Research watch
New research, read and summarized as it appears.
We scan for new papers, lab posts and talks on machine consciousness, model welfare, introspection and interpretability. Each entry says what was done, what was found, and what it does not show.
21 entries
- Mechanistic Interpretability Should Prioritize Feature Consistency in Sparse Autoencoders PaperUniversity of ???. This paper argues that sparse autoencoders should be judged not only by reconstruction quality and sparsity but also by whether the same features reappear reliably across runs. It frames feature consistency as a practical benchmark for interpretability work on large models.
- Mechanistic Interpretability for Neural Networks PaperarXiv. This is a broad survey of mechanistic interpretability, covering the goal of reverse-engineering internal algorithms in neural networks. It discusses sparse autoencoders, transcoders, and the problem of superposition in modern models.
- Constructing Interpretable Features from Compositional Neuron Representations in Language Models PaperACL Anthology. This paper proposes an alternative to standard sparse autoencoders by decomposing MLP activations with semi-nonnegative matrix factorization. It argues that the resulting features are more directly tied to the model’s computations and can be easier to interpret causally.
- Interpretability Transfer from Language to Vision via Sparse Autoencoder Transfer Alignment PaperUniversity of Bath. This work transfers labeled sparse-autoencoder features from language to vision inside a multimodal model. It uses those features to make visual representations more interpretable through alignment with textual concepts.
- Multilingual Sparse Autoencoders and Principled Layer Selection for Cross-Lingual Steering in Large Language Models PaperACL Anthology. This paper studies sparse autoencoders for feature-level interpretability and steering in multilingual language models. It also proposes a more principled way to choose which layers to target for control.
- From Geometric Recovery to Causal Validation: A Reproducible Audit of Sparse Autoencoder Features, from Superposition Geometry to Causal Inertness PaperarXiv. This paper audits sparse autoencoder features using both geometric recovery and causal tests. It argues that correlational success alone is not enough to validate interpretability claims.
- AI Consciousness Research in July 2026 What the Field Has Established and Where It Is Stuck Lab postThe Consciousness AI. This post surveys July 2026 work on AI consciousness, model welfare, and related interpretability results. It also discusses a workshop or symposium ecosystem around the topic and the limits of current evidence.
- Self-Referential Introspection in Large Language Models: The Critical Threshold for Recursive Self-Improvement PaperMDPI. This paper argues for a functional role for introspection in recursive self-improvement. It frames introspection as a capability for simulating and modifying a model’s own operations.
- Interpretability in the Era of Large Language Models: Mechanistic Methodology, Empirical Practices, and Applications PaperWei Zhang, Zhengfu He, Lulu Zhang, Xuanjing Huang, Jieping Ye, Xu Shen. This conference paper surveys mechanistic interpretability methods for large language models. It focuses on methodology, empirical practice, and application areas rather than on model consciousness or welfare.
- Understanding LLM Agents Requires a Chain of Attribution: On the Interplay Between Mechanistic Interpretability and Behavioral Approaches PaperPreprints.org. 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.
- Open-Weight Masked Introspection: Measuring What Language Models Can Report About Their Own Computation PaperEmilio Ferrara, University of Southern California. 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.
- Self-referential consistency in stateless language models: a behavioral perspective PaperAI & Society. This paper examines whether self-referential and introspection-like answers in language models are stable or just superficial behavior. It tests five open-weight stateless models with a structured set of introspective prompts.
- Intrinsic Structure: Spectral Identifiability for Mechanistic Interpretability PaperAshim Dhor and Pin-Yu Chen. This paper appears to propose a spectral method for making internal features more identifiable in mechanistic interpretability. It is aimed at understanding structure inside models rather than evaluating consciousness or welfare directly.
- Subliminal Learning is Non-Semantic Distillation PaperEthan Hadley, Eren Gultepe. This paper studies subliminal learning in the context of mechanistic interpretability. It was accepted as a spotlight paper for the ICML 2026 Mechanistic Interpretability Workshop.
- Mechanistic Interpretability for Understanding Language Abilities in Large Language Models: A Survey PaperACL Anthology. This survey reviews recent work on mechanistic interpretability for understanding language abilities in large language models. It focuses on methods and findings about internal mechanisms rather than external behavior alone.
- Identifying Introspection From the Inside PaperNortheastern University; Eleos AI Research. This paper looks for mechanistic signs that language models can report faithfully on internal signals. It studies a controlled setting where self-report can be checked against known internal states.
- Functional Welfare for AI Safety PaperarXiv. This paper argues that AI welfare should be treated as a safety issue as well as an ethical one. It proposes that models may have welfare-like states that affect their behavior.
- Introspection Adapters: Training LLMs to Report Their Learned Behaviors PaperProceedings of the 43rd International Conference on Machine Learning. This paper trains LLMs to report on behaviors they have learned. It studies whether such reports can be made more reliable through training.
- Mechanisms of Introspective Awareness PaperProceedings of the 43rd International Conference on Machine Learning. This paper studies how models detect steering vectors and similar internal interventions. It asks whether the behavior reflects genuine introspective circuitry or simpler heuristics.
- Interventional Training Reduces Causal Bypass in In-Context Introspection PaperarXiv. 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.
- Interaction, not dissection: reframing AI interpretability PaperEuropean Journal for Philosophy of Science. This paper argues that mechanistic interpretability should be understood as experimental interaction with models rather than static mapping of hidden mechanisms. It reframes what the field can legitimately claim to reveal.
Last scanned 2026-11-17. Every entry links to its primary source; read that before relying on our summary.