Theme of inquiry
How do architectural decisions and system constraints affect LLM agent performance?
A question within its area, explored through 5 lines of inquiry below — each a family of specific questions the research asks.
51 specific questions
- Can parallel reasoning chains outperform longer sequential chains with the same compute?
- Can parallel thinking outperform sequential thinking under the same token budget?
- What makes parallel thinking more efficient than sequential chains?
- Why does parallel thinking outperform sequential thinking under fixed token budgets?
- When does sequential chain-of-thought dramatically beat parallel voting approaches?
- Can parallel independent reasoning outperform sequential iterative refinement?
- What advantages emerge from running 13 times more parallel reasoning chains with the same budget?
25 specific questions
- Why do continuously consolidated agent memories eventually degrade below no-memory baseline?
- Why does memory consolidation degrade agent performance below baseline?
- What makes memory consolidation fragile compared to raw trajectory storage?
- Why does uniform memory consolidation sometimes degrade below the no-memory baseline?
- When does memory consolidation help agents instead of hurting performance?
- Why does LLM memory consolidation regress below no-memory baselines?
- Why does consolidating more state sometimes hurt performance below the no-memory baseline?
32 specific questions
- Can KV cache pruning serve as an alternative to consolidation?
- Can KV blocks be safely shared between different concurrent requests without computation errors?
- Can KV cache offloading generalize beyond vLLM and SGLang inference engines?
- Can memory primitives become first-class design objects like computation sparsity?
- Why does KV cache memory waste grow worse with longer sequences and larger models?
- When should architects prioritize consolidation compute over larger context windows?
- How do memory hierarchies and compression reduce context management demands?
18 specific questions
- Which components of LLM routers actually drive performance differences?
- How much worse are convergent routing methods than theoretical oracle routing?
- Why do routers fail when only a few models have the right answer?
- How do pre-training and distillation enable minimal routing signals to work?
- Can clustering-based routing match neural network routers in cost and quality?
- What makes routing a governance mechanism rather than just a predictor?
- Why do commercial LLM routers fail to beat lightweight baselines?
27 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?
- Should model routing decisions account for prompt-tier dependencies?
- Does model selection matter more than model improvement for query routing?
- How do routing and test-time compute scaling work together as optimization axes?
- Can routing policies remain meaningful over behaviorally homogeneous model pools?