LLM Circuit Finder: Duplicating Specific Layers in Transformer Models Improves Reasoning Performance Without Training
By
xlayn
An everything bagel for the brain. Substantive, layered, well-seasoned.
Summary
The article describes a GitHub project called 'llm-circuit-finder' that implements a method for discovering and exploiting 'reasoning circuits' within transformer-based large language models. The author replicated Ng's RYS method and found that duplicating specific layers in models like Qwen2.5-32B and Devstral-24B significantly improves reasoning performance without any training or weight changes. For Qwen2.5-32B, duplicating 3 specific layers boosted reasoning by 17%, while for Devstral-24B, duplicating layers 12-14 improved logical deduction scores from 0.22 to 0.76 on the BBH benchmark. The approach involves routing hidden states through the same circuit twice, and the toolkit includes tools for finding these reasoning circuits. The project was completed using two AMD GPUs in one evening.
Key quotes
· 5 pulledDuplicate 3 layers. No training. Logical deduction goes from 0.22 → 0.76.
This toolkit finds and exploits 'reasoning circuits' hidden inside transformer models.
The idea: certain contiguous blocks of layers act as indivisible cognitive units.
I replicated Ng's RYS method and found that duplicating 3 specific layers in Qwen2.5-32B boosts reasoning by 17%
no training, no weight changes, just routing hidden states through the same circuit twice
You might also wanna read
Researchers Develop Method to Predict Real-Time Progress in Reasoning Language Models
This research paper investigates whether real-time progress prediction is feasible for reasoning language models that use long latent chains
arxiv.org·4d ago
HSIR: New Method Improves Self-Improvement Training for Large Reasoning Models
This research paper identifies two key problems in self-improvement training for Large Reasoning Models (LRMs): data imbalance (too many sim
arxiv.org·5d ago
Study Reveals How RL and SFT Differently Teach Transformers Chain-of-Thought Reasoning on Sparse Boolean Functions
This research paper analyzes how transformers learn Chain-of-Thought (CoT) reasoning capabilities through Reinforcement Learning (RL) with p
arxiv.org·3d ago
RTP-LLM: Alibaba's High-Performance Inference Engine for Large Language Model Deployment
This paper presents RTP-LLM, a high-performance inference engine developed by Alibaba for industrial-scale deployment of Large Language Mode
arxiv.org·2d ago
Feedback Distillation: A New Training Method for Improving LLM Reasoning in Theorem Proving
This paper introduces Feedback Distillation, a novel training method for reasoning models that improves upon standard GRPO (Group Relative P
arxiv.org·6h ago
Bridge-Garden Theory Explains Why Mixing Hard and Soft Labels Improves Knowledge Distillation for LLMs
This research paper investigates knowledge distillation (KD) for language models, specifically why mixing hard labels (sampled tokens) and s
arxiv.org·4d ago