Lakshya A Agrawal
@lakshyaaagrawal.bsky.social
PhD @ucberkeleyofficial.bsky.social | Past: AI4Code Research Fellow @msftresearch.bsky.social | Summer @EPFL Scholar, CS and Applied Maths @IIITDelhi | Hobbyist Saxophonist
https://lakshyaaagrawal.github.io
Maintainer of https://aka.ms/multilspy
https://lakshyaaagrawal.github.io
Maintainer of https://aka.ms/multilspy
Reposted by Lakshya A Agrawal
Stop what you are doing and try out GEPA now!
"GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning" presents such elegant ideas by a collection of amazing researchers!
Here is a tldr of how it works:
"GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning" presents such elegant ideas by a collection of amazing researchers!
Here is a tldr of how it works:
October 21, 2025 at 3:03 PM
Stop what you are doing and try out GEPA now!
"GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning" presents such elegant ideas by a collection of amazing researchers!
Here is a tldr of how it works:
"GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning" presents such elegant ideas by a collection of amazing researchers!
Here is a tldr of how it works:
Reposted by Lakshya A Agrawal
GEPA (Genetic-Pareto) is a sample-efficient prompt optimization method for compound AI systems that works by reflectively evolving prompts using natural language feedback instead of traditional scalar rewards.
October 21, 2025 at 3:03 PM
GEPA (Genetic-Pareto) is a sample-efficient prompt optimization method for compound AI systems that works by reflectively evolving prompts using natural language feedback instead of traditional scalar rewards.
Reposted by Lakshya A Agrawal
In each iteration, GEPA samples system rollouts (including reasoning traces, tool outputs, and any diagnostic text), reflects on them via an LLM to identify issues or propose improvements, and updates specific module prompts accordingly based on the feedback.
October 21, 2025 at 3:03 PM
In each iteration, GEPA samples system rollouts (including reasoning traces, tool outputs, and any diagnostic text), reflects on them via an LLM to identify issues or propose improvements, and updates specific module prompts accordingly based on the feedback.
Reposted by Lakshya A Agrawal
To ensure diversity and avoid local optima, GEPA maintains a pool of candidates and uses Pareto-based selection, which keeps all non-dominated strategies discovered so far and stochastically proposes new prompt variants, enabling robust generalization with far fewer rollouts than reinforcement
October 21, 2025 at 3:03 PM
To ensure diversity and avoid local optima, GEPA maintains a pool of candidates and uses Pareto-based selection, which keeps all non-dominated strategies discovered so far and stochastically proposes new prompt variants, enabling robust generalization with far fewer rollouts than reinforcement
Reposted by Lakshya A Agrawal
GEPA: prompt optimization can exceed RL performance
They used Qwen3-8B (which was not trained for math, coding, agency, etc.) and show that GEPA performed better than RL rollouts
paper: arxiv.org/abs/2507.19457
github: github.com/gepa-ai/gepa
DSPy docs: dspy.ai/api/optimize...
They used Qwen3-8B (which was not trained for math, coding, agency, etc.) and show that GEPA performed better than RL rollouts
paper: arxiv.org/abs/2507.19457
github: github.com/gepa-ai/gepa
DSPy docs: dspy.ai/api/optimize...
1. GEPA Overview - DSPy
The framework for programming—rather than prompting—language models.
dspy.ai
October 22, 2025 at 11:55 AM
GEPA: prompt optimization can exceed RL performance
They used Qwen3-8B (which was not trained for math, coding, agency, etc.) and show that GEPA performed better than RL rollouts
paper: arxiv.org/abs/2507.19457
github: github.com/gepa-ai/gepa
DSPy docs: dspy.ai/api/optimize...
They used Qwen3-8B (which was not trained for math, coding, agency, etc.) and show that GEPA performed better than RL rollouts
paper: arxiv.org/abs/2507.19457
github: github.com/gepa-ai/gepa
DSPy docs: dspy.ai/api/optimize...
Reposted by Lakshya A Agrawal
Automating Agentic Prompts: A new algorithm called GEPA, developed by researchers at UC Berkeley, Stanford, and other institutions, improves the performance of agentic systems by automatically refining their prompts.
October 23, 2025 at 7:30 PM
Automating Agentic Prompts: A new algorithm called GEPA, developed by researchers at UC Berkeley, Stanford, and other institutions, improves the performance of agentic systems by automatically refining their prompts.
Reposted by Lakshya A Agrawal
AGI is just around the corner!
I'm learning to use DSPy with GEPA (Genetic-Pareto) prompt optimization. In GEPA a larger "teacher" LLM adjusts the prompt for a smaller "student" LM to perform a specific task as well as possible. The teacher will try many different prompts and evaluate the […]
I'm learning to use DSPy with GEPA (Genetic-Pareto) prompt optimization. In GEPA a larger "teacher" LLM adjusts the prompt for a smaller "student" LM to perform a specific task as well as possible. The teacher will try many different prompts and evaluate the […]
Original post on sigmoid.social
sigmoid.social
September 30, 2025 at 6:56 AM
AGI is just around the corner!
I'm learning to use DSPy with GEPA (Genetic-Pareto) prompt optimization. In GEPA a larger "teacher" LLM adjusts the prompt for a smaller "student" LM to perform a specific task as well as possible. The teacher will try many different prompts and evaluate the […]
I'm learning to use DSPy with GEPA (Genetic-Pareto) prompt optimization. In GEPA a larger "teacher" LLM adjusts the prompt for a smaller "student" LM to perform a specific task as well as possible. The teacher will try many different prompts and evaluate the […]
Reposted by Lakshya A Agrawal
Reposted by Lakshya A Agrawal
Just what I was looking for. Thank you for sharing, looking forward to the read.
September 28, 2025 at 12:58 AM
Just what I was looking for. Thank you for sharing, looking forward to the read.
Reposted by Lakshya A Agrawal
propose and test prompt updates, and combine complementary lessons from the Pareto frontier of its own attempts.
arxiv.org/abs/2507.19457
arxiv.org/abs/2507.19457
GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning
Large language models (LLMs) are increasingly adapted to downstream tasks via reinforcement learning (RL) methods like Group Relative Policy Optimization (GRPO), which often require thousands of rollo...
arxiv.org
September 28, 2025 at 12:28 AM
propose and test prompt updates, and combine complementary lessons from the Pareto frontier of its own attempts.
arxiv.org/abs/2507.19457
arxiv.org/abs/2507.19457
Reposted by Lakshya A Agrawal
DSPy folks love GEPA, so here's a GEPA paper for anyone who wants to learn more.
Given any AI system containing one or more LLM prompts, GEPA samples system-level trajectories (e.g., reasoning, tool calls, and tool outputs) and reflects on them in natural language to diagnose problems,
Given any AI system containing one or more LLM prompts, GEPA samples system-level trajectories (e.g., reasoning, tool calls, and tool outputs) and reflects on them in natural language to diagnose problems,
September 28, 2025 at 12:28 AM
DSPy folks love GEPA, so here's a GEPA paper for anyone who wants to learn more.
Given any AI system containing one or more LLM prompts, GEPA samples system-level trajectories (e.g., reasoning, tool calls, and tool outputs) and reflects on them in natural language to diagnose problems,
Given any AI system containing one or more LLM prompts, GEPA samples system-level trajectories (e.g., reasoning, tool calls, and tool outputs) and reflects on them in natural language to diagnose problems,
Reposted by Lakshya A Agrawal
September 28, 2025 at 11:13 AM
Reposted by Lakshya A Agrawal
..make adapting large models more practical—especially when compute or data is limited. It’s like giving AI a way to learn from its own “thinking out loud,” turning natural language into a powerful tool for self-improvement.
Links:
Paper on arXiv: https://arxiv.org/abs/2507.19457 ..
(6/7)
Links:
Paper on arXiv: https://arxiv.org/abs/2507.19457 ..
(6/7)
September 28, 2025 at 11:13 AM
..make adapting large models more practical—especially when compute or data is limited. It’s like giving AI a way to learn from its own “thinking out loud,” turning natural language into a powerful tool for self-improvement.
Links:
Paper on arXiv: https://arxiv.org/abs/2507.19457 ..
(6/7)
Links:
Paper on arXiv: https://arxiv.org/abs/2507.19457 ..
(6/7)
Reposted by Lakshya A Agrawal
..code on the fly.
What’s cool here is the shift from treating AI tuning as a blind search for a higher score to a reflective process that leverages the AI’s native strength: language. By evolving prompts through thoughtful reflections, GEPA unlocks smarter, faster learning that could..
(5/7)
What’s cool here is the shift from treating AI tuning as a blind search for a higher score to a reflective process that leverages the AI’s native strength: language. By evolving prompts through thoughtful reflections, GEPA unlocks smarter, faster learning that could..
(5/7)
September 28, 2025 at 11:13 AM
..code on the fly.
What’s cool here is the shift from treating AI tuning as a blind search for a higher score to a reflective process that leverages the AI’s native strength: language. By evolving prompts through thoughtful reflections, GEPA unlocks smarter, faster learning that could..
(5/7)
What’s cool here is the shift from treating AI tuning as a blind search for a higher score to a reflective process that leverages the AI’s native strength: language. By evolving prompts through thoughtful reflections, GEPA unlocks smarter, faster learning that could..
(5/7)
Reposted by Lakshya A Agrawal
..fewer attempts than traditional reinforcement learning methods. On several tough tasks like multi-step question answering and instruction following, GEPA consistently outperforms both standard reinforcement learning and previous prompt optimizers. It even shows promise for optimizing..
(4/7)
(4/7)
September 28, 2025 at 11:13 AM
..fewer attempts than traditional reinforcement learning methods. On several tough tasks like multi-step question answering and instruction following, GEPA consistently outperforms both standard reinforcement learning and previous prompt optimizers. It even shows promise for optimizing..
(4/7)
(4/7)
Reposted by Lakshya A Agrawal
..strategies by mixing and matching what works best.
GEPA treats AI prompt tuning like a conversation with itself, iterating through generations of prompts that learn from detailed feedback written in words, not just numbers. This lets it learn much more efficiently—up to 35 times..
(3/7)
GEPA treats AI prompt tuning like a conversation with itself, iterating through generations of prompts that learn from detailed feedback written in words, not just numbers. This lets it learn much more efficiently—up to 35 times..
(3/7)
September 28, 2025 at 11:13 AM
..strategies by mixing and matching what works best.
GEPA treats AI prompt tuning like a conversation with itself, iterating through generations of prompts that learn from detailed feedback written in words, not just numbers. This lets it learn much more efficiently—up to 35 times..
(3/7)
GEPA treats AI prompt tuning like a conversation with itself, iterating through generations of prompts that learn from detailed feedback written in words, not just numbers. This lets it learn much more efficiently—up to 35 times..
(3/7)
Reposted by Lakshya A Agrawal
..what went wrong and how to fix it? That’s the idea behind a new approach called GEPA. Instead of relying solely on those sparse reward signals, GEPA has AI inspect its own attempts using natural language reflections. It diagnoses errors, proposes prompt fixes, and evolves smarter..
(2/7)
(2/7)
September 28, 2025 at 11:13 AM
..what went wrong and how to fix it? That’s the idea behind a new approach called GEPA. Instead of relying solely on those sparse reward signals, GEPA has AI inspect its own attempts using natural language reflections. It diagnoses errors, proposes prompt fixes, and evolves smarter..
(2/7)
(2/7)
Reposted by Lakshya A Agrawal
What if language itself could teach AI to get better, faster?
Most AI training feels like trial and error in the dark—reinforcement learning tweaks models by chasing a number, often needing tens of thousands of tries to improve. But what if the AI could actually *talk to itself* about..
(1/7)
Most AI training feels like trial and error in the dark—reinforcement learning tweaks models by chasing a number, often needing tens of thousands of tries to improve. But what if the AI could actually *talk to itself* about..
(1/7)
September 28, 2025 at 11:13 AM
What if language itself could teach AI to get better, faster?
Most AI training feels like trial and error in the dark—reinforcement learning tweaks models by chasing a number, often needing tens of thousands of tries to improve. But what if the AI could actually *talk to itself* about..
(1/7)
Most AI training feels like trial and error in the dark—reinforcement learning tweaks models by chasing a number, often needing tens of thousands of tries to improve. But what if the AI could actually *talk to itself* about..
(1/7)
Reposted by Lakshya A Agrawal
gepa 0.0.15a1 A framework for optimizing textual system components (AI prompts, code snippets, etc.) using LLM-based reflection and Pareto-efficient evolutionary search.
Origin | Interest | Match
Origin | Interest | Match
gepa
A framework for optimizing textual system components (AI prompts, code snippets, etc.) using LLM-based reflection and Pareto-efficient evolutionary search.
pypi.org
September 24, 2025 at 1:09 AM
Reposted by Lakshya A Agrawal
gepa 0.0.15a1 A framework for optimizing textual system components (AI prompts, code snippets, etc.) using LLM-based reflection and Pareto-efficient evolutionary search.
Interest | Match | Feed
Interest | Match | Feed
Origin
pypi.org
September 24, 2025 at 1:09 AM
Reposted by Lakshya A Agrawal
gepa 0.0.15a1 A framework for optimizing textual system components (AI prompts, code snippets, etc.) using LLM-based reflection and Pareto-efficient evolutionary search.
Interest | Match | Feed
Interest | Match | Feed
Origin
pypi.org
September 24, 2025 at 1:09 AM
Reposted by Lakshya A Agrawal
gepa 0.0.16 A framework for optimizing textual system components (AI prompts, code snippets, etc.) using LLM-based reflection and Pareto-efficient evolutionary search.
Interest | Match | Feed
Interest | Match | Feed
Origin
pypi.org
September 24, 2025 at 2:02 AM
Reposted by Lakshya A Agrawal
gepa 0.0.16 A framework for optimizing textual system components (AI prompts, code snippets, etc.) using LLM-based reflection and Pareto-efficient evolutionary search.
Interest | Match | Feed
Interest | Match | Feed
Origin
pypi.org
September 24, 2025 at 2:02 AM
Reposted by Lakshya A Agrawal
New research released today from Databricks shows how its GEPA (Generative Evolutionary Prompt Adaptation) technique improves prompt optimization by an order of magnitude.
venturebeat.com/ai/the-usd10...
venturebeat.com/ai/the-usd10...
venturebeat.com
September 25, 2025 at 9:56 PM
New research released today from Databricks shows how its GEPA (Generative Evolutionary Prompt Adaptation) technique improves prompt optimization by an order of magnitude.
venturebeat.com/ai/the-usd10...
venturebeat.com/ai/the-usd10...
Reposted by Lakshya A Agrawal
🚀 #GEPA: Automatic #Prompt Optimization by @databricksinc.bsky.social: gpt-oss-120b beats Claude Sonnet 4 (+3%) at ~20x lower cost. Completes with DSPy SIMBA/MIPROv2
📜 MIT lic
🔗 Link in first 💬⤵️
Repost 🔁 #AI #LLM #RAG #PromptEngineering #ContextEngineering
📜 MIT lic
🔗 Link in first 💬⤵️
Repost 🔁 #AI #LLM #RAG #PromptEngineering #ContextEngineering
September 27, 2025 at 1:11 PM
🚀 #GEPA: Automatic #Prompt Optimization by @databricksinc.bsky.social: gpt-oss-120b beats Claude Sonnet 4 (+3%) at ~20x lower cost. Completes with DSPy SIMBA/MIPROv2
📜 MIT lic
🔗 Link in first 💬⤵️
Repost 🔁 #AI #LLM #RAG #PromptEngineering #ContextEngineering
📜 MIT lic
🔗 Link in first 💬⤵️
Repost 🔁 #AI #LLM #RAG #PromptEngineering #ContextEngineering