4.05
Cross-modal retrieval, such as searching images with text or text with images, is often trained on data where some samples lack one modality, and the class labels contain errors. Existing methods treat labels as the semantic reference, so a wrong label makes the filled-in content wrong too. SPIRN gives each class a fixed, mutually orthogonal reference vector that is not updated during training.
4.05
Representing a text by its yes or no answers to a list of questions gives an embedding whose every dimension can be read, but with hundreds or thousands of questions, people cannot examine them all. This paper builds a hierarchy over the questions so that people only need to read a few broad ones to interpret an embedding, with task performance largely maintained. The questions are produced like this. A large text corpus is first clustered into a few broad groups, and an LLM contrasts the groups to write questions that tell them apart. Each group is then clustered into smaller groups and questions are written again, so shallow questions are broad and deep ones are specific and more numerous. Each deep question is attached to a broader shallow question (for illustration, a question about running sits under one about exercise). The paper also finds that, for the same number of questions, shallow questions give better embeddings than deep ones.
3.90
RAG assumes the model reasons over the relations stated in its context rather than falling back on associations memorized in pretraining. This paper tests exactly that in isolation, with the facts placed directly in the prompt and no retrieval. The multiple-choice questions start from Wikidata relations and come at five levels, remember, understand, apply, analyze, and evaluate, and were run on dozens of models in six languages. At the analyze level the model must rule out near-miss relations by checking them against the stated fact. At the evaluate level it must judge which statement is the relation's defining criterion, and checking against the fact cannot reject the distractors. Accuracy falls off a cliff from the apply level to the analyze level in every open-weight model series, by as much as 47.2 points, but not for either Gemini model. The evaluate level is a bottleneck too: Gemini 3.1 Pro averages 99.5% on the other four levels but only 60.7% on evaluate. A stronger model does not guarantee reliable reasoning over the evidence you give it.
3.75
RAG and deep research systems repeatedly generate queries, retrieve documents, read them, and decide what to search next, yet there is no common way to evaluate these autonomous search policies. This paper offers a simple framework. It treats the search environment as a set of ranked lists, and a policy has only two actions: keep judging documents further down an existing list, or open a new list by generating or reformulating a query. Each action brings one document to inspect, the information in it updates the search state, and a reward is recorded for the task objective, such as finding a relevant document, so many existing retrieval strategies and objectives fit into one description. Experiments on high-recall retrieval show that both the underlying retrieval method and the deep research policy affect results materially.
3.75
Personalized RAG has to find a user's preferences in very long conversation histories, which is hard for two reasons. The evidence is scattered across conversations, so no single fragment is complete, and preferences depend on time and situation, so they are easily mixed up. Weaver first has an LLM split the current request into interdependent subquestions, and the information a subquestion still lacks is a gap. For example, a request to recommend an activity can split into which activity settings the user prefers and how they feel in crowds. Each time it finds a clue, it uses that clue to locate the next missing piece of evidence instead of retrieving everything at once. It then reads the evidence together to decide which preferences apply to the current situation and which conflict, and updates what is still missing. It reaches 86.42% accuracy on PersonaMem, 16.01 points above the strongest baseline, and 91.10% on PrefEval.
3.75
GraphRAG retrieval strategies are usually tied to their own graph-construction pipeline, and few people test whether they still work when the graph is built differently. This paper finds that retrieval need not be co-designed with the graph. When answering a question, ArcRAG starts from the entities the question names, assembles on the spot the node labels, relation descriptions, and property names reachable from them, plans multi-hop traversals against that interface, executes them as Cypher, and repairs them on failure. The same retrieval configuration works on a graph with typed nodes and a closed relation set, and on an untyped graph produced by open information extraction. On the open graph, MuSiQue EM is 38.0, which is 6.0 below the structured graph. On the structured graph it is 44.0, above every system reported in the paper's table. Construction still shapes absolute accuracy.
3.75
Small models are limited in what they can understand when doing RAG, and having a large model read documents online is too expensive. CacheGraft works offline and only once per corpus passage. It borrows the existing Cache-to-Cache fusion method: a trained fuser network merges, layer by layer, the KV cache a larger model builds from the passage into the cache the small model builds from the same passage. The result lives in the small model's own KV space and is stored. At query time, the relevant passage is chosen by embedding similarity, the small model loads its fused cache directly, and the question is read as text, so no large model is needed online. The test covers five QA benchmarks, with either the top retrieved passage or the gold passage as context. Answers are scored by two LLM judges and by whether they contain the correct answer. Answer quality beats the text-only small model in 29 of 30 comparisons, by up to about 8 points, and the fuser was never trained on retrieval data. It is worth a look where only a small model can be deployed.
3.75
When a prompt contains irrelevant content, such as statements from unrelated domains or random tokens, the model gets answers wrong. This is common in RAG, tool use, and long context, yet the cause had not been explained. This paper traces an interference path inside the model and finds a few attention heads, which it calls intrusion heads. They write irrelevant content into the prediction position and promote distractor-related content, and the correct answer receives weaker support as a result. The intrusion heads are largely disjoint from the heads that integrate useful context. Replacing the corrupted last-token representation with its clean version recovers 97.6% of the distractor-induced accuracy drop on average, but that needs an extra clean forward pass and cannot be deployed. The deployable fix is to knock out about ten intrusion heads, roughly 1% of all attention heads, which recovers 57%.
3.45
Legal judgment prediction needs to retrieve statutes and precedents, but existing legal search agents often use the case description directly as the query, missing latent legal issues and key details. LegalEvolver trains a model to write queries that carry more legal information. It first samples diverse retrieval trajectories at high temperature. It then rewrites their queries, either adding key case facts such as conduct and consequences or, depending on which step failed, turning to the elements of a charge or sentencing thresholds. The successful trajectories are kept for supervised fine-tuning. Next it compares successful and failed trajectories on the same case and has a strong LLM summarize scoring criteria for what makes a query good. During reinforcement learning the criteria score each query, together with the reward for a correct judgment, and they are updated with new trajectories after each round. Criteria guidance beats direct SFT plus RL, and evolving criteria beat fixed ones.
3.45
Searching in professional domains, such as lawyers looking for prior cases or researchers reading background material for a paper, differs a lot from web search. Documents, queries, and sessions are all long, and users' needs are very specific. This is a survey monograph on domain-specific retrieval. It first lays out the most common domain tasks, datasets, and applications, then discusses the challenges these traits bring along with bias and fairness issues in models, and finally describes the models and methods developed for them, including query handling, user models, and user interaction. One conclusion is that many methodological advances in domain-specific retrieval transfer to other domains and even to open-domain web search.