Theme of inquiry
How should model design balance architectural complexity and computational cost?
A question within its area, explored through 6 lines of inquiry below — each a family of specific questions the research asks.
63 specific questions
- Does including full context always degrade memory retrieval quality in practice?
- How does externalized state affect the long-context bottleneck in language models?
- How do recurrent memory systems handle ultra-long context differently than attention?
- Can context compression preserve what matters without introducing bias?
- What makes sparse attention more reliable for long-context retrieval?
- How does context budget create tradeoffs between memory and skills?
- How does accumulated context history degrade iteration quality in long-horizon tasks?
71 specific questions
- Do smaller models optimized for benchmarks perform worse than their scores suggest in real tasks?
- Why do benchmark scores rise while reasoning quality declines?
- Do reasoning benchmarks predict real performance in long delegated workflows?
- Should benchmarks measure trace length or whether constraints were actually satisfied?
- Can benchmark scores on verifiable tasks transfer to unseen problems outside the training domain?
- Why do estimates for task-level performance differ so much from full job automation timelines?
- How can benchmark accuracy scores mask the absence of interpretable reasoning structure?
42 specific questions
- Can inference budgets be allocated adaptively based on prompt difficulty?
- How should inference budgets adapt based on prompt difficulty?
- How should we allocate compute between reasoning and retrieval iterations?
- What mechanisms drive test-time compute allocation in reasoning tasks?
- How should inference compute budget be allocated across different prompt difficulties?
- What makes inference budgets allocate adaptively per prompt difficulty?
- Can inference budgets be allocated differently based on prompt difficulty?
48 specific questions
- Can looped architectures achieve reasoning abilities that fixed-depth models cannot?
- How does hierarchical recurrence compare to selective layer looping for computational depth?
- Can looping enable reasoning capabilities that fixed-depth transformers fundamentally cannot achieve?
- Why does recursion on latent states improve generalization more than scale?
- Why does recursion on latent state drive generalization better than hierarchy?
- Why does reapplying the same computation stages improve model performance?
- Can latent recurrence achieve the depth that standard transformers cannot?
42 specific questions
- Can routing ever beat model improvement as a performance lever?
- Can multiple small models outperform a single large model with good routing?
- What makes routing a better investment than training larger models?
- Does model selection matter more than model improvement for query routing?
- Should model routing decisions account for prompt-tier dependencies?
- How does routing decide between models before generation happens?
- Can routing policies remain meaningful over behaviorally homogeneous model pools?
59 specific questions
- How does test-time compute substitute for model parameter scaling?
- Does test-time compute actually substitute for having larger model parameters?
- Can test-time compute scaling substitute for larger model parameters?
- Where does inference compute stop substituting for model capacity?
- Do models excel at reasoning depth or memory breadth when scaling test time compute?
- Does inference-time compute improve pretraining data efficiency in practice?
- What patterns emerge across test-time scaling and reasoning architectures?