dynamics, leaving the system's capability boundaries rigid and unknown. To address this, we propose the In-Situ Self-Evolving paradigm. This approach treats sequential task interactions as a continuous stream of experience, enabling the system to distill short-term execution feedback into long-term, reusable capabilities without access to ground-truth labels. Within this framework, we identify tool evolution as the critical pathway for capability expansion, which provides verifiable, binary feedback signals. Within this framework, we develop Yunjue Agent, a system that iteratively synthesizes, optimizes, and reuses tools to navigate emerging challenges. To optimize evolutionary efficiency, we further introduce a Parallel Batch Evolution strategy. Empirical evaluations across five diverse benchmarks under a zero-start setting demonstrate significant performance gains over proprietary baselines. Additionally, complementary warm-start evaluations confirm that the accumulated general knowledge can be seamlessly transferred to novel domains. Finally, we propose a novel metric to monitor evolution convergence, serving as a function analogous to training loss in conventional optimization. We open-source our codebase, system traces, and evolved tools to facilitate future research in resilient, self-evolving intelligence.