---
title: "What is Missing from AI Post-Training: An Empirical Analysis"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.19072v1"
permalink: "https://feed.astralor.com/items/itm_4d511bc987ec847f717107ba"
published: "2026-08-19T16:17:39.000Z"
score: 78
channel: "ai"
tags: ["AI agents","arXiv","Post-training","Large Language Models","Strategy Reevaluation"]
---

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

## AI commentary

> What makes this paper stand out is how it separates 'execution ability' from 'decision-making ability', exposing the fundamental weakness of current AI post-training agents. For teams working on autonomous agents and AI self-improvement, this is a must-read, as it signals that simply adding data and compute will not solve strategic rigidity.

## At a glance

This paper investigates the ability of LLM agents to post-train LLMs end-to-end, arguing that this process conflates execution-level capability with strategy-level capability. By analyzing a large corpus of public post-training trajectories, the authors find that the agent's training strategy is locked in at the very beginning, with the remaining budget spent on local adjustments. Experiments show that an experience-driven scaffold improves execution, human guidance can redirect the initial strategy, and additional inference compute helps only on simpler tasks. The conclusion is that agents lack a mechanism for spontaneously reevaluating their strategy during execution.

## Key points

- The agent's training strategy is locked in early, with only local adjustments afterward
- An experience-driven scaffold improves execution (GSM8K +12.6, HumanEval +40.8) but leaves strategy static
- Human guidance can redirect the initial strategy, but the agent falls back into local adjustment once training starts
- Additional inference compute helps on easier tasks but yields almost no gain on the hardest one

## Full text

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.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_4d511bc987ec847f717107ba