ArXiv for AI: personas generate, review, and publish research papers autonomously

[joelteply]

Vision

Continuum personas don't just code — they do science. The Academy trains them, the tools let them experiment, and the collaboration system lets them peer-review. The missing piece: a pipeline to generate, iterate, and publish research papers.

Why This Matters

The system already has:

• Academy: structured learning with evaluation
• Plasticity compaction: novel technique, published model on HF, needs a paper (Research paper: Plasticity Compaction — benchmarks + HuggingFace publication #391)
• Collaborative training: multi-persona learning, needs writeup (Research paper: Academy Collaborative Training — full session data + HuggingFace publication #392)
• Synthetic citizens: the whole thesis, needs a master document (Research paper: Synthetic Citizens — master thesis on HuggingFace #393)
• Tool use: personas can read code, run experiments, analyze data
• Decision system: peer review via ranked voting

What's missing: a pipeline from "interesting finding" → "structured paper" → "peer review" → "arXiv submission"

The Pipeline

1. Discovery: Persona identifies an interesting pattern (training results, benchmark comparison, architectural insight)
2. Hypothesis: Formulates a testable claim ("MoE expert pruning preserves 95% of domain-specific capability at 1/8 the parameters")
3. Experiment: Runs benchmarks using jtag commands (academy-session, coding challenges, inference quality)
4. Analysis: Collects results, generates tables/charts
5. Draft: Generates LaTeX or Markdown paper with proper structure (abstract, intro, method, results, conclusion)
6. Peer Review: Other personas review via decision/propose — vote on claims, flag weak evidence, suggest experiments
7. Revision: Incorporate feedback, re-run experiments if needed
8. Publication: Push to arXiv via API, publish to HuggingFace as model card supplement

Immediate Papers (ready to write)

Paper 1: Plasticity Compaction (#391)

• Claim: Training-informed head pruning + utilization-aware mixed-precision quantization
• Evidence: continuum-ai/qwen2.5-coder-14b-compacted (27GB→8.9GB, 3x compression)
• Needs: HumanEval/MBPP benchmarks, comparison to GPTQ/AWQ/standard quantization

Paper 2: MoE Expert Surgery (#439 — NEW)

• Claim: Individual expert extraction from MoE models produces domain-specialist models at fraction of original size
• Evidence: (needs experiments on Qwen3.5-35B-A3B)
• Needs: Benchmarks showing extracted code expert vs full MoE vs dense equivalent

Paper 3: Synthetic Citizens (#393)

• Claim: Persistent AI personas with genome, academy, and self-improvement form a self-sustaining development ecosystem
• Evidence: This entire system — 14+ personas, tool use, collaborative voting, code review
• Needs: Longitudinal data on persona improvement over time

Beyond CS — The Bigger Vision

Continuum isn't limited to coding. The same architecture works for:

• Drug discovery: personas analyzing molecular structures, running docking simulations, reviewing results
• Geology/paleontology: analyzing fossil morphology data, comparing geological formations
• Physics: exploring simulation parameter spaces, reviewing theoretical derivations
• Biology: protein folding analysis, genomics pipeline orchestration
• Chemistry: reaction pathway optimization, materials science modeling

The personas ARE the researchers. The Academy trains them on domain knowledge. The tools let them run experiments. The collaboration system lets them peer-review. ArXiv is just the output format.

Technical Implementation

• research/draft command: generate structured paper from data + template
• research/review command: submit paper for peer review (via decision system)
• research/benchmark command: run standardized benchmarks with result capture
• LaTeX generation from Markdown (pandoc or custom)
• arXiv API integration for submission
• HuggingFace model card generation from paper results

Related

• Research paper: Plasticity Compaction — benchmarks + HuggingFace publication #391, Research paper: Academy Collaborative Training — full session data + HuggingFace publication #392, Research paper: Synthetic Citizens — master thesis on HuggingFace #393 (specific papers)
• MoE surgery: extract individual experts for targeted training + tiny deployment #439 (MoE surgery — paper-worthy technique)
• Personas autonomously code the system: propose → implement → test → PR #411 (self-improving system — the meta-paper)
• Academy pipeline (Academy: no full training session proven end-to-end #377) — generates the data papers are based on