Frontier Lab Notes

Research Memo Template for LLM Papers

Use this for every serious paper. The goal is to move from "I read it" to "I can reason from it."

---
tags:
  - llm/paper
status: reading
paper:
authors:
year:
link:
---

# Paper: Title

## One-Sentence Claim

What does the paper claim that was not obvious before?

## Bottleneck

Which bottleneck does this attack?

- Data
- Architecture
- Optimization
- Scaling
- Systems
- Inference
- Post-training
- Evaluation
- Safety

## Mechanism

Explain the method without marketing language.

## Equations / Algorithm

Copy only the minimum necessary equations, then explain each term.

## Experimental Setup

- Model sizes:
- Data:
- Compute:
- Baselines:
- Metrics:
- Hardware:

## Main Evidence

What result actually supports the claim?

## Missing Baselines

What comparison would make the paper more convincing?

## Confounders

Could the result be caused by:

- More data?
- Better data?
- More compute?
- Different optimizer?
- Different eval prompt?
- Contamination?
- Inference settings?

## Scaling Guess

Will this get better, worse, or irrelevant at frontier scale?

## Systems Guess

Will this be easy or painful to train and serve?

## Failure Modes

How might this break?

## Falsification Experiment

What result would make me stop believing the paper's main claim?

## Reproduction Plan

What can I reproduce with:

- 1 GPU:
- 8 GPUs:
- Lab-scale compute:

## What I Learned

## Follow-Up Papers

How To Use It

  1. Fill it quickly after first read.
  2. Revise after implementing or reproducing.
  3. Add links to the relevant concept notes.
  4. Keep the falsification experiment honest.

Related