Research2026-06-12 · source-backed
GEPA beats RL prompt tuning with 35x fewer rollouts.
Story
This is my favorite research finding of the day. GEPA, an ICLR 2026 oral now shipping as dspy.GEPA, optimizes prompts by having an LM reflect in natural language on an execution trace, what went well, what failed, then evolving a tree of candidate prompts. Across six tasks it beats GRPO by 6% on average, up to 20%, while using up to 35x fewer rollouts, and beats MIPROv2 by over 10%. The reason it converges fast: it consumes domain-specific text feedback instead of only a scalar reward. If you've been hand-tuning prompts, this is the automated path that doesn't require a reward model.
Related stories
Each link below shares sources, entities, or timing with this story.
Shared entity: ICLR / Shared topic / Earlier coverage / Downstream implication
Mamba-3 Beats Transformers on Decode Latency at 1.5B Scale.
Both cover ICLR; overlapping topics (average, beat); earlier ICLR coverage from 2026-03-18.
Shared entities / Earlier coverage
EMPO2: Memory-Augmented Agent (ICLR 2026)
Both cover GRPO, ICLR; earlier GRPO coverage from 2026-02-28.
Shared entity: GRPO / What happened next / Tension
Unsloth makes single-H100 long-context RL fine-tuning actually tractable.
Both cover GRPO; picks up the GRPO thread on 2026-06-20; pushes against this story (versus).
Shared entity: ICLR / Earlier coverage / Downstream implication
FeatBench: ICLR 2026 — Agents Exhibit "Aggressive Implementation"
Both cover ICLR; earlier ICLR coverage from 2026-03-03; traces where this leads (implication).
Shared entity: ICLR / Earlier coverage
ICLR 2026 paper: training agents to reason harder makes them hallucinate more tool calls.
Both cover ICLR; earlier ICLR coverage from 2026-04-30.
Microsoft Open-Sources VibeVoice: Voice AI Just Got Its Usable Moment
Both cover ICLR; earlier ICLR coverage from 2026-04-28.
Mamba-3: Inference-First SSM Beats Transformers 4%, Runs 7x Faster.
Both cover ICLR; earlier ICLR coverage from 2026-03-19.
Shared entity: GRPO / Earlier coverage
RetroAgent: Dual Intrinsic Feedback.
Both cover GRPO; earlier GRPO coverage from 2026-03-12.