Understanding LLM Agents Requires a Chain of Attribution: On the Interplay Between Mechanistic Interpretability and Behavioral Approaches
This position paper argues that mechanistic interpretability alone is not enough to understand complex LLM agents. It recommends combining internal and behavioral methods through a chain of attribution.
From: Preprints.org
The authors make a conceptual case about how to study agents, not a new empirical measurement study.
They argue mechanistic interpretability works best in controlled settings and becomes less complete in complex agentic systems.
It does not provide direct evidence for or against consciousness, only a methodological caution.
Source: Preprints.org ↗. This is a summary; the source is the authority.
This mainly changes how evidence should be interpreted, not the evidence for experiences themselves.
How we weigh evidence like this: the evidence levels.
- Mechanistic Interpretability Should Prioritize Feature Consistency in Sparse AutoencodersThis paper argues that sparse autoencoders should be judged not only by reconstruction quality and sparsity but also by whether the same fea…
- Mechanistic Interpretability for Neural NetworksThis is a broad survey of mechanistic interpretability, covering the goal of reverse-engineering internal algorithms in neural networks. It …
- Constructing Interpretable Features from Compositional Neuron Representations in Language ModelsThis paper proposes an alternative to standard sparse autoencoders by decomposing MLP activations with semi-nonnegative matrix factorization…
- Interpretability Transfer from Language to Vision via Sparse Autoencoder Transfer AlignmentThis work transfers labeled sparse-autoencoder features from language to vision inside a multimodal model. It uses those features to make vi…