Theme of inquiry
How do systems prioritize structural meaning over statistical patterns?
A question within its area, explored through 7 lines of inquiry below — each a family of specific questions the research asks.
11 specific questions
- Can filtering unknown examples during fine-tuning prevent hallucination increases?
- Why does model confidence fail to detect hallucinations on rare entity pairs?
- Does cross-example gradient contamination explain finetuning-induced hallucination patterns?
- Why does model confidence fail to detect hallucinations about rare entities?
- Can scaling up contradictory training data overcome unpredictable override effects?
- Can pretraining-frequency signals alone prevent RAG systems from confabulating about common knowledge?
- Why do interventions for hallucination or automation bias fail to address capability misattribution?
38 specific questions
- Does transformer attention architecture systematically bias models toward sycophancy?
- Why does transformer attention architecture reinforce sycophancy and agreement?
- Does transformer attention architecture inherently bias models toward sycophancy?
- Why do transformer attention patterns show positional and sequential bias across tasks?
- How does transformer attention structurally bias models toward prominent and repeated content?
- What role does attention structure play in creating position bias?
- How does transformer attention bias toward repeated and context-prominent content?
41 specific questions
- What makes sparse attention more reliable for long-context retrieval?
- Why do longer sequences tolerate higher sparsity than shorter ones?
- How do retrieval heads achieve sparse attention naturally in transformers?
- Does attention linearity alone explain the efficiency gains over standard transformers?
- Which attention heads are essential for maintaining factuality in sparse models?
- Can retrofitted sparse attention ever match jointly-trained sparse attention?
- Why does attention quality degrade as context length increases?
19 specific questions
- How do mechanistic features compare to natural language for interpretability?
- Can mechanistic interpretability explain explanation-execution disconnection?
- How much do mechanistic interpretability findings reflect true reasoning architecture?
- Can interventions on model components prove mechanism without explaining encoding?
- What makes representation engineering better than mechanistic interpretability for detecting hidden objectives?
- How does mechanistic interpretability complement learning mechanics in explaining deep learning?
- Can mechanistic interpretability tools decode the biases alignment training conceals?
79 specific questions
- Do long-term memory modules outperform consolidation into fast weights?
- Can precomputed inferences be stored in memory modules between model interactions?
- What persistent memory architectures best support storing precomputed inferences across sessions?
- Does including full context always degrade memory retrieval quality in practice?
- Why do accumulated memory systems sometimes hurt continual learning?
- Can data pruning strategies exploit the finite nature of memorization capacity?
- How do adaptive memory modules compare to feedback-based working memory for long context?
32 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?
- What role does visual perception play alongside accessibility tree information?
- Should GUI perception happen inside or outside the foundation model?
22 specific questions
- Can fixing hallucination address AI's structural epistemic problem?
- How do external safeguards like retrieval augmentation prevent hallucination?
- Does inevitable LLM hallucination make detection metric validity critical?
- Why do language models hallucinate even with perfect training?
- How does interleaving reasoning with action prevent hallucination in language models?
- Can architectural changes reduce hallucination without external retrieval or verification?
- Why is hallucination the wrong term for all LLM false outputs?