Understanding the Challenges in Iterative Generative Optimization with LLMs

• URL: http://arxiv.org/abs/2603.23994v1
• Date: Wed, 25 Mar 2026 06:49:24 GMT
• Title: Understanding the Challenges in Iterative Generative Optimization with LLMs
• Abstract summary: We argue that to set up a learning loop, an engineer must make hidden'' design choices.<n>We investigate three factors that affect most applications: the starting artifact, the credit horizon for execution traces, and trials and errors into learning evidence.<n>We conclude that the lack of a simple, universal way to set up learning loops across domains is a major hurdle for productionization and adoption.
• Score: 19.536425405805957
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Generative optimization uses large language models (LLMs) to iteratively improve artifacts (such as code, workflows or prompts) using execution feedback. It is a promising approach to building self-improving agents, yet in practice remains brittle: despite active research, only 9% of surveyed agents used any automated optimization. We argue that this brittleness arises because, to set up a learning loop, an engineer must make ``hidden'' design choices: What can the optimizer edit and what is the "right" learning evidence to provide at each update? We investigate three factors that affect most applications: the starting artifact, the credit horizon for execution traces, and batching trials and errors into learning evidence. Through case studies in MLAgentBench, Atari, and BigBench Extra Hard, we find that these design decisions can determine whether generative optimization succeeds, yet they are rarely made explicit in prior work. Different starting artifacts determine which solutions are reachable in MLAgentBench, truncated traces can still improve Atari agents, and larger minibatches do not monotonically improve generalization on BBEH. We conclude that the lack of a simple, universal way to set up learning loops across domains is a major hurdle for productionization and adoption. We provide practical guidance for making these choices.

Related papers

• The Optimizer Is the Agent: Reasoning-Driven Search across Prompts, Programs, and ML Workflows [72.9051182568928]
  We present ReASearch, a unified framework for reasoning-driven optimization.<n>With a shared agent loop and domain-specific tools, ReASearch instantiates the exact same scaffold to optimize prompts, programs, and ML.<n>We observe that complex search behaviors emerge naturally from the agent's reasoning process.
  arXiv  Detail & Related papers  (2026-08-07T02:15:39Z)
• FrontierOR: Benchmarking LLMs' Capacity for Efficient Algorithm Design in Large-Scale Optimization [61.43300970020897]
  Large language models (LLMs) are increasingly used for optimization modeling and solver-code generation.<n>Existing benchmarks are limited to small or simplified examples far below real-world scale and complexity.<n>We introduce FrontierOR, among the first benchmarks to evaluate LLM-based efficient algorithm design for realistic large-scale optimization problems.
  arXiv  Detail & Related papers  (2026-05-24T20:10:42Z)
• MLS-Bench: A Holistic and Rigorous Assessment of AI Systems on Building Better AI [92.82821272460423]
  MLS-Bench is a benchmark for evaluating whether AI systems can invent generalizable and scalable ML methods.<n>We study the effects of test-time scaling, adaptive compute allocation, and context provision on agents' discovery performance.
  arXiv  Detail & Related papers  (2026-05-09T04:29:46Z)
• Execution-Verified Reinforcement Learning for Optimization Modeling [49.171122807323634]
  Execution-verified learning framework treats a mathematical programming solver as a deterministic, interactive verifier.<n>Experiments on NL4OPT, MAMO, IndustryOR, and OptiBench across Gurobi, OR-Tools, and COPT show that EVOM matches or outperforms process-supervised SFT.
  arXiv  Detail & Related papers  (2026-04-01T03:39:11Z)
• ML-Tool-Bench: Tool-Augmented Planning for ML Tasks [23.54937738755734]
  We introduce a benchmark for evaluating tool-augmented machine learning agents.<n>Our benchmark goes beyond traditional tool-use evaluation by incorporating an in-memory named object management.<n>Our approach improves over ReAct by 16.52 percentile positions, taking the median across all Kaggle challenges.
  arXiv  Detail & Related papers  (2025-11-29T23:59:40Z)
• Increasing LLM Coding Capabilities through Diverse Synthetic Coding Tasks [41.75017840131367]
  Large language models (LLMs) have shown impressive promise in code generation.<n>We present a scalable synthetic data generation pipeline that produces nearly 800k instruction-reasoning-code-test quadruplets.
  arXiv  Detail & Related papers  (2025-10-27T10:54:25Z)
• Tutoring LLM into a Better CUDA Optimizer [0.0]
  We focus on the capabilities of the most recent reasoning models to generate optimized code for predefined, well-known tasks.<n>Our objective is to determine which types of code optimizations and parallel patterns the LLMs can perform by themselves and whether they can be improved by tutoring.
  arXiv  Detail & Related papers  (2025-10-19T17:09:15Z)
• Do AI models help produce verified bug fixes? [62.985237003585674]
  Large Language Models are used to produce corrections to software bugs.<n>This paper investigates how programmers use Large Language Models to complement their own skills.<n>The results are a first step towards a proper role for AI and LLMs in providing guaranteed-correct fixes to program bugs.
  arXiv  Detail & Related papers  (2025-07-21T17:30:16Z)
• OPT-BENCH: Evaluating LLM Agent on Large-Scale Search Spaces Optimization Problems [19.586884180343038]
  OPT-BENCH is a benchmark designed to evaluate Large Language Models (LLMs) on large-scale search space optimization problems.<n> OPT-Agent emulates human reasoning when tackling complex problems by generating, validating, and iteratively improving solutions through historical feedback.
  arXiv  Detail & Related papers  (2025-06-12T14:46:41Z)
• Improving LLM Reasoning through Scaling Inference Computation with Collaborative Verification [52.095460362197336]
  Large language models (LLMs) struggle with consistent and accurate reasoning. LLMs are trained primarily on correct solutions, reducing their ability to detect and learn from errors. We propose a novel collaborative method integrating Chain-of-Thought (CoT) and Program-of-Thought (PoT) solutions for verification.
  arXiv  Detail & Related papers  (2024-10-05T05:21:48Z)
• Trace is the Next AutoDiff: Generative Optimization with Rich Feedback, Execution Traces, and LLMs [19.89948665187903]
  We study a class of optimization problems motivated by automating the design and update of AI systems like coding assistants, robots, and copilots. We provide a Python, Trace, that efficiently converts a workflow optimization problem into an OPTO instance using PyTorch-like syntax.
  arXiv  Detail & Related papers  (2024-06-23T21:05:31Z)
• MLAgentBench: Evaluating Language Agents on Machine Learning Experimentation [96.71370747681078]
  We introduce MLAgentBench, a suite of 13 tasks ranging from improving model performance on CIFAR-10 to recent research problems like BabyLM. For each task, an agent can perform actions like reading/writing files, executing code, and inspecting outputs. We benchmark agents based on Claude v1.0, Claude v2.1, Claude v3 Opus, GPT-4, GPT-4-turbo, Gemini-Pro, and Mixtral and find that a Claude v3 Opus agent is the best in terms of success rate.
  arXiv  Detail & Related papers  (2023-10-05T04:06:12Z)
• Large Language Models as Optimizers [106.52386531624532]
  We propose Optimization by PROmpting (OPRO), a simple and effective approach to leverage large language models (LLMs) as prompts. In each optimization step, the LLM generates new solutions from the prompt that contains previously generated solutions with their values. We demonstrate that the best prompts optimized by OPRO outperform human-designed prompts by up to 8% on GSM8K, and by up to 50% on Big-Bench Hard tasks.
  arXiv  Detail & Related papers  (2023-09-07T00:07:15Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.