Duplicate 3 layers in a 24B LLM, logical deduction .22→.76. No training
I replicated David Ng's RYS method (https://dnhkng.github.io/posts/rys/) on consumer AMD GPUs (RX 7900 XT + RX 6950 XT) and found something I didn't expect. Transformers appear to have discrete "reasoning circuits" — contiguous blocks of 3-4 layers that act as indivisible cognitive units. Duplicate the right block and the model runs its reasoning pipeline twice. No weights change. No training. The model just thinks longer. The results on standard benchmarks (lm-evaluation-harness, n=50): Devstral-24B, layers 12-14 duplicated once: - BBH Logical Deduction: 0.22 → 0.76…
In plain words
This project demonstrates a technique for improving large language model reasoning by duplicating specific layer blocks without retraining. By duplicating contiguous groups of 3-4 layers in models like Devstral-24B, the approach enables models to run their reasoning processes multiple times using unchanged weights. Testing shows significant improvements on logical reasoning and mathematical tasks, with logical deduction accuracy jumping from 0.22 to 0.76 and code generation improving from 0.72 to 0.78. The method works on consumer AMD GPUs and requires no model fine-tuning.
written from the facts on this page · September 2026
From the sources
In the maker’s words, at launch
I replicated David Ng's RYS method (https://dnhkng.github.io/posts/rys/) on consumer AMD GPUs (RX 7900 XT + RX 6950 XT) and found something I didn't expect. Transformers appear to have discrete "reasoning circuits" — contiguous blocks of 3-4 layers that act as indivisible cognitive units. Duplicate the right block and the model runs its reasoning pipeline twice. No weights change. No training. The model just thinks longer. The results on standard benchmarks (lm-evaluation-harness, n=50): Devstral-24B, layers 12-14 duplicated once: - BBH Logical Deduction: 0.22 → 0.76 - GSM8K (strict): 0.48 → 0.64 - MBPP (code gen): 0.72 → 0.78 - Nothing degraded Qwen2.5-Coder-32B, layers 7-9 duplicated once: - Reasoning probe: 76% → 94% The weird part: different duplication patterns create different cognitive "modes" from the same weights. Double-pass boosts math. Triple-pass boosts emotional reasoning. Interleaved doubling (13,13,14,14,15,15,16) creates a pure math specialist. Same model, same VRAM, different routing. The circuit boundaries are sharp — shift by one layer and the effect disappears or inverts. Smaller models (24B) have tighter circuits (3 layers) than larger ones (Ng found 7 layers in 72B). Tools to find circuits in any GGUF model and apply arbitrary layer routing are in the repo. The whole thing — sweep, discovery, validation — took one evening. Happy to answer questions.
Does the same job
all alternatives →- HIHow I topped the HuggingFace open LLM leaderboard on two gaming GPUsMar 2026 · dnhkng.github.io · ▲495
I found that duplicating a specific block of 7 middle layers in Qwen2-72B, without modifying any weights, improved performance across all Open LLM Leaderboard benchmarks and took #1. As of 2026, the top 4 models on that leaderboard are still descendants. The weird finding: single-layer duplication does nothing. Too few layers, nothing. Too many, it gets worse. Only circuit-sized blocks of ~7 layers work. This suggests pretraining carves out discrete functional circuits in the layer stack that only work when preserved whole. The whole thing was developed on 2x RTX 4090s in my basement. I'm…
- IBI built a tiny LLM to demystify how language models workApr 2026 · github.com · ▲915
Built a ~9M param LLM from scratch to understand how they actually work. Vanilla transformer, 60K synthetic conversations, ~130 lines of PyTorch. Trains in 5 min on a free Colab T4. The fish thinks the meaning of life is food. Fork it and swap the personality for your own character.
- SUSpeeding up LLM inference 2x times (possibly)2024 · asciinema.org · ▲419
Here's a project I've been working on for the last few months. It's a new (I think) algorithm, that allows to adjust smoothly - and in real time - how many calculations you'd like to do during inference of an LLM model. It seems that it's possible to do just 20-25% of weight multiplications instead of all of them, and still get good inference results. I implemented it to run on M1/M2/M3 GPU. The mmul approximation itself can be pushed to run 2x fast before the quality of output collapses. The inference speed is just a bit faster than Llama.cpp's, because the rest of implementation…
- ABAutoThink – Boosts local LLM performance with adaptive reasoning2025 · ▲397
I built AutoThink, a technique that makes local LLMs reason more efficiently by adaptively allocating computational resources based on query complexity. The core idea: instead of giving every query the same "thinking time," classify queries as HIGH or LOW complexity and allocate thinking tokens accordingly. Complex reasoning gets 70-90% of tokens, simple queries get 20-40%. I also implemented steering vectors derived from Pivotal Token Search (originally from Microsoft's Phi-4 paper) that guide the model's reasoning patterns during generation. These vectors encourage behaviors like numerical…
- IBI built a toy TPU that can do inference and training on the XOR problem2025 · tinytpu.com · ▲134
We wanted to do something very challenging to prove to ourselves that we can do anything we put our mind to. The reasoning for why we chose to build a toy TPU specifically is fairly simple: - Building a chip for ML workloads seemed cool - There was no well-documented open source repo for an ML accelerator that performed both inference and training None of us have real professional experience in hardware design, which, in a way, made the TPU even more appealing since we weren't able to estimate exactly how difficult it would be. As we worked on the initial stages of this project, we…
More ai this month
the category →
I trained a 125M-parameter transformer to autocomplete piano performances in real time (~108 notes/sec on an iPhone 15). The idea is basically GitHub Copilot or Tabnine, except instead of prompting it with code, you prompt it by playing a few notes on a MIDI piano. The model then continues what you played, entirely on-device. The app is free if anyone wants to try it. Happy to answer questions about the model, training, Core ML, or the many things that didn't work.
AI · 17d ago · simedw.com
Astute▲585Automate your B2B brand going viral, with new media creators
AI · 18d ago · company-app.joinastute.com
Hey HN, Henry from Cactus here! We previously released Cactus Needle, a 14MB agentic LLM for tool call, device use, and structured extraction for phones, wearables, smart homes, small robots and microcontrollers. We got really great feedback here, and have now incorporated the suggestions to release Needle 2. The whole model is a single 14MB binary that runs a full session in 28MB of RAM; 45m parameters at 2bit compression. Needle hits 500 tokens/sec decode speed on a Raspberry Pi 5, sits between 400-1,500 tokens/sec on VR devices like Meta Quest 3S and Apple Vision Pro, and ranges…
AI · 27d ago · cactuscompute.com
Launched alongside, March 2026
the whole month →
Switch from ChatGPT to Claude with import memory feature
AI · Mar 2026 · claude.com
