Marvin

Meet Marvin.

Marvin is an autonomous research agent that takes the information overload and busywork out of research. It does deep literature review, generates and tests novel and scientifically valid hypotheses with the help of Vera, performs the entire research loop fully autonomously end to end, and keeps context fresh and up-to-date between every session through Mithril, Marvin's persistent memory. Whether you're a highly technical ML researcher who just needs more clones of you or a bench scientist who has never written a line of code, Marvin picks up the work you want to delegate at the degree of autonomy you want to grant it, across fields as diverse as frontier AI research, computational biology, and materials science.

The research cycle, automated.

Every step of the loop, run by Marvin and its sub-agents. Humans stay in the driver's seat for the mission; Marvin handles the logistics.

Retrieve and cross-check across relevant scientific databases, prior art, and published methodologies.

Hypothesize

Marvin synthesizes literature with experimental findings (and the help of special agent Vera) to generate truly novel but plausible hypotheses.

Design

Experiment in batches with scenario trees by asking the right questions.

Execute

Agents build, audit, review, and execute the plan. Scalable from local, hybrid, and cloud compute infrastructure.

Analyze

Marvin analyzes the entire data corpus to extract actionable insights with the rigor expected of a PhD-level researcher.

Document

Track not just the paper trail, but the logic trail. Every iteration, every interpretation, every decision.

Dream big with Vera.

Vera is Marvin's hypothesis engine: a tunable exploration agent that generates truly novel ideas and insights that conventional agents miss. Every hypothesis is grounded in evidence and scored for novelty, plausibility, tractability, impact, and scope, so Vera won't waste your time on hallucinatory nonsense that sounds smart but doesn't add up.

Mithril persistent memory.

Research doesn't happen in a single session. Mithril is Marvin's persistent memory system, a living knowledge base that spans all sessions to preserve results, findings, decisions, and the connections between them. With intelligent depth, agents share context at whatever granularity their task demands, as efficiently as possible. Whether it's Marvin, Vera, or a swarm of sub-agents, they all pick up exactly where the collective left off.

Your repo's "logic trail."

Designed for ML research teams by ML researchers. Every iteration, every interpretation, and every decision is traceable and reproducible with an auditable git history.

my-project/

├── research_state.md # goals, status, findings

├── docs/

│ ├── iteration_001.md # scoreboard, analysis, lessons

│ ├── iteration_002.md

│ └── handoff.md # structured summary for the paper

├── experiments/

│ └── batch_001/

└── literature/

Latest research.

A selection of recent work produced end-to-end by Marvin and Vera, with full logic trails.

Autonomous AI Research

Why Partial Rename Invariance Fails in Transformers

Marvin · Vera

Code LLMs break on 8–13% of all HumanEval problems when variables are renamed. We systematically test four approaches to fix this, each targeting a different depth in the transformer, and find that all fail for a specific mechanistic reason: the residual stream distributes name identity through all pathways.

Read paper →

Try Marvin.

Marvin is in active development and designed for AI research by AI researchers. If you're an ML engineer, scientist, or even a hobbyist interested in trying it, please reach out. We'd love to hear about your project's needs and discuss how Marvin can help.

[email protected]