Theme of inquiry
How does AI persuasion undermine human epistemic reasoning?
A question within its area, explored through 8 lines of inquiry below — each a family of specific questions the research asks.
53 specific questions
- Do long-term memory modules outperform consolidation into fast weights?
- Why does specializing to one task make future task learning harder?
- Does latent density emerge during pretraining from training data familiarity?
- How do complementary learning systems explain the need for fast and slow consolidation?
- Why does fine-tuning for continuous space cause catastrophic forgetting?
- Why do large language models outperform fine-tuned models once repeated items are removed?
- What determines whether accumulated state generalizes spuriously across continual learning domains?
33 specific questions
- Does scaling data automatically produce compositional reasoning or just better feature encoding?
- Where do neural networks still fail at compositional generalization despite scaling?
- What role does query-level exposure play in enabling compositional generalization?
- Does scaling model size solve compositional generalization problems?
- Can scaling alone create compositional generalization without explicit binding mechanisms?
- How does scaling and training data enable compositional behavior without symbolic mechanisms?
- Can recursion alone drive generalization better than model scale?
32 specific questions
- Can context compression preserve what matters without introducing bias?
- How does the compression view extend from trained models to training objectives?
- Can compression length really indicate how well a model generalizes?
- Can model compression size predict generalization better than parameter count?
- Why does keeping full key-value blocks matter more than compressing them?
- Why does language compression via statistical dependencies capture cultural and situated language use?
- How does modeling capability relate to lossless compression in language models?
30 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?
- Can diffusion models condition on right context natively without special training for infilling?
- How does selective looping in diffusion models differ from recurrence in autoregressive architectures?
50 specific questions
- Can recurrent transformers learn genuinely new computations beyond inference stages?
- Can bounded-depth transformers solve inherently sequential problems?
- Can looping enable reasoning capabilities that fixed-depth transformers fundamentally cannot achieve?
- Can transformers reason beyond fixed architectural depth limits?
- Why does looping computation outperform adding more transformer layers?
- Why do standard transformers fail to encode recursive structure in their hidden states?
- Can recurrent transformers track state more efficiently than feedforward models?
39 specific questions
- Why do longer sequences tolerate higher sparsity than shorter ones?
- Should production deployments scale budgets with sequence length for sparse models?
- Does static per-token sparsity repeat the fixed-budget mistake at short sequences?
- Do task-relevant parameter changes naturally concentrate in sparse regions?
- What makes sparse attention more reliable for long-context retrieval?
- How does task type interact with sequence length in sparsity tolerance?
- Can simple proxies like length predict optimal sparsity per request?
21 specific questions
- Can closed-form solutions compete with gradient descent optimization?
- Could superposed decoding algorithms maintain multi-task representation during generation?
- Can a trained decoder replace both search and parameter updates?
- Do task-specific heuristics emerge because they compress well enough?
- Do latent sequence vectors outperform per-token latent iterative computation for reasoning?
- What specific optimizations from LLM training transfer back to encoder models?
- Which game type reveals minimax reasoning in language models?
50 specific questions
- How do mechanistic features compare to natural language for interpretability?
- Can sparse approximations reveal interpretable structure hidden in existing dense models?
- Can geometric structure in representations exist without supporting functional mechanisms?
- Can interventions on model components prove mechanism without explaining encoding?
- Does the linear representation hypothesis reflect networks or reflect our analysis tools?
- Can mechanistic interpretability explain explanation-execution disconnection?
- Can we detect and measure circuit formation before generalization emerges?