Pith. sign in

REVIEW 25 cited by

Automatic Prompt Optimization with "Gradient Descent" and Beam Search

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2305.03495 v2 pith:AUOJIZBD submitted 2023-05-04 cs.CL cs.AIcs.LG

classification cs.CLcs.AIcs.LG
keywords promptgradientautomaticdatadescentoptimizationbeamediting
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
read the original abstract

Large Language Models (LLMs) have shown impressive performance as general purpose agents, but their abilities remain highly dependent on prompts which are hand written with onerous trial-and-error effort. We propose a simple and nonparametric solution to this problem, Automatic Prompt Optimization (APO), which is inspired by numerical gradient descent to automatically improve prompts, assuming access to training data and an LLM API. The algorithm uses minibatches of data to form natural language "gradients" that criticize the current prompt. The gradients are then "propagated" into the prompt by editing the prompt in the opposite semantic direction of the gradient. These gradient descent steps are guided by a beam search and bandit selection procedure which significantly improves algorithmic efficiency. Preliminary results across three benchmark NLP tasks and the novel problem of LLM jailbreak detection suggest that Automatic Prompt Optimization can outperform prior prompt editing techniques and improve an initial prompt's performance by up to 31%, by using data to rewrite vague task descriptions into more precise annotation instructions.

Discussion (0). Sign in to comment.

Forward citations

Cited by 25 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Harness-R1: Learning to Edit Executable Runtime Harnesses from Agent Failure Trajectories

    cs.AI 2026-08 conditional novelty 7.0 of 10

    A dedicated harness-editor policy trained with RL on the realized outcomes of executable patches raises frozen-agent success by 9.3 points across WebShop, ALFWorld, and DBBench.

  2. ToolGrad: Efficient Tool-use Dataset Generation with Textual "Gradients"

    cs.CL 2025-08 unverdicted novelty 7.0 of 10

    ToolGrad inverts tool-use dataset generation: build valid tool-call chains first, synthesize queries second, yielding lower cost and near-100% pass rates.

  3. Predicting Empirical AI Research Outcomes with Language Models

    cs.AI 2025-06 conditional novelty 7.0 of 10

    A fine-tuned, retrieval-augmented language model predicts which of two AI research ideas will perform better empirically, beating human experts and frontier models on a new contamination-controlled benchmark.

  4. Asking Questions the Right Way: A Multi-Agent Conversational System for Prompt Formulation in Complex Task Resolution

    cs.MA 2026-08 conditional novelty 6.0 of 10

    An eight-agent question-asking system that front-loads intent clarification produced more complete prompts, higher-rated outputs, and single-turn task completion in a four-person pilot, with unstable effect sizes.

  5. MemoHarness: Agent Harnesses That Learn from Experience

    cs.AI 2026-07 conditional novelty 6.0 of 10

    MemoHarness learns from past executions to adapt an LLM agent's control layer per test case, improving task success over fixed harnesses on shell, code, and reasoning benchmarks.

  6. Heuristic Learning for Active Flow Control Using Coding Agents

    cs.LG 2026-07 conditional novelty 6.0 of 10

    Under equal simulation budgets, coding-agent search finds explicit flow controllers that match or beat the best DRL baseline on 10 of 13 active-flow-control benchmarks.

  7. MASPOB: Bandit-Based Prompt Optimization for Multi-Agent Systems with Graph Neural Networks

    cs.LG 2026-03 conditional novelty 6.0 of 10

    MASPOB combines a GNN surrogate, LinUCB-style uncertainty, and coordinate ascent to optimize prompts in fixed-topology multi-agent LLM systems, beating AFlow and MIPRO on average across six benchmarks.

  8. Visual Persuasion: What Influences Decisions of Vision-Language Models?

    cs.CV 2026-02 conditional novelty 6.0 of 10

    Naturalistic edits to background, lighting, and staging systematically shift the choices of frontier vision-language models, and a new optimization-plus-interpretation framework (CVPO) discovers and explains the visua...

  9. Knowing How to Edit: Reliable Evaluation Signals for Diagnosing and Optimizing Prompts at Query Level

    cs.AI 2025-11 reject novelty 6.0 of 10

    An execution-free evaluator that predicts prompt-quality metrics guides per-query prompt rewriting, but the reported consistent gains are not supported by the paper's own tables.

  10. CaPulse: Detecting Anomalies by Tuning in to the Causal Rhythms of Time Series

    cs.LG 2025-08 conditional novelty 6.0 of 10

    A single-pass instruction rewriter trained on MCTS-synthesized preference data improves LLM alignment win rates over the BPO baseline.

  11. TRPrompt: Bootstrapping Query-Aware Prompt Optimization from Textual Rewards

    cs.CL 2025-07 conditional novelty 6.0 of 10

    TRPrompt trains an 8B prompt model directly on natural-language textual rewards and reports the highest accuracies on GSMHard and MATH among the compared methods.

  12. Stabilizing Black-Box Prompt Optimization with Textual Regularization and Signal Aggregation

    cs.LG 2025-07 conditional novelty 6.0 of 10

    TRAS adds success-based textual regularization and Monte Carlo signal aggregation to black-box prompt optimization, improving accuracy and reducing instruction loss when moving prompts across models.

  13. Direct Behavior Optimization: Unlocking the Potential of Lightweight LLMs

    cs.CL 2025-06 conditional novelty 6.0 of 10

    A gradient-free Monte Carlo tree search over JSON key-step plans produces few-shot demonstrations that let LLaMA3-8B and LLaMA3.2-3B outperform GPT-3.5 on most of seven BIG-Bench Hard tasks.

  14. Visual Large Language Models Exhibit Human-Level Cognitive Flexibility in the Wisconsin Card Sorting Test

    cs.AI 2025-05 conditional novelty 6.0 of 10

    With chain-of-thought prompting and text inputs, GPT-4o, Gemini-1.5 Pro, and Claude-3.5 Sonnet reach or exceed human-level set-shifting on the WCST, but not with visual inputs or direct answers.

  15. LAD-VF: LLM-Automatic Differentiation Enables Fine-Tuning-Free Robot Planning from Formal Methods Feedback

    cs.RO 2025-09 conditional novelty 5.0 of 10

    LAD-VF iteratively rewrites LLM prompts using formal verification failure rates as a loss, lifting safety-spec compliance from about 60% to over 90% without model fine-tuning.

  16. Enhancing Reasoning Capabilities of Small Language Models with Blueprints and Prompt Template Search

    cs.LG 2025-06 conditional novelty 5.0 of 10

    Small language models achieve higher accuracy on math, coding, and logic benchmarks when their prompts contain LLM-generated reasoning blueprints and a per-model, per-task searched template.

  17. Cog-TiPRO: Iterative Prompt Refinement with LLMs to Detect Cognitive Decline via Longitudinal Voice Assistant Commands

    cs.CL 2025-05 conditional novelty 5.0 of 10

    Voice assistant commands from 15 older adults, analyzed with LLM-refined linguistic features plus acoustic and temporal modeling, detect mild cognitive impairment with 73.8% accuracy in a pilot study.

  18. WST: Weak-to-Strong Knowledge Transfer via Reinforcement Learning

    cs.LG 2025-08 reject novelty 4.0 of 10

    WST uses RL to train a small model to generate instructions that boost a larger model's benchmark performance, with reported gains of 98% on MATH-500 and 134% on HH-RLHF.

  19. Prompt Smart, Pay Less: Cost-Aware APO for Real-World Applications

    cs.LG 2025-07 conditional novelty 4.0 of 10

    APE-OPRO, a hybrid of APE and OPRO, achieves similar weighted F1 to OPRO at roughly 18% lower API cost on a 2,500-product commercial classification benchmark.

  20. BuildEvo: Designing Building Energy Consumption Forecasting Heuristics via LLM-driven Evolution

    cs.AI 2025-07 reject novelty 4.0 of 10

    BuildEvo evolves interpretable Python heuristics for building energy forecasting via LLM-driven search, reporting lower error on five BDG2 buildings than the baselines tested.

  21. Audit, Alignment, and Optimization of LM-Powered Subroutines with Application to Public Comment Processing

    cs.CL 2025-07 conditional novelty 4.0 of 10

    A framework for auditable, type-checked LM subroutines with bandit prompt optimization and self-critique is applied to NEPA public comment processing; the baseline evaluation shows high quote precision but low recall.

  22. SI-Agent: An Agentic Framework for Feedback-Driven Generation and Tuning of Human-Readable System Instructions for Large Language Models

    cs.AI 2025-07 reject novelty 4.0 of 10

    The paper proposes a multi-agent loop (instructor, follower, feedback) to auto-generate human-readable system prompts, claiming good benchmark performance and readability, but the supporting experiments are not reprod...

  23. Tournament of Prompts: Evolving LLM Instructions Through Structured Debates and Elo Ratings

    cs.AI 2025-05 conditional novelty 4.0 of 10

    DEEVO evolves better LLM prompts by debating outputs and selecting survivors with Elo ratings, without requiring labeled data or a hand-written fitness function.

  24. Research on Driving Scenario Technology Based on Multimodal Large Lauguage Model Optimization

    cs.CV 2025-05 conditional novelty 3.0 of 10

    On a private driving-scenario test set, a pipeline combining dynamic prompts, synthetic data, distillation with LoRA, and AWQ quantization raises average accuracy of a 7B vision-language model from 0.542 to 0.894.

  25. Inverse Reinforcement Learning Meets Large Language Model Post-Training: Basics, Advances, and Opportunities

    cs.LG 2025-07 unverdicted novelty 1.0 of 10

    A tutorial reviewing LLM alignment through the lens of inverse reinforcement learning, arguing that neural reward models learned from human data are central to post-training.

Pith tools