AI SEARCH · 2026-11-11

AI Search — Paper Daily, 2026-11-11

Today's retrieval track has one main thread: neither similarity nor a ranking score equals what the user actually wants. Yuanbao-Reranker makes negatives order among themselves, and the Jev calibration study finds that models that rank well are still too optimistic in their probabilities, though two parameters correct much of it. Another paper shows that dense retrieval leans toward answers the model already knows, and TRIAD finds that spreading a request over several turns makes multimodal retrieval worse almost everywhere. On the engineering side, Lightfoot turns exact GraphRAG deletion into a database operation, and NIFT-SQL trains table-finding models on each customer's own questions.

4.45
LEAD STORY · Tencent

1. Yuanbao-Reranker: Document-Grounded Synthesis and Preference Distillation for Text Reranking

Rerankers fine-tuned pointwise on relevant and not-relevant labels have a gap: every negative for a query carries the same label, so the model never learns which negatives sit closer to the positive. Tencent's Yuanbao Search pairs each positive with three negatives of different difficulty and adds a distillation loss: the teacher's score gap between two documents acts as a lower bound, and the student stops being penalized once it reaches that distance.

4.45

2. Beyond Ranking: Are Decision Models Calibrated Relevance Scorers?

Decision models such as Jev generate no text and output a probability for each relevance level directly, and the vendor promises those probabilities are calibrated, yet earlier independent checks never compared them with graded human judgments. Checked against TREC human judgments, only 66% of the documents it scores above 0.9 are actually relevant, and the other models tested are too optimistic as well.

4.15

4. PatternBloom: Empowering Agentic RAG with Externalized RL-Distilled Graph Patterns

RAG agents trained with reinforcement learning can learn to fetch better content, yet the steps for reasoning through a given type of question are worked out from scratch every time, and methods that retrieve only content plateau as questions get deeper. Asking which country the directors of two different films were born in follows the same steps: film, director, country. PatternBloom abstracts high-scoring reasoning runs into templates that keep only entity types (the paper calls the store Graph Pattern Memory). There are 27 of them, kept outside the model and looked up first so evidence is gathered along them. On five datasets unseen in training, the 7B version averages 49.0 EM, while Search-R1, the best trained method of the same 7B size, reaches 38.1. Reasoning paths that recur across multi-hop questions can themselves be the thing you retrieve. Teams building agent search can turn the query steps of common questions into templates people can read and edit, though these experiments all use Wikipedia corpora.

4.15

5. Search-by-Answer Bias in Retrieval Models

Dense retrieval should find passages by the relation a query asks about, but retrieval models carry world knowledge, and this paper shows they lean toward answers the model already knows, which can penalize rarer or newer answers. The authors build sentence pairs, one kind changing only the answer entity, the other pairing a right-relation sentence holding an alternative answer with a wrong-relation sentence holding the familiar answer. Across eight retrievers, the higher the probability a model gives an answer, the higher it scores passages containing that answer. On examples where the model already favors the canonical answer, Qwen3-Embedding-8B picks the familiar sentence 94.6% of the time when only the answer differs, against 50% with no bias, and ranks the wrong-relation sentence first 65.0% of the time. For queries with several valid answers or answers that go stale, retrieval is easily pulled toward what the model remembers. The test sentences come from templates, but on QAMPARI, whose questions have many valid answers, the retrievers also favor the known answer.

4.15

6. CoRA-Bench: Evaluating Evidence Preservation by Context Selectors Under Token Budgets

Filtering content before long-context inference saves tokens, but it can also throw away the evidence the answer needs, and truncation, retrieval and compression have lacked a common way to be compared. CoRA-Bench compares truncation, BM25, dense retrieval, reranking and two compression methods under the same token budgets, computing evidence recall and answer accuracy on the same selected text, and no method wins everywhere. On NoLiMa, where evidence and question share little wording, dense retrieval reaches 0.908 evidence recall against 0.467 for BM25, and on BABILong, with explicit surface cues, BM25 leads instead. Keeping evidence is not the same as answering correctly either: under verbatim scoring Provence, which alters the text, keeps almost no complete evidence on LongMemEval, yet its answer accuracy is 66.60%. When choosing a context selector, look at evidence recall and answer accuracy separately, and teams doing long-document QA or memory retrieval can pick a method by whether their evidence relies on literal matches or semantic association.

4.15

7. TRIAD: Multimodal Retrieval across Abstraction, Constraints, and State

In multimodal retrieval for agents, top results can look very similar to the request and still fail to satisfy it, and those results often feed straight into the next reasoning step or tool call with nobody screening them. TRIAD tests twenty model configurations on three kinds of tasks, hierarchical concepts, positive and negative constraints, and multi-turn dialogue, and no system leads on all six subsets. Multi-turn results are the clearest: when the same final request is spread over several turns, mAP falls in 39 of 40 comparisons, and the loss concentrates in sessions where the user revises a condition, such as asking for orange and then changing to blue. Keeping only the last turn does not fix it and is worse overall than keeping the full history. Adding explicit reasoning is no cure-all either, with effects that vary by model and task. Multi-turn retrieval has to keep the conditions still in force and drop the values that were replaced, so teams building conversational search should test sessions where users change their minds separately.

4.15

8. NIFT-SQL: Customer-Supervised Table Scoring for Enterprise Text-to-SQL Retrieval

The first step of enterprise text-to-SQL is finding the tables, yet schemas often offer only cryptic table and column names, a query needs every table it joins, and generic retrieval struggles here. NIFT-SQL trains separately on each customer's example questions, has a small cross-encoder judge for every table whether it is needed, and fuses that with hybrid search over a retriever fine-tuned the same way. On LiveSQLBench-Large, with nearly a thousand tables, the share of questions with every needed table in the top fifteen is 21.8% for hybrid search without fine-tuning and 85.6% for NIFT-SQL, while the fine-tuned retriever alone reaches 70.4%. The lead is not stable: on a newer BEAVER release, trained only on synthetic questions, it loses to the fine-tuned retriever on real questions. With labeled questions in hand, a fine-tuned retriever is itself a cheap and strong baseline, and the scorer adds the most when only a few tables fit in the context.

4.15

9. Two-Sided Conformal Evidence Sets for Retrieval-Augmented Generation

Passages retrieved for RAG can support a claim or contradict it, yet existing calibration methods mostly handle only the supporting side, leaving refuting evidence uncovered. This paper uses conformal prediction, which sets a threshold from a batch of labeled samples and gives a coverage guarantee, to produce one evidence set for support and one for refutation. Each side is guaranteed to contain at least one true piece of evidence at a chosen probability, provided relevant evidence was retrieved at all. Swapping the scorer shows the difference most clearly: with a weaker model, evidence coverage stays close to target while verdict accuracy falls from 0.898 to 0.516. When the data distribution shifts, coverage breaks too, by up to 12.7 percentage points on FEVER, and labeled data from the deployment setting is needed to recalibrate. Finding all the evidence and reaching the right verdict are two different things, so RAG evaluations should report them separately, and calibration cannot make up for a weak scorer.

4.15

10. Calibrating Document-Local Boundaries for Contextual Chunk Retrieval

Late Chunking encodes a whole document before splitting it, which preserves context but makes similar chunks of the same document harder to tell apart, such as one player's scoring records from different seasons. On InSeNT, a baseline fine-tuned with same-document negatives, 82.21% of the failed queries have another chunk of the same document scoring above the correct one. DcLBC from Alibaba's Qwen team changes two things. It scales a query's scores up more when the correct chunk leads the highest-scoring wrong chunk of the same document by less, and it tells the attention layers during encoding where the chunk boundaries are. Neither change touches the embedding dimension or the retrieval procedure. Split into four tiers by same-document competition, the NDCG@10 gain concentrates in the hardest tier at 5.46 points, while the two easiest tiers barely move. Many long-document retrieval errors come from telling apart chunks of the same document, so passage retrieval teams should count this kind of error separately.