AI SEARCH · 2026-11-18

AI Search — Paper Daily, 2026-11-18

Today's retrieval track asks whether the first-stage search is bringing in what actually needs to be read. Community image-text search has to write intent into retrieval, long documents have to recover figures and cross-page relations together, and multi-hop question answering still has to follow links on pages already in hand. A search agent can be led astray by copied echoes, and a visual query may already be warped or blurry before it reaches the retriever. Elsewhere, people are asking whether embeddings should run the same layers again, whether spurious attributes should be erased first, whether a hyperbolic space can keep hierarchy, and whether a smaller biomedical encoder can beat a larger one after retrieval-oriented pretraining. A blog on reranking adds a reminder: vector search can land near the right topic without ranking the right chunk first, and rescoring candidates against the question is often easier to measure than swapping in a larger model.

4.73

2. EchoSpec: Training Search Agents to Weigh Evidence Above Echoes

Tsinghua University📖 Deep dive →

In real corpora, one claim is copied and paraphrased across many pages, while the primary source that settles a question often appears once. Search agents trained with reinforcement learning treat those echoes as if they were independent sources: eight paraphrased false-claim echoes overturn 58% to 73% of previously correct answers.

4.47

4. LMEG-Doc: Layout-Grounded Multi-View Evidence Graphs for Multimodal Long-Document Retrieval

Capital OneUSC📖 Deep dive →

When evidence in a long multimodal document is spread across pages and modalities, ranking each element only by its own similarity to the question can leave relevant figures and tables low in the list. LMEG-Doc first builds an evidence graph from layout and context, then uses a vision-language model to add relations across distant pages, and propagates multimodal retrieval scores along the graph so connected supporting evidence is recovered together.

4.47

5. Fix Before Search: Benchmarking Agentic Visual Query Pre-processing in Multimodal Retrieval-augmented Generation

ICML 2026Michigan State UniversityJilin University

In multimodal retrieval-augmented generation, text queries are often rewritten first, but visual queries are still treated as clean, fixed inputs. Real visual queries are frequently warped, degraded, or ambiguous, and retrieval can then fail badly. V-QPP-Bench casts visual query pre-processing as an agent decision: the model must diagnose the defect itself and call perceptual tools to repair the query. The evaluation covers 46,700 imperfect queries and a range of multimodal retrieval pipelines. Visual defects sharply lower retrieval recall and end-to-end performance; oracle pre-processing can recover near-perfect results, but off-the-shelf multimodal models still struggle to choose tools and predict their parameters, while supervised fine-tuning lets compact models match or surpass larger proprietary ones. A visual query should not be assumed clean; repairing it before search can matter more than swapping the retriever. For image-text retrieval-augmented generation, query pre-processing deserves its own evaluation and training, rather than an assumption that the input has no noise.

4.25

6. When Does Reusing Layers Help Multimodal Embeddings?

AmazonUC Merced

Can a compact universal multimodal embedding model produce better search vectors by running the same layers again? A higher retrieval score alone does not show what the extra pass contributes. If information only travels forward, an image placed before the query stays query-independent even after the layers are repeated. The authors compare eleven loop designs and training variants, and they open a path from the query to the image, including an alternating order that places the query first on the next pass. On every dataset of Massive Multimodal Embedding Benchmark version 2, a standard added loop scores 0.46 points below its trained control. An intermediate generative stage turns the effect into a 0.36-point gain, and a model pretrained with recurrence is 4.23 points above its own loop-removed control. Running the same layers again does not help by itself; what matters is whether image states actually use the query. For image-text search embeddings, ask first whether the extra compute makes the image sensitive to the query, then decide whether recurrence belongs in pretraining.

4.13

7. When Does Label-Free Deconfounding of Document Embeddings Work?

Document embeddings encode spurious attributes that can distort semantic comparisons, yet supervised erasure often needs metadata that is not available. The authors systematically compare label-free deconfounding across corpora and embedding models. Label-free methods improve semantic utility over raw embeddings, and simple geometric approaches recover much of the supervised benefit. On SwiLTra, metadata-free clustering erasure matches metadata-supervised erasure at the reported precision, with a linear leakage AUC of 0.577 and retrieval R@10 of 0.969. Utility can rise even when confounder information remains, and stronger suppression does not necessarily raise utility; lower linear recoverability can also coexist with substantial nonlinear recoverability. Whether deconfounding worked has to be judged on both residual leakage and semantic utility, not on either measure alone. For retrieval or comparison with document vectors, geometric erasure is worth trying without labels, but the two outcomes must be scored separately.

4.13

8. HypRAG: Hyperbolic Dense Retrieval for Retrieval Augmented Generation

ICML 2026Yale UniversityHKUST (Guangzhou)NetApp

Dense retrievers for retrieval-augmented generation still live mostly in Euclidean space, but language is hierarchical from broad topics to specific entities. Euclidean embeddings struggle to keep that structure, so semantically distant documents can look spuriously close and raise the risk of hallucination. HypRAG performs dense retrieval in the Lorentz model of hyperbolic space, with a fully hyperbolic transformer and a hybrid that maps pretrained Euclidean embeddings into that space, and it aggregates sequences with an outward Einstein midpoint so the hierarchy does not collapse during pooling. On RAGBench, the hybrid model gains up to 29% over Euclidean baselines in context relevance and answer relevance; in the fully hyperbolic model, radius grows by 20.2% from general to specific concepts, a structure Euclidean embeddings do not show. The geometry assumed by the retrieval space decides whether hierarchical relations can be kept. For retrieval-augmented generation, a space that can express hierarchy can help more than simply scaling up a Euclidean retriever.

3.86

9. Bioformer-RetroMAE: retrieval-oriented pretraining of a compact biomedical language model for literature search

Dense retrieval is increasingly used for biomedical literature search, but BERT-sized encoders are costly to index at corpus scale and to encode queries under tight latency, and masked language modeling does not directly train the vector that a bi-encoder uses for a whole text. Bioformer-RetroMAE continually pretrains a compact eight-layer Bioformer with retrieval-oriented RetroMAE on about 32 million PubMed titles and abstracts, then fine-tunes a bi-encoder with BM25 hard negatives. On BioASQ Task 12B test questions, retrieving from all PubMed abstracts, it outperforms identically fine-tuned BioBERT and PubMedBERT bi-encoders on every metric, with about two-fifths of their parameters. Relative to PubMedBERT, mean average precision rises by 37.0% and Recall@100 by 29.1%. Retrieval-oriented pretraining can let a smaller biomedical encoder beat larger masked-language models in full-corpus search. For literature search or biomedical retrieval-augmented generation, changing the pretraining objective may cost less than moving to a larger encoder.

3.63

10. Reranking in RAG: How to Fix Bad Retrieval with a Reranker

OpenRouter

Vector search is good at finding chunks on the right topic, but it often ranks near misses first, such as a similar plan limit that is not the one the user asked about. If the correct passage is not among the few chunks sent onward, generation can still be wrong even when the answer sits in the document collection. Reranking takes the candidates from vector search and lets a cross-encoder rescore each chunk directly against the question. On a 274-question synthetic benchmark, vector search with a Gemini embedding model placed the exact right chunk first for 81% of questions; a dedicated reranker applied to the top thirty candidates raised that share to 97%. The authors note that this is a small synthetic set, so gains on real documents still have to be measured on one's own corpus. When retrieval is wrong, a question-aware rerank of the shortlist can come before swapping in a larger embedding model. For retrieval-augmented generation, retrieving about thirty candidates and sending the top three to the generator is a path that can be measured first.