alainnothere

alainnothere / llm-circuit-finder

Public

I replicated Ng's RYS method and found that duplicating 3 specific layers in Qwen2.5-32B boosts reasoning by 17% and duplicating layers 12-14 in Devstral-24B improves logical deduction from 0.22→0.76 on BBH — no training, no weight changes, just routing hidden states through the same circuit twice. Tools included. Two AMD GPUs, one evening.

48
2
100% credibility
Found Mar 19, 2026 at 48 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

A collection of tools to scan language models, identify key reasoning circuits, and create enhanced versions by duplicating those circuits for better performance on benchmarks.

How It Works

1
🔍 Discover the boost trick

You hear about a simple way to make your AI helper much better at thinking and solving problems by repeating its best thinking steps.

2
📥 Grab your AI brain

You pick a ready-to-use AI model file from your collection and get the free helper tools to start experimenting.

3
🧠 Hunt for magic thinking spots

You run a quick scan that tests puzzles and conversations to pinpoint the exact group of thinking steps that make your AI shine brightest.

4
✨ Build the smarter version

You create a new copy of your AI that loops through those top thinking spots extra times, keeping everything else the same.

5
🧪 Test the upgrade

You challenge the new AI with math riddles, logic games, and people-smarts questions to see the big jumps in smarts.

🚀 Smarter AI unlocked

Your AI now crushes reasoning tasks like never before, getting scores up by double digits without any extra training or hassle.

Sign up to see the full architecture

4 more

Sign Up Free

Star Growth

See how this repo grew from 48 to 48 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is llm-circuit-finder?

This Python toolkit replicated Ng's RYS method on GitHub to uncover "reasoning circuits" in LLMs—hidden blocks of transformer layers that boost capabilities when duplicated. Users get CLI tools to sweep models for optimal circuits (like layers 12-14 in Devstral-24B or 7-9 in Qwen2.5-32B), edit GGUF files via llama.cpp for rerouting hidden states twice, and probe results on BBH or custom benchmarks—no training or weight changes needed. Run it on two AMD GPUs in one evening to find circuits that improve logical deduction from 0.22 to 0.76.

Why is it gaining traction?

It delivers massive gains like 17% reasoning boosts or +245% on BBH logical deduction by simply duplicating layers, without fine-tuning or extra VRAM hacks. Developers dig the quick sweep CLI for any GGUF model, plus eval tools comparing lm-evaluation-harness runs, all validated on real benchmarks. The "replicated nghia la gi" (what replication means) and định nghĩa replicated aspects shine: exact repro of findings with visualization heatmaps.

Who should use this?

LLM tinkerers hacking local inference on AMD GPUs, researchers probing transformer internals without retraining, or teams evaluating Devstral-24B/Qwen2.5 for reasoning tasks like BBH causal judgment. Ideal for nights spent circuit-finding to squeeze more from open models before deploying.

Verdict

Grab it for weekend experiments—48 stars and solid README make it approachable, but 1.0% credibility score flags early maturity with no tests. Promising circuit finder if you're AMD-equipped; fork and contribute for production use.

(198 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.