PaperlayerView exact revision on arXiv

Source-linked AI summary

What is Missing from AI Post-Training AI: An Empirical Analysis

Joy Jia Yin Lim, Xin Huang, Hao Peng, Yaxi Lu, Xin Cong, Zhong Zhang, Maosong Sun, Yankai Lin

arXiv:2608.19072v1cs.AIcs.CLcs.LG

TL;DR

LLM post-training agents can execute training strategies, but it remains unclear whether they can revise those strategies as evidence accumulates. This paper analyzes post-training trajectories and tests experience, human guidance, and additional reasoning compute, finding that strategies lock in early and agents lack spontaneous strategy reevaluation during execution.

  • Problem

    The paper asks whether LLM post-training agents can revise high-level training strategies from evidence, rather than merely execute within a selected strategy.

  • Method

    The paper analyzes publicly released post-training trajectories and evaluates escalating interventions targeting experience, human guidance, and reasoning compute.

  • Results

    Across tasks, strategies lock in at the beginning; experience improves execution by +12.6 points on GSM8K and +40.8 on HumanEval but leaves strategy static, while guidance and compute do not sustain revision.

  • Takeaways & Limitations

    Agents’ post-training ceiling is set by their initial strategy, indicating a need for mechanisms that make strategy revision an explicit, rewarded action.

  • Takeaways & Limitations

    The released trajectories do not expose a uniform wall-clock timestamp for when strategy lock-in occurs.

Abstract

from arXiv · show

Large language model (LLM) agents can now post-train an LLM end-to-end. They can write code, launch training, evaluate checkpoints, and improve downstream performance, raising the prospect of AI-for-AI. We argue that this picture conflates two distinct capabilities: execution-level capability, iterating within a selected training strategy; and strategy-level capability, revising the high-level judgment as experimental evidence accumulates. Analyzing a large corpus of publicly released post-training trajectories, we find that across different tasks, the agent's training strategy is locked in at the very beginning, and the entire remaining budget is spent on local adjustments within the selected strategy. We then examine three natural explanations--missing experience, missing guidance, and insufficient reasoning--with escalating interventions. Extensive experiments show that (1) an experience-driven scaffold improves execution across the board (+12.6 points on GSM8K and +40.8 on HumanEval) but leaves the strategy static; (2) human guidance effectively redirects the initial strategy, yet the agent falls back into local adjustment loops once training starts; and (3) additional inference compute pays off on easier tasks but yields almost no gain on the hardest one. In conclusion, what agents lack is neither experience, guidance, nor reasoning compute, but a mechanism for spontaneously reevaluating their strategy during execution.

1 INTRODUCTION

LLM agents can execute end-to-end post-training pipelines, but their high-level training strategies typically lock in before experimentation and remain unchanged during subsequent local optimization. The study argues that current AI post-training systems therefore improve mainly within an initially selected strategy rather than spontaneously revising that strategy.

  • Conceptual framing: The paper distinguishes execution-level capability—iterating within a selected training strategy—from strategy-level capability—revising high-level judgments as evidence accumulates.Post-training provides a standardized testbed because it has established infrastructure and evaluation benchmarks, while frontier agents can already execute the pipeline end to end.
  • Empirical finding: Agents lock in a training strategy before writing code or running experiments, then spend the remaining budget on local adjustments and incremental gains.This pattern persists across task differences: the same agent selects highly similar strategies across tasks, while different agents anchor on different defaults.
  • Interventions: The study tests missing experience, guidance, and extended reasoning as explanations for strategy lock-in through escalating interventions, including persistent experiment journals, skill libraries, and evaluator-agent diagnoses.These interventions target the agent’s ability to retain observations, access procedural knowledge, and interpret experimental outcomes during post-training.
  • Conclusion: The central conclusion is that automated AI R&D closes its hypothesis–experiment loop only locally: agents adjust experiments and generate new results without revising the underlying strategy.Consequently, the ceiling of current AI post-training AI is set by the quality of the initial strategy rather than by continued optimization.
  • Empirical finding: 180.7% of the Claude Code trajectories converge on full-parameter supervised fine-tuning (SFT), while 89.6% of Codex CLI trajectories converge on parameter-efficient fine-tuning (PEFT).

2 PRELIMINARIES

The paper separates execution-level pipeline implementation from strategy-level revision, then analyzes agent trajectories using explicit experiment-counting and evaluation protocols. The study covers 1,338 trajectories across diverse benchmarks, models, and agent configurations.

  • Conceptual distinction: The paper distinguishes training strategy from training pipeline, separating high-level plans governing paradigm, data, and stages from their concrete implementation.This distinction organizes the analysis and interventions into strategy-level and execution-level capabilities.
  • Strategy-Level Capability: Strategy-level capability selects and revises the strategy by changing paradigms, stages, data-source types, or budget allocation as evidence accumulates.It concerns updating high-level judgment rather than making local pipeline adjustments.
  • Data and protocols: Evaluation reports pass@1 across benchmarks, while AIME 2025 uses pass@8 because its 30-problem size makes one-problem differences potentially evaluation variance.Small AIME score differences are interpreted qualitatively alongside trajectory-level evidence.

3 ANALYSIS: COMPETENT EXECUTORS WITH LIMITED STRATEGY

Released trajectories show that frontier agents competently execute LLM post-training—building pipelines, diagnosing failures, and improving performance within a chosen strategy—but usually lock that strategy during the pre-update planning phase and do not revise it thereafter.

  • Execution-level capability: Agents average 3.82 trainings and 13.80 evaluations per trajectory, and almost all frontier agents complete the pipeline while improving over the base model on every benchmark.The execution-level results cover data preparation, training, evaluation, and checkpoint submission.
  • Execution-level capability: Agents perform substantive within-strategy repairs, including changing generation templates, concentrating data on function completion, repairing EOS handling, staging data construction, and reducing learning rates progressively.These interventions turned failing runs into scorable improvements, indicating that execution-level capability is not the dominant limitation.
  • Strategy lock-in: 80.7% of Claude Code trajectories anchor on full-parameter SFT and 89.6% of Codex CLI trajectories on PEFT across seven benchmarks and four base models, with strategy changes counted across training paradigms, data sources, or stage structures.Different agents therefore lock into systematically different default strategies on the same tasks.
  • Strategy lock-in: The strategy locks within a short pre-execution window, remains fixed across training, feedback, and additional budget, and differs systematically across agents rather than tasks.Here, “beginning of a run” means the pre-update planning phase because released trajectories lack uniform wall-clock timestamps.

4 WHAT IS MISSING FROM AI POST-TRAINING AI

The interventions show that experience improves execution without revising strategy, guidance redirects strategy only before training, and extra reasoning compute helps easier tasks but saturates on harder ones. The central missing capability is spontaneous strategy reevaluation during execution.

  • Strategy-level limitation: The experience scaffold makes evidence clearer and execution more reliable, yet negative evidence triggers local adjustments rather than revisions such as switching from SFT to RL.The main agent adopts execution-level suggestions but none of the strategy-level suggestions, even when evaluators and the journal identify plateaus or formatting failures.
  • Human guidance: Human guidance redirects the initial strategy, but the agent later iterates within that strategy without further gains.After an early peak, it adjusts hyperparameters and checkpoint-resumption choices rather than investigating alternative causes, showing that sustained strategy revision remains necessary.
  • Reasoning compute: On AIME 2025, spending 7.9× the baseline tokens yields only one additional problem solved in the best run, indicating a ceiling for compute scaling on hard tasks.By contrast, extra evaluation and diagnostic passes produce large gains on GSM8K and HumanEval, where the cost–performance trade-off is favorable.
  • Summary: Experience improves execution across tasks but leaves the training strategy unchanged; guidance redirects strategy before training but later erodes, while extra compute helps easier tasks and plateaus on harder ones.Together, these results identify strategy reevaluation during execution—not experience, guidance, or reasoning compute—as the missing capability.

5 WHAT THIS IMPLIES FOR AUTOMATED AI R&D

Automated AI R&D closes the execution loop but leaves the strategy loop open: agents diagnose and repair locally without actually revising their initial strategy. Consequently, run quality is bounded by the initial strategy, while effective remedies must trigger strategy revision rather than merely increase model size or compute.

  • 5 WHAT THIS IMPLIES FOR AUTOMATED AI R&D: Agents close local execution loops but leave the strategy-level loop open, following largely linear trajectories without acting on diagnosed alternatives.Figure 7 depicts effective post-training as a globally closed loop, whereas agents operate in local execution loops.
  • 5 WHAT THIS IMPLIES FOR AUTOMATED AI R&D: The initial strategy sets the run’s upper bound because additional iterations and compute refine the chosen strategy without recovering from a wrong choice.An agent can iterate efficiently inside the wrong strategy for ten hours.
  • 5 WHAT THIS IMPLIES FOR AUTOMATED AI R&D: Agents can depart from their default strategy when guided, but they do not spontaneously initiate that departure during execution.In the human-guided run, the agent understood the rationale, implemented an unfamiliar strategy, and independently extended the guidance.
  • 5 WHAT THIS IMPLIES FOR AUTOMATED AI R&D: The remedy is to reward reopening committed choices when evidence warrants and make strategy revision an explicit interaction decision point.These changes target initiation rather than requiring a larger model.

6 RELATED WORK

Related work frames AI systems as increasingly capable of automating end-to-end research and AI R&D workflows. Existing efforts span multiple automation paradigms and benchmark domains, from literature review and experimentation to coding, debugging, and training.

  • AI for Research: AI systems increasingly automate research workflows spanning literature review, hypothesis generation, experimentation, validation, and reporting.Representative systems include template-based automation, template-free tree search, and human-guided workflows.
  • AI for Research: Representative AI-for-research systems use template-based automation (Lu et al., 2024), template-free tree search (Yamada et al., 2025), or human-guided workflows.Later work adds hypothesis search and cross-run experience (Tang et al., 2026; Liu et al., …).
  • Automated AI R&D: Automated AI R&D agents operate over datasets, code, training, evaluation protocols, and checkpoints across benchmarks for experimentation, engineering, debugging, scientific coding, replication, and research engineering.The cited benchmark landscape includes machine learning experimentation, Kaggle-style engineering, interactive debugging and training, scientific coding, paper replication, expert-level scientific tasks, and research engineering.

7 CONCLUSION · A TRAJECTORY ANALYSIS DETAILS · A.1 CORPUS AND SCOPE

The paper distinguishes executing a committed post-training strategy from revising that strategy based on evidence, finding that current agents are competent executors but rarely revise their initial commitment. Its trajectory corpus covers 1,338 runs across seven benchmarks, four base models, five interfaces, and 20 agent-model configurations, while comparisons remain observational.

  • 7 CONCLUSION: The paper separates executing an established strategy from revising that strategy as evidence accumulates.It tests whether the gap reflects missing experience, guidance, or reasoning compute through three escalating interventions.
  • A TRAJECTORY ANALYSIS DETAILS: The analysis examines publicly released post-training trajectories to characterize how frontier agents execute strategies and whether they revise them.Across these trajectories, the training strategy is locked in at the beginning of post-training.
  • 7 CONCLUSION: Agents’ post-training loop closes at execution but remains open at strategy revision, so run ceilings depend on initial strategy quality.The paper identifies spontaneous strategy revision when evidence warrants it—not additional resources or raw capability—as the missing ingredient.
  • A.1 CORPUS AND SCOPE: The corpus contains 1,338 trajectories spanning seven benchmarks, four base models, five agent interfaces, and 20 agent-model configurations.Each trajectory uses a 10-hour budget on one NVIDIA H100 80GB GPU, and 900 trajectories launch at least one model update.
  • A.1 CORPUS AND SCOPE: The corpus is observational because agent models, interfaces, prompts, and benchmarks were not independently randomized.Therefore, the comparisons describe recurring behavior rather than causal effects of any particular agent.
  • A.1 CORPUS AND SCOPE: Table 4 defines the benchmark coverage of the agent post-training trajectory corpus.The corpus spans seven benchmarks, as described in the accompanying corpus summary.

A.2 STRATEGY ANNOTATION · A.3 METRICS DEFINITION · A.4 AUDIT OF STRATEGY-CHANGING TRAJECTORIES

The analysis conservatively separates execution-level adjustments from strategy revisions and finds that recognized strategy changes are rare. An audit shows that switching can help or hurt, so effective adaptation requires evidence-based testing and selection rather than switching alone.

  • A.2 STRATEGY ANNOTATION: Verified experiments count only executed commands that start model parameter updates, excluding scripting, data preparation, evaluation, adapter merging, and other non-updating operations.Each experiment is annotated with training strategy, data-source type, and stage structure; Full SFT and PEFT are distinct from the first executed training.
  • A.2 STRATEGY ANNOTATION: Data-source labels cover curated, self-generated, or mixed data, but provenance is identifiable for only 1,801 of 4,344 strategy-labeled experiments, making these changes lower bounds.On-policy RL rollouts remain part of the same objective, and ordinary continuation with a new learning rate is not a new stage.
  • A.2 STRATEGY ANNOTATION: 2.1% of 3,557 recognized adjacent experiment pairs changed strategy: 35 objective-family, 38 data-source, and one stage change.The analysis recognized changes in only one coarse dimension per pair.
  • A.2 STRATEGY ANNOTATION: Initial strategies differed systematically: Claude Code began with Full SFT in 80.7% of recognized cases, versus 89.6% PEFT for Codex CLI.Claude Code had the higher Full-SFT share in all 28 matched benchmark–base-model cells, while Codex CLI had the higher PEFT share; the observational design cannot isolate the cause.
  • A.3 METRICS DEFINITION: The metrics define execution-level actions as changes made while the strategy state remains fixed, whereas a strategy revision changes any component of s_i,t = (k_i,t, d_i,t, g_i,t).Persistence and change rates are computed over adjacent experiment pairs with recognized strategy states.
  • A.4 AUDIT OF STRATEGY-CHANGING TRAJECTORIES: Among 16 trajectories with objective-family changes, 14 began with SFT, 11 switched at least twice, and 15 of 35 transitions returned to SFT.The first switch occurred at median normalized progress 0.40, while all switches had median progress 0.67.
  • A.4 AUDIT OF STRATEGY-CHANGING TRAJECTORIES: Audited strategy switches produced both gains and regressions, so the evidence supports testing and selecting alternatives rather than switching for its own sake.Because trajectories retain their original comparators and sample sizes, the cases do not establish a pooled treatment effect.

B CONTROLLED INTERVENTION PROTOCOL · C EXPERIENCE-DRIVEN FRAMEWORK

The controlled protocol fixes model, benchmark, budget, prompting, and evaluation conditions while testing an experience-driven framework that supplies three resources but leaves training decisions to the main agent. Evaluation uses pass@1 for GSM8K and HumanEval, pass@8 for AIME 2025, and interprets small AIME differences alongside trajectories.

  • B CONTROLLED INTERVENTION PROTOCOL: The experiments use Qwen3-1.7B-Base on GSM8K, HumanEval, and AIME 2025, with autonomous baselines from Claude Code, GLM-5.2, and Codex CLI.All interventions use Claude Code with Opus 4.6, and each configuration has three independent 10-hour runs on four NVIDIA A800 GPUs.
  • B CONTROLLED INTERVENTION PROTOCOL: Trace-level case studies cover 16 strategy-changing trajectories while preserving each log’s original comparator and sample size.The table reports values drawn from trajectory logs rather than introducing a separate experimental comparison.
  • B CONTROLLED INTERVENTION PROTOCOL: Within each comparison, the base model, benchmark, hardware budget, system prompt, and evaluator are fixed; human guidance adds plan review before training.The human-guidance condition retains the full experience-driven framework while adding this review step.
  • B CONTROLLED INTERVENTION PROTOCOL: AIME 2025 is evaluated with pass@8 rather than pass@1 because its 30 problems make each solved problem worth 3.33 percentage points.All AIME submissions use eight completions per problem with identical evaluator and decoding configurations, and small differences are interpreted with trajectories.
  • C EXPERIENCE-DRIVEN FRAMEWORK: Checkpoint lineage trees track SFT or GRPO/RL objectives, identify best or submitted checkpoints, and mark reverts to earlier checkpoints with dashed edges.The visualization covers GSM8K, HumanEval, and AIME 2025 under the experience-driven framework.
  • C EXPERIENCE-DRIVEN FRAMEWORK: The experience-driven framework supplies three resources while leaving all training decisions to the main agent.Figure 9 presents the framework’s conceptual overview, while Table 9 enumerates its components.

C.1 EXPERIMENT JOURNAL … D.1 PLAN REVIEW

The detailed interventions externalize experimentation through journals, distilled skills, evaluator feedback, and iterative human plan review. These mechanisms improve planning and execution support, but evaluator feedback does not reliably induce strategy-level revisions, while human review redirects the initial plan before autonomous training.

  • C.1 EXPERIMENT JOURNAL: The experiment journal is an append-only, typed record of plans, observations, lessons, evaluation results, and analyses that captures pipeline-relevant conclusions without necessarily revising strategy.On HumanEval, it records that the SFT plateau is confirmed and further SFT should be abandoned because of diminishing returns.
  • C.2 SKILL LIBRARY: The skill library distills 908 documents totaling approximately 937K words from open-source training projects into a compact 60-page knowledge wiki.Its sources include documentation, training recipes, and issue threads from projects such as verl, TRL, OpenRLHF, NeMo-RL, and slime.
  • C.2 SKILL LIBRARY: Despite consulting existing skills 18, 20, and 60 times per run on GSM8K, HumanEval, and AIME 2025 respectively, the agent creates no new skills, including in the human-guidance run.This shows substantial knowledge consumption without reusable-knowledge production.
  • C.3 EVALUATOR AGENT: On HumanEval, seven of nine evaluation cycles recommend code-execution-reward RL, yet the agent launches 14 SFT variants and no RL; on AIME 2025, it ignores three SFT warm-up suggestions and remains GRPO-only.The remaining evaluator suggestions concern execution-level details such as rewards, hyperparameters, data, formatting, and checkpoint management.
  • C.4 OTHER RELATED WORK: Experience-driven systems build on verbal reflections, natural-language lessons, executable skill libraries, runtime experience, lifelong learning, and newer co-evolving memory, skill, and policy approaches.The cited literature includes Shinn et al. (2023), Zhao et al. (2024), Wang et al. (2023), Zhou et al. (2026), Ma et al. (2025), and Xu et al. (2026a).
  • D.1 PLAN REVIEW: Plan review uses repeated iterations in which the agent submits a complete research plan and the human returns a fixed-format JSON decision before training begins.The first iteration revises an approximately 10K–30K-example, 2–3-epoch SFT proposal; the second accepts the reduced plan with one clarification.
  • D HUMAN GUIDANCE DETAILS: Human plan review redirects the initial strategy from extensive SFT toward a formatting-only warm-up and GRPO, after which the agent autonomously skips SFT and assigns the full budget to GRPO.The review also requires an explicit reasoning-degradation check and benchmark-aligned evaluation format.

D.2 LATER EXECUTION · E CROSS-AGENT ANALYSIS

In later execution, the human-guided run peaks within the first few training versions, while subsequent checkpoints regress without prompting checkpoint restoration or strategy reconsideration. Later interventions adjust hyperparameters after resuming earlier checkpoints, but do not test ongoing human guidance or reliably preserve the best state.

  • D.2 LATER EXECUTION: The representative human-guided run reaches its best checkpoint in the first few training versions, after which later versions mostly score zero.The early peak is not exceeded during the remaining trajectory.
  • D.2 LATER EXECUTION: The agent neither restores the best checkpoint nor treats regression as a trigger to reconsider its training strategy.Performance degradation therefore does not induce state recovery or high-level strategy revision.
  • D.2 LATER EXECUTION: Later checkpoints do not surpass the best observed intermediate result in representative experience-driven and human-guided AIME 2025 trajectories.Chronological in-run diagnostics show that later checkpoints fail to exceed the best intermediate result.
  • D.2 LATER EXECUTION: Later experiments mainly adjust the entropy coefficient and learning rate when resuming from earlier checkpoints.These interventions operate through local hyperparameter changes rather than a changed objective.
  • D.2 LATER EXECUTION: The journal records instability and failed retries, but this evidence does not change the objective or reliably preserve the best state.Observed failures remain documented rather than producing robust recovery or objective revision.
  • D.2 LATER EXECUTION: The intervention tests initial plan revision only, not ongoing human guidance during execution.Its design therefore cannot establish whether continued human input would redirect later training decisions.

E.1 GPT-5.2 (CODEX CLI) · E.2 GLM-5.2 (CLAUDE CODE)

GPT-5.2 executes frequent, technically capable local adjustments but rarely changes strategy, whereas GLM-5.2 makes larger data and compute changes while its underlying training strategy remains fixed. Both agents improve some benchmarks, yet their outcomes depend strongly on task, scale, and implementation details.

  • E.1 GPT-5.2 (CODEX CLI): GPT-5.2 spent 61.8% of the six annotated trajectories’ 28.40 hours training, with AIME v1 allocating 87.6% of wall time to training without reliable gain.Evaluation consumed 12.9% of the combined time, while inspection and debugging consumed 9.7%.
  • E.1 GPT-5.2 (CODEX CLI): GPT-5.2 diagnoses scripts, protocols, EOS, and template problems effectively, but negative evidence typically triggers nearby data, template, hyperparameter, or checkpoint variants rather than a new strategy.The trajectories are therefore characterized as high-feedback, local-action behavior.
  • E.2 GLM-5.2 (CLAUDE CODE): GLM-5.2’s strongest loops use rejection-filtered data, selected checkpoints, and schedule changes: GSM8K uses three rejection-sampling rounds, while HumanEval benefits from staged expansion and lower-learning-rate continuation.Intermediate GSM8K checkpoints outperform final training steps.
  • E.2 GLM-5.2 (CLAUDE CODE): On AIME, correcting generation config.json raises both trained runs from 1/30 and 2/30 under a 2048-token cap to 7/30 with 16,000 tokens, while the training strategy itself never changes.Rejection-filtered self-generated data changes the data source, but is classified as non-strategy revision under the paper’s criterion.
Loading 2608.19072v1…