LLM Circuit Finder boosts local LLM reasoning
Building upon David Ng's RYS method, LLM Circuit Finder provides scripts to discover and exploit reasoning circuits within transformer models by duplicating specific hidden state layers. The toolkit automates finding optimal layer blocks for any GGUF model locally, achieving measurable capability jumps without retraining.
This project represents a fascinating shift from weight-based fine-tuning to architectural "surgery," treating transformer layers as functional neuroanatomy that can be rearranged to create specific cognitive profiles. It achieves substantial capability jumps purely by routing hidden states through existing layers multiple times, acting as zero-shot architecture hacks. The entire discovery and modification process can be run on consumer GPUs in a single evening, democratizing advanced model optimization. However, the method highlights an inherent cognitive trade-off, as boosting mathematical and causal reasoning can come at the expense of instruction following and precise code generation. Because this approach modifies the model's architecture rather than its weights, it remains orthogonal to fine-tuning and can be stacked with traditional methods for compounded performance gains.
DISCOVERED
19d ago
2026-03-23
PUBLISHED
19d ago
2026-03-23
RELEVANCE
AUTHOR
Github Awesome