Technical Summary: What Is Missing from AI Post-Training AI

Problem Statement

Recent advances in Large Language Model (LLM) agents have enabled them to autonomously post-train other LLMs, writing code, launching experiments, and managing pipelines. This capability raises the prospect of "AI-for-AI," where agents recursively improve AI systems. However, current discussions often conflate two distinct capabilities: execution-level capability (iterating within a selected training strategy, e.g., tuning hyperparameters, fixing bugs) and strategy-level capability (revising the high-level judgment about what to try next as evidence accumulates, e.g., switching from Supervised Fine-Tuning to Reinforcement Learning).

The paper argues that the prevailing view of automated AI R&D assumes a closed loop where agents propose, execute, interpret, and revise. The authors posit that while agents are competent at execution, they lack the strategy-level capability to spontaneously reevaluate and revise their initial training strategy once the process begins. This "strategy lock-in" acts as a ceiling for automated post-training, where the quality of the initial strategy determines the outcome rather than the number of iterations or compute resources.

Methodology

The study employs a mixed-method approach combining large-scale empirical analysis of existing trajectories with controlled, escalating interventions.

1. Empirical Analysis of Public Trajectories

The authors analyzed a corpus of 1,338 publicly released agent trajectories from PostTrainBench, covering seven benchmarks (including GSM8K, HumanEval, AIME 2025, and GPQA), four base models, and 20 distinct agent configurations (e.g., Claude Code, Codex CLI).

• Annotation Protocol: Trajectories were annotated to distinguish between execution changes (hyperparameter tuning, data formatting, bug fixing) and strategy changes (altering the training paradigm, data-source type, or stage structure).
• Metric: The study measured the frequency of strategy changes versus execution iterations and analyzed the convergence of initial strategies across different agents and tasks.

2. Escalating Interventions

To test whether the lack of strategy revision is due to missing experience, guidance, or reasoning compute, the authors conducted three controlled experiments using Qwen3-1.7B-Base on GSM8K, HumanEval, and AIME 2025. All interventions built upon a Claude Code (Opus 4.6) baseline.

• Intervention 1: Experience-Driven Scaffold. The agent was augmented with:
  • An Experiment Journal to persist observations and lessons across iterations.
  • A Skill Library distilled from open-source training documentation (e.g., verl, TRL).
  • An Evaluator Agent to provide concrete diagnoses and actionable suggestions.
• Intervention 2: Human Guidance. Before training began, a human reviewer revised the agent's proposed strategy. Once approved, the agent executed the run autonomously using the experience-driven framework.
• Intervention 3: Reasoning Compute Scaling. The experience-driven framework was run with significantly increased inference tokens (2–8× the baseline) to test if additional deliberation would trigger strategy revision.

Key Findings

1. Agents are Competent Executors but Locked into Default Strategies

Analysis of the 1,338 trajectories revealed a robust pattern of strategy lock-in:

• Initial Convergence: Agents converge on a narrow set of default strategies immediately, before writing code or running experiments. For instance, 80.7% of Claude Code trajectories anchored on Full-Parameter SFT, while 89.6% of Codex CLI trajectories anchored on Parameter-Efficient Fine-Tuning (PEFT).
• Static Strategy: Once training begins, the strategy remains fixed. Across 3,557 adjacent training pairs, only 2.1% involved a strategy change. The remaining budget was spent on local adjustments (learning rates, data mixtures) within the selected paradigm.
• Task Agnosticism: The lock-in reflects the agent's prior rather than the task; the same agent converges to similar strategies across different benchmarks, while different agents diverge on the same task.

2. Experience Improves Execution but Not Strategy

The experience-driven scaffold significantly improved execution-level performance:

• Performance Gains: It yielded +12.6 points on GSM8K and +40.8 points on HumanEval compared to autonomous baselines.
• Behavioral Shift: Agents utilized the journal and skill library to debug, repair implementations, and tune hyperparameters more effectively.
• Strategy Stagnation: Despite the evaluator agent explicitly suggesting strategy changes (e.g., switching from SFT to RL, or adding an SFT warm-up), the main agent never adopted these suggestions. The agent continued to iterate within the locked-in strategy, even when evidence indicated a plateau or failure.

3. Human Guidance Redirects Strategy but Fails to Sustain It

Human intervention successfully altered the initial strategy:

• Initial Success: A human reviewer redirected an agent from a pure SFT pipeline to an RL-focused pipeline, resulting in a higher initial performance (13.33% pass@8 on AIME 2025 vs. 3.33% for the baseline).
• Subsequent Collapse: Once training commenced, the agent reverted to local adjustment loops. It failed to investigate new causes for performance drops or revise the strategy further, eventually plateauing. The human guidance improved the starting point but could not sustain strategy-level adaptation during execution.

4. Additional Compute Yields Diminishing Returns on Hard Tasks

Scaling inference compute (tokens) improved performance on easier tasks (GSM8K, HumanEval) by enabling denser local search and better execution. However, on the hardest task (AIME 2025), the framework consumed 7.9× more tokens than the baseline but achieved no significant gain (solving only one additional problem, within evaluation variance). The extra compute was spent refining the locked-in strategy rather than discovering a new one.

Significance and Claims

The paper concludes that the bottleneck in AI post-training AI is not a lack of resources (experience, guidance, or compute) but a lack of a mechanism for spontaneous strategy reevaluation.

• The "Open Joint": The automated R&D loop is closed at the execution level (agents produce results, diagnose, and repair) but remains open at the strategy level. Agents record diagnoses and even write code for alternatives but fail to take the step of switching strategies.
• Capability vs. Spontaneity: The authors argue that the capability to revise a strategy exists (demonstrated when humans guided the agent to successfully implement a non-default strategy). What is missing is the spontaneity to initiate this revision when evidence warrants it.
• Implication for AI R&D: The ceiling of current automated post-training is set by the quality of the initial strategy, not by the number of iterations. To advance, the field must move beyond simply scaling models or adding more context; instead, it requires training signals and interaction protocols that explicitly reward the reopening of committed choices and make strategy revision an explicit decision point.

In summary, frontier agents are highly effective executors of established pipelines but systematically deficient in the high-level judgment required to pivot when those pipelines fail. The path forward lies in engineering mechanisms that trigger strategy revision, rather than merely optimizing execution within a fixed paradigm.