LEMVO / THINKING RESEARCH AGENT

Lemvo

A research agent that thinks and acts.

From questions to findings to papers. Claude Code + Codex CLI.

Understand the question, investigate the literature, build methods, choose experiments, and learn what to do next. Research Orchestrator coordinates 114 skills through two nested loops: the inner loop implements and tests methods; the outer loop interprets findings and revises hypotheses and routes. Carry ideas into discoveries, then shape the paper with one shared narrative.

ChooseActLearnWrite

Start researching Watch the workflow

✳ Claude Code />_ Codex CLI One repository. Two CLI engines.

RESEARCH IN MOTION01 — ∞

research-lab / sessionILLUSTRATIVE

›Investigate this question, choose a route, and move it forward

⎇ research / new-hypothesis

research(init)
Understand the question · read evidence · form hypotheses
✓
research(protocol)
Choose experiments · state predictions · establish a baseline
✓
research(results)
Build the method · measure outcomes · record findings
✓
research(reflect)
Interpret results · revise the method · choose a direction
✓

findings.md · research that remembers▌

Experiments move results. Reflection moves direction.

114 / skills

A shared research skill library

2 / engines

Claude Code · Codex CLI

4 / directions

DEEPEN · BROADEN · PIVOT · CONCLUDE

1 + 1 / loops

Experiment + reflection

Counted from repository SKILL.md files · 2026-10-14View evidence ↗

01 / THE RESEARCH ORCHESTRATOR

Decide what comes next. Then make it happen.

Research Orchestrator chooses a route from your goal, existing assets, and project state, connecting literature, methods, experiments, and writing. The inner loop tests ideas; the outer loop makes sense of the results and turns findings into the next action.

RESEARCH ORCHESTRATORRead the workflow ↗

START / BOOTSTRAP

Frame the question. Start working.

GATE 0

Clarify the goal → evidence-backed directions

GATE 1

Define the scope → an actionable plan

Build on the question and assets you already have. Start a probe when ready.

OUTER LOOP / SYNTHESIS & DIRECTION

Synthesize→Ask WHY→New hypotheses

INNER LOOP / EXPERIMENT & VALIDATION

Protocol→Experiment→Measure→Learn

↶keep / revert / record

↳findings.mdMemory across sessions↺

OUTPUT / FINALIZE

Make discovery durable.

≡Evidence & synthesis

⌁Figures & progress reports

▤Paper drafts & archives

Deliver code, results, and findings. Develop a paper when the argument is ready.

Choose a route, not a rigid pipeline.

QUESTION → DISCOVERY

Start with a question or an idea

Literature & ideation → Build a method → Test hypotheses

SOTA → IMPROVEMENT

Build on a strong existing method

Select an implementation → Bound reproduction → Iterate in place

FORGE → AUTONOMY

Forge a package for standalone runs

Fix evaluation & budget → Build the package → Run autonomously

You set the research goal. Lemvo chooses how to experiment: iterating on an existing model or building a self-running experiment package to automate method changes, evaluation, and result tracking.

After the results, where next?

02 / ONE COORDINATED RESEARCH WORKFLOW

From understanding a question to making a discovery.

Literature reveals questions worth answering. Experiments turn ideas into methods. Writing connects findings into an argument. Shared research state and narrative plans keep the stages connected.

RESEARCH ORCHESTRATORCoordinated work · Shared research state

01Research & ideation 02Methods & experiments 03Writing & review

114 SKILLSResearch workflows and domain methods, called on demand

01 RESEARCH + IDEATION

Explore broadly. Read deeply. Think from evidence.

Map the field, then investigate the papers that matter. Turn consensus, disagreements, and open questions into candidate ideas, and use structured ideation and evaluation to find a question worth testing.

BREADTH

Map approaches, compare research directions, and identify unresolved gaps.

DEPTH

Use ArXiv MCP to retrieve and search full text. Keep versions and sources for the claims that matter.

IDEATION

Transfer ideas across fields, invert assumptions, and reframe problems. Turn feasible advances and original ideas into testable methods.

Explore the ideation methods ↗

FROM PAPER EVIDENCE TO A TESTABLE IDEA

01search_papersDiscover relevant work

02download_paperRetrieve text & version

03search_paper_textFind & check the claim

↳report.provenance.mdTRACEABLE

Evidence → candidate ideas → evaluationbrainstorming · creative-thinking · idea-evaluator

HigherFasterStrongerCheaperBroader

Deliver: a sourced research brief, candidate directions, and an actionable research question.

02 EXPERIMENT FORGE + AUTORESEARCH

Make every experiment answer a research question.

Experiments improve models, test hypotheses, and investigate failures. Research Orchestrator chooses a route for the question: build on an existing implementation, or use Experiment Forge to deliver a Karpathy-style autonomous task package. Autoresearch runs the unattended modify, measure, keep-or-revert inner loop.

CHOOSE EXPERIMENTS FOR THE QUESTION

Test hypothesesUnderstand mechanismsImprove models

Optimize accuracy, loss, or efficiency; use controls and diagnostics to distinguish explanations. Turn a diagnosed issue into a concrete method change, test it again, and retain the progress.

Explore the autonomous task specification ↗

OPTION / FORGEOptional standalone task package

▣prepare.pyLOCKED DATA + EVAL

⌘train.pyEDITABLE MODEL + TRAINING

≡program.mdGOAL, BUDGET & SCOPE

▤pyproject.tomlLOCKED DEPENDENCIES

↳results.tsvRUN LEDGER

INNER LOOP / AUTORESEARCHRun the autonomous loop

Baseline firstPropose & modifyTrain & evaluate

Keep improvements · Revert regressions

Time budgetRegression guardNoise rechecksGit record

Each trial leaves code, measurements, and decisions. The outer loop synthesizes them into findings and chooses whether to deepen, broaden, revise, or conclude.

03 PAPER NARRATIVE + PAPER PRODUCTION

A clear research story. A compelling paper.

Paper Narrative starts with your contribution and shapes one scientific story across the question, insight, method, and evidence. The abstract captures the discovery; the introduction builds the problem; sections and figures advance the argument; the conclusion answers the opening question and leaves a new understanding.

Research findings feed one shared narrative plan across the abstract, introduction, figures, and conclusion. As discoveries evolve, the story and manuscript stay aligned. Scientific figures, citation checks, and dual review carry the paper through to delivery.

Explore Paper Narrative ↗ Explore the paper-production pipeline ↗

01Core finding & argument structureresearch findings → argument

02Narrative, introduction & figure orderpaper-narrative → intro-drafter

03Abstract, body & conclusionshared narrative → paper-writing

04Evidence checks + writing reviewverifier + pre-submission-reviewer

SIX SCIENTIFIC NARRATIVE LENSES

Find the story in your contribution.

Six narrative routes, selected and combined for the research.

Root-cause scalpel

Move from a failure pattern to a mechanism and a targeted design.

Counter-intuitive reframing

Revisit a default assumption and establish a more useful perspective.

Theory illuminates practice

Turn formal results into a new understanding the reader can follow.

Benchmarks expose failure

Make overlooked problems measurable and interpretable.

Societal value

Connect technical choices and application value to a real need.

A unifying principle

Connect multiple problems and results through one common principle.

Leave the reader with what this research establishes—and why it matters.

⎇

Commit the protocol before running the experiment.

Keep protocols and results in separate commits: a durable record that the plan preceded the run. Failed experiments and negative results belong here, too.

protocol → commit → run

BUILT FOR CONTINUOUS RESEARCH

Research that keeps its thread—and moves forward.

Project memory, launch conditions, and cumulative budgets connect every turn. Turn preparation into a probe, diagnosis into a change, and results into the next action.

01 / REMEMBER

Continue from the progress you have

Save the question, findings, decisions, and next action in project files. Read the state and findings before continuing, keeping the existing contract, ledger, and best result.

state · findings · decisions

02 / START SMALL

Ready to run? Start small.

Launch the smallest real probe when the implementation runs, measurement is valid, and budget permits. Track preparation costs and specific blockers to move toward an observable result.

ready → probe → observation

03 / LEARN & REVISE

Turn diagnosis into a change

Fix measurement or implementation issues where they occur. Revise or replace an ineffective candidate and test again. Informative negative results help choose the next step.

diagnose → revise → retest

04 / STAY IN SCOPE

Preserve the budget and the work

Carry resource records through continuation, pivots, and handoffs. At an agreed boundary, explicitly continue, narrow, or conclude with code, results, and findings preserved.

continue · narrow · conclude

Research records connect experiments and writing: establish the findings, then shape a mature contribution into a paper.

Start with a clear question ↗

A SMALL QUESTION. THE WHOLE WORKFLOW.

用一个小问题，看懂双循环。

让小模型更准确地识别手写数字。跟随一次教学演示，看调研、实验、反思与写作如何接成完整流程。

RESEARCH ORCHESTRATOR

BOOTSTRAP

01调研与构思
02选择实验路线
03双循环实验
04写作与评审

OUTER LOOP下一步，该研究什么？

QUESTION

怎样让一个小模型更准确地识别手写数字？

INNER LOOP这个想法，怎样做得更好？

改模型train.py

训练与测量固定评估

保留或回滚keep / revert

先明确课题，再开始实验

WHY?

跳出单次实验，看清下一步。

综合实验记录、检查错误、修订假设，再回到内循环。

共享研究记忆findings.mdresults.tsv

01 / 14

BOOTSTRAP

从一个简单问题开始。

目标很清楚：让模型更准确地识别手写数字。先确认数据、算力预算与评价标准。

当前调用

research-orchestrator

本次测量 · 示意—

当前保留 · 示意—

READY

这一步留下research-question.md

流程进度01 / 14

内循环：构建与验证方法。修改、测量、保留进展，再验证。

外循环：更新研究方向。解释发现，把诊断变成下一步。

流程演示，指标为示意数据。

03 / A RESEARCH TOOLKIT THAT GOES DEEP

114 skills. A method for every stage.

From AlphaFold2 to DeepSpeed, 114 research skills span literature, experiments, writing, and specialist methods. Orchestrator calls on the right expertise as the project needs it.

12 curated topic groups · Open any skill to read its source

One library of 114 research skills, adapted for Claude Code and Codex CLI. Counted separately from claude/skills and codex/skills.

04 / TWO ENGINES. ONE RESEARCH MINDSET.

Your engine. A shared research mindset.

One Lemvo repository, with dedicated Claude Code and Codex CLI implementations. Choose your engine and bring the same research workflow to your terminal.

Explore Lemvo on GitHub · Star ↗

✳

LEMVO / CLAUDE CODE

Lemvofor Claude Code

114 skills and 4 specialist subagents, with native Claude Code collaboration. Choose guided or unattended research; keep the same project memory when you switch.

Install for Claude Code↓

LEMVO / CODEX CLI

Lemvofor Codex

The same 114 skills, adapted for Codex CLI. Choose guided or unattended research; the resident controller connects experiments, results, and the next round.

Install for Codex CLI↓

05 / YOUR NEXT QUESTION STARTS HERE

Your next discovery.
Starts in your terminal.

Install Lemvo, open your CLI in a project folder, and tell it what you want to research.

Source available · Local workflows · Your compute

BASH

Install the research skills

git clone https://github.com/Yichen-ZJU/lemvo.git
cd lemvo/claude
bash install.sh

Set up ArXiv MCP full-text retrieval

Install uv first, then run these commands from the cloned repository. Start Codex once to create its configuration.

uv tool install arxiv-mcp-server
claude mcp add --scope user arxiv -- "$HOME/.local/bin/arxiv-mcp-server"

Restart the CLI and check that arxiv tools loaded. Adjust the command if you use a custom installation path.

uv ↗

OPEN YOUR CLI. TELL IT YOUR DIRECTION.

I want to research X.

Lemvo brings in Research Orchestrator to clarify the goal, create research-state.yaml, and coordinate the two loops: test and improve methods inside; interpret results and choose the next direction outside.

Choose your mode: guided or unattended

Choose once at the start, switch as you go. Guided pauses at stage handoffs for your input. Unattended advances across stages through the resident controller; switching preserves results and running experiments.

Controller setup and mode switching ↗

Sources and demonstration notes＋

Skill counts checked on 2026-10-14 at Lemvo 9f2a5e0: claude/ and codex/ each contain the same 114 skill names, with engine-specific implementations preserved. The terminal and handwritten-digit animations illustrate the workflow. The 12 topic groups are editorial categories for this site.

Autonomous work uses your configured CLI, compute, and budget. The skill catalogue is shared; execution and continuation are adapted to each engine. Long-running drivers are configured as needed; existing launchers are not replaced automatically.

Claude · 源码 ↗Codex · 源码 ↗