AIAny
Icon for item

Beyond Final Scores: A Systematic Evaluation of Agents for Long-Horizon AI Research and Development

Systematically evaluates LLM-driven autonomous agents on long-horizon AI research tasks using rule-based within-run metrics (Solution Framing, Execution, Feedback Control). Focuses on experience reuse and harness effects across 36 tasks and seven frontier models, finding agents act more like engineering optimizers than autonomous researchers.

Introduction

Most agent benchmarks report final scores, which hide where improvements originate and whether accumulated experience actually helps. This paper argues that diagnosing research-loop behavior requires within-run, rule-based metrics and controlled comparisons; applying this view reveals where current agents succeed, fail, and transfer experience (for better or worse).

Key Findings
  • Within-run decomposition matters: measuring Solution Framing, Execution, and Feedback Control uncovers distinct process bottlenecks that final scores mask, so one model’s strong final result can stem from very different internal behaviors.
  • Experience reuse is unstable: agents sometimes leverage prior runs to improve decisions but can also propagate misleading artifacts across tasks, so selective memory mechanisms are needed to realize consistent gains.
  • Harness design affects reliability more than raw capability: shared tooling and iteration policies change performance stability and reproducibility, so engineering the harness is as important as model choice.
  • Methodological novelty is rare: the best solutions typically adapt or combine established techniques rather than invent new algorithms, implying current systems function more like optimizers than independent researchers.
Who it's for and tradeoffs

Great fit if you care about building or evaluating autonomous ML/engineering agents and need diagnostic metrics beyond final-task scores. The paper gives concrete, testable metrics and an experimental protocol (36 long-horizon tasks, 3 rollouts each) that help pinpoint where to intervene (training, inference-time selection, memory policies, harness design). Look elsewhere if you only need single-turn benchmarks or short-horizon agent traces—the framework targets sustained iterative improvement and requires more complex verifiers and wall-clock budgets.

Method and scope

The study evaluates seven frontier models across 36 expert-curated long-horizon tasks with rule-based verifiers and three complementary capability axes (Solution Framing, Execution, Feedback Control). It reports avg@3 and best@3 per model-task pair and includes controlled intra- and inter-task comparisons to measure experience transfer and harness effects. The protocol emphasizes deterministic, verifier-driven metrics over subjective LLM judgments to enable fine-grained diagnosis and actionable recommendations for agent design.

Information

  • Websitearxiv.org
  • OrganizationsAffiliation: Meituan, Affiliation: University of Chinese Academy of [email protected]
  • AuthorsYiwei Li, Wanli Yang, Hexiang Tan, Xiangzhou Huang, Zhengyu Chen, Ziran Li, Borun Chen, Shanglin Lei, Huaisheng Zhu, Hao Tian
  • Published date2026/08/13

Categories

More Items

Provides a year-scale multimodal benchmark and evaluation framework for on-device long-term memory in personal assistants, built from real mobile user trajectories. Tests memory construction, retrieval, updating, temporal reasoning, and implicit preference inference, and includes a knowledge-grounded synthesis pipeline to form coherent long-horizon trajectories.

Treats agent self-improvement as natural selection over a population of harnesses (prompts, tools, skills, control flow), evolving a frozen-model agent by selecting harness edits that extend capability without regressing others. Uses a preserve-and-extend contract, lineage archive, and verifier-driven fitness (no gold solutions) to recombine complementary edits and transfer across benchmarks.

Supports multimodal scientific understanding, long-horizon agentic workflows and scientific tool interaction using a unified pipeline of multimodal pretraining, supervised fine-tuning and scalable multi-task reinforcement learning. Distinctive features include time-series modules for signal forecasting and a separate Memory Decoder that enables rapid domain specialization without changing the frozen 397B backbone.