AI SEARCH · 2026-11-10

AI Search — Paper Daily, 2026-11-10

Today's retrieval track has three threads. The first is whether multi-hop and multimodal retrieval needs a graph built in advance. EAGR stores links between passages as metadata in a vector database, TrioRAG skips the graph and retrieves with three signals, and SkinMind turns a dermatology knowledge graph into tools that the model probes step by step. The second is what retrieval should keep, where FondueQA finds that a bridge document can be more useful than the one holding the answer. The third is base models and evaluation: SPARSEUP shows how to train sparse retrievers stably on new backbones, and two more papers build dedicated features and benchmarks for Vietnamese statutory retrieval and for hiring search.

4.35

1. SPARSEUP: Towards Better Recipes for Training Modern Learned Sparse Retrievers

Sparse retrievers like SPLADE, moved onto newer backbones such as ModernBERT, often train poorly, and RoBERTa does not train at all. The authors trace this to the starting point: most of the new backbones' vocabulary scores are already positive, so at the start of training almost every term is active, the document representation is dense, and the gradient that tells terms apart weakens. The fix has three steps. Before training, one constant is subtracted from all scores, set at a high percentile of the score distribution estimated on a small sample of documents. Each input position may activate at most 12 terms. Case and word-boundary variants of the same word are merged. The final model reaches 56.4 average nDCG@10 on BEIR, only 1.5 points below a dense model trained on the same backbone and data. Anyone moving sparse retrieval to a new backbone can start with these three steps.

4.05

2. FondueQA: A Benchmark for Evaluating Parametric Compensation in Multi-Hop QA

Ask "What is the nationality of the director of Titanic?" If retrieval can return only one document, which one helps more? Intuition says the one holding the answer, the document stating that director James Cameron is Canadian. The authors find that keeping the bridge document, the one stating that Cameron directed the film, often works better, especially for smaller models: once the model knows whom to look up, it can often fill in the missing fact from its own memory. To run such comparisons, they built a multi-hop QA set, FondueQA, where each needed fact appears in exactly one designated document, so documents can be removed one at a time. The more popular the entity behind a missing fact, the more easily models fill it in. But adding a few plausible competing entities to the context can cut some models' accuracy by about 15 percentage points. In RAG, besides how much is retrieved, it matters which link of the reasoning chain is kept.

4.00

3. MARS: Training Multilingual Search Agents for Retriever-Free Evidence Acquisition

Multilingual retrieval-augmented generation usually builds an index first and then ranks with a trained retriever, and a slight change in how an entity is spelled across languages can send the search to the wrong place. This work lets the model issue executable search commands against the raw corpus and trains it with reinforcement learning. It first searches the corpus to verify each entity's native spelling, then reuses those spellings in later queries. Each search command is treated as a set of terms. If the intersection-over-union of two sets exceeds a chosen threshold, the commands are treated as near-duplicates and merged. only a trajectory that finally answers correctly receives a reward scaled by the fraction of commands that remain distinct. On MKQA and MLQA, average accuracy rises 3.2 points over GrepSeek and replies become 29% shorter. Repeated searches fall while cross-lingual generalization still holds.

3.95

4. EAGR: Evidence-Guided Adaptive Graph Retrieval

Ask who someone's mother-in-law is, and a first retrieval rarely finds the answer: you first need the passage about their spouse, then follow the spouse to her mother. Where to search next only becomes clear once intermediate evidence arrives. Common approaches either build a knowledge graph in advance or have an LLM rewrite the query at every step. The first needs a separate store to build and maintain, the second adds a round of work per step. At ingestion, EAGR has an LLM extract descriptors such as entities from each passage and stores them as ordinary metadata in the vector database, so passages sharing a descriptor count as neighbors. At query time, dense retrieval finds a few starting passages. The LLM reads the evidence so far, picks the descriptors worth following, fetches neighboring passages through the metadata, and stops once the evidence suffices. Averaged over three multi-hop QA datasets, LLM-judged accuracy is 67.1%, against 59.8% for vanilla RAG and 63.2% for the strongest graph-based baseline.

3.95

5. Orthogonal Norm Growth in Two-Tower Retrieval: A Popularity Mechanism for Unified Discovery

When search and recommendation share the same item vectors, their updates can jointly push some vectors to grow longer. In unnormalized dot-product retrieval, items with larger norms are easier to retrieve. The authors prove that, under stated sufficiency conditions, a cosine-loss gradient moves tangent to the equal-norm sphere, so each nonzero update increases the squared norm by the squared length of that tangential step. Training also updates encoder parameters, so the actual move need not stay tangent. On MovieLens 32M and Amazon Books, this tangent accumulation comes with a correlation between frequency and norm. Ranking with the raw unnormalized dot product retrieves more head items than first normalizing the item vectors. A control that, after each item-id vector update, retracts the immediate norm change from that step weakens this exposure, but recommendation quality also drops, so the control is a diagnosis rather than a fix the authors recommend.

3.80

6. Whispers from the Neighbors: Rethinking Caption Consensus as Inference-Time Evidence for Vision-Language Retrieval

An image is rarely described only once: retrieval benchmarks often attach several captions to the same picture, yet dual-encoder models still score each image-text pair on its own. This paper asks how much a frozen encoder gains at inference if it also looks at the other captions of the same image. It averages and normalizes those caption vectors into a consensus, then adds a weighted combination of the consensus-to-query text similarity and the image-to-query similarity. When the captions and the queries come from the same annotators, text-to-image R@1 rises by 3.4 to 16.8 points across nine frozen encoders. If an off-the-shelf captioner writes the captions instead, the gain shrinks to 1.9 to 7.2 points. The captions help match a candidate to how people describe the image, rather than improving image-text alignment itself.

3.60

7. Differentially Private GraphRAG: Protecting Relation Connectivity and Semantics

Graph retrieval-augmented generation links entity relations in documents into a knowledge graph so a model can reason over those links. With only black-box access, an attacker may still reconstruct the graph, recovering both whether two entities are connected and what the edge means. Before indexing, this work extracts triples, generates candidate predicates, and uses the exponential mechanism to sample a replacement phrasing or a null outcome. Utility is the negative semantic distance between a candidate phrasing and the original relation, so closer meanings are more likely to be drawn; a null draw deletes the edge and has a fixed utility that does not depend on the original wording. Across three graph-RAG systems and two targeted reconstruction attacks, relation-recovery F1 falls 26.98% to 44.35% versus the strongest baseline. Question-answering F1 drops only 2.98% to 8.97%.

3.50

8. Less Is More: Graph-free Multimodal RAG via Multi-signal Late Fusion

For multimodal cross-document QA, many systems first extract entities from the whole corpus and build a joint text-image knowledge graph, then retrieve over the graph. Building the graph is expensive and slow, and any document change means extracting again. TrioRAG builds no graph. It puts each page's text and page image into one multi-vector index and retrieves with three signals: the question text, the image attached to the question, and a question rewritten by a vision-language model that looks at the image. The three rankings are then merged. On three benchmarks it matches or beats two graph-based systems, and its total time including indexing is 12 to 65 times shorter. The authors also built AutoQA, an automotive QA set whose images come from the web and are not in the corpus. There, image retrieval alone reaches only 19.3% document recall, and the two text signals carry retrieval. For multimodal RAG over frequently updated documents, this graph-free setup is worth trying first.

3.40

9. Improving Contextual Retrieval for Long Documents using Semantic Chunking and Hybrid Retrieval Strategies

When an answer has to join evidence spread across long documents, ordinary retrieval-augmented generation tends to fragment the context and struggles with multi-hop reasoning. This work keeps both the explicit relations in a knowledge graph and the semantic coverage of dense vector retrieval. At inference it ranks entities by how densely they are connected and starts graph traversal from the most informative ones, following paths that link those entities. Each path is scored from the semantic similarity between its edge labels and the query terms, how often that relation appears across document segments, and a penalty on overly long or looping chains. Only if no path is found, or this combined score is too low, does it fall back to encoding the query as a vector and retrieving with dense search plus keyword matching. On MultiHop-RAG, the hybrid method reaches 55.8% exact match, compared with 51.5% for vector retrieval alone and 48.7% for the graph alone.

3.15

10. SkinMind: Equipping Agentic RAG with Human-Inspired Graph Probing for Reliable Dermatology AI

Dermatology questions often need guideline text, relations between diseases, and lesion images at the same time, and a single retrieval pass over a text corpus rarely gathers all of that. SkinMind first builds a joint text-image knowledge graph from dermatology guidelines and textbooks, with over a million nodes, then treats it as an environment to probe step by step. The orchestrating LLM has eight tools. It can find relevant entities, inspect node details, follow relations and paths, and finally pull source passages and images as evidence. What to query at each step, and at what granularity, depends on the clues gathered so far. The authors average scores over six dermatology visual QA benchmarks: accuracy on five multiple-choice sets, and an LLM judge comparing answers with references on one open-ended set. With one orchestrator model, the average rises from 59.20 with direct answering to 67.90. The other orchestrator model gains about as much, and SkinMind consistently beats other RAG methods. The graph is opened up as tools for step-by-step probing rather than used only as a retrieval index.

3.05

11. Structure and Lineage Evidence for Vietnamese Statutory Article Retrieval on DRiLL

In Vietnamese legal QA, answering one question often means finding several statutory Articles together, such as a base provision and a later Article that amends it. Retrieval by wording similarity alone easily misses amendments and companion Articles. On DRiLL, a Vietnamese statutory Article retrieval dataset, the authors add several kinds of structural features to each candidate Article: which law it belongs to, which other candidates come from the same law, what it cites, and what has amended it. LightGBM then combines these with LLM relevance judgments to pick the final answer. Averaged over five random seeds, the official F2 score rises from 71.77% to 72.91%, a small gain that is clearest on questions needing two Articles. Among the added features, same-law and citation evidence contribute most. For anyone who wants to use citation and amendment links in legal retrieval, the paper offers a concrete feature list.

3.05

12. A Taxonomy-Driven Synthetic Dataset for Retrieval Effectiveness and Fairness Evaluation in Hiring

Retrieval models for hiring are hard to benchmark publicly, because real resumes and job requirements involve privacy and commercial secrets. Starting from a large set of real freelancer profiles and a taxonomy of occupations and skills, the authors use Gemini to synthesize multilingual candidate profiles and project briefs. Each profile gets a counterfactual twin that is identical except that gendered word forms are switched between masculine and feminine. Testing embedding models on it shows that general MTEB rankings do not carry over well: Arctic Embed scores lower on the general multilingual board, yet its NDCG@10 on project-brief retrieval is 0.07 higher than several instruction-tuned models. Model size also shows no systematic link to fairness, and models of similar size can react very differently to the gender switch. When choosing a model for hiring search, test effectiveness and fairness together in your own setting.