Related work and contribution boundaries

Status: targeted primary-source review, consolidated 2026-09-18. These methods motivate experiments; their published performance has not been reproduced here. Our evaluation framework is a project proposal, not a set of metrics claimed to originate in the papers below.

GEPA: reflective search over prompt programs

GEPA, paper v2, Section 3 and Algorithms 1–2, uses execution feedback to propose prompt changes and maintains candidate programs. Its method includes minibatch comparisons and selection using performance across individual evaluation instances. The Pareto mechanism is not simply a frontier between reward and monetary cost.

The relevant lesson is to evaluate proposed changes through execution, retain alternatives and account for search effort. GEPA does not supply a universal numerical measure of a prompt edit's semantic quality or a literal text derivative. Our action-flip and distribution probes would help inspect how a Jev edit changes behavior; online return remains a separate endpoint.

The official adapter interface was inspected at the linked revision: evaluate, make_reflective_dataset and propose_new_texts are relevant integration points. We have not integrated or benchmarked this adapter. A future comparison must record deviations from the paper and library, evidence access, evaluation instances and total resource use.

TextGrad: textual feedback and validation

TextGrad, Sections 2 and 3.3, frames language feedback as textual gradients and uses it to revise text variables. Its prompt optimization example uses validation to decide whether an update improves results. The feedback is language, not a numerical derivative of Atari return.

For this project, separate the teacher's critique from the measured gate. A fluent diagnosis is a candidate explanation whose proposed intervention still needs a behavioral and online outcome test.

Reflexion and Voyager: experience and memory

Reflexion uses linguistic feedback and episodic memory to improve subsequent attempts without updating the underlying model's weights. It motivates a memory ablation, including the possibility that accumulated advice becomes misleading outside the conditions in which it was obtained.

Voyager studies a Minecraft agent with iterative feedback and a reusable skill library. Reusing code skills differs from revising Jev questions, but both raise testable questions about retaining useful experience and transferring it to new tasks. Neither paper alone establishes our Atari optimizer's effectiveness.

RLPrompt: reward-driven discrete prompt optimization

RLPrompt optimizes discrete prompts using a reward-driven learned policy. It is relevant prior art for prompts as optimization variables. Its parameter-training mechanism differs from our initial fixed-weight teacher proposing edits to an external question program. Calling both approaches "reinforcement learning" does not make their update rules equivalent.

TypeSafe: question-feature discovery

TypeSafe's autoresearch feature-discovery cookbook describes adding, revising and dropping question-derived features evaluated in a supervised prediction workflow with CatBoost. It is direct prior art for treating Jev question definitions as searchable artifacts.

Our proposed distinction is the sequential control setting: actions change future observations and rewards, creating credit assignment, policy evaluation and exploration questions. That distinction motivates experiments; it is not by itself evidence of a novel algorithm or a better optimizer.

Intended contribution

Seek evidence about which teacher-optimizer components produce repeatable, resource-efficient question improvements and under which conditions. Candidate contributions include a controlled operator/situation pattern, a useful credit or representation mechanism, transfer evidence or a reproducible negative finding.

A higher score after one hand-written revision cannot establish such a claim. The roadmap specifies the comparisons and evidence needed before turning these ideas into paper conclusions.