Trace2Tower: Constructing Multi-Level Skill Towers via Transition-Aware Feature Trajectory Induction

Background and Context

Large language model agents are increasingly dependent on execution trajectories to learn and master skills within complex interactive tasks. However, current mainstream paradigms are fundamentally constrained by shallow trajectory retrieval mechanisms and flat skill summarization methods. These limitations ignore the critical temporal dependencies and result-conditioned topological structures inherent in agent behavior. Consequently, agents struggle to extract deep, generalizable behavioral patterns from historical experiences, which restricts their generalization capabilities in long-horizon tasks. To address this core pain point, researchers have proposed the Trace2Tower framework, a transition-aware feature trajectory induction method. This framework represents a significant shift from simple trajectory copying or shallow matching to a novel distillation mechanism that transforms chaotic raw trajectories into structured, robust skill hierarchies. This evolution enables agents to abstract high-level task strategies from low-level step interactions, marking a crucial advancement from flat to hierarchical and structured skill learning.

The technical implementation of Trace2Tower relies on a sophisticated graph construction and spectral decomposition pipeline. The framework first abstracts raw step-level interactions into standard "event" units that capture key semantic information from agent-environment interactions. It then constructs a unified graph structure constrained by three core elements: semantic compatibility ensures logical coherence between events, transition dynamics capture the probabilistic characteristics of state transfers, and outcome evidence evaluates the validity of behavioral paths. This structural approach allows for a more nuanced understanding of how specific actions lead to desired outcomes, moving beyond simple correlation to causal understanding within the agent's operational context.

Deep Analysis

At the heart of Trace2Tower is a novel contrastive spectral decomposition algorithm designed to isolate stable behavior patterns that are highly aligned with successful outcomes. This algorithm applies strict mathematical constraints to suppress shortcut behaviors that may appear effective in the short term but are prone to failure over longer durations. By leveraging this decomposition, the framework effectively filters out noise and unreliable strategies, retaining only the most robust patterns. These extracted patterns are then organically populated into a dynamic "skill tower," which serves as the agent's internal knowledge base. This tower is not a static repository but a living structure that evolves with the agent's experience, ensuring that the most relevant and effective skills are always readily accessible for decision-making processes.

The dynamic skill tower is organized into three distinct hierarchical levels: action templates at the bottom, procedural routines in the middle, and macro-level task strategies at the top. This multi-level architecture allows for efficient reuse of context-specific experiences. The entire system is continuously refined through validator-guided feedback, creating a closed-loop learning system. This feedback mechanism ensures that the skill library adapts to new information and corrects errors in real-time. The integration of semantic compatibility, transition dynamics, and outcome evidence within the graph structure provides a comprehensive view of the agent's operational environment, enabling more accurate predictions and more effective skill selection during task execution.

Industry Impact

The introduction of Trace2Tower has profound implications for the open-source community, industrial applications, and subsequent research directions. For the open-source community, the framework offers an interpretable and structured paradigm for agent skill management. It encourages researchers to re-examine the agent learning process from topological and spectral analysis perspectives, potentially leading to new methodologies for understanding and improving agent cognition. By providing a clear structure for how skills are organized and retrieved, Trace2Tower lowers the barrier to entry for developing sophisticated agents, fostering a more collaborative and innovative research environment focused on robust and scalable AI systems.

In terms of industrial落地, the framework's efficient experience reuse capabilities and low invalid action rate make agents more feasible for resource-constrained real-world scenarios. This is particularly relevant for service robots and automated workflow systems that require long-term interaction and complex decision-making. The ability to quickly learn from a small number of successful trajectories and transfer these skills to new tasks significantly reduces training costs and inference latency. This efficiency is critical for commercial viability, as it allows companies to deploy agents that are not only intelligent but also cost-effective and reliable in dynamic operational environments.

Outlook

The dynamic skill tower concept proposed by Trace2Tower provides a new technical path for building general agents with lifelong learning capabilities. Future research can explore the construction of larger-scale skill libraries, cross-domain skill transfer, and more complex validator designs based on this framework. These advancements could further push large language model agents toward higher levels of cognition and autonomous decision-making. By addressing the limitations of shallow retrieval and flat summarization, Trace2Tower sets a new standard for agent skill learning, offering a robust foundation for solving complex, long-horizon tasks in the real world. The framework's emphasis on topological structure and spectral analysis opens up new avenues for research into the fundamental mechanisms of agent intelligence.

Experimental results on challenging benchmarks such as ALFWorld and WebShop demonstrate the framework's superior performance. On ALFWorld, Trace2Tower achieved an 87.31% success rate with an average of 10.35 steps and only 0.26 invalid actions. On WebShop, it reached a 50.67% exact success rate. These results significantly outperform existing baselines in task mastery and efficient contextual experience reuse. As the field of AI continues to evolve, frameworks like Trace2Tower will play a crucial role in shaping the next generation of intelligent agents, enabling them to operate with greater autonomy, efficiency, and reliability in increasingly complex environments. The transition from flat to hierarchical skill structures represents a fundamental shift in how we approach agent learning, promising more capable and adaptable AI systems in the future.