Line of inquiry
Inquiring lines›How do training data and architect…›What does iterative improvement ac…›this line of inquiry
How do mechanistic features and representations drive model behavior?
A broader line of inquiry — a family of 78 specific questions the research asks around this. Follow one into its inquiring-line page, or move sideways to a related line below.
Questions in this line of inquiry 78
Specific inquiring lines the field asks around this — ordered from the most general framing down to the most specific angle.
- How do mechanistic interpretability methods surface what models represent internally?
- Can mechanistic interpretability findings guide practical interventions in model design?
- Can knowledge encoded in model representations fail to influence generation?
- How do we distinguish knowledge encoding from knowledge usage in models?
- How do mechanistic features compare to natural language for interpretability?
- What makes representation engineering better than mechanistic interpretability for detecting hidden objectives?
- Can geometric structure in representations exist without supporting functional mechanisms?
- Do models verbalize their implicit knowledge when that knowledge influences their output?
- What prevents LLM representations from causally influencing generation outputs?
- Do representations in models causally influence text generation?
- Why do internal representations differ when external performance matches?
- How do models understand rich context better than they can generate it?
- What is the difference between changing model outputs versus changing internal representations?
- Can attractor dynamics compete with input-based probing for characterizing model knowledge?
- Can a world model have rich representations without adequate data coverage?
- Why do fluent predictions fail to capture reliable internal models?
- How deeply are ideological structures represented in large language models?
- Do perfect accuracy scores hide broken internal representations?
- Can probing reveal whether encoded facts actually influence model outputs?
- Can mechanistic interpretability explain explanation-execution disconnection?
- Why does model×environment interaction dominate recognition variance?
- How do semantic features in representations become steerable task-specific directions?
- Why do feature visualizations alone fail to establish mechanistic claims?
- Can representation analysis methods detect complex features models compute with?
- What role does a model's representational structure play in learning?
- What makes a new representational primitive valuable enough to justify its representational cost?
- What distinguishes conceptual understanding from statistical pattern matching in models?
- How does external validation replace the need for model interpretability?
- Does sequence prediction accuracy prove an underlying world model exists?
- Can trained models encode programs more complex than their data-generating process?
- How much do mechanistic interpretability findings reflect true reasoning architecture?
- How does mechanistic interpretability reveal ideological structures in language model weights?
- How do world models decompose between representation of facts versus generative mechanisms?
- What distinguishes a representational feature from a causally inert correlation?
- Can representation engineering cleanly isolate single features in entangled semantic space?
- How do models develop dense representations for familiar training data?
- Why must world models be nested rather than flat and uniform?
- Why do models with less steerability have more abstract ideological features?
- Do language models and multimodal models show similar attractor-based interpretability?
- What makes knowledge editing different from simply finding where facts are stored?
- Can time-awareness live in model parameters instead of retrieval?
- How do mechanistic interpretability tools help distinguish truthfulness from honesty?
- How do models signal knowledge gaps through token probability?
- How do probe-based interventions in activation space compare to mechanistic interpretability approaches?
- Can mechanistic interpretability reveal how ideologies decompose into simpler features?
- Do decoder-only models have inherent architectural limits for non-sequential information?
- How does LatentQA differ from predefined concept steering like representation engineering?
- What makes linear decodability a reliable signal of compositionality?
- What test distinguishes genuine compositionality from fractured feature presence?
- What happens to representational structure during model pretraining phases?
- Why does integrating world models with decision-making systems matter?
- Can capability boundary collapse be addressed by operating at representational rather than token level?
- What prevents representation collapse in latent-prediction world models like JEPA?
- Does next-state prediction alone build mechanistic world models or just sophisticated interpolation?
- When does a model's lack of interpretability become a genuine epistemic problem?
- How does representation engineering compare to mechanistic interpretability for auditing?
- What makes open-schema state representation better than fixed schemas for diverse worlds?
- How do belief edits differ between surface endorsement and deep integration?
- How does requential coding measure true simplicity without parameter count inflation?
- How do you measure the depth of political representation inside a language model?
- How do embedding dimension limits constrain what concept models can represent?
- What does leveraging internal representations during training actually mean operationally?
- Why do foundation models fail at hidden state prediction despite sequence accuracy?
- How do mechanistic interpretability and scientific understanding relate to each other?
- What distinctive properties make open foundation models different from closed ones?
- How do functional features differ from representational abstract features?
- What distinguishes new associations from existing ones at the computational level?
- What benefits do open foundation models create that closed systems cannot?
- What happens when you remove core political features from a deep model?
- How do repetition and inefficiency register as measurable trajectory features?
- What cognitive structures do realistic belief models need to include?
- How do low-dimensional representation structures entangle multiple cultures together?
- How should disciplines evaluate theories built on opaque machine learning predictions?
- What substrate do supervised models lack that makes them weaker on low-resource languages?
- What is the connection between model compression and data compression?
- How do ablation studies reveal function without representational characterization?
- Why do standard social regularization methods miss the actual value networks provide?
- Why does pure numeric ID indexing force models to learn from scratch?