DSPy compiles short declarative programs into LM pipelines that self-optimize and outperform both standard few-shot prompting and expert-written chains on math, retrieval, and QA tasks.
hub Canonical reference
Large Language Models as Optimizers
Canonical reference. 80% of citing Pith papers cite this work as background.
45
Pith papers citing it
Background
80% of classified citations
abstract
Optimization is ubiquitous. While derivative-based algorithms have been powerful tools for various problems, the absence of gradient imposes challenges on many real-world applications. In this work, we propose Optimization by PROmpting (OPRO), a simple and effective approach to leverage large language models (LLMs) as optimizers, where the optimization task is described in natural language. In each optimization step, the LLM generates new solutions from the prompt that contains previously generated solutions with their values, then the new solutions are evaluated and added to the prompt for the next optimization step. We first showcase OPRO on linear regression and traveling salesman problems, then move on to our main application in prompt optimization, where the goal is to find instructions that maximize the task accuracy. With a variety of LLMs, 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. Code at https://github.com/google-deepmind/opro.
hub tools
citation-role summary
background 4
method 1
citation-polarity summary
representative citing papers
A meta-agent uses failure analysis to evolve a task agent's instructions for coordinating lexical, semantic, and multimodal retrievers, leading to up to 19.6 point gains on document QA benchmarks.
Open LLMs function as structural priors for MIMO controller tuning by proposing asymmetric structures on coupled plants, reaching better penalized cost with fewer evaluations than pure optimization or classical methods.
DICE formalizes multi-agent LLM coordination as discounted incomplete-information Markov games and introduces Heterogeneous Quantal Response Equilibrium (HQRE) to achieve unique stable equilibria with bounded regret, demonstrated via prompt-control and fine-tuning algorithms on eleven benchmarks.
PRISM automates continuous prompt creation, simulation-based testing, diagnosis, and repair for enterprise LLM agents, cutting authoring time to under 30 minutes while reaching 99% reliability and catching drift within 24 hours.
Fast-Slow Training uses context optimization as fast weights alongside parameter updates as slow weights to achieve up to 3x better sample efficiency, higher performance, and less catastrophic forgetting than standard RL in continual LLM learning.
TSCG compiles JSON tool schemas into token-efficient structured text, raising tool-use accuracy for small LLMs from 0% to 84.4% on benchmarks while cutting tokens by 52-57%.
AgentFlow uses a typed graph DSL covering roles, prompts, tools, topology and protocol plus a runtime-signal feedback loop to optimize multi-agent harnesses, reaching 84.3% on TerminalBench-2 and discovering ten new zero-days in Chrome including two critical sandbox escapes.
Massive activations are constant large values in LLMs that function as indispensable bias terms and concentrate attention probabilities on specific tokens.
LIVE uses language to generate task-centric vision embeddings at inference, reducing hallucinations by 34 points on MMVP, outperforming larger VLMs on VQA, and generalizing to unseen tasks.
Gemini 3.1-Pro with Ukrainian minimal-edits + few-shot prompting reaches F0.5=69.22 on Ukrainian GEC, closing over 90% of the gap to fine-tuned SOTA at 73.14.
MemoPilot trains memory updates for LLM agents via multi-turn GRPO on RPS and poker, achieving top Elo scores and outperforming baselines including DeepSeek-V3.2.
An LLM-based self-evolving agent discovers a traveling-wave controller with body-frame guidance and yaw feedback that generalizes to unseen targets for an underactuated fluid swimmer.
HiSME is a lightweight hierarchical meta-evolving approach that learns meta-skills from traces to refine both skills and evolving strategies, producing higher-quality skill libraries than pure skill evolving on agent benchmarks.
DEI shows a heterogeneous four-LLM ensemble achieving 124% higher QD-Score and 28% higher coverage than single-model baselines on Core War at equal compute budget.
Training-free prompt optimization methods, including five new education-focused ones, surpass the strongest RL-trained baseline across five conditions on two OOD suites while showing distinct teaching behavior patterns.
A universal LLM optimizer for text artifacts achieves SOTA results on six tasks including tripling ARC-AGI accuracy and cutting cloud costs by 40% via cross-task transfer and side information.
NCCE reframes context engineering as instance-level recommendation via bootstrapped anchor contexts and a co-evolving neural collaborative filtering router that assigns specialized contexts per input.
FitText embeds evolutionary retrieval of tool descriptions into the agent loop, yielding 2.7-10.6 point NDCG@5 gains on ToolRet and 26.7-point pass-rate gains on StableToolBench.
AloLab, an iterative meta-agent prompt optimizer, raises structured output accuracy for 7-9B models from 0% to 84-87% on GSM8K while preserving near-native inference speed.
ContraPrompt extracts optimization rules from dyadic differences in reasoning traces on identical inputs and organizes them into input-aware decision trees, outperforming GEPA on four benchmarks with gains up to 8.29 pp.
Frontier-Eng is a new benchmark for generative optimization in engineering where agents iteratively improve designs under fixed interaction budgets using executable verifiers, with top models like GPT 5.4 showing limited success.
Pioneer Agent automates the full lifecycle of adapting and continually improving small language models via diagnosis-driven data synthesis and regression-constrained retraining, delivering gains of 1.6-83.8 points on benchmarks and large lifts in production-style tasks.
Reflective Context Learning unifies context optimization for agents by recasting prior methods as instances of a shared learning problem and extending them with classical primitives such as batching, failure replay, and grouped rollouts, yielding improvements on AppWorld, BrowseComp+, and RewardBene
citing papers explorer
-
Hybrid Retriever Evolution for Multimodal Document Reasoning Agents
A meta-agent uses failure analysis to evolve a task agent's instructions for coordinating lexical, semantic, and multimodal retrievers, leading to up to 19.6 point gains on document QA benchmarks.
-
Structure from Reasoning, Numbers from Search: On-Premise Open LLMs as Structural Priors for Coupled MIMO Controller Tuning
Open LLMs function as structural priors for MIMO controller tuning by proposing asymmetric structures on coupled plants, reaching better penalized cost with fewer evaluations than pure optimization or classical methods.
-
DICE: Entropy-Regularized Equilibrium Selection for Stable Multi-Agent LLM Coordination
DICE formalizes multi-agent LLM coordination as discounted incomplete-information Markov games and introduces Heterogeneous Quantal Response Equilibrium (HQRE) to achieve unique stable equilibria with bounded regret, demonstrated via prompt-control and fine-tuning algorithms on eleven benchmarks.
-
PRISM: Prompt Reliability via Iterative Simulation and Monitoring for Enterprise Conversational AI
PRISM automates continuous prompt creation, simulation-based testing, diagnosis, and repair for enterprise LLM agents, cutting authoring time to under 30 minutes while reaching 99% reliability and catching drift within 24 hours.
-
Learning, Fast and Slow: Towards LLMs That Adapt Continually
Fast-Slow Training uses context optimization as fast weights alongside parameter updates as slow weights to achieve up to 3x better sample efficiency, higher performance, and less catastrophic forgetting than standard RL in continual LLM learning.
-
TSCG: Deterministic Tool-Schema Compilation for Agentic LLM Deployments
TSCG compiles JSON tool schemas into token-efficient structured text, raising tool-use accuracy for small LLMs from 0% to 84.4% on benchmarks while cutting tokens by 52-57%.
-
Synthesizing Multi-Agent Harnesses for Vulnerability Discovery
AgentFlow uses a typed graph DSL covering roles, prompts, tools, topology and protocol plus a runtime-signal feedback loop to optimize multi-agent harnesses, reaching 84.3% on TerminalBench-2 and discovering ten new zero-days in Chrome including two critical sandbox escapes.
-
Language-Instructed Vision Embeddings for Controllable and Generalizable Perception
LIVE uses language to generate task-centric vision embeddings at inference, reducing hallucinations by 34 points on MMVP, outperforming larger VLMs on VQA, and generalizing to unseen tasks.
-
How Far Can Prompting Go for Minimal-Edit Ukrainian Grammatical Error Correction?
Gemini 3.1-Pro with Ukrainian minimal-edits + few-shot prompting reaches F0.5=69.22 on Ukrainian GEC, closing over 90% of the gap to fine-tuned SOTA at 73.14.
-
From Player to Master: Enhancing Test-Time Learning of LLM Agents via Reinforcement Learning over Memory
MemoPilot trains memory updates for LLM agents via multi-turn GRPO on RPS and poker, achieving top Elo scores and outperforming baselines including DeepSeek-V3.2.
-
Self-Evolving Scientific Agent Discovers Generalizable Physically-Reasoned Fluid Control
An LLM-based self-evolving agent discovers a traveling-wave controller with body-frame guidance and yaw feedback that generalizes to unseen targets for an underactuated fluid swimmer.
-
You Live More Than Once: Towards Hierarchical Skill Meta-Evolving
HiSME is a lightweight hierarchical meta-evolving approach that learns meta-skills from traces to refine both skills and evolving strategies, producing higher-quality skill libraries than pure skill evolving on agent benchmarks.
-
DEI: Diversity in Evolutionary Inference for Quality-Diversity Search
DEI shows a heterogeneous four-LLM ensemble achieving 124% higher QD-Score and 28% higher coverage than single-model baselines on Core War at equal compute budget.
-
LLMs Are Already Good Tutors: Training-Free Prompt Optimization for Pedagogical Math Tutoring
Training-free prompt optimization methods, including five new education-focused ones, surpass the strongest RL-trained baseline across five conditions on two OOD suites while showing distinct teaching behavior patterns.
-
optimize_anything: A Universal API for Optimizing any Text Parameter
A universal LLM optimizer for text artifacts achieves SOTA results on six tasks including tripling ARC-AGI accuracy and cutting cloud costs by 40% via cross-task transfer and side information.
-
Contexting as Recommendation: Evolutionary Collaborative Filtering for Context Engineering
NCCE reframes context engineering as instance-level recommendation via bootstrapped anchor contexts and a co-evolving neural collaborative filtering router that assigns specialized contexts per input.
-
FitText: Evolving Agent Tool Ecologies via Memetic Retrieval
FitText embeds evolutionary retrieval of tool descriptions into the agent loop, yielding 2.7-10.6 point NDCG@5 gains on ToolRet and 26.7-point pass-rate gains on StableToolBench.
-
When Correct Isn't Usable: Improving Structured Output Reliability in Small Language Models
AloLab, an iterative meta-agent prompt optimizer, raises structured output accuracy for 7-9B models from 0% to 84-87% on GSM8K while preserving near-native inference speed.
-
ContraPrompt: Contrastive Prompt Optimization via Dyadic Reasoning Trace Analysis
ContraPrompt extracts optimization rules from dyadic differences in reasoning traces on identical inputs and organizes them into input-aware decision trees, outperforming GEPA on four benchmarks with gains up to 8.29 pp.
-
Frontier-Eng: Benchmarking Self-Evolving Agents on Real-World Engineering Tasks with Generative Optimization
Frontier-Eng is a new benchmark for generative optimization in engineering where agents iteratively improve designs under fixed interaction budgets using executable verifiers, with top models like GPT 5.4 showing limited success.
-
Pioneer Agent: Continual Improvement of Small Language Models in Production
Pioneer Agent automates the full lifecycle of adapting and continually improving small language models via diagnosis-driven data synthesis and regression-constrained retraining, delivering gains of 1.6-83.8 points on benchmarks and large lifts in production-style tasks.
-
Reflective Context Learning: Studying the Optimization Primitives of Context Space
Reflective Context Learning unifies context optimization for agents by recasting prior methods as instances of a shared learning problem and extending them with classical primitives such as batching, failure replay, and grouped rollouts, yielding improvements on AppWorld, BrowseComp+, and RewardBene
-
Self-Optimizing Multi-Agent Systems for Deep Research
Multi-agent deep research systems self-optimize prompts through self-play to match or outperform expert-crafted versions.
-
Two-Stage Prompt Optimization for Few-Shot Relation Extraction: From Reasoning-Guided Search to Gradient-Guided Refinement
A two-stage prompt optimization framework combining reasoning-guided search with gradient-guided refinement via GradPO reaches state-of-the-art on FS-TACRED using Qwen3-4B.
-
VirtualMLE: A Virtual ML Engineer that Optimizes Sequential Recommenders
VirtualMLE deploys an LLM agent with execution-reflection-memory to tune sequential recommenders, reaching competitive quality on Amazon benchmarks with fewer trials and transferring heuristics across datasets.
-
Exploring Autonomous Agentic Data Engineering for Model Specialization
LLMs functioning as autonomous agents can curate and optimize training data end-to-end, yielding up to 57.29% performance gains on specialized tasks via iterative adaptation guided by post-training metrics.
-
Prompt Optimization for LLM Code Generation via Reinforcement Learning
A PPO agent with hybrid actions and test-driven rewards optimizes prompts for code LLMs, raising strict Pass@1 scores on MBPP+, HumanEval+, and APPS over prior methods.
-
FORGE: Self-Evolving Agent Memory With No Weight Updates via Population Broadcast
FORGE is a staged population protocol that evolves prompt-injected memory (Rules, Examples, or Mixed) for ReAct agents via reflection and broadcast, yielding 1.7-7.7× gains over zero-shot and 29-72% over Reflexion on CybORG CAGE-2.
-
Prompting Policies for Multi-step Reasoning and Tool-Use in Black-box LLMs with Iterative Distillation of Experience
Iterative distillation of experience trains prompting policies that boost black-box LLM performance on reasoning and tool-use tasks from 55-74% to 90-91%.
-
Evolutionary Ensemble of Agents
EvE co-evolves code solvers and guidance states via synchronous races and Elo updates, discovering a rescale-then-interpolate mechanism that enables example-count generalization in ICON.
-
A Control Architecture for Training-Free Memory Use
A training-free control architecture with uncertainty-based routing, confidence-selective acceptance, and evidence-based memory governance improves arithmetic reasoning by +7 points on SVAMP and ASDiv benchmarks.
-
Agent Mentor: Framing Agent Knowledge through Semantic Trajectory Analysis
Agent Mentor analyzes semantic trajectories in agent logs to identify undesired behaviors and derives corrective prompt instructions, yielding measurable accuracy gains on benchmark tasks across three agent setups.
-
AI Training Manager: Bounded Closed-Loop Control of Adaptive Training Recipes
An LLM-based bounded controller adapts ML training parameters from structured telemetry to correct overfitting and exploration issues, shown on TinyStories and robotic RL tasks.
-
Towards Robust Argumentative Essay Understanding via TIDE: An Interactive Framework with Trial and Debate
TIDE integrates trial and debate mechanisms to improve criteria-based prompt optimization for argumentative essay tasks including automated scoring, component detection, and relation identification.
-
MedThink: Enhancing Diagnostic Accuracy in Small Models via Teacher-Guided Reasoning Correction
MedThink, a two-stage teacher-guided reasoning correction distillation framework, boosts small language models' medical diagnostic accuracy by up to 12.7% on benchmarks and achieves 56.4% on a gastroenterology dataset.
-
Neural at ArchEHR-QA 2026: One Method Fits All: Unified Prompt Optimization for Clinical QA over EHRs
A DSPy-based per-stage prompt optimization pipeline with self-consistency achieves second place among full participants in the ArchEHR-QA 2026 EHR QA shared task.
- Prompt Optimization Is a Coin Flip: Diagnosing When It Helps in Compound AI Systems