💡 The Core Idea
A self-evolving agentic loop where an LLM iteratively rewrites its output guided solely by external quality scores — no access to model internals needed.
The oracle only needs to return a score. It does not need to be differentiable, open-source, or even documented internally. This makes the loop applicable to grammar checkers, readability formulas, factuality verifiers, and AI detectors — all with the same code, zero modification.
⚡ Why Prompt-Only? The Generalization Argument
Existing gradient-based and token-level methods are fundamentally tethered to a single oracle type. Here is why.
| Oracle Type | Differentiable? | Gradient Methods Work? | Our Loop Works? |
|---|---|---|---|
| 🤖 AI Detection | Only with proxy | ✓ with retraining | ✓ |
| 📖 Readability Flesch-Kincaid, etc. |
No — rule-based formula | ✗ cannot apply | ✓ |
| ✍️ Grammar LanguageTool, etc. |
No — rule-based | ✗ cannot apply | ✓ |
| 🔍 Factuality | No — LLM-as-judge | ✗ prohibitively expensive | ✓ |
| ⚖️ Bias / Toxicity | Marginally | Requires model access | ✓ |
Prompt-only is not a simplified substitute — it is the only general-purpose architecture that works across the full space of real-world quality oracles.
🔬 Connection to Automated Scientific Discovery
This is an instance of a paradigm shift happening across AI: using agentic loops with external feedback signals when gradient flow is unavailable.
When the evaluation function is non-differentiable — which is the norm, not the exception, in real-world settings — prompt-based feedback loops are the architecturally natural choice. This work applies that paradigm to text quality optimization and is the first to study it systematically across multiple oracle types.
❓ Research Questions
- RQ1 Does a prompt-only agentic loop successfully reduce AI detection scores without gradient access or training?
- RQ2 What are the convergence dynamics — how many iterations are needed, does quality degrade as scores improve?
- RQ3 Does the same unmodified loop successfully optimize text against non-detection oracles (readability, grammar, toxicity)?
- RQ4 Can the loop simultaneously satisfy multiple quality constraints using a combined feedback signal?
- RQ5 Does loop performance generalize across text domains (news, essays, Wikipedia, medical)?
- RQ6 What is the performance cost of oracle-agnosticism vs. gradient methods on the detection task specifically?
🧪 Experiment Plan
Eight progressive experiments, from minimal to comprehensive. Each builds on the previous. Run E1 first — everything depends on it.
🎯 Result-Agnostic Design
The study is designed to be publishable regardless of outcome direction. The contribution is the systematic study itself.
💻 What the Loop Looks Like in Practice
All LLM calls use the Claude CLI — no API key setup, no model weights, no GPU required for the core loop.
The oracle is a black box that returns a score. Swap DetectionOracle for ReadabilityOracle or GrammarOracle — the loop code does not change.
📄 Paper Positioning
Framed as a text quality framework paper — not an AI detection evasion paper.
"Bypassing AI detection systems" — dual-use concerns, likely desk-rejected at top NLP venues.
"Oracle-agnostic text quality optimization" — in the tradition of Self-Refine, FunSearch, AI Scientist. AI detection is one application among several.
Target venues: EMNLP 2026, ACL 2026, or NAACL 2026 (main track).