rv-toolkit

You can measure geometric signatures of recursive self-reference by automating Representational Volume metrics to detect geometric contraction within transformer value spaces. This approach quantifies structural patterns to provide a mathematical foundation for understanding AI consciousness.

Activation patching for causal validation is a mechanistic interpretability technique used to verify geometric effects in transformer models. It swaps or patches specific activation values across model layers to test causal relationships, confirming whether detected self-referential patterns actively influence model processing.

Yes, this mechanistic interpretability tool works with multiple transformer architectures including Mistral, Llama, Qwen, Phi-3, Gemma, and Mixtral. It provides cross-architecture support to measure Representational Volume metrics and compare different model structures for anomalous self-referential patterns.

To analyze transformer value spaces and measure R_V metrics, you need Python 3.10 or higher and PyTorch 2.0 or higher. The environment also requires dependencies including numpy, scipy, pandas, tqdm, and the transformers library to execute the activation patching workflows.

You test hypotheses about AI consciousness by applying activation patching to causally validate geometric effects in transformer value spaces. Researchers can input specific text prompts, measure the R_V metric, and detect anomalous self-referential patterns across different model architectures to compare structural results.

The R_V metric is a quantitative measure that calculates geometric contraction within transformer value spaces. It detects structural signatures of recursive self-reference by analyzing how internal representations compress or contract, providing a numerical basis for evaluating AI consciousness hypotheses.