Line of inquiry
Inquiring lines›What determines success when AI in…›How should we design retrieval sys…›this line of inquiry
When should retrieval be triggered to augment language model reasoning?
A broader line of inquiry — a family of 120 specific questions the research asks around this. Follow one into its inquiring-line page, or move sideways to a related line below.
Questions in this line of inquiry 120
Specific inquiring lines the field asks around this — ordered from the most general framing down to the most specific angle.
- Should retrieval be triggered always or only for difficult questions?
- What causes the retrieval-augmented generation to fail in practice?
- How should retrieval systems decide when to fetch new information?
- What role does document reranking play alongside decisions about whether to retrieve?
- Can long-context models replace retrieval-augmented generation systems?
- When does long-context LLM reasoning fail where structured retrieval succeeds?
- Why does search-augmented generation still not solve the verification problem?
- What makes proactive tool retrieval better than single-round semantic matching?
- Could eliminating retrieval entirely work better than shifting the burden?
- Do retrieval systems help language models overcome their inability to estimate idea value?
- How does uncertainty-gated retrieval compare to continuous retrieval efficiency?
- Why do retrieval-augmented generation systems fail to detect knowledge conflicts?
- Can adaptive retrieval triggered by model uncertainty improve RAG reliability?
- Can generator feedback backpropagate through the entire retrieval pipeline?
- How does query decomposition reduce retrieval costs at inference?
- Does retrieval iteration improve accuracy only after strong first-stage ranking?
- Should retrieval be triggered by model uncertainty or fixed intervals?
- Are uncertainty estimation and external feature signals complementary for retrieval?
- How do confidence thresholds compare to learned policies for triggering retrieval?
- Why does standard RAG succeed for evidence-based but fail for debate questions?
- How much does retrieval budget improve when triggered by dual signals instead of fixed intervals?
- Does tail distribution collapse in training predict retrieval failure patterns?
- Why does query routing matter for retrieval-augmented systems?
- Can retrieval systems decide when to retrieve instead of always querying?
- How does retrieval-augmented generation create topically redundant content patterns?
- How do case memory and Q-function updates enable better retrieval decisions over time?
- Does retrieval quality depend more on access structure or write gating?
- How do retrieved documents in RAG systems compound input length problems?
- What would instruction-following retrieval enable that query-only systems cannot?
- Can relevance judgments based on query resemblance miss essential factual connections?
- Can learned verifiers detect structural near-misses that pooled retrievers miss?
- Can retrieval of correct information guarantee it will shape model behavior?
- Do single-step retrieval systems with sophisticated synthesis qualify as deep research?
- How can inference-time retrieval avoid the domain boundary problem?
- Can retrieval augmented generation systems defend against corpus poisoning without retraining?
- Can retrieval strategies drive both draft refinement and new research question generation?
- How does reflection-based query refinement differ from single-pass retrieval strategies?
- Does full conversation history improve or degrade multi-turn retrieval accuracy?
- Are retrieval heads the mechanistic explanation for needle-in-haystack performance failures?
- Why does adaptive document allocation improve over fixed k selection?
- What makes process-level supervision better than outcome-only rewards for RAG training?
- Why do conversational systems struggle more than static retrieval with ambiguous queries?
- How does response content compare to model confidence as a retrieval trigger?
- Can factually wrong generated documents still improve retrieval accuracy?
- Can task-aware ranking replace similarity scoring in other RAG systems?
- Does filtering passages before generation improve large model answer quality?
- Why do pretrained retrievers struggle with ambiguous or implicit queries?
- Does grep-style corpus search outperform dense retrieval on entity-heavy questions?
- Does retrieval augmented generation actually improve language model answer diversity?
- What makes web retrieval more effective than static knowledge bases?
- How does retrieval-augmented generation extract structured properties from domain descriptions?
- Why does single-round retrieval fail on multi-step tasks across different domains?
- How do retrieval systems handle feedback expressed as negations rather than preferences?
- When do queries fail to capture relevance patterns effectively?
- Can retrieval behavior be compressed into a small parametric decoder?
- How do time-based and entity-based queries differ from semantic similarity retrieval?
- Do Doc2Query approaches suffer from the same misaligned-target problem?
- Can precision and recall metrics work without a ground truth?
- How does semantic mismatch between user language and API documentation degrade tool retrieval?
- What makes reranking during retrieval better than catching failures at plan time?
- What makes legal and medical queries particularly vulnerable to structural near-misses?
- How does gist-first lookup compare to pure retrieval or context stuffing?
- How does overthinking in early turns degrade later retrieval rounds?
- How does time-partitioned routing compare to retrieval-augmented temporal grounding?
- What execution cost does computing relevance scores add to grep traversal?
- Does RL pruning of documents differ fundamentally from rationale-driven evidence selection?
- What threshold combinations for uncertainty and rarity signals maximize RAG performance?
- Can temporal ranking improve retrieval without modifying the underlying video model?
- Why does domain-specific terminology require customization of vector search and generation?
- Can reranking candidate summaries improve perspective representation better than prompting?
- Why does explicit reasoning degrade passage reranking performance?
- How do pseudo-relevance labels enable training without ground truth relevance judgments?
- How do comparison and debate questions differ in their aspect retrieval needs?
- What makes task similarity the right retrieval key for confidence calibration?
- Why do longer queries benefit less from clarification questions?
- Can adaptive elbow detection replace fixed top-k limits in evidence retrieval?
- How does merging retrieval and generation shift the computational bottleneck in dialogue systems?
- Does uncertainty trigger retrieval better than fixed-interval tool calls?
- Can token-level verification catch structural mismatches that pooled relevance scores miss?
- What hidden costs might fine-tuning retrieval models introduce on out-of-distribution queries?
- Why does long-form generation need different retrieval than factoid questions?
- How should retrieval triggers use model uncertainty instead of fixed intervals?
- What are the 27 external features that predict retrieval need?
- Why does GraphRAG prioritize corpus completeness while LogicRAG prioritizes query adaptivity?
- What makes prerequisite filtering more reliable than semantic similarity matching?
- Could real-time search systems avoid era sensitivity in legal reasoning?
- Why do fixed-size document chunks break complex procedural question answering?
- Why does retrieval quality sometimes conflict with final answer quality?
- Can semantic query expansion overcome vocabulary mismatch in corrupted text?
- How does retrieval-augmented training reduce domain specialization cliff failures?
- Can task-agnostic compression of documents remain broadly useful for later queries?
- Can the same description-then-retrieve pattern work for domain adaptation without target data?
- How should temporal metadata indexing differ from semantic indexing?
- What language skills matter most for entity extraction from retrieval context?
- How can affordance become a primary retrieval signal instead of a filter?
- How does representation-level reranking address residual gaps after decomposition?
- Why do question types determine retrieval and decomposition strategy in QA?
- How does temporal grounding in retrieval compare to architectural approaches?
- How does MaxSim reranking differ from structural verification at the token level?
- Why does RAGU retrieve more complete context than HippoRAG 2 despite smaller model size?
- How should a reranker adjust both document order and retrieval count dynamically?
- How can frame sampling and ranking improve temporal understanding in long-video retrieval?
- Why do standard RAG systems struggle with pronouns and demonstratives?
- How much does using full PDF text improve over abstract-only retrieval?
- Why do external feature triggers outperform uncertainty on complex questions?
- Can other RAG hyperparameters like chunk size be learned through generator feedback?
- What documents improve answers beyond surface query similarity?
- Why does production retrieval augmented generation underperform in real deployments?
- Can a single meeting summary format serve both scanning and reference needs?
- What design tradeoffs exist between pure ID and pure text indexing?
- What concrete failures happen when RAG ignores temporal relevance?
- How severely do minimal corpus modifications damage RAG accuracy in practice?
- How does ranking-aligned summarization compare to aspect-controlled generation methods?
- Why do RAG systems fail when demo queries work correctly?
- How should query augmentation strategies be properly evaluated against baselines?
- How many books does it take before raw navigation collapses completely?
- What classifier accuracy is needed to assign memory roles reliably at retrieval time?
- Why does bidirectional RAG amplify the risk of corpus poisoning attacks?
- How can gradients flow through discrete document selection?
- How much of the incomplete candidate data came from Wikipedia versus live retrieval?