Safety3h ago
Safety
Testing Anthropic's Neural Language Autoencoders on Qwen
Anthropic released Neural Language Autoencoders to translate model activations into natural language. A researcher applied these tools to Qwen 2.5 7B to extract its internal multiplication algorithm. The process uses round-trip validation to ensure faithfulness. This approach offers a concrete method for practitioners to interpret hidden states without relying on vague approximations.