Toward Ultra-Long-Horizon Agentic Science: Cognitive Accumulation for Machine Learning Engineering

Abstract: The advancement of artificial intelligence toward agentic science is currently bottlenecked by the challenge of ultra-long-horizon autonomy, the ability to sustain strategic coherence and iterative correction over experimental cycles spanning days or weeks. While LLMs have demonstrated prowess in short-horizon reasoning, they are easily overwhelmed by execution details in the high-dimensional, delayed-feedback environments of real-world research, failing to consolidate sparse feedback into coherent long-term guidance. Here, we present ML-Master 2.0, an autonomous agent that masters ultra-long-horizon machine learning engineering (MLE) which is a representative microcosm of scientific discovery. By reframing context management as a process of cognitive accumulation, our approach introduces Hierarchical Cognitive Caching (HCC), a multi-tiered architecture inspired by computer systems that enables the structural differentiation of experience over time. By dynamically distilling transient execution traces into stable knowledge and cross-task wisdom, HCC allows agents to decouple immediate execution from long-term experimental strategy, effectively overcoming the scaling limits of static context windows. In evaluations on OpenAI's MLE-Bench under 24-hour budgets, ML-Master 2.0 achieves a state-of-the-art medal rate of 56.44%. Our findings demonstrate that ultra-long-horizon autonomy provides a scalable blueprint for AI capable of autonomous exploration beyond human-precedent complexities.

• The paper introduces a cognitive accumulation framework using Hierarchical Cognitive Caching to refine and abstract agent memory over extended horizons.
• The paper demonstrates ML-Master 2.0’s state-of-the-art performance with a 56.44% medal rate and over 63% success in tasks compared to human median benchmarks.
• The paper validates that each memory tier is critical, as ablation of any layer significantly degrades performance and strategic convergence.

Ultra-Long-Horizon Agentic Science via Cognitive Accumulation in ML-Master 2.0

Motivation and Problem Statement

Contemporary AI efforts in agentic science—where autonomous agents orchestrate long-term, iterative research processes—face critical bottlenecks in ultra-long-horizon autonomy. Real-world scientific inquiry is defined by prolonged, high-dimensional trial-and-error cycles and delayed feedback, which overwhelm naive context aggregation strategies in current LLM-based agents. The challenge is to maintain strategic coherence and consolidate sparse, valuable insights—rather than simply extending context windows—over sustained temporal horizons in environments like OpenAI's MLE-Bench.

Cognitive Accumulation: Architecture and Formalization

The ML-Master 2.0 agent reinterprets long-term context management as a process of cognitive accumulation: not just retaining historical data, but evolving context via continual refinement, stabilization, and abstraction. This paradigm is operationalized through Hierarchical Cognitive Caching (HCC), which structurally differentiates agent memory into three tiers:

• Evolving Experience ($\mathcal{L}_1$): High-fidelity traces for immediate decision-making and debugging.
• Refined Knowledge ($\mathcal{L}_2$): Abstracted, validated insights derived from completed exploration phases.
• Prior Wisdom ($\mathcal{L}_3$): Persistently stored, task-agnostic strategic knowledge transferable across domains.

This multi-tiered structure mirrors computer system cache hierarchies, effectively segregating volatile, frequently accessed information from durable, high-level knowledge.

Figure 1: The ML-Master 2.0 framework emphasizing hierarchical caching (HC) and context migration (CM) for scalable long-horizon context management.

Context Migration: Dynamic Knowledge Evolution

Context migration orchestrates the dynamical flow of information between cache tiers via three main operations:

• Context Prefetching: At initialization, semantic embeddings retrieve relevant prior wisdom from $\mathcal{L}_3$ to inform early exploration.
• Context Hit Policy: During event-driven interaction, raw experience in $\mathcal{L}_1$ is favored, with fallbacks to $\mathcal{L}_2$ summaries to prevent context saturation.
• Context Promotion: Phase-level traces are abstracted into knowledge units via LLM-driven summarization and consolidated further into wisdom artifacts for inter-task transfer.

  Figure 2: Example of context migration in a complex MLE task, showing promotion from raw traces to knowledge and then wisdom.

Experimental Evaluation and Empirical Performance

Evaluation on the MLE-Bench benchmark demonstrates that ML-Master 2.0 achieves a state-of-the-art medal rate of 56.44%, outperforming previous open-source and proprietary methods both in aggregate and at all difficulty levels. The agent maintains robust valid submission rates and demonstrates a strong capability to exceed the median performance of human participants in over 63% of tasks.

Ablation studies confirm the synergistic necessity of all hierarchical caching layers: removal of $\mathcal{L}_1$ drastically reduces valid submissions, omission of $\mathcal{L}_2$ degrades medal rates, and exclusion of $\mathcal{L}_3$ impairs initial convergence and cross-task transfer, confirming the critical role of structured cognitive accumulation.

Figure 3: Context length growth controlled by HCC in the “random-acts-of-pizza” task, demonstrating effective containment of peak context tokens and successful medal acquisition.

Theoretical and Practical Implications

The ML-Master 2.0 paradigm advances agentic science by providing a scalable solution for ultra-long-horizon autonomy. By abstracting agent memory as a multiscale, adaptive process, it enables agents to synthesize and reuse knowledge over weeks-long exploration cycles—addressing the combinatorial explosion of execution details and preventing strategic drift. Theoretical implications suggest a departure from static memory architectures, advocating for dynamic, life-cycled experience promotion regulated by value and stability.

Practically, the framework allows for robust autonomous MLE agents that can generalize across heterogeneous scientific and engineering tasks, potentially offering scaffolding for future AGI systems capable of continual, self-refining research.

Future Directions

Prospective work will involve:

• Meta-learning enhancements: Dynamic adjustment of cache promotion thresholds and embedding strategies for improved transferability.
• Extending to other scientific domains: Application of cognitive accumulation in physical experimentation, simulation-based research, and multi-modal scientific workflows.
• Integrating external memory systems: Interfacing HCC with operating-system-level memory architectures and persistent out-of-band storage for lifelong learning.

Conclusion

ML-Master 2.0 establishes a new paradigm for agentic science by operationalizing cognitive accumulation through hierarchical caching and structured context migration. Its strong empirical performance validates the critical role of memory differentiation and abstraction for ultra-long-horizon autonomy in real-world scientific discovery. This framework offers scalable foundational principles for future autonomous research agents capable of surpassing human precedent in complexity, robustness, and strategic continuity.