AI SEARCH · 2026-10-24
AI Search — Paper Daily, 2026-10-24
Today's 10 papers cover several stages of a question-answering pipeline. On retrieval, one paper analyzes the queries where hybrid search does worse than vector search alone, Lean Finder V2 uses reinforcement learning to generate short queries for training theorem search, and SurveyGraph retrieves papers for surveys along a graph of method relations. In the RAG pipeline, one paper measures how many chunk embeddings must be recomputed after a knowledge base is revised, and CANOPY trims multimodal results to the granularity needed. Another paper replaces sequential decoding in hierarchical RAG with one forward pass, and GBOR-RAG pulls evidence subgraphs from a gene graph. On the answering side, JCA has the model check each condition in the question. On the industry side come Alibaba Cloud's full-stack AI Search optimization and Tencent Cloud TokenHub's web search API.
4.85
Semantic Retrieval
Hybrid retrieval fuses the results of term-matching sparse retrieval such as BM25 with those of dense vector retrieval into a single ranking, and it always wins on average. Yet query by query, 13–48% of queries rank worse than the better of the two retrievers on that query. This paper keeps the books differently: it compares nDCG@10 before and after fusion query by query, and treats no fusion as a legitimate option. The authors call this per-query loss the fusion tax. The losses concentrate on queries where only one retriever finds a relevant document, and the other retriever then contributes only noise. Switching to a larger dense encoder or adding a cross-encoder reranker after fusion does not make this tax disappear.
4.55
RAG Pipeline
Chunking schemes for RAG are almost always compared on one static corpus, yet the documents in a knowledge base keep changing. Each chunk is embedded on its own, so if an edit moves a chunk boundary, the chunks on either side become new embedding inputs even when not a single word in them changed. The authors take adjacent revisions of Chinese and English Wikipedia articles and allow an old vector to be reused only when a chunk's complete embedding input is byte-identical to one in the old version. They compare three chunking methods: non-overlapping fixed-size chunking, paragraph-first recursive chunking and content-defined chunking, which hashes a sliding window of a few words and cuts wherever the hash meets a preset condition. Fixed-size chunking recomputes 57–68% of the new version's embedding input, yet the three methods differ in evidence recall by less than 2 percentage points within each language. Fixed-size chunking shifts everything after the insertion point, so none of the downstream windows can match the old version, which is why it recomputes the most.
4.25
Query Understanding & Rewriting
People writing proofs and coding agents alike must first find suitable theorems in Mathlib, the mathematics library of the Lean proof assistant. Real search queries are short, they leave out much of the intent, and the same phrase may correspond to several similar theorems. Lean Finder V2 tackles the problem from both ends. On the model side, the authors use reinforcement learning to train a query generator that writes, for each theorem, short queries the current retriever ranks poorly but whose meaning still matches. These queries train the retriever, the new retriever then scores the generator, and the two alternate for two rounds. On the input side, the authors automatically optimize a rewriting prompt: candidate prompts rewrite a batch of synthetic queries, the retriever's rank for the target theorem is checked, and only versions that improve the rank are kept. Before retrieval, it rewrites the user's short query. On MathlibQR, made up of human-written short queries, Lean Finder V2 reaches Recall@10 of 0.893, against 0.780 for the previous best retriever-plus-reranker system.
4.00
RAG Pipeline
When multimodal RAG compresses retrieved results before handing them to the model, how much to keep cannot be fixed in advance; sometimes a whole table is needed, sometimes one row, sometimes one sentence. CANOPY decides the granularity per item and per question, and cuts the evidence tokens handed to the model by 14.2% to 27.7% at comparable answer accuracy. It organises each retrieved result into nested regions, such as a whole table, a few rows or a single row, and scores each region's relevance to the question with a trained encoder. A smaller region replaces its parent only if it scores at least as high, otherwise the larger context is kept. A separate LLM checks which fact the evidence still lacks and, if one is missing, searches again and compresses the new results the same way. When compressing evidence, how much to keep depends on the question, and sometimes a whole row is more useful than a sentence.
3.95
Multimodal Matching
Alibaba Alibaba Cloud says that after adding FalconSeek, its in-house C++ query kernel, to Elasticsearch, the geometric mean query time on Rally Big5, a set of typical ES queries, got about 6.87 times faster. ES still handles the API, writes and cluster management, and only the most expensive query steps, such as inverted-index traversal, sorting, aggregation and vector recall, run in the C++ kernel, while unsupported queries fall back to the original ES execution. Vector search first finds candidates on the HNSW graph using compressed vectors and then rereads the original vectors to rerank them, handling more than 80 thousand queries per second at about 0.98 recall on a ten-million-vector test set. On the model side, the multimodal document embedding model they open-sourced ranks first on the MMEB VisDoc leaderboard. These are all the vendor's own tests, and the article itself notes that the compared systems ran on different machine specs. Search that feeds context to agents needs the engine and the model optimized together.
3.90
QA & Deep Research
Tencent An LLM's knowledge goes out of date, and questions like what the weather is today require going online. Tencent Cloud's TokenHub turns web search into a tool the model calls itself: once it is switched on, the model decides whether to search and what search terms to write, then builds its answer from the results and marks the cited sources. Developers only choose between the lite and standard versions and how much search content to inject per turn; the model decides how many searches to run. This is a product API document with no evaluation of results.
3.80
QA & Deep Research
CMU When retrieved passages mention a wrong answer that meets only some of the conditions more often than the right one, the model is easily pulled the wrong way. Checking every candidate against the same set of conditions and then comparing them together sharply reduces this error. On AmbigQA questions where retrieval favours a wrong answer, Qwen3-30B answers 86.4% correctly with step-by-step reasoning and 93.5% with this method. It works in three steps. First the question is broken into conditions every candidate must meet; for a question about which actress said a famous line, the answer must be an actress, not a character she played. Then each candidate is checked against the retrieved passages, with a verdict and a reason for each condition. Finally all candidates' reports are compared side by side and the one that meets every condition is chosen. When a question carries qualifiers, checking each one before comparing is more reliable than answering directly.
3.60
RAG Pipeline
In hierarchical RAG, an LLM picks a few sections from a candidate list to read next, usually by writing out its choice token by token, but the choice can be computed at once. The picked sections almost always match serial decoding, and selection latency falls by 35% to 47%. Each candidate section is bound to a single-token label, and at the position where the next label would be written, the logits of those labels alone give each section's score. One pass over the prompt picks the first section by score. When three sections are picked greedily, once the first is fixed only a few prefixes are ever needed: the one with just the first section, and the first section plus one candidate second section. All these prefixes are packed into one input where each can see only the prompt and its own earlier part. One forward pass gives every section's score under every prefix: the prefix with just the first section yields the second, and the prefix ending in that second section yields the third.
3.55
When surveys are written automatically, building a graph of which paper extends or improves which, then finding literature along that graph, makes citations better grounded, but the relations in the graph can themselves be wrong. With a graph built from LLM function-calling papers and Qwen writing the survey, the share of cited sentences actually supported by the papers they cite is 2.8 percentage points higher than in a version with the same papers and outline but no graph relations. SurveyGraph links methods, tasks, datasets and other entities into a graph, follows its relations to find related papers for each section, then drafts and revises; relations with too little evidence can be filtered out before retrieval. A human check of 108 relations found only 68.5% fully or partly supported; one claimed that a paper on scientific tool use extends API-Bank, but API-Bank is a benchmark, not an earlier method.
3.30
RAG Pipeline
Running enrichment analysis on one knowledge graph built from several gene-set databases gives almost the same results as the standard method, runs much faster, and shows which genes link terms across databases. Enrichment analysis checks whether a set of genes with clearly changed expression falls mostly within a known pathway or functional term, and the standard method runs once per database, producing several unconnected long lists. Here the terms of several databases and the genes they annotate are built into one graph, where shared genes connect terms from different databases. Each term is first scored by whether the changed genes are clearly over-represented in it, then a trade-off between collecting high-scoring terms and adding few edges yields a small subgraph connected through shared genes. High-scoring terms too costly to connect are dropped, and an LLM writes a biological interpretation from the subgraph. On 42 real gene-expression queries, median query time is 37 times faster, and the mean Pearson correlation of term scores with the standard method is 0.997.
EMNLP 2026