Line of inquiry
Inquiring lines›How can we train models to be prop…›What architectural and training in…›this line of inquiry
How do sparse and dense representations support compositionality in language models?
A broader line of inquiry — a family of 85 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 85
Specific inquiring lines the field asks around this — ordered from the most general framing down to the most specific angle.
- Does latent density emerge during pretraining from training data familiarity?
- How do sparse weight patterns affect model interpretability?
- How does scaling and training data enable compositional behavior without symbolic mechanisms?
- How does representational density emerge from training data familiarity?
- Can fractured representations explain why models fail at systematic generalization?
- Do dense models already achieve high interpretability without sparse training?
- Can scaling alone create compositional generalization without explicit binding mechanisms?
- Can sparse approximations reveal interpretable structure hidden in existing dense models?
- Where do neural networks still fail at compositional generalization despite scaling?
- Why does gradient descent discover compositional structure without explicit pressure?
- Does sparsity enforce compositional structure or merely amplify existing modularity?
- Does the linear representation hypothesis reflect networks or reflect our analysis tools?
- Does recursive generalization explain why neurons develop single interpretable functions?
- How can interpretability methods account for shifting representational density across task conditions?
- How do neural networks decompose tasks into modular subnetworks that transfer?
- Why does weight sparsity reduce superposition and force disentangled representations?
- Do task-relevant parameter changes naturally concentrate in sparse regions?
- Does representational density emerge from training data exposure during pretraining?
- Can neural networks represent symbolic structures without explicit mechanisms?
- What inductive bias would force models to learn Newtonian mechanics instead of shortcuts?
- Can architectural changes reduce polysemanticity without explicit sparsity training?
- How does representation sparsity change when inputs fall outside the training distribution?
- How would weight sparsity change what representation analysis methods can detect?
- How does weight sharing compound the advantages of deeper model designs?
- Can we detect and measure circuit formation before generalization emerges?
- What role does a model's representational structure play in learning?
- How do LLM activations sparsify differently under out-of-distribution inputs?
- Can geometric structure in representations exist without supporting functional mechanisms?
- What are fractured entangled representations in neural networks?
- How do semantic features in representations become steerable task-specific directions?
- Can world models form from aggregated partial information across training distributions?
- How do sparse networks trade capability for human-understandable circuits?
- How do latents at the same hierarchy level become more correlated than tokens?
- Can fractured entangled representations hide undetected by standard analysis methods?
- Could activation sparsity signal task difficulty and guide routing decisions?
- Can training order and structure shape what networks retain and learn?
- How do models develop dense representations for familiar training data?
- Can sparsity patterns reliably indicate how well a model knows its input?
- What makes a feature abstract versus concrete in neural network activations?
- Why do multimodal models fail on rare and underrepresented concepts?
- How much do measurement thresholds affect estimates of neuron interpretability rates?
- What makes linear decodability a reliable signal of compositionality?
- Can representation engineering cleanly isolate single features in entangled semantic space?
- Do KANs maintain their advantages in deep architectures and large-scale training?
- What makes a new representational primitive valuable enough to justify its representational cost?
- What role does query-level exposure play in enabling compositional generalization?
- What inductive biases help networks segregate entities from raw inputs?
- What happens to representational structure during model pretraining phases?
- How do neural networks decompose complex tasks into modular subnetworks?
- Does task superposition explain how models learn from multiple in-context trajectories?
- Can dense models partially address modality friction without full expert specialization?
- Can we predict which tasks will decompose into modular subnetworks?
- How do sparse mixture-of-experts models resolve modality capacity competition?
- Can dense models match specialized architectures by mixing data better?
- Why do some neurons detect multiple unrelated concepts in early models?
- How does modality-specific sparsity enable capacity flexibility that dense models cannot provide?
- How do sparse circuits compare to the modular subnetworks that emerge naturally?
- Does compositional generalization emerge suddenly or improve smoothly with scale?
- What prevents representation collapse in latent-prediction world models like JEPA?
- How can neural networks be interpretable by design rather than post-hoc?
- What role does recursive generalization play in semantic drift during learning?
- What makes multimodal conditioning effective when features are decomposed to the right granularity?
- Why does scaling data and model size improve compositional generalization?
- How does representation-level reranking address residual gaps after decomposition?
- Can activation sparsity patterns guide the selection of in-context learning demonstrations?
- What solvable idealized settings reveal fundamental phenomena in realistic deep learning?
- Can curvature measurements predict task difficulty without behavioral labels?
- Can generative reconstruction preserve latent manifold structure better than geometric compression?
- How do encode-decode contractive biases create stable attractors in latent space?
- What limits the extrapolation of learned operations like rotation and reflection?
- What makes regularization an implicit factor in embedding geometry?
- What non-linear patterns do autoencoders discover that matrix factorization misses?
- Can autoencoders act as associative memory systems like Hopfield networks?
- How does joint backpropagation differ from training separate ensemble models?
- Is the superposition hypothesis necessary to explain neural network behavior?
- How do classical mechanics and statistical mechanics provide methodological templates for learning theory?
- What non-parametric methods could replace latent factors for inductive learning?
- What makes a self-supervised pruning metric work without labels at scale?
- Why are polysemantic features concentrated in early neural network layers?
- What physical structure does a Gaussian-regularized latent space actually encode?
- How do gradients flowing through both branches simultaneously reshape each component's role?
- What does leveraging internal representations during training actually mean operationally?
- Do generic kernel-decay assumptions alone explain coarse-to-fine spectral ordering?
- How do neural networks extend contextual bandits beyond linear reward assumptions?
- Why did early neural networks need massive computing power for simple image tasks?