Interventional Training Reduces Causal Bypass in In-Context Introspection
This paper studies whether models can introspect on low-level internal signals without shortcutting the task. It tests a training method meant to make self-reports causally grounded.
From: arXiv
The authors train many introspectors across several internal signal types and multiple language models.
They find that standard supervised fine-tuning can lead to causal bypass, while interventional training improves grounding.
The results show better access to internal signals, but not that models have conscious access to their own states.
Source: arXiv ↗. This is a summary; the source is the authority.
This improves evidence that some self-reports can be causally grounded, but it does not imply phenomenology or welfare.
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…