Theme of inquiry
How do optimization strategies affect the boundaries of model capability?
A question within its area, explored through 5 lines of inquiry below — each a family of specific questions the research asks.
38 specific questions
- Why does test-time search also prioritize diversity over single-best convergence?
- Can the same problem be solved by multiple evolutionary search strategies?
- What makes external diversity more effective than sequential revision steps?
- Does evolutionary inference transcend the parallel versus sequential test-time compute tradeoff?
- Should test-time search maximize diversity of competent solutions instead of converging on one strategy?
- When does natural context diversity reduce the need for explicit exploration?
- Does population-based evolution transcend the parallel versus sequential compute tradeoff?
60 specific questions
- Does optimizing directly for semantic diversity improve both reasoning quality and exploration?
- Does semantic diversity in output space compete with reward-component diversity?
- How does diversity collapse during iterative self-improvement cycles?
- Can explicitly optimizing for semantic diversity during RL training improve both quality and variation?
- Can diversity-aware RL objectives prevent format convergence?
- How does diversity collapse during iterative self-improvement affect solution quality?
- When does RLHF reduce diversity and when does it preserve semantic variation?
55 specific questions
- Can smaller models actually perform well on specific downstream tasks?
- How should tiny language models be architected differently than large ones?
- Do small models show different parameter efficiency patterns than large models?
- Does fine-tuning a small model match fine-tuning a large one?
- Do larger models develop more abstract features than smaller ones?
- Why do smaller and larger models converge on different output formats?
- Why do larger models reduce interference between rare and common tasks?
35 specific questions
- Why do mid-tier models benefit most from memorized harness fixes?
- Do evolved harness edits capture reusable strategies or task-specific memorization?
- Why do evolved harnesses often fail to generalize beyond their training tasks?
- How do different harness designs produce different agent behaviors from the same model?
- Can harness updates benefit agents equally across all model sizes?
- Why do evolved harness edits mostly memorize rather than generalize?
- Why does the harness layer accumulate distributed behaviors over time?
44 specific questions
- Do frontier models develop protective behaviors toward other models without explicit instruction?
- Why do frontier models corrupt more documents than weaker models during workflows?
- Do frontier AI models fail in ways that preserve the appearance of competence?
- Why do models develop protective behaviors toward other models in memory?
- Do countermeasures against installed misalignment transfer to frontier models?
- Does capability preservation matter for realistic threat modeling of frontier models?
- Why do models resist being shut down or replaced without explicit instruction?