Laplace’s Scissors

Laplace's Scissors is an advanced tool designed for open LLMs (Large Language Models) that implements output matrix rotations for meaningful edits while safeguarding successful answers. This repository leverages Rust, Triton, and CUDA/HIP to effectively address the challenges of weak output directions, rectify flawed outputs, and analyze deviations across model scaling.

Key Features

• Rotation Edits: Compute precise rotation edits for the original output matrix of an LLM to maintain the integrity of verified answers while proposing improvements to others.
• Diagnostic Tools: Utilize the probe command to identify directions with low discrimination, assess cancellation effects across layers, and trace weight coordinates contributing to the model responses.
• Answer Repair: The diagnose command identifies potential plane or joint rotations to fix erroneous outputs in line with expected answer strings, ensuring the preservation of successful generations through GPU validations.
• Scaling Analysis: Implement scaling policies that evaluate deviations from reference benchmarks across different model checkpoints, enabling in-depth analysis of model behavior.

Workflow Overview

The process begins with fetching a model, inspecting its structure, and tracing weak discrimination directions using:

cargo run --release -- fetch Qwen/Qwen3-0.6B  
cargo run --release -- inspect Qwen/Qwen3-0.6B --out "$LAPLACE_CACHE/graph.json"  

Subsequently, users can diagnose potential rotations and propose fixes by comparing expected outputs:

cargo run --release -- diagnose Qwen/Qwen3-0.6B ...  

Finally, you can generate output using the original or modified model:

cargo run --release -- generate Qwen/Qwen3-0.6B ...  

Command Overview

Conclusion

Laplace's Scissors provides a powerful infrastructure for enhancing LLM outputs through innovative editing techniques. By integrating diagnostic capabilities and rigorous analysis workflows, this tool ensures the generation of high-quality responses while maintaining the necessary safeguards for existing successful outputs.
For full functionality, review additional documentation such as DESIGN.md for computation contracts and core design principles.