SEER

Automated Survey Generation via Recursive Backtracking Multi-Deep-Research-Agent Systems

中文文档

SEER is an evidence-grounded multi-agent system for producing structured scientific surveys. Instead of treating a survey as a one-pass search-plan-write task, it recursively builds an outline tree, lets parent agents repair completed child subtrees, writes leaf sections in parallel with traceable evidence, and assembles the result into a citation-normalized LaTeX document.

Project framework. Recursive backtracking changes survey generation from a frozen, single-pass pipeline into a hierarchy of local research tasks with shared literature memory and parent-level structural adjudication.

🔔 News:

• 🥳 2026/07/30: We have released SEER, a recursive-backtracking framework for evidence-grounded scientific survey generation. 🔥
• 🥳 2026/07/30: We have released the SEER-based auto-survey-agent Codex skill for self-contained, evidence-grounded long-form literature surveys. 🔥

Highlights

• Recursive multi-agent planning: each outline node follows a ReAct-style search, browse, propose, critique, and finish loop.
• Backtracking structural repair: after child subtrees complete, the parent can promote, demote, move, merge, or remove nodes to resolve redundancy and hierarchy conflicts.
• Evidence-grounded parallel writing: leaf agents write disjoint sections with local evidence sets and citation provenance, then pass through citation normalization and quality checks.
• Task-oriented context management: scoped per-task context is paired with a shared PaperDB for papers, evidence, and concept references.
• End-to-end assembly: SEER supports outline generation, restructuring, drafting, multimodal markers, citation repair, and chunkwise LaTeX/PDF export.
• Plug-and-play Codex skill: 🧩 We released the bundled auto-survey-agent Codex skill, so the evidence-grounded long-form survey workflow can be installed and used directly in Codex.

Architecture

SEER has four cooperating layers: recursive outline agents, a parent adjudication loop, parallel draft agents, and final document assembly. The shared literature memory keeps validated knowledge reusable without carrying every intermediate trace into every task context.

Codex Skill

The latest self-contained auto-survey-agent skill is included with this repository. It provides domain-aware survey planning, PaperDB/BM25 retrieval helpers, artifact validation, multimodal figure rendering, PDF export, and quality gates for book-length literature surveys. It is packaged separately from the SEER runtime and contains no credentials.

Install or update the skill from the repository root in PowerShell:

New-Item -ItemType Directory -Force "$env:USERPROFILE\.codex\skills" | Out-Null
Copy-Item -Recurse -Force .\skills\auto-survey-agent "$env:USERPROFILE\.codex\skills"

Open a new Codex task and request a literature survey to make the installed skill available. To use the included utilities directly, initialize a run with the appropriate domain profile:

$skillDir = "$env:USERPROFILE\.codex\skills\auto-survey-agent"
python "$skillDir\scripts\skill_runner.py" init-run --topic "agent memory" --domain-profile computing-engineering --out-dir .\survey_runs
python "$skillDir\scripts\skill_runner.py" bm25 --papers-jsonl .\survey_runs\<run>\papers.jsonl --query "agent memory retrieval reflection" --top-k 12
python "$skillDir\scripts\skill_runner.py" quality-gates --run-dir .\survey_runs\<run>

Available domain profiles are natural-sciences, health-medicine, social-sciences, humanities, computing-engineering, interdisciplinary, and auto. After assembling survey-final.md, install the optional rendering dependencies and produce a marker-free Markdown report and PDF:

python -m pip install matplotlib reportlab pypdf
python "$skillDir\scripts\skill_runner.py" render-report --run-dir .\survey_runs\<run>
python "$skillDir\scripts\skill_runner.py" check-report --report .\survey_runs\<run>\export\survey-rendered.md --require-rendered

Repository Layout

seer/
  agents/       Recursive outline, restructuring, and leaf-writing agents
  core/         Runtime configuration, environment loading, and token accounting
  pipeline/     CLI orchestration, drafting, citations, artifacts, and LaTeX export
  prompts/      Prompt templates and language policies
  retrieval/    PaperDB, retrieval, and search adapters
  schemas/      Shared outline data structures
  tools/        Web/paper search, browsing, image, and multimodal tools
  utils/        LLM clients plus browsing and Markdown helpers
  resources/    Packaged LaTeX templates
assets/
  Figures/      Paper figures in PDF and README-ready PNG formats
  SEER_camera.pdf
run.py          Backward-compatible checkout entry point

Installation

SEER requires Python 3.10 or newer and uses OpenAI-compatible model endpoints.

python -m venv .venv
source .venv/bin/activate
pip install -e .

For a non-editable checkout, install the same runtime dependencies with pip install -r requirements.txt.

Configuration

Create a local .env from .env.example and set the clients that your deployment uses. No credentials are committed to this repository.

SEARCH_API_KEY=...
SEARCH_BASE_URL=https://your-openai-compatible-endpoint/v1
SEARCH_MODEL=your-search-model
WRITER_API_KEY=...
WRITER_BASE_URL=https://your-openai-compatible-endpoint/v1
WRITER_MODEL=your-writer-model

SEARCH_* configures the outline/search client and WRITER_* configures section drafting. CITATION_JUDGE_* and PAPER_SEARCH_OPENAI_* are optional overrides for citation grounding and page summarization. Search adapters can use Jina, Serper, visit-pages, or a DeepXiv-compatible endpoint; set only the corresponding variables in .env.example.

Run

Generate an outline:

python -m seer --stage outline --topic "agent memory" --max-depth 3

Run the complete outline, draft, and LaTeX pipeline:

python -m seer --stage pipeline --topic "agent memory" --max-depth 3

Resume drafting from an existing run, or export its LaTeX/PDF artifacts:

python -m seer --stage draft --outline-result outputs/auto_outline_runs/<run_dir>
python -m seer --stage latex --outline-result outputs/auto_outline_runs/<run_dir>

python run.py ... remains supported for existing scripts. The installed seer console command is also available after pip install -e ..

Citation

Please cite the accompanying paper by title:

SEER: Automated Survey Generation via Recursive Backtracking
Multi-Deep-Research-Agent Systems.

License

Unless a file states otherwise, the source code and packaged skills in this repository are licensed under the Apache License 2.0. The accompanying paper and figures are included for reference; their copyright and reuse remain subject to their authors and any applicable publication terms.

Release Notes

This repository intentionally contains source code, packaged prompts and LaTeX resources, the paper, and its figures only. Runtime credentials, generated outputs, logs, benchmarks, evaluation code, tests, and experimental artifacts are excluded.