Open-Weight Masked Introspection: Measuring What Language Models Can Report About Their Own Computation
This paper studies whether language models can report on their own internal computation using a masked introspection setup. It presents a benchmark and reports that current open-weight models do not reliably introspect in the tested setting.
From: Emilio Ferrara, University of Southern California
The paper defines an evaluation framework for asking models about their own internal states and computation while controlling what information is available to them.
Its main result is negative for the tested models: they do not give dependable self-reports about internal computation under the benchmark conditions.
That result does not establish that introspection is impossible in principle, and it does not directly speak to consciousness or subjective experience.
Source: arXiv ↗. This is a summary; the source is the authority.
This weakens any claim that current open-weight models reliably access or report their own internal states, but it does not bear directly on whether they have experiences.
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…