EvoQuant

An autonomous research agent for quantitative science.

EvoQuant is a self-evolving AI research agent specialized in quantitative investment research. It runs the full research loop autonomously — digesting research reports into structured knowledge, navigating a local papers library, generating and ranking research ideas, and executing real factor experiments on offline market data with rigorous IC-style evaluation.

Where general-purpose "AI scientist" frameworks target broad academic discovery, EvoQuant is purpose-built for the quant research workflow: alpha factor research, alpha generation methodology, and portfolio strategy research, with reproducible experiment runtimes and quantitative metrics (IC / ICIR / RANKIC / coverage) as first-class citizens.

🎯 Why EvoQuant?

Most autonomous research agents assume an open ecosystem — public papers with reference implementations on GitHub. Quantitative research rarely works that way: much of the field's methodology lives in broker research reports (研报) — unstructured PDFs that almost never ship with open-source code.

EvoQuant is designed for exactly this gap:

• 📖 Local knowledge extraction — Research report PDFs dropped into rawpaper/ become a structured, searchable private knowledge base (wiki/): every record carries title / source / strategy / method / experiment / result fields with inline evidence citations, extracted under strict no-fabrication rules.
• 🔧 From-scratch reproduction — When no reference code exists (the industry norm), EvoQuant reads reports at the L1 "able-to-reimplement" depth, re-implements the described method as an executable Research Artifact, and validates it on real offline market data with IC / ICIR / RANKIC metrics — rather than trusting the numbers printed in the PDF.
• 💡 Innovation on top — Anchor-first ideation: inherit ≥70% of an anchor report's method, contribute a focused ≤30% innovation delta, and let an ELO tournament (Final = Novelty + Relevance + Clarity − Difficulty) decide which ideas are worth running.

✨ Features

• 🤖 Multi-Agent Team — 6 sub-agents (plan, research, code, debug, analyze, write) working in concert.
• 🧠 Self-Evolving Memory — Observations auto-distilled each turn and self-linked into a knowledge graph that grows across sessions; research cycles feed ideation and experimentation memory.
• 🔬 Quant Research Pipeline — Report extraction → literature grounding → anchor-first ideation with ELO tournament ranking → experiment execution with IC metrics.
• 📊 Experiment Runtime — A self-contained executor for "Research Artifacts": offline dataset discovery, panel building, train/val/test splitting, and extensible metric registry.
• 🌐 Multi-Provider — Anthropic, OpenAI, Google, MiniMax, NVIDIA — one config to switch.
• 📱 Multi-Channel — CLI/TUI as the hub; Telegram, Slack, Feishu, WeChat, Discord and more — one agent session.
• 🖥️ WebUI — Workspace-panel web app via --ui webui.
• ⏰ Scheduled Tasks — Cron-style recurring research runs that operate unattended and report back.
• 🔌 MCP & Skills — Plug in MCP servers or install additional skills from GitHub on the fly.

🏗️ Architecture

EvoQuant is built on a DeepAgents / LangGraph core: a main agent orchestrates specialized sub-agents, middleware and tools around a persistent state graph.

flowchart TB
    subgraph Surfaces["🎛️ Surfaces"]
        CLI["CLI / TUI"]
        WEB["WebUI"]
        CH["Channels<br/>(Telegram · Slack · Feishu · …)"]
        CRON["Scheduler<br/>(cron-style tasks)"]
    end

    subgraph Core["🧠 Agent Core (LangGraph gateway)"]
        MAIN["Main Agent<br/>plan · orchestrate · summarize"]
        SUB["Sub-Agents<br/>research · code · debug · analyze · write"]
        MEM["Memory<br/>observation graph<br/>(distill · link · recall)"]
        MW["Middleware<br/>context editing · adaptive tools"]
    end

    subgraph Capabilities["⚙️ Capabilities"]
        SKILLS["Skills<br/>(SKILL.md knowledge packs)"]
        TOOLS["Tools + MCP"]
        LLM["LLM Providers<br/>(multi-provider registry)"]
    end

    subgraph Workspace["📁 Research Workspace (workdir)"]
        KB["knowledge base<br/>(markdown papers · JSONL)"]
        EXP["experiments/<br/>(panels · artifacts · results)"]
        CODE["code-repo/<br/>(offline data packages)"]
    end

    Surfaces --> Core
    MAIN --> SUB
    MAIN <--> MEM
    MW --> MAIN
    Core --> Capabilities
    SKILLS --> Workspace
    SUB --> Workspace

Key ideas

The quant research loop

flowchart LR
    A["📄 quant-paper-extractor<br/>PDF → markdown → JSONL"] --> B["📚 local-paper-navigator<br/>paper search · reading"]
    B --> C["💡 research-ideation<br/>literature tree · ELO tournament"]
    C --> D["🧪 quant-experiment-runtime<br/>panel · entry point · IC metrics"]
    D --> E["🧠 memory<br/>(what worked / what failed)"]
    E -.-> C

1. Ingest — quant-paper-extractor converts quant research report PDFs into structured JSONL records (strategy, method, experiment, result) for the local knowledge base.
2. Ground — local-paper-navigator searches the papers library by keyword/abstract/full-text, disambiguates queries, and reads papers with an L1/L2/L3 strategy.
3. Ideate — research-ideation builds challenge-insight trees, generates anchor-first ideas, refines them in persona-driven tracks, and ranks them with an ELO tournament (Final = Novelty + Relevance + Clarity − Difficulty).
4. Execute — quant-experiment-runtime discovers offline datasets under code-repo/, builds panels, runs a Research Artifact through its Python entry point, and evaluates IC / ICIR / RANKIC / coverage.
5. Evolve — outcomes feed persistent memory, so the next cycle starts from what worked and avoids known dead ends.

📄 Feeding the knowledge base

The papers library lives at the repo root papers/ — mounted at /papers/, served through the paper_search / paper_read / paper_section tools:

papers/
  raw/{paperId}.pdf       ← drop research report PDFs here (any filename)
  markdown/{paperId}.md   ← auto-created: full-text markdown per report
  cards/{paperId}.jsonl   ← auto-created: structured JSONL knowledge records
  context_brief.md / index.jsonl / manifest.jsonl

Three steps: drop PDFs into papers/raw/ → tell the agent 「入库」 → start a new session (the paper tools mount at agent startup). The CLI prints a hint at startup whenever papers/raw/ holds PDFs that have not been ingested yet.

• quant-paper-extractor runs the pipeline on 「入库」: PDF → markdown → cards → refresh, hash-keyed and incremental — re-running only processes new files.
• This local papers library is the primary knowledge source of the autonomous research loop: research-ideation grounds every idea in reports retrieved from it via local-paper-navigator — by design, generic web search is never used to find papers.

📦 Skills

All skills live under EvoQuant/skills/ and are self-contained (SKILL.md + references/ + assets/ + scripts/).

Quant Research Core

Research Workflow