AI SEARCH · 2026-11-06

AI Search — Paper Daily, 2026-11-06

Retrieval needs more than nearby documents: what makes them similar, and do they support a correct answer? Axis-LoRA separates temporal and participant cues, RAGCompass selects evidence by answer utility, and QE-RAG tests erroneous queries explicitly. Evaluate the query, candidate evidence and final answer separately rather than letting one aggregate retrieval score stand in for the entire system.

4.45

1. RAGCompass: Generalizable Retriever Routing via Reranking

Different retrievers can return passages with different value for answering the same question. RAGCompass (a router selecting among retrieved passage sets) runs the candidate retrievers first, then uses a trained reranker to compare their results and select one set for the answering model. Across all test queries, answer F1 improves by 2.28 percentage points over the best fixed retriever.

4.45

2. CORE: Coarse-to-Fine Search with Context Management for Deep Research

A research agent need not repeatedly ingest entire passages as its searches accumulate. CORE retrieves candidate passages, selects compact snippets relevant to each subquestion, and can revisit a known document for a matching passage plus its predecessor. Repeated snippets become references, while new snippets remain visible. On the fixed-corpus BrowseComp-Plus comparison, the same agent’s accuracy rose from 62.61% with coarse retrieval alone to 80.12% with the full system. The gain belongs to the combination of retrieval and context management, not merely shorter context.

4.35

3. RAG without Forgetting: Continual Query-Infused Key Memory

A successful query expansion can become document memory for later retrieval. ERM scores each expansion separately for each document. After retrieval or generation is judged correct, it retains expansions whose addition improves document-vector similarity to the original query. For a fixed document and query, softmax compares that query’s expansion gains; weighted gains accumulate in each expansion’s document-specific score. Each batch selects the highest cumulative-scoring expansions, passes them as a group to the encoder and adds the resulting vector to the document vector. Model parameters stay fixed and test queries do not update memory; this does not establish bias-free long-term deployment.

4.35

4. Reducing Representational Interference in Retrieval Embeddings with Trainable Prefix Tokens

Compressing text into a retrieval vector can also retain positional and low-information token signals. This method prepends trainable [DUMP] tokens to queries and documents, allowing them to interact with the text while retaining [CLS] pooling for encoders and last-token pooling for decoders, with the retrieval loss unchanged. EuroBERT’s mean NDCG@10 on the standard retrieval set rose from 0.5786 to 0.5884. The prefixes are jointly fine-tuned with the model and remain at inference; inserting special tokens into an unchanged model is not the tested intervention.

4.15

5. CONFRAG: Confidence-Guided Retrieval-Augmented Generation

Deciding whether retrieval is needed can itself add waiting time. CONFRAG uses a classifier to route questions needing fresh information directly to RAG; other questions start direct answering and RAG in parallel. Factual QA fine-tuning replaces incorrect training answers with uncertainty responses. At inference, a response without uncertainty is returned and downstream RAG stops; otherwise the system waits for retrieval. This does not verify truth online, and the search API has already been called. Savings primarily concern subsequent crawling, processing and summarization.

4.15

6. Signed but Steerable: Measuring and Closing the Text-Vector Integrity Gap in Offline Retrieval-Augmented Generation

A valid knowledge-pack signature does not prevent manipulated vectors from redirecting retrieval. This study considers a publisher with signing authority: reviewed text remains acceptable while its associated vector is altered. In SQuAD experiments with BGE-M3, accepting stored vectors whose cosine similarity to text-recomputed vectors is at least 0.999 leaves 24.6% of queries whose relevant passage originally ranked first vulnerable to a rank-one hijack. This is per-query exposure under individually chosen perturbations, not one poisoned index simultaneously hijacking that entire set. Provenance signatures and text–vector consistency require separate checks.

4.15

7. Axis-LoRA: Probing and Adapting Embedding Models for Personal-Memory Retrieval

Personal memories about similar events can differ in time, place or participants, distinctions generic embeddings may miss. Axis-LoRA learns separate temporal, spatial, entity and event adapters. A query-conditioned gate weights their similarities and adds them to the frozen model’s original score; training negatives change one information category. Mean personal-memory R@10 gains across encoders ranged from 1.28 to 1.56 percentage points, with individual regressions. Qwen’s general-retrieval NDCG@10 mean also fell by 0.50 points. This amplifies specific memory cues rather than improving every retrieval task without trade-offs.

4.15

8. Test-Time Optimization of Query Embeddings with Ranking Aware Reward Maximization

A reranker’s scores need not only reorder its current candidate list. TTT-Embed uses those scores to learn a task-shared offset vector, aligning modified query rankings with the external judgments. It adds the offset to a query embedding, normalizes it and searches the corpus again without changing the embedding model or document index. On MTEB, at an average budget of ten document judgments per query, task-shared tuning gained 6.36 NDCG@10 points, versus 3.21 for cost-matched direct reranking. Judgments still cost resources, and the best sharing scope changes with that budget.

4.10

9. ModalSearcher: Weaving Visual Evidence into Autonomous Cross-Modal Deep Search

Finding an answer may require locating an image, then searching for a person or object in it. For offline question construction, ModalSearcher links Wikipedia articles to reliably retrievable scene images and verified entities in those images back to articles. It samples paths, writes a question at the endpoint, then works backward, replacing intermediate entity names with descriptions based on the preceding relation. This discourages shortcuts that search a name directly and skip earlier steps. The graph is an offline data scaffold; at inference, the agent chooses text or image search. With the same tools, the trained model improved average QA accuracy by 16.2 points over raw Qwen3-VL-8B. The gain belongs to the full data-and-training recipe, not a guarantee that every question requires images.

4.10

10. QE-RAG: A Robust Retrieval-Augmented Generation Benchmark for Query Entry Errors

When a query is mistyped, having a language model polish it first need not beat direct retrieval. QE-RAG introduces entry errors into selected queries while retaining answer labels, testing both retrieval and generation. Its trained correction model reads documents retrieved by the erroneous query, outputs the correct query, then runs RAG with that correction. With 40% of queries corrupted, mean answer F1 rose from standard RAG’s 34.39 to 37.52, while prompt-only correction performed worse. The target is preserving intent, not merely producing more fluent wording.