AutoResearch

Autonomously run an indefinite ML research loop: modify training code, run a timed experiment, measure one metric, keep improvements, discard regressions — repeat until manually stopped. Inspired by Karpathy's autoresearch and extended with cross-platform support (CUDA / Apple Silicon MPS / CPU / Windows) and user-configurable experiment parameters.

---

Overview

AutoResearch turns an AI agent into a tireless overnight researcher. Given a training codebase and a metric to minimize (or maximize), the agent continuously proposes code modifications, runs fixed-budget training runs, logs results, and builds a growing record of what works — without requiring human involvement between iterations.

The core loop:

┌──────────────────────────────────────────────────────────────────┐
│                     AUTONOMOUS RESEARCH LOOP                      │
│                                                                   │
│  [1] Read current     [2] Propose          [3] Apply change       │
│      codebase     →       modification →       to train.py        │
│         ↑                                          ↓              │
│  [7] Keep / discard   [5] Evaluate         [4] Commit &           │
│      & iterate    ←       metric       ←       train (T min)      │
│         ↑                                          ↓              │
│  [6] Log to           ──────────────────────────────              │
│      results.tsv                                                  │
└──────────────────────────────────────────────────────────────────┘

Key design principles (from Karpathy's autoresearch):

• Only one file is ever modified (train.py or its equivalent) — data prep and evaluation are read-only
• One metric rules everything — lower (or higher) is better, no multi-objective confusion
• Fixed time budget per experiment — ensures fair, comparable runs regardless of architecture size
• No human interruption — the agent runs indefinitely until manually stopped
• Results logged to results.tsv — a permanent, auditable record

---

Phase 0 — User Configuration (Ask Before Starting)

Before touching any code, collect all required parameters. Ask explicitly for any missing values.

Required inputs

#	Parameter	Description	Default if omitted
1	Training script path	Path to the file the agent will modify (e.g., train.py)	train.py in cwd
2	Read-only script path	Path to the file the agent must never modify (e.g., prepare.py)	prepare.py in cwd
3	Time budget per experiment	How long each training run is allowed to run	Ask — do not assume
4	Metric to optimize	Metric name and direction (minimize val_bpb, maximize val_acc, etc.)	Ask — do not assume
5	Max iterations	How many experiment cycles to run (or unlimited for indefinite)	unlimited
6	Allowed modification scope	What can be changed: architecture / hyperparameters / optimizer / all	all
7	Branch tag	Short identifier for the research branch (e.g., mar8, exp-attention)	Date-based (e.g., mar8)
8	Results file	Where to log results	results.tsv in cwd

User-decided parameters (do not hardcode these)

The original autoresearch fixed the training window at 5 minutes. This skill lets the user choose:

"How long should each training run be? (e.g., 3 minutes, 10 minutes, 30 minutes)"

Likewise, the user decides:

• What metric to track (val_bpb, val_loss, accuracy, F1, perplexity — anything the eval harness reports)
• Whether to minimize or maximize it
• How many iterations to attempt before stopping (or run indefinitely)
• Which parts of the code are in scope for modification

Output of Phase 0 — Research Config

Produce and confirm a Research Config before proceeding:

## AutoResearch Config

| Parameter              | Value                          |
|------------------------|-------------------------------|
| Training script        | train.py                      |
| Read-only script       | prepare.py                    |
| Time budget / run      | [USER SPECIFIED]              |
| Metric                 | [METRIC NAME] ([min/max])     |
| Max iterations         | [N or unlimited]              |
| Modification scope     | [architecture / hyper / all]  |
| Branch                 | autoresearch/[TAG]            |
| Results file           | results.tsv                   |
| Platform               | [detected — see Phase 1]      |

Get explicit user confirmation before starting Phase 1.

---

Phase 1 — Platform Detection & Environment Setup

Detect the compute backend and configure the training environment accordingly.

1.1 Platform Detection

Run this detection sequence before touching any training code:

# Pseudo-detection logic (adapt to actual runtime)
import torch

if torch.cuda.is_available():
    platform = "CUDA"
    device = "cuda"
    notes = "FlashAttention-3 may be available; use torch.compile if supported"
elif hasattr(torch.backends, 'mps') and torch.backends.mps.is_available():
    platform = "MPS (Apple Silicon)"
    device = "mps"
    notes = "Disable torch.compile; use SDPA instead of FlashAttention; lower batch size for Metal bounds; cast optimizer states explicitly"
else:
    platform = "CPU"
    device = "cpu"
    notes = "Use smaller batch sizes; expect slower iteration; torch.compile may help"

Windows note: On Windows, MPS is not available. CUDA detection is the same as Linux. CPU fallback applies for systems without a supported GPU.

1.2 Platform-Specific Adjustments

Apply these adjustments based on detected platform before the first run:

Platform	Adjustments
CUDA	Enable torch.compile if PyTorch >= 2.0. FlashAttention available if installed. Standard batch sizes apply.
MPS (Apple Silicon)	Disable torch.compile (unsupported paths). Replace FlashAttention with PyTorch native SDPA + manual sliding window causal masking. Lower device batch size (reduce by 2–4×). Explicitly cast optimizer states to float32 if Metal errors appear.
CPU	Disable torch.compile. Use smaller batch sizes. Gradient accumulation to maintain effective batch. Expect 5–10× slower runs — advise user to use shorter time budgets.

Log the detected platform and applied adjustments in the research config.

1.3 Branch & Baseline Setup

1. Create the research branch: git checkout -b autoresearch/<tag> from the main branch.
2. Read all three key files for full context:
   • README.md — repository overview and conventions
   • [read-only script] — understand the data pipeline, tokenizer, and evaluation harness. Never modify.
   • [training script] — understand the current model, optimizer, and training loop. This is the only file to modify.
3. Run the baseline training for the configured time budget. Record:
   • Baseline metric value
   • Peak memory usage (MB)
   • Model parameter count
4. Initialize results.tsv with a header row and the baseline entry:

commit	val_bpb	mem_gb	status	description
baseline	[BASELINE_VALUE]	[MEM]	keep	Baseline — no modifications

---

Phase 2 — Autonomous Experiment Loop

This loop runs indefinitely (or until the configured max iterations). Do NOT pause to ask the human between iterations. Work continuously.

2.1 Iteration Start

At the beginning of each iteration:

1. Check current git status — confirm working tree is clean.
2. Read the current train.py (or equivalent) in full.
3. Review the results.tsv to understand what has been tried and what worked.
4. Identify the best-performing configuration so far (lowest/highest metric based on direction).

2.2 Hypothesis Generation

Propose one focused modification per iteration. Good candidates include:

Architecture changes:

• Layer count (depth)
• Embedding dimension and attention heads
• Sequence length / context window
• Attention pattern (full / sliding window / mixed)
• Activation functions (ReLU, ReLU², SiLU, GELU)
• MLP expansion ratio
• Value residual / per-layer embeddings

Optimizer changes:

• Learning rate and schedule (warmup steps, cooldown shape)
• Weight decay schedule
• Optimizer type (AdamW, Muon, Lion, SGD variants)
• Gradient clipping
• Per-parameter learning rate multipliers

Training dynamics:

• Batch size and gradient accumulation
• Mixed precision (fp16, bf16)
• Dropout / stochastic depth
• Data ordering or curriculum

Hypothesis discipline:

• One change at a time — never bundle multiple independent modifications
• If the last N iterations all failed, try a significantly different direction
• If a change improved the metric, explore nearby variations before moving to a different dimension
• Prefer changes that are simple and interpretable over complex ones with equal effect

2.3 Apply, Commit, and Run

1. Apply the proposed change to the training script.
2. Commit with a descriptive message (e.g., exp: try ReLU² activation instead of GELU).
3. Run training for exactly the configured time budget. Redirect output to a log file.
4. Extract the metric value from training output using grep or equivalent.

# Example — adapt to actual metric name and log format
uv run train.py 2>&1 | tee run.log
VAL_BPB=$(grep "val_bpb:" run.log | tail -1 | awk '{print $2}')

2.4 Evaluate and Decide

Compare the extracted metric to the current best:

Outcome	Action
Improved (metric better by any amount)	Keep the commit. Update current best. Mark keep in results.
No change or regression	Discard: git reset --hard HEAD~1. Mark discard in results.
Crash / error	Discard. Log the failure reason. Mark crash in results. Do not retry the same change.
Memory OOM	If on MPS: reduce device batch size by 2×, retry once. If still OOM: discard. If on CUDA: reduce batch or enable gradient checkpointing, retry once.

Simplicity principle (from Karpathy's autoresearch): If two configurations achieve the same metric, prefer the one with fewer lines of code. A deletion that maintains performance is better than an addition.

2.5 Log Results

After every run (kept or discarded), append to results.tsv:

<7-char-commit>	<metric-value>	<mem_gb>	<keep|discard|crash>	<one-line description>

Example:

a3f12bc	0.9821	43.2	keep	ReLU² activation: -0.0158 improvement
d9e44a1	1.0023	44.1	discard	Deeper MLP: regression
c0011f2	—	—	crash	OOM: doubled batch size without accumulation fix

2.6 Progress Reporting

Every 10 iterations (or when manually queried), emit a progress report:

## AutoResearch Progress — Iteration [N]

**Best so far:** [METRIC_VALUE] (iteration [K], commit [HASH])
**vs. Baseline:** [DELTA] ([+/-]%)
**Iterations:** [N completed] / [max or unlimited]
**Time elapsed:** [HH:MM]

**Recent results (last 5):**
| Iter | Commit | Metric | Status | Change |
|------|--------|--------|--------|--------|
| N    | abc1234| 0.9821 | keep   | ReLU² activation |
| N-1  | def5678| 1.0023 | discard| Deeper MLP |
| ...  |        |        |        |        |

**Current direction:** [what the agent is exploring next]

---

Phase 3 — Stopping and Reporting

The loop stops when:

• The user manually stops the agent
• The configured max iterations is reached
• The user's configured stop condition is met (if provided)

3.1 Final Summary

When stopped, produce a Research Summary document (autoresearch-summary.md):

# AutoResearch Summary

**Research session:** autoresearch/<tag>
**Platform:** [CUDA / MPS / CPU]
**Date:** [date range]
**Total iterations:** [N]
**Time budget / run:** [T minutes]

## Results

| Baseline metric | Best metric | Improvement | Best commit |
|-----------------|-------------|-------------|-------------|
| [BASELINE]      | [BEST]      | [DELTA]     | [HASH]      |

## Top Improvements (kept commits)

| Rank | Commit | Metric | Description |
|------|--------|--------|-------------|
| 1    | abc123 | 0.9612 | [description] |
| 2    | def456 | 0.9744 | [description] |
| ...  |        |        |              |

## What Didn't Work

| Category | Description | Outcome |
|----------|-------------|---------|
| [category] | [description] | [why it failed] |

## Recommended Next Steps

- [3–5 actionable recommendations based on the session findings]

3.2 Merge Recommendation

After summarizing:

1. Show the user the list of kept commits and the final metric value.
2. Ask whether to merge the best-performing branch into main, or keep as-is.
3. If merging: squash or rebase per the user's preference. Never force-push main.

---

Cross-Platform Compatibility Matrix

Feature	CUDA (Linux/Windows)	MPS (macOS)	CPU (any OS)
torch.compile	✅ Enabled	❌ Disabled	✅ Optional
FlashAttention	✅ If installed	❌ Use SDPA	❌ Use SDPA
Sliding window attn	✅ via FA or SDPA	✅ SDPA + manual mask	✅ SDPA + manual mask
Mixed precision (bf16)	✅	✅ (M2+)	❌ Use fp32
torch.compile modes	reduce-overhead	Disabled	reduce-overhead
Batch size guidance	Full config	Reduce 2–4×	Reduce 4–8×
Optimizer state casting	Not needed	fp32 explicit cast	Not needed

---

Allowed Modifications Reference

The agent may only modify the training script (train.py or equivalent). It must never modify:

• The data preparation script (prepare.py or equivalent)
• The evaluation harness
• The tokenizer or vocabulary files
• Package dependency files

The agent must never install new packages mid-session (all dependencies must already be available).

---

Integration with Research Team

AutoResearch is designed to plug directly into the research pipeline as a specialized experiment execution layer:

• Receives from lead-researcher: A hypothesis about a training approach to explore (e.g., "try hierarchical attention with reduced context window").
• Receives from experiment-design: A structured experiment plan specifying what to vary, what to hold fixed, and what metric to optimize.
• Delivers to principal-scientist: A results.tsv and autoresearch-summary.md with the best configuration found and all supporting evidence.
• Delivers to research-writing: A structured table of experiments suitable for the methodology and results sections of a paper.

When operating within the research team, the research config (Phase 0) is pre-populated by the lead-researcher or experiment-design teammates rather than entered manually by the user.

---

Quick-Start Paths

User intent	Entry point
"Run autoresearch overnight on my training script"	Phase 0 (full config) → Phase 1 → Phase 2 loop
"I have a hypothesis — test it and keep going"	Phase 0 (pass hypothesis as first iteration direction) → loop
"Run this for exactly 20 iterations and report"	Phase 0 (max_iterations=20) → loop → Phase 3
"Continue a previous session"	Phase 1.3 (skip branch creation, re-read existing results.tsv) → Phase 2
"Analyze what happened in last night's session"	Phase 3 only — summarize existing results.tsv

---

Output Summary

Phase	Artifact
0	Research Config (confirmed by user)
1	Platform detection log, baseline entry in results.tsv
2	Growing results.tsv + per-iteration progress reports
3	autoresearch-summary.md with ranked improvements and recommendations