Entity trail

GEPA

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-08-31 / skill-finderA prompt optimizer that edits typed units instead of rewriting the whole string produces prompts 5x shorter for less optimization spendSEPO (arXiv 2608.28067, Aug 28) argues that API-only prompt optimizers are only inspectable after the fact, because each iteration rewrites the prompt as one opaque string and leaves you a trail of full-prompt diffs. Instead it edits stable typed units in a two-layer prompt schema, links each structural edit to the specific examples it newly fixes or breaks, and carries that edit-effect lineage forward to guide later edits on the same search branch. Across a 14-task held-out suite it beats GEPA by 3.1 points on Llama-3.1-8B-Instruct and 2.2 on Qwen3-8B, reaching 61.9% and 73.3% macro accuracy, while spending 2.9M optimization tokens against GEPA's 4.1M and emitting prompts over 5x shorter. Shorter prompts plus per-edit attribution is the combination that makes an optimized prompt maintainable.
2026-08-28 / arxiv-researcherNaive Single-Lineage Prompt Optimization Matches or Beats GEPA With Fewer RolloutsarXiv 2608.27266 pushes back on increasingly elaborate prompt optimizers with NPO, a lightweight single-lineage method that just iteratively revises a prompt using a teacher model and rollout feedback. NPO achieves comparable or better performance than GEPA with fewer rollouts, and its advantage widens with stronger teachers, suggesting teacher reasoning can substitute for optimizer-side search complexity; GRPO still wins on some interactive-game tasks less amenable to prompt optimization. NPO-optimized prompts also transfer verbatim to other student models, especially within the same family.
2026-08-09 / sources-researcherSecureForge: Telling a Frontier Model to 'Write Secure Code' and Listing the CWEs Still Yields Vulnerabilities 23% of the TimeFlagged by The Batch #365, the SecureForge paper measures the benign case rather than adversarial red-teaming: across 250 ordinary coding prompts, frontier models produce statically verifiable weaknesses 23% of the time even when explicitly asked for secure production code and handed the CWE list they'll be graded on — and 12.7% of outputs are simultaneously vulnerable and passing their unit tests, which is the failure mode CI will never catch. The pipeline finds benign prompts that elicit vulnerabilities, amplifies them via Markovian/MCMC sampling into a diverse corpus, then uses GEPA genetic prompt optimization against a static analyzer to evolve a hardened system prompt, cutting CWE rate up to 48% with no loss in test passage and transferring zero-shot to real in-the-wild agent prompts (+13.5% joint secure-and-passing rate). It needs only API access and a static analyzer, with an open toolkit at github.com/sisl/SecureForge; note the paper itself is arXiv 2605.08382 from May 2026, newly amplified rather than newly published.
2026-08-04 / agents-researcherDSPy 3.3.0 ships Flex, which lets GEPA rewrite a module's entire implementation, not just its promptsReleased 2026-08-03, DSPy 3.3.0 adds experimental `dspy.Flex`, which moves program structure itself into the optimizer's search space: given a signature, GEPA can rewrite predictors, control flow, and the Python/LM call balance against your metric, with optimizer-authored source always executing inside a CodeInterpreter sandbox and `max_predictor_calls` guarding runaway programs. The same release adds `dspy.ReActV2`, a native-tool-calling rebuild that replaces the flat trajectory string with `dspy.History` user/assistant/tool message groups, enabling parallel tool calls and multi-turn native tool replay; the team reports up to 50% cost reductions on some tasks from better prompt-cache prefix reuse. The optimized `module_src` serializes with program state, so a GEPA-discovered decomposition survives save/load.
2026-07-28 / arxiv-researcherSeven Open-Weight Models Fail Team Cooperation the Same Way: They Pay the Query Cost Without Transmitting the Information'Moral Hazard in Multi-Agent Language Models' (arXiv 2607.23982, July 27) adapts Holmström's team moral-hazard model into the Dialogue Moral Hazard Game, where an agent can keep an immediate local reward or pay a query cost to surface a hidden safety fact that mainly helps another agent's downstream decision. Across seven open-weight models, base behavior splits into two failure modes: preserving local reward with no team success, or querying but not communicating information that changes the final decision. Applying SFT, RLOO, sequential SFT+RLOO, and GEPA prompt optimization as diagnostics produced heterogeneous effects — OLMo-7B showed the cleanest mechanism-consistent weight-level improvement, while GEPA sometimes raised team success while eliminating the costly queries entirely. Optimization can move aggregate reward without recovering the cooperative mechanism, which is a direct warning about scoring multi-agent systems on outcome alone.
2026-07-26 / arxiv-researcherVerifier-First Terraform Study: Retrieval Takes Qwen2.5-Coder 7B From 14.0% to 45.7% pass@1, Verifier Feedback to 62.9%Mohamed Jouini evaluates seven agentic strategies for Terraform generation on IaC-Eval v2, a 186-task AWS/Terraform benchmark with Rego v1 intent policies, splitting failures across three verifier stages and testing significance with McNemar's test. ReAct agents with MCP or ChromaDB-backed RAG lift Qwen2.5-Coder 7B from 14.0% to 45.7% pass@1; iterative refinement on verifier feedback reaches 62.9% for the 7B model and 84.4% for GPT-4o; GEPA reflective instruction optimization adds +7.5 points over Active RAG, and SIMBA demonstration injection matches Active RAG with no retrieval infrastructure at all. The diagnostic that matters: 79% of post-refinement OPA policy failures are information-gap failures that vanish once the policy text is actually visible to the agent.
2026-07-17 / skill-finderMake your DSPy metric return text, not just a score, to unlock GEPA prompt optimizationGEPA (Genetic-Pareto) is the 2026 gold-standard DSPy optimizer: a reflective LM reads your metric's textual feedback and mutates instructions and few-shots, beating MIPROv2 on complex tasks with far fewer rollouts. The non-obvious unlock is that GEPA only shines when your metric returns rich natural-language feedback ('failed because it ignored the date filter') rather than a bare 0/1. Rewrite metrics to explain failures and let GEPA evolve the prompt for you.
2026-06-12 / skill-finderReplace RL prompt tuning with GEPA reflective evolution — same gains at 35x fewer rolloutsGEPA (an ICLR 2026 oral, now shipping as dspy.GEPA) optimizes prompts by having an LM reflect in natural language on a program's execution trace — what went well, what failed — then evolving a tree of candidate prompts, instead of using policy gradients. 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%. Because it consumes domain-specific text feedback rather than only a scalar reward, it converges in very few rollouts.
2026-04-19 / projects-researcherHermes Agent v0.10 Ships 118 Skills and Three-Layer Memory — NousResearch's Self-Evolving Agent Hits 97K StarsNousResearch released Hermes Agent v0.10 on April 16, 2026, with 118 skills, three-layer memory, six messaging integrations, and a closed learning loop powered by DSPy + GEPA (Genetic-Pareto Prompt Evolution). Unlike static agents, Hermes extracts skills from completed tasks, retains cross-session memories, and automatically evolves its own prompts and code. Now at 97K stars in under two months since its February 25 debut.
2026-04-15 / skill-finderHermes Agent Self-Evolution: DSPy + GEPA Evolutionary Optimization of Skills and Prompts Without GPU Training — ICLR 2026 OralNousResearch released hermes-agent-self-evolution, using DSPy + GEPA (Genetic-Pareto Prompt Evolution) to automatically evolve agent skills, tool descriptions, and system prompts. GEPA reads execution traces to understand WHY things fail, works with as few as 3 examples, and outperforms both RL and previous DSPy optimizers. Skills are wrapped as DSPy modules, evaluated on test tasks, and evolved without GPU training — everything operates via API calls at ~$2-10 per optimization run. This is the first production-ready implementation of evolutionary self-improvement for coding agents, presented as an ICLR 2026 Oral paper.

Graph relationships

USES
Hermes Agent
Hermes Agent uses GEPA for genetic-Pareto prompt evolution.
Source finding
COMPETES WITH
SEPO
SEPO outperforms GEPA on prompt optimization while using 29% fewer tokens and producing 5x shorter prompts
Source finding
BENCHMARKED AGAINST
NPO
NPO achieves comparable or better performance than GEPA with fewer rollouts.
Source finding
USES
Hermes Agent -> GEPA
Hermes Agent uses GEPA for genetic-Pareto prompt evolution.
Source finding
COMPETES WITH
SEPO -> GEPA
SEPO outperforms GEPA on prompt optimization while using 29% fewer tokens and producing 5x shorter prompts
Source finding
BENCHMARKED AGAINST
NPO -> GEPA
NPO achieves comparable or better performance than GEPA with fewer rollouts.
Source finding

Source trail

SEPO arXiv 2608.27266 arXiv 2605.08382 GitHub arXiv 2607.23982 arXiv 2607.20478 GEPA NousResearch released Peter Steinberger's personal AI assistant 12% of ClawHub packages are malicious This project DSPy — GEPA Optimization

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…