Theme of inquiry
What determines model reasoning capability at inference and training time?
A question within its area, explored through 7 lines of inquiry below — each a family of specific questions the research asks.
141 specific questions
- Does scaling reasoning capability create tradeoffs with instruction following?
- What evidence shows that reasoning chains encode token-level functional structure?
- Can minimal reasoning steps match verbose reasoning accuracy?
- Do reasoning models perform genuine logical evaluation or pattern matching?
- What sparse mechanistic structures drive reasoning traces in language models?
- Why does extended reasoning fail for search and knowledge retrieval tasks?
- Can recursive sub-calls decompose reasoning across multiple context chunks?
41 specific questions
- Can causal steering change verbalization without changing internal representation?
- Does this reasoning steering method work consistently across all model sizes?
- Can activation-space steering vectors replicate thinking model performance without retraining?
- Can activation steering causally control evaluation framing effects across downstream tasks?
- Can targeted activation steering surface latent reasoning in base models?
- Can activation patching reveal which reasoning steps actually matter?
- Can steering vectors prove that representations are genuinely organized?
27 specific questions
- Can reasoning learned from language modeling actually transfer to knowledge-intensive domains?
- What makes knowledge-rich specialized domains structurally different from general reasoning tasks?
- Why does general reasoning not transfer to knowledge-intensive medical domains?
- Can mathematical reasoning improvements transfer across problem subdomains?
- Which domains need knowledge injection versus reasoning-focused training?
- How does cross-domain reasoning transfer differ from domain-specific knowledge transfer?
- Why do medical and math domains need different types of model improvements?
46 specific questions
- Do base models contain latent reasoning that minimal training can unlock?
- Can minimal training signals unlock latent reasoning capability in base models?
- Does the base model already contain latent reasoning capability?
- Can latent reasoning scale test-time compute without verbalized tokens or special training?
- Why might latent reasoning capture types of thinking that verbalized CoT cannot?
- Do base models truly possess latent reasoning capability?
- Can models possess latent reasoning capability that training signals fail to unlock?
124 specific questions
- Why do instruction following and reasoning capability trade off in training?
- Can fine-tuning ever teach semantic inference instead of amplifying training shortcuts?
- Why do reasoning gains resist clear attribution to specific training changes?
- Can reasoning fine-tuning improve both capability and instruction compliance together?
- Why does instruction tuning hurt knowledge-intensive tasks more than reasoning tasks?
- How does optimizing for accuracy during training degrade downstream reasoning quality?
- Does fine-tuning models for specific tasks destroy their ability to reason?
91 specific questions
- What mechanisms drive test-time compute allocation in reasoning tasks?
- How should we allocate compute between reasoning and retrieval iterations?
- Can test-time compute allocation shift from solutions to strategies?
- How should inference budgets adapt based on prompt difficulty?
- Can inference budgets be allocated adaptively based on prompt difficulty?
- Does test-time compute actually substitute for having larger model parameters?
- Where does inference compute stop substituting for model capacity?
21 specific questions
- Why does training data format shape reasoning strategy more than domain content?
- How much does training data format influence reasoning strategy versus domain content?
- How does training format shape reasoning strategy more than content?
- Does training data format shape reasoning strategy more than domain content?
- Why does training data format shape reasoning strategy more than content?
- Does training data format shape model reasoning more than domain content?
- Can training format itself shape what reasoning strategy a model learns?