Pairing language models with external knowledge sources — search indexes, vector stores, databases — to ground answers in retrieved evidence rather than parametric memory.
Retrieval-Augmented Generation: Find the Right Evidence, Shape the Context, Audit the AnswerRetrieval-augmented generation (RAG) separates finding evidence from generating an answer. That makes knowledge easier to update, but it also creates a chain of failure points: the query may omit the decisive constraint, query and item representations may not align, the retrieved documents may form a redundant or conflicting set, and the generator may produce a correct answer without relying on its context.Start with the failure you observe. Read PACE for hidden user-specific blockers; CoGR for two-sided text matching; SetwiseEvalKit for evidence-set composition; KBMR when visual similarity confuses entity identity; and the cache-quantization and Enoki papers when you need to test and localize failures of grounding.
PACE: Towards Surfacing Hidden Conflicts in User RequestsStart here when topically similar retrieval misses hidden blockers. It explains PaceMaker's counter-view planning, hybrid search, graph traversal, ablations, and costs; the benchmark evaluates feasibility decisions rather than completed assistant actions.PACE uses synthetic personas and about 2,035 atomic facts per knowledge base; it does not test stale real-world memories or end-to-end execution.
What are you trying to solve?
Search for decisive evidence, not just similar textA request can omit the fact that determines the answer. In personalized systems, a restaurant request may require finding an allergy or calendar conflict with little lexical overlap. PACE reframes retrieval as hypothesis-driven search: propose possible conflict dimensions, issue counter-view queries, traverse related facts, and filter for evidence that changes the decision.
PACE: Towards Surfacing Hidden Conflicts in User RequestsStart here when topically similar retrieval misses hidden blockers. It explains PaceMaker's counter-view planning, hybrid search, graph traversal, ablations, and costs; the benchmark evaluates feasibility decisions rather than completed assistant actions.PACE uses synthetic personas and about 2,035 atomic facts per knowledge base; it does not test stale real-world memories or end-to-end execution.
When does ordinary semantic-similarity retrieval fail?It fails when the request describes the intended action but not the hidden condition that blocks it. PACE demonstrates this with persona knowledge bases, where decisive evidence can concern timing, personal constraints, or external state rather than the request's surface topic.PACE: Towards Surfacing Hidden Conflicts in User Requests
How can an assistant search for implicit user-specific conflicts?First hypothesize several ways the request could conflict with the user's circumstances, then generate counter-view queries for those angles. PaceMaker combines those views with hybrid retrieval, expands from promising facts through a document graph, and filters the pool before deciding feasibility.PACE: Towards Surfacing Hidden Conflicts in User Requests
Align matching, then optimize the evidence setRAG quality depends on both candidate matching and the composition of the final context. CoGR learns compact keyword representations on the query and item sides for inverted-index matching. SetwiseEvalKit evaluates documents together, exposing redundancy, conflict, complementarity, coverage, and other interactions that independent relevance scores miss.
It Takes Two to Match: Co-Evolving Generative Retriever with Reinforcement LearningUseful when query rewriting has plateaued and item-side representations may be the bottleneck. CoGR co-trains compact keyword generators for both sides while retaining inverted-index retrieval; it does not evaluate a RAG reader or generated answers.Results cover an internal app-marketplace dataset and the public WANDS product-search benchmark.
Beyond Relevance-Centric Retrieval: Rubric-Oriented Document Set Selection and RankingRead this to shift evaluation from isolated relevance scores to the quality of the evidence bundle delivered to a generator. It supplies a rubric benchmark and training-free selector, though the abstract does not give implementation details.Evidence covers short- and long-form scenarios and comparisons among 12 rerankers.
When is query expansion alone an incomplete retrieval strategy?When item representations also constrain matching. CoGR trains query- and item-side generators to emit keywords in an aligned space, then matches them through an inverted index. This directly changes both sides of candidate retrieval rather than using an LLM only to rewrite queries.It Takes Two to Match: Co-Evolving Generative Retriever with Reinforcement Learning
Why evaluate a retrieved set rather than only its individual documents?Documents interact: two relevant items may repeat each other, while individually weaker items may provide complementary evidence. SetwiseEvalKit diagnoses such interactions with multilevel rubrics, and Rubric4Setwise converts the criteria into training-free signals for selecting context for generation.Beyond Relevance-Centric Retrieval: Rubric-Oriented Document Set Selection and Ranking
Retrieve entity identity, not just visual resemblanceIn knowledge-based visual question answering, visually similar entities can require different external facts, while the same entity can vary greatly across images. KBMR maps images into a semantic embedding space and uses continuous entity-consistency weights to train against noisy supervision, targeting identity alignment rather than surface resemblance alone.
Beyond Visual Similarity: Entity-Aligned Retrieval for Knowledge-Based Visual Question AnsweringUse this when image similarity retrieves the wrong long-tail entity or misses varied appearances of the same one. KBMR targets entity-aligned embeddings with continuous semantic supervision; the abstract does not describe behavior outside knowledge-based visual QA.The reported setting is Wikipedia-scale retrieval for knowledge-based visual question answering.
When can visual similarity retrieve the wrong knowledge?When distinct entities look alike or one entity appears in varied forms. In those cases, a CLIP-style similarity signal can favor appearance over identity. KBMR is relevant when downstream answers depend on retrieving the correct long-tail entity from Wikipedia-scale knowledge.Beyond Visual Similarity: Entity-Aligned Retrieval for Knowledge-Based Visual Question Answering
How does KBMR handle noisy entity supervision?It uses an MLLM-based semantic discriminator to assign continuous entity-consistency weights, then applies those weights through semantic distillation, hard-negative sampling, and soft supervision. The abstract reports gains in retrieval Recall@1 and end-to-end VQA accuracy over CLIP baselines.Beyond Visual Similarity: Entity-Aligned Retrieval for Knowledge-Based Visual Question Answering
Audit grounding after compression and localize failuresServing optimizations can preserve headline accuracy while weakening the link between an answer and retrieved context. A single factuality score also cannot show which text failed. Use cache stress tests to detect degraded evidence use, then claim- and span-level verification to identify the unsupported content; this is an evaluation sequence, not a tested joint system.
Faithfulness Is Not Free: Auditing Offline KV-Cache Quantization in Retrieval-Augmented GenerationUseful before deploying precomputed, quantized document KV caches. It shows why accuracy and faithfulness need separate checks and identifies retrieval noise and context size as stressors; it does not compare other models or compression families.Tests use Qwen2.5-7B-Instruct with INT8 and INT4 caches on RGB and HotpotQA.
Enoki: Efficient Multi-Level Hallucination DetectionRead this when a binary hallucination flag is insufficient and debugging requires both the unsupported claim and its exact span. Enoki shares an extraction representation across verification and localization.Evaluation covers claim-, span-, and entity-level hallucination detection, including the aligned EnokiQA dataset.
Is answer accuracy enough to validate quantized retrieval caches?No. In Qwen2.5-7B-Instruct experiments on RGB and HotpotQA, INT8 document-cache quantization was near-lossless, while INT4 reduced accuracy and made more than 90% of measured faithfulness changes negative among answers that remained correct. Harm increased with noisy retrieval and more chunks.Faithfulness Is Not Free: Auditing Offline KV-Cache Quantization in Retrieval-Augmented Generation
How can a detector verify claims and locate unsupported spans without separate alignment?Enoki extracts text-anchored relational facts, verifies them against evidence, and projects unsupported facts back to their source spans. Its shared representation links claim-level judgments to span- and entity-level localization while supporting LLM-, encoder-, or rule-based extraction.Enoki: Efficient Multi-Level Hallucination Detection