Research map / Large language models / AI and society
Revolve: Optimizing AI Systems by Tracking Response Evolution in Textual Optimization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes REVOLVE, a textual-optimization method that augments TextGrad's first-order textual gradient with a similarity term over consecutive responses, framed as a simulated second-order update, and evaluates it on prompt, solution, and code optimization using LLM backends.
Problem
Automatic optimization of LLM-based systems via textual gradients relies on immediate feedback, which can stagnate or fluctuate when responses evolve slowly or irregularly.
Approach
Augment the TextGrad first-order textual gradient with a similarity term between consecutive responses (loosely analogous to a second derivative), and inject past-iteration responses into the optimizer and gradient prompts so the system makes gradual, measured updates.
Key results
- On solution optimization with Llama 3.1 8B Instruct, REVOLVE reaches 57.1% on MMLU-Machine Learning after final iteration, a 20.72% relative gain over TextGrad (47.3%).
- On LeetCode Hard with Llama 3.1 8B Instruct, REVOLVE achieves a completion rate of 0.31 versus 0.24 for TextGrad (29.17% relative improvement).
- On Object Counting with Llama 3.1 8B Instruct, REVOLVE reaches 83.0% accuracy versus 77.0% for TextGrad (7.8% relative improvement).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Track response evolution across iterations | TextGrad prompt optimization · Object Counting, Llama 3.1 8B Instruct | accuracy: 77.0 -> 83.0 | Helped |
| Track response evolution across iterations | TextGrad prompt optimization · GSM8K, Gemini 1.5 Pro | accuracy: 93.3 -> 93.0 ± 0.3% | Hurt |
| Track response evolution across iterations | TextGrad prompt optimization · GSM8K, Llama 3.1 8B Instruct | accuracy: 84.6 -> 84.6 ± 0.0% | No clear effect |
| Track response evolution across iterations | TextGrad solution optimization · Google-proof QA, Llama 3.1 8B Instruct | accuracy: 23.7 -> 28.3 ± 0.4% | Helped |
| Track response evolution across iterations | TextGrad solution optimization · MMLU-Machine Learning, Llama 3.1 8B Instruct | accuracy: 47.3 -> 57.1 ± 0.6% | Helped |
| Track response evolution across iterations | TextGrad solution optimization · MMLU-College Physics, Llama 3.1 8B Instruct | accuracy: 60.3 -> 66.4 ± 0.5% | Helped |
| Track response evolution across iterations | TextGrad code optimization · LeetCode Hard, Llama 3.1 8B Instruct | completion rate: 0.24 ± 0.005 -> 0.31 ± 0.006 | Helped |
| Use a prompt variant instructing more diverse responses instead of the REVOLVE prompt (ablation) | Prompt optimization · Object Counting | accuracy: 77.0 -> 80.0 | Helped |
| Use the REVOLVE prompt rather than the diversity-focused variant (ablation) | Prompt optimization · Object Counting | accuracy: 80.0 -> 83.0 | Helped |
Surprising findings
- TextGrad performance deteriorates below the initial state on MMLU benchmarks, which the authors attribute to first-order instability.
- Gemini 1.5 Pro with GPT-4o as optimizer shows no gain (REVOLVE ties or trails TextGrad), which the authors attribute to the optimizer being weaker than the target model.
Experimental setup
Datasets: Big Bench Hard Object Counting GSM8K GPQA Diamond MMLU Machine Learning MMLU College Physics LeetCode Hard
Benchmarks: Big Bench Hard Object Counting GSM8K GPQA Diamond MMLU Machine Learning MMLU College Physics LeetCode Hard
Metrics: string-based exact match accuracy completion rate runtime per iteration GPU memory usage
Baselines: Zero-shot CoT TextGrad M-TextGrad Reflexion Zero-shot
Builds on: TextGrad Momentum-enhanced TextGrad Chain-of-thought prompting Reflexion
Closest papers by meaning
- metaTextGrad: Automatically optimizing language model optimizers (NeurIPS 2025 · similarity 0.89)
- ThetaEvolve: Test-time Learning on Open Problems (ICML 2026 · similarity 0.88)
- Retroformer: Retrospective Large Language Agents with Policy Gradient Optimization (ICLR 2024 · similarity 0.88)
- REvolve: Reward Evolution with Large Language Models using Human Feedback (ICLR 2025 · similarity 0.88)
- ToolSearcher: Optimizing Tool Selection at Scale via Reinforcement Learning (NeurIPS 2026 · similarity 0.88)
- AgentGrad: Intervention-guided Prompt Optimization for Multi Agent Systems (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.