Paper Lantern MCP server boosts coding agents with 2M+ CS papers
Paper Lantern is an MCP server that gives coding agents access to 2M+ CS research papers, returning ranked techniques with implementation steps, hyperparameters, and failure modes — and achieved 3.2% lower val loss than a web-search-only baseline in an autoresearch experiment.
Score breakdown
Coding agents using Paper Lantern can retrieve and apply specific, peer-reviewed ML techniques — including hyperparameters and failure modes — that web search alone misses, directly improving the quality of agentic research and training runs.
- 01Paper Lantern is an MCP server that searches 2M+ CS research papers for coding agents.
- 02It returns ranked techniques with implementation steps, hyperparameters, and failure modes for a given problem.
- 03Tested on Karpathy's autoresearch framework, comparing Claude Opus 4.6 + web search vs. the same agent + Paper Lantern.
Paper Lantern is an MCP server that connects coding agents to a corpus of 2M+ computer science research papers. Rather than returning raw documents, it takes a natural-language problem description from the agent and responds with ranked techniques, implementation steps, hyperparameters, and known failure modes — making research findings directly actionable within an agentic coding loop.
The baseline was Claude Opus 4.6 paired with web search; the experimental condition added Paper Lantern to the same agent.
The team evaluated Paper Lantern using Karpathy's autoresearch framework, where an agent proposes LLM architecture or training config optimizations, runs 5-minute training experiments, measures validation loss, and keeps or discards each change. The baseline was Claude Opus 4.6 paired with web search; the experimental condition added Paper Lantern to the same agent. Two concrete examples illustrate the difference: when both agents halved batch size, the Paper Lantern agent retrieved a 2022 paper prescribing a learning rate scaling of `1/sqrt(2)`, applied it correctly, and continued to improve — while the web-search agent saw worse loss and moved on without diagnosing the learning rate. The Paper Lantern agent also implemented AdaGC (adaptive gradient clipping, arxiv 2502.11034, published Feb 2025) on its first attempt with no manual tuning, a technique the baseline never tried.
After iterative short runs, both final configs were trained for 2 hours. The config discovered by the Paper Lantern agent achieved a 3.2% lower val loss. The project is open source, with experiment code on GitHub at `paperlantern-ai/autoresearch-experiment`, and can be installed with `npx paperlantern@latest`.
Key facts
- 01Paper Lantern is an MCP server that searches 2M+ CS research papers for coding agents.
- 02It returns ranked techniques with implementation steps, hyperparameters, and failure modes for a given problem.
- 03Tested on Karpathy's autoresearch framework, comparing Claude Opus 4.6 + web search vs. the same agent + Paper Lantern.
- 04The Paper Lantern agent found a config with 3.2% lower val loss after a 2-hour training run vs. the baseline.
- 05Paper Lantern agent correctly applied a 1/sqrt(2) learning rate scaling from a 2022 paper when halving batch size; the baseline agent did not.
- 06Paper Lantern agent implemented AdaGC (adaptive gradient clipping, arxiv 2502.11034, Feb 2025) on first try with no tuning.
- 07