ICA Lens: Interpreting Language Models Without Training Another Dictionary

Researchers introduce ICALens, a new method for interpreting language model representations using independent component analysis (ICA) instead of expensive sparse autoencoders (SAEs). The approach efficiently recovers interpretable directions without requiring large neural dictionary training, achieving competitive performance on standard benchmarks while offering a faster, more accessible alternative for LLM analysis.

ICALens addresses a significant bottleneck in AI interpretability research by reviving independent component analysis as a practical tool for understanding language model behavior. Rather than training computationally expensive sparse autoencoders—the current standard approach—the method leverages classical statistical techniques optimized for GPU efficiency to identify non-Gaussian directions in model activations that correlate with interpretable features. This represents a meaningful shift in how researchers can explore and audit model internals.

The research builds on decades-old mathematical foundations but applies them with modern infrastructure and LLM-specific improvements that prior work lacked. Previous ICA implementations proved brittle on transformer activations, forcing researchers into the SAE training paradigm. ICALens solves this through optimized FastICA pipelines, stability recipes, and better diagnostics tailored to neural network representations. Testing across multiple small models (GPT-2 Small, Gemma 2 2B, Qwen 3.5 2B) shows ICA recovers interpretable directions efficiently and matches or exceeds SAE performance on several metrics.

For the broader AI ecosystem, this work democratizes interpretability research by reducing computational barriers. Training large SAEs requires significant resources, limiting who can conduct rigorous interpretability studies. ICALens offers researchers and developers a lightweight alternative that enables rapid exploration before committing to expensive dictionary training. This accelerates the feedback loop between model understanding and control, particularly valuable as models scale and interpretability becomes increasingly critical for safety and trust. The method suggests the field may have overspecialized on neural approaches when classical statistical methods remain competitive, encouraging researchers to reconsider conventional wisdom about model analysis tooling.

• →ICALens provides an efficient, GPU-accelerated alternative to sparse autoencoders for finding interpretable directions in language models without training large neural dictionaries.
• →The method achieves competitive or superior performance on SAEBench benchmarks while requiring significantly less computational overhead than standard SAE approaches.
• →ICA-based analysis removes barriers to interpretability research by enabling rapid exploration of model representations for resource-constrained teams and researchers.
• →The work demonstrates that classical statistical techniques remain underutilized in modern deep learning and warrant reconsideration alongside neural approaches.
• →ICALens combines optimized infrastructure with LLM-specific stability improvements that address historical brittleness of prior ICA implementations on transformer activations.