AI SEARCH · 2026-10-29
AI Search — Paper Daily, 2026-10-29
The AI search track has 10 papers today, and the common thread is which stage of the retrieval pipeline is most worth replacing. Google's EmbeddingGemma 2 encodes text, images, video and audio directly with one small model, and an instruction-following retrieval paper turns original positives into the hardest negatives to build training data. On the RAG side, LTRR picks a retriever per question, Tessera re-splits web pages into small header-completed units before reading, QESChunker uses question-and-answer records to decide where to cut chunks, and DISCo finds a set of documents that together cover the query. On agents, RouteDRAG learns when to start the agent, and another paper removes scores inflated by heavy connectivity on a document graph. Two papers rewrite queries: GRAPE uses retrieval rank as the reward, and an e-commerce challenge winner anchors task-style queries in catalog wording.
4.70
LEAD STORY ·
GoogleGoogle To search text, images, video and audio with one embedding model, the usual route is to transcribe audio and caption images first, then hand everything to a text model. That costs extra compute and throws away tone of voice and visual detail. Google's EmbeddingGemma 2 is a 744M-parameter model that maps these inputs directly into one vector space, and each modality encoder can be loaded only when needed. Finding text passages from a spoken question shows the difference best: the same model transcribing speech to text first reaches an MRR@10 of 57.59, while reading the audio directly reaches 69.54. A small multimodal model need not give up text and code; its code retrieval score is almost ten points above its predecessor's.
4.45
Instruction-following retrieval asks one query to rank documents differently under different instructions: once an instruction adds a requirement, the original positive should drop if it fails that requirement. Training data for this is scarce, and existing methods keep the positive fixed and have an LLM invent a negative passage that violates the instruction. This paper does the reverse. It picks an existing negative that meets some extra condition, writes a short instruction that makes it the positive, and the original positive naturally becomes the hardest negative under that instruction. InstructIR gives each query several user instructions and scores the worst top-ten nDCG among them. With only 6,716 such examples built from MS MARCO, the small Qwen3-Embedding's score rises from 49.29 to 67.06, beating Promptriever, a model more than ten times larger.
4.45
RAG systems usually wire in one fixed retriever, yet different questions suit different retrievers, and some are better answered with no retrieval at all. This paper treats choosing a retriever as a ranking problem. For each training question it records how much each retriever improves answer quality over no retrieval, then trains a ranking model to order the retrievers, with no retrieval as one of the candidates. Over six retrievers built from sparse and dense retrieval plus different rerankers, a pairwise XGBoost router lifts the answer-correctness score from 0.5776 for the best single retriever to 0.5900, while picking right on every question would reach 0.7948. Routing gains are real but far from the ceiling, and the answer-quality metric used for training labels matters a lot.
4.45
In classic retrieval, documents compete for the top ranks. In multi-hop QA or text-to-SQL, a single document is often incomplete and the answer must be assembled from several. In dense retrieval terms, the task is to find a small set of documents whose token vectors together cover every token vector of the query. Coverage here means every token vector of the query finds a close token vector somewhere in the chosen documents. This paper shows that coverage is submodular, meaning each added document contributes less new coverage than the one before. So it can probe a vector index round by round, each round finding the document that fills the most of what is still uncovered, and greedily assemble a near-optimal set in time sublinear in corpus size. The authors build a new kind of vector database, DISCo, on this idea, and it beats the alternatives on the coverage versus latency tradeoff. For RAG that needs several pieces of evidence, this is a better-matched retrieval goal than top-k.
4.45
Retrievers like CLIP degrade noticeably on Chinese queries, long-form queries, or combined image-and-text queries, yet retraining the retriever and re-embedding the whole corpus is too expensive. This paper leaves the retriever untouched and trains only an LLM that rewrites the query, using GRPO with retrieval rank as the reward for each rewrite. The authors find that if the reward only looks at how similar the rewritten query is to the correct result, the model learns to raise the scores of all candidates at once, irrelevant ones included. Rewarding where the correct result ranks within the whole corpus removes the problem. Across multilingual, long-form and composed image retrieval benchmarks, Recall@10 improves by 4.9% on average. Systems with a frozen retriever and a large corpus can start with query rewriting.
4.35
Ask Wikipedia who produced Willy Wonka and the Chocolate Factory, and the answer sits in the producer row of the page's infobox. Text extraction drops the two names in that cell, the reader model sees only the director, and it answers wrong. Tracing answer evidence from page source to reader input over all of English Wikipedia, this paper finds such loss is common: for 7.6% of questions the evidence never enters the index intact, and 81% of those losses are table cells, table headers or captions. Tessera leaves retrieval alone. Just before reading, it re-splits the top-ranked pages into small units such as one table row or one list item, adds the title, section and headers back, and packs the relevant ones into the same token budget. Answer accuracy rises by 5 to 6 points on both text and screenshot pipelines.
4.35
Most RAG chunking is rule of thumb: fixed length, by paragraph, or by semantic similarity, and none aims directly at whether a question's evidence can be retrieved complete. This paper uses existing questions and their evidence to decide where each document is cut, aiming to put as many evidence passages as possible into chunks that the question can actually retrieve. Without overlap the objective splits chunk by chunk, so dynamic programming finds the optimal cut exactly. With overlap the problem becomes NP-hard, and a greedy algorithm gives an approximation. The test uses two Chinese policy corpora and three public English datasets, fixes the same hybrid retriever, and compares the share of questions whose evidence is retrieved complete. The non-overlapping version ranks first in 9 of 10 comparison settings and the overlapping version in all 10, with larger margins when less context is returned. It suits document collections that are queried again and again and already have question-and-answer records.
4.10
Many e-commerce queries describe a task rather than a product. A user types make lasagna and actually needs a baking dish, ingredients and kitchenware, but product titles contain none of those words, so BM25 struggles to retrieve them. This winning solution to an e-commerce challenge first aligns the query with category and entity terms in the catalog, then has an LLM generate five complementary rewrites in one pass. It finally selects by protocol. The automatic setting gets the single highest-quality rewrite. The interactive setting gets up to four: among the groups of four out of five rewrites, it compares average quality first and then the largest overlap in retrieved results, preferring less overlap. A rewrite's quality is the rank-weighted average of reranker relevance scores on the top five products its BM25 search returns. Task-style query rewrites must be anchored in catalog wording, or the product words an LLM invents may not be searchable.
4.05
In multimodal QA over long documents, agent-style repeated retrieval and reasoning is not always useful: some questions are answered right after one retrieval, and extra steps can turn a right answer wrong. The authors first show experimentally that one more tool call can improve, preserve or degrade the answer. RouteDRAG therefore learns two things: whether a question should go to the agent at all, and, while the agent runs, whether each step should call another tool and when to stop. On two benchmarks it beats short-path-only execution by 3.47 percentage points on average while sending only 79.3% of questions to the agent, and it is more accurate than sending every question to the agent while making fewer answering calls. Whether to start the agent is itself worth learning.
4.05
When a search agent looks for evidence on a document graph, candidates are often scored with Personalized PageRank. Well-connected documents are naturally easy to reach, so their scores are inflated even when their evidence is no more useful, and they crowd the agent's limited observation window. This paper scores each candidate by one strongest path only, the one from the initially retrieved documents whose edge weights multiply to the largest value. It then divides by the square root of a background probability: how often a random walk that restarts uniformly from any document on the same graph visits the candidate. So well-connected documents need stronger paths to rank high, and the paths go to the agent too. On BrowseComp-Plus and TREC-COVID it gets the best answer accuracy and nDCG@10 at the lowest generation cost. Against the strongest baselines, tool calls fall by more than half and total generation cost by more than 40% at most.
SIGIR 2026
ICLR 2026
ICML 2026