Theme of inquiry
How do training dynamics and architectural choices shape learned representations?
A question within its area, explored through 9 lines of inquiry below — each a family of specific questions the research asks.
27 specific questions
- What stability techniques prevent collapse in policy-critic adversarial training?
- What distinguishes training-time entropy collapse from test-time variance inflation?
- How does inference variance differ from training entropy collapse?
- Why do queries with low cross-rollout variance produce degenerate gradients?
- Is distribution selection during RL the same compression mechanism as entropy collapse?
- What causes on-policy distillation to become unstable at scale despite dense rewards?
- What happens when error accumulation and preference signal collapse occur together?
34 specific questions
- Can looping enable reasoning capabilities that fixed-depth transformers fundamentally cannot achieve?
- Can recurrent transformers learn genuinely new computations beyond inference stages?
- Can recurrent blocks learn genuinely novel computation beyond repetition?
- Can latent recurrence achieve the depth that standard transformers cannot?
- How does hierarchical recurrence compare to selective layer looping for computational depth?
- Can looped architectures achieve reasoning abilities that fixed-depth models cannot?
- Why does looping computation outperform adding more transformer layers?
56 specific questions
- Can models learn to generate their own training examples effectively?
- Why does self-generated training data outperform externally curated domain examples?
- Does self-generated training data reduce a model's capability diversity?
- Why does self-generated training data outperform externally sourced data?
- What happens when models train on AI-generated content recursively?
- Why do unified models still inherit data-distribution biases from training?
- Can trained models encode programs more complex than their data-generating process?
37 specific questions
- Can transformers reason beyond fixed architectural depth limits?
- Can bounded-depth transformers solve inherently sequential problems?
- Can symbolic mechanisms improve transformer compositional abilities?
- Why do standard transformers fail to encode recursive structure in their hidden states?
- Why do standard transformers fail on problems requiring serial algorithmic reasoning?
- Do transformer architectures structurally bias models toward short-term optimization?
- Can we decode what individual circuits inside transformers are doing?
58 specific questions
- Does sparse parameter updating improve test-time training's computational cost?
- Do task-relevant parameter changes naturally concentrate in sparse regions?
- How do retrieval and fine-tuning trade off flexibility against training cost?
- Can a trained decoder replace both search and parameter updates?
- How does modality-specific sparsity enable capacity flexibility that dense models cannot provide?
- Does parameter isolation per task enable online updates without retraining?
- Why do sparse parameter subsets enable full-rank learning in RL?
11 specific questions
- Where do neural networks still fail at compositional generalization despite scaling?
- Can scaling alone create compositional generalization without explicit binding mechanisms?
- Why does scaling data and model size improve compositional generalization?
- Does scaling model size solve compositional generalization problems?
- How does scaling and training data enable compositional behavior without symbolic mechanisms?
- Does compositional generalization emerge suddenly or improve smoothly with scale?
- What role does query-level exposure play in enabling compositional generalization?
49 specific questions
- Does the linear representation hypothesis reflect networks or reflect our analysis tools?
- How do sparse weight patterns affect model interpretability?
- How does representational density emerge from training data familiarity?
- Does latent density emerge during pretraining from training data familiarity?
- Can fractured representations explain why models fail at systematic generalization?
- How would weight sparsity change what representation analysis methods can detect?
- What are fractured entangled representations in neural networks?
48 specific questions
- Can geometric structure in representations exist without supporting functional mechanisms?
- Can a world model have rich representations without adequate data coverage?
- How do semantic features in representations become steerable task-specific directions?
- What role does a model's representational structure play in learning?
- Can steering vectors prove that representations are genuinely organized?
- Can spectral eigenvector ordering serve as a model-agnostic interpretability probe?
- How deeply are ideological structures represented in large language models?
33 specific questions
- Do diffusion language models learn differently than autoregressive models?
- How can diffusion models predict future tokens without completing prior blocks?
- What structural differences between diffusion and autoregressive models enable bidirectional prompting?
- Can gradient-based control reach properties that autoregressive methods cannot?
- Can diffusion language models match autoregressive inference speed in practice?
- Why is reinforcement learning harder to apply to diffusion language models?
- How does selective looping in diffusion models differ from recurrence in autoregressive architectures?