AI SEARCH · 2026-11-17
AI Search — Paper Daily, 2026-11-17
Today's retrieval track asks whether the retrieval object actually matches the task a user is trying to finish, and whether the result can be inspected. One paper turns query rewriting into a searchable strategy, so prompts and decoding parameters evolve with retrieval feedback. Another breaks medical dense vectors into interpretable sparse features that can explain why a document was returned. Public-interest search over governance documents insists on binding answers to the exact text and edition, and a digital-library study embeds both the figures and the text of scanned pages. Other work writes reading progress as a spoiler boundary, first assigns a finance agent's failures to the layer that should be fixed, and measures how much local paper search depends on the model and on how specific the query is. Medication prediction from health records retrieves a population prior instead of similar patients, and visual question answering first judges whether a reference is clear before deciding how to answer.
4.30
In retrieval-augmented generation, rewriting the user question is supposed to make retrieval more accurate, but retrieval feedback is often noisy, prompt text and decoding parameters such as temperature must be tuned together, and a fixed strategy breaks down on semantically complex questions. The authors treat rewriting as a non-differentiable black-box problem: prompt templates and decoding parameters become rewriting agents that can be updated iteratively, with a global search to escape local optima, local refinements, and a memory pool that keeps strong strategies to resist noise.
4.05
Medical question answering first has to retrieve relevant passages from large document collections, but the dense vectors in common use are almost impossible to explain, and fine-tuning on one corpus often loses performance on a new clinical collection. The authors use sparse autoencoders to turn dense text vectors into interpretable features, training both to reconstruct the original vectors and, with a contrastive loss, to keep query-document relevance.
3.80
Governments are writing rules for artificial intelligence, often defining what counts as an automated decision or a harm before the people affected have a chance to question them. The authors build a public-interest system that turns thousands of governance documents, from cities to national governments, into passages that keep their page and edition, and uses bounded agentic search to verify each cited statement. A resident can look up the city's rules on AI in hiring, or upload a national strategy, read it as an affected person would, and draft a revision. On new questions that require finding the relevant documents, macro-average recall reaches 83.0%, compared with 73.7% without iterative search. In a user study, most proposed changes asked how a provision would actually apply. Retrieval has to be bound to the exact text and edition of a public document before people can use it to discuss and revise the rules. For public-policy retrieval, finding the right documents and giving checkable citations matters more than a fluent answer alone.
3.70
In large digital libraries, many holdings are scanned pages that mix text and figures. A full-text inverted index can barely see diagrams, tables, and layout, so text from optical character recognition alone misses key information. The authors embed each page image as a vector that carries both textual and visual semantics, retrieve with a vector database, and compare similarity functions. On the full benchmark, cosine similarity reaches a mean reciprocal rank of 0.801, clearly better than 0.418 for Euclidean distance, and it is especially suited to questions about diagrams or tables and to visually dense pages with little text. Retrieval precision itself is not high, so better reranking is still needed. When the corpus grows from one hundred pages to a little more than four thousand, latency rises only about twenty times, and throughput can exceed 3,500 queries per second. For mixed image-and-text collections, a text index is not enough; vector retrieval must also capture visual semantics. For scholarly discovery systems, the point is not only a stronger vision model, but an index that can run at library scale.
3.70
When a language model sits beside a reader, sending the whole book in at once makes it easy to reveal later plot points, and the text often has to be uploaded to the cloud. The authors build an offline reading assistant in which retrieval and generation are confined to the reader's current place in the text, and combine vector similarity, keyword matching, and summary embeddings so the answer stays with the current progress. On comprehension questions and spoiler checks from three English novels, mid-sized local models can reach comprehension quality close to larger counterparts. The strongest of them reach a BERT F1 of about 0.74 to 0.75, at about 30 seconds per item. A progress boundary is used to block unread content without giving up retrieval quality on what has already been read. A reading companion should move forward with the reader, not finish the book first and then answer. For teams that want local reading support, writing progress as a boundary that both retrieval and generation must respect can protect privacy without spoiling the story.
3.55
A single leaderboard score for a finance agent you did not build rarely says whether the system is stable, who should fix a failure, or whether a change can be trusted. The authors take an untuned reference agent and repeatedly run the same questions; the number of rubric criteria passed moves from 105 to 123, and the bill also varies clearly. Under a taxonomy frozen in advance, most failures belong to the agent's instructions, then to its tools and the model, and none to its data. An optimization loop that reads traces, changes one thing per round, and keeps it only past a fixed threshold cuts cost on the held-out questions to 0.34 times the baseline, with accuracy still within noise. A gain under an accuracy objective does not pass a test written before the run was opened. An agent's score must first survive repeated runs, before one asks what to change and whether the change is real. This is a case study of one agent and a small test set, but it shows auditors how to measure noise and assign the responsible layer first.
3.55
If scientific literature search can run with a locally deployed language-model assistant, teams can send less data and infrastructure to a hosted provider. It is still unclear, though, whether open-weight models can actually retrieve papers under practical compute limits. The authors compare inference latency on a Mac Studio and on a two-GPU server, then have eight models search for 70 papers in seven repositories, with queries written at three levels of expertise. Retrieval success depends strongly on which model is chosen and how specific the query is: only 6.2% of basic phrasings succeed, while 48.2% of expert phrasings do. The server is usually faster, but the Mac's unified memory can load larger models that did not fit in the GPU configuration tested here. For local paper search, the model and how specific the query is often decide success more than a slightly faster platform. For research groups that want their own literature assistant, testing models against their own query habits is more reliable than looking only at parameter counts or official demos.
3.45
Semantic search for an object or scene in a large satellite mosaic usually starts by training a detector for that class, and a new class means doing the work again. The authors build a system that needs no further training for a new class: a frozen self-supervised vision model represents each tile, positional bias is removed, features are centered on the area mean, and matching uses cosine similarity. A query can be a small neighborhood around a click, or a whole tile for scene search, and finer boundaries are computed once and then cached. On a test set of a little more than five thousand tiles, mean pixel-level F1 is 0.39, with a large spread across classes. A click returns a rendered overlay in about 6 to 10 milliseconds. Even without a detector trained for this class, a clickable semantic index can first surface the candidates. For interactive search over large remote-sensing mosaics, this replaces training a detector first with building a reusable feature index and searching from a click.
3.15
Predicting a medication set from electronic health records is a structured problem over a fixed vocabulary of drug classes, yet many clinical retrieval-augmented systems first fetch similar patients and then ask a language model to infer the medications. What is retrieved is a patient record, while the target is a discrete set of drug labels, so the two are not aligned. The authors instead use a tabular population prior estimated from condition-medication co-occurrence in the training data, listing candidate drug classes by co-occurrence strength, and replace only the patient retriever in an otherwise unchanged pipeline. On a little more than a thousand held-out admissions, F1 rises from 0.303 to 0.338, with no extra language-model calls and no patient index at inference time. The prior is already predictive before a language model is involved. For set prediction over a fixed vocabulary, the retrieval object should be candidates from that vocabulary, not a detour through similar patients. For clinical medication prediction, population statistics are easier to audit and often a better fit to the task.
3.15
Vision-language models are often overconfident, giving a firm answer even when the visual evidence is not enough. A typical case is referential ambiguity in visual question answering: several objects in the image could be what the question refers to. The model should then ask a clarifying question or list the plausible referents, and should answer directly only when the reference is clear. Current models struggle with this distinction and often default to a single reading. The authors use activation steering at inference time to elicit awareness of ambiguity, and find that steering alone has limited, model-dependent effects, while prompting is usually more effective. Combining the two increases clarification and enumeration, but also triggers them on unambiguous cases. They therefore use a linear classifier to judge whether the input is ambiguous and to choose the steering direction, raising the share of inputs handled appropriately from 59.0% to 77.0% on average across models. A model should first decide whether the question is clear, and only then choose to answer, to ask, or to list the possibilities. For multimodal question answering, this is a way to reduce overconfidence at inference time without retraining.