This commit is contained in:
Oleg Maslov
174
docs/research/PAPER-2-CONCEPT_gepa-evolution.md
Normal file
174
docs/research/PAPER-2-CONCEPT_gepa-evolution.md
Normal file
@@ -0,0 +1,174 @@
|
||||
# Paper 2 — Closed-Loop Prompt Evolution: Can Smaller Models Match Flagships?
|
||||
|
||||
**Status:** CONCEPT — section skeleton + key claims. Full write-up after v2 experiment ($200 budget approved).
|
||||
**Target:** arXiv cs.AI preprint → ICLR / NeurIPS workshop
|
||||
**Authors:** Marko Markovic (Egzakta Group)
|
||||
**Drafted:** 2026-04-15
|
||||
|
||||
---
|
||||
|
||||
## Thesis
|
||||
|
||||
A closed-loop prompt-evolution system — combining GEPA (Genetic-Pareto prompt optimization) with schema-level behavioral constraints — can produce prompts for smaller open-weight models (Gemma 4 31B) that match or exceed the performance of flagship commercial models (Opus 4.6) on domain-specific tasks, at 10-50x lower inference cost.
|
||||
|
||||
---
|
||||
|
||||
## Section Skeleton
|
||||
|
||||
### 1. Abstract (~250 words)
|
||||
- Problem: organizations pay flagship-model prices for tasks where a tuned smaller model suffices
|
||||
- Approach: closed-loop system that traces agent behavior → builds eval datasets → evolves prompts → gates deployments → deploys to production
|
||||
- v1 result: 108.8% C/A ratio (Gemma 4 + evolved prompt vs raw Opus 4.6, n=10, 4 blind judges)
|
||||
- v2 design: n=60, 3 domains, 4-judge multi-vendor pool, bootstrap CI, pre-committed negative-result publication
|
||||
- **Placeholder:** insert v2 numbers after experiment
|
||||
|
||||
### 2. Introduction
|
||||
- The cost problem: flagship models at $15-75/M tokens for tasks that don't require frontier reasoning
|
||||
- The prompt-engineering bottleneck: human prompt engineers are expensive and iterate slowly
|
||||
- Our approach: automated prompt evolution with production-grade safety gates
|
||||
- **Key insight:** the evolution system runs on production traces, not synthetic data — the prompts are evolved against real user behavior
|
||||
|
||||
### 3. Background and Related Work
|
||||
- **GEPA** (Agrawal et al., arXiv:2507.19457, ICLR 2026 Oral): Genetic-Pareto prompt evolution, beats RL (GRPO) by +6% avg / +20% max with ≤35x fewer rollouts, beats MIPROv2 by >10%. Integrated into DSPy 3.0.
|
||||
- **ACE** (Zhang et al., arXiv:2510.04618, Stanford/SambaNova): Agentic Context Engineering — closest public analog to our EvolveSchema approach for behavioral-spec mutation
|
||||
- **Reflection 70B** (Shumer, Sept 2024): cautionary tale — "small beats big" claims require extreme rigor
|
||||
- **DSPy** (Khattab et al.): programmatic prompt optimization framework
|
||||
- **TextGrad** (Yuksekgonul et al.): gradient-based prompt optimization
|
||||
- Our positioning: we're not a prompt optimizer — we're a closed-loop production system that happens to use prompt evolution as one component
|
||||
|
||||
### 4. System Architecture
|
||||
- **6-stage pipeline (ASCII diagram):**
|
||||
```
|
||||
Traces → Dataset → Compose → Gates → Deploy → Monitor
|
||||
↑ ↓
|
||||
└────────── Feedback Loop ───────────┘
|
||||
```
|
||||
- **TraceRecorder:** captures every agent turn (model, tokens, tools, human action)
|
||||
- **EvalDatasetBuilder:** builds train/test splits from production traces
|
||||
- **LLM-as-Judge:** multi-vendor blind judging with rubric
|
||||
- **GEPA + EvolveSchema composition:** GEPA mutates the prompt, EvolveSchema mutates behavioral rules, feedback separation ensures they don't interfere
|
||||
- **Constraint gates:** 4 categories (safety, quality, cost, behavioral) with configurable thresholds
|
||||
- **Evolution deploy:** atomic persona override + behavioral-spec override with backup/rollback
|
||||
- **Boot-time merge:** accepted overrides merge into active behavioral spec on server start
|
||||
|
||||
### 5. The Evolution Pipeline in Detail
|
||||
|
||||
#### 5.1 Trace Collection
|
||||
- Automatic via `wireAgentLoopCallbacks` — every tool call, every model response
|
||||
- Stored in `execution_traces` table with session scoping
|
||||
|
||||
#### 5.2 Dataset Construction
|
||||
- Group traces by tool patterns (e.g., "save_memory → search_memory" sequences)
|
||||
- Score by outcome quality (human approval rate, task completion)
|
||||
- Split: 50/50 train/test with reproducible seed
|
||||
|
||||
#### 5.3 GEPA Iterative Evolution
|
||||
- Population of prompt candidates
|
||||
- Pareto frontier across objectives (quality, cost, length)
|
||||
- Iteration cap: 500 iterations OR cost ceiling
|
||||
- Early-abort: plateau detection (50 consecutive iterations below threshold)
|
||||
|
||||
#### 5.4 EvolveSchema Integration
|
||||
- Behavioral-spec mutations: add/remove/modify rules
|
||||
- Feedback separation: GEPA feedback on prompt quality vs EvolveSchema feedback on behavioral compliance
|
||||
- Compose: merge prompt mutations and schema mutations into a single candidate
|
||||
|
||||
#### 5.5 Constraint Gates
|
||||
- **Safety gate:** injection scan score must not degrade
|
||||
- **Quality gate:** judge scores must meet minimum threshold
|
||||
- **Cost gate:** token count must not exceed budget
|
||||
- **Behavioral gate:** all critical behavioral rules must pass
|
||||
|
||||
#### 5.6 Deploy and Monitor
|
||||
- Atomic file writes with backup/rollback
|
||||
- Hot-reload via event emission (`behavioral-spec:reloaded`)
|
||||
- Production monitoring via the same TraceRecorder
|
||||
|
||||
### 6. Experimental Design
|
||||
|
||||
#### 6.1 v1 (completed, preliminary)
|
||||
- **Setup:** 10 coder questions, Gemma 4 31B + Waggle-evolved prompt vs raw Opus 4.6
|
||||
- **Judges:** 4 blind judges (Opus, Sonnet, Haiku, GPT-4o)
|
||||
- **Result:** C/A ratio 108.8% — evolved-Gemma beat raw-Opus on per-judge mean
|
||||
- **Limitations:** n=10, one domain, eval=train, Opus-as-judge bias, no CI
|
||||
- **Placeholder:** v1 results table
|
||||
|
||||
#### 6.2 v2 (designed, budget approved, not yet run)
|
||||
- **Dataset:** 60 examples (30 train + 30 test), 3 domains (writer/analyst/researcher), 10 per domain per stratum
|
||||
- **Arms:**
|
||||
- A: raw Opus 4.6 (flagship reference)
|
||||
- B: Gemma 4 + human-engineered prompt (100 tokens)
|
||||
- C: Gemma 4 + GEPA-evolved prompt (from training run)
|
||||
- **Judges:** 4-judge multi-vendor pool:
|
||||
- Sonnet 4.6 (Anthropic)
|
||||
- Haiku 4.5 (Anthropic)
|
||||
- GPT-5 (OpenAI)
|
||||
- Gemini 2.5 Pro (Google)
|
||||
- **Rotation:** each test example judged by all 4, randomized arm-letter assignment
|
||||
- **Statistics:** bootstrap 95% CI + permutation test at alpha=0.05
|
||||
- **Hypotheses:**
|
||||
- H1: C/A >= 0.95 for 2+ of 3 domains
|
||||
- H2: C/A >= 1.00 for 1+ domain (replicates v1 headline)
|
||||
- H3: C/A < 0.90 for 2+ domains → publish negative (pre-committed)
|
||||
- **Budget:** $200 hard cap ($80 training + $80 evaluation + $40 buffer)
|
||||
- **Placeholder:** v2 results table, per-domain breakdown, judge agreement, CI intervals
|
||||
|
||||
### 7. Results (TO BE COMPLETED)
|
||||
- **Placeholder:** v2 primary results table
|
||||
- **Placeholder:** per-domain breakdown
|
||||
- **Placeholder:** inter-judge agreement (Krippendorff's alpha)
|
||||
- **Placeholder:** bootstrap CI visualization
|
||||
- **Placeholder:** training curve (score per iteration)
|
||||
- **Placeholder:** prompt length analysis (tokens added vs quality gained)
|
||||
- **Placeholder:** cost analysis (evolution cost vs inference savings)
|
||||
|
||||
### 8. Discussion
|
||||
- What the results mean for the cost-quality tradeoff
|
||||
- When evolution works (domain-specific, well-defined tasks) vs when it doesn't (open-ended reasoning)
|
||||
- The role of judge diversity in credibility
|
||||
- Production implications: how often to re-evolve, drift detection
|
||||
|
||||
### 9. Limitations
|
||||
- v1 sample size (n=10) is underpowered
|
||||
- Single-organization traces (Waggle users, not a general population)
|
||||
- GEPA is domain-specific — global prompts may not generalize across all task families
|
||||
- Judge-based evaluation inherits judge biases
|
||||
- No human evaluation (judges are all LLMs)
|
||||
- Evolution cost is amortized but non-trivial ($80-150 per run)
|
||||
|
||||
### 10. Reproducibility
|
||||
- Split seed committed to repo
|
||||
- GEPA config, judge prompts, dataset — all published
|
||||
- Raw results JSON available
|
||||
- Cost tracking via CostTracker
|
||||
|
||||
### 11. Conclusion
|
||||
- Closed-loop evolution is a viable alternative to paying flagship prices
|
||||
- The system is production-grade, not a research prototype
|
||||
- Pre-committed negative-result publication builds trust
|
||||
|
||||
---
|
||||
|
||||
## Key Claims (must be defensible with data)
|
||||
|
||||
| # | Claim | Evidence needed | Status |
|
||||
|---|-------|----------------|--------|
|
||||
| 1 | Evolved Gemma 4 matches Opus 4.6 (C/A >= 0.95) | v2 experiment (30 test, 4 judges) | PLANNED |
|
||||
| 2 | Evolution cost ($80-150) is amortized over thousands of inferences | Cost analysis | PLANNED |
|
||||
| 3 | Multi-vendor judges agree on ranking (alpha > 0.6) | Inter-judge agreement stats | PLANNED |
|
||||
| 4 | Constraint gates prevent quality regressions | Gate pass/fail rates from training | PLANNED |
|
||||
| 5 | System handles negative results gracefully | Pre-committed H3 publication | DESIGNED |
|
||||
| 6 | Evolution converges within 500 iterations | Training curve analysis | PLANNED |
|
||||
|
||||
---
|
||||
|
||||
## What Needs to Happen Before Full Write-Up
|
||||
|
||||
1. **Pre-flight checklist** — verify API keys for all 4 judges + Gemma endpoint
|
||||
2. **Write-up skeleton** — abstract + methods with [X.X%] placeholders (this document)
|
||||
3. **Run v2 training** — GEPA on 30 train examples ($80 budget)
|
||||
4. **Run v2 evaluation** — 30 test x 3 arms x 4 judges ($80 budget)
|
||||
5. **Statistical analysis** — bootstrap CI, permutation test, inter-judge agreement
|
||||
6. **Fill in results** — tables, charts, per-domain breakdown
|
||||
7. **External reviewer pass** — one ML peer reads methods + results
|
||||
8. **Publish** — arXiv preprint + reproducibility repo
|
||||
Reference in New Issue
Block a user