Theme of inquiry
How do systems prioritize meaningful information across modalities and scales?
A question within its area, explored through 5 lines of inquiry below — each a family of specific questions the research asks.
85 specific questions
- Do transformer architectures structurally bias models toward short-term optimization?
- How do retrieval heads achieve sparse attention naturally in transformers?
- Why do transformer attention patterns show positional and sequential bias across tasks?
- How do attention patterns and circuits function as algorithmic representations?
- Can symbolic mechanisms improve transformer compositional abilities?
- Can recurrent transformers learn genuinely new computations beyond inference stages?
- How do induction heads learn to overwrite computational representations?
34 specific questions
- Can hallucination be understood as a structural feature rather than a bug?
- Can fixing hallucination address AI's structural epistemic problem?
- Why do language models hallucinate even with perfect training?
- Why do hallucination rates differ between vendor AI products and student-used models?
- How do external safeguards like retrieval augmentation prevent hallucination?
- Can architectural changes reduce hallucination without external retrieval or verification?
- Does inevitable LLM hallucination make detection metric validity critical?
54 specific questions
- How does semantic association differ from mechanistic causal reasoning?
- Can activation patching reveal which reasoning steps actually matter?
- Can external actions provide causal necessity that language models lack?
- Why do causal reasoning directions succeed while temporal reasoning directions fail?
- Do LLMs show stronger reasoning about causality than about temporal ordering?
- How might human-LLM teams reinforce each other's causal reasoning mistakes?
- What are collider structures and why do they reveal reasoning errors?
43 specific questions
- Should GUI agents use intermediate structured representations instead of raw pixels?
- Why does explicit screen parsing outperform pure vision in GUI agents?
- Can specialized perception components replace end-to-end vision in GUI agents?
- Can screen perception be effectively decoupled from planning in GUI agents?
- Why does identifying UI element types and locations enable downstream task learning?
- Should GUI perception happen inside or outside the foundation model?
- What role does visual perception play alongside accessibility tree information?
53 specific questions
- Why does LLM memory consolidation regress below no-memory baselines?
- What makes memory consolidation fragile compared to raw trajectory storage?
- Why does uniform memory consolidation sometimes degrade below the no-memory baseline?
- Why do external memory consolidation systems fail worse than naive in-context learning on continual tasks?
- Why do accumulated memory systems sometimes hurt continual learning?
- Why do accumulated memories hurt continual learning more than no memory?
- Why do continuously consolidated agent memories eventually degrade below no-memory baseline?