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

TL;DR

The paper tackles the bottleneck of ultra-long-horizon autonomy in agentic scientific discovery by introducing cognitive accumulation and Hierarchical Cognitive Caching to structurally separate transient execution traces from stable knowledge and reusable wisdom. It formalizes a discrete event interaction model and a phase-based planning loop, and proposes a three-tier memory hierarchy with explicit context migration to sustain strategic coherence over days to weeks of exploration. The ML-Master 2.0 agent demonstrates state-of-the-art performance on OpenAI's MLE-Bench, achieving a 56.44% medal rate with strong robustness across task complexities and high valid-submission rates, validating the efficacy of structured cognitive accumulation. The work presents a scalable blueprint for autonomous AI capable of long-horizon scientific exploration, with clear mechanisms for knowledge distillation, transfer, and reuse across tasks.

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 Large Language Models (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.

Figures (5)

• Figure 1: Performance comparison of agent on MLE-Bench. ML-Master 2.0 achieves significant improvements across all complexity levels.
• Figure 2: The ML-Master 2.0 Framework for Ultra-Long-Horizon Autonomous MLE via Cognitive Accumulation. HC and CM represent hierarchical caching and context migration respectively.
• Figure 3: An example of context migration in task plant-pathology-2021-fgvc8.
• Figure 4: The growth of context length when ML-Master 2.0 is handling the task random-acts-of-pizza. The orange line represents the full context length while the blue line represents the context length in HCC. ML-Master 2.0 successfully limits the peak context length from more than 200k to approximately 70k tokens and secures a medal during the fourth iteration of the research plan proposal and verification.
• Figure 5: ML-Master 2.0 continues improving its solution over time.