AI SEARCH · 2026-11-13

AI Search — Paper Daily, 2026-11-13

Today's main thread is that retrieval no longer stops at finding relevant passages and must also check whether the evidence holds up. PCG-RAG pins down which entity a question means before querying its attributes, and HiddenPathQA finds that question answering that walks a knowledge graph hop by hop goes astray when questions do not name the relations. ARMOR trains RAG models to answer correctly amid many noisy documents, and VisProve asks models to explain how each piece of evidence was used. DeepResearch RewardBench shows LLM judges often grade reports on dimensions outside the criteria, while SNAPRAG can extract the knowledge base behind a RAG system bit by bit. Alibaba's 1688, for its part, moves user behavioral evidence into search relevance tiers.

4.40

1. Beyond Query-Item Matching: Personalized Behavioral Relevance Modeling for E-commerce Search via Rubric-Guided Reinforcement Learning

E-commerce search usually has a relevance model sort candidates into tiers by text match between query and item, then hands them to a CTR model, so an item the user clearly prefers can get little exposure just because it landed in a low tier. This work from Alibaba's 1688 moves part of personalization into the relevance step: an LLM reads the user's same-day, recent and long-term search and purchase records, judges whether the item is backed by behavioral evidence, and if so promotes it one relevance tier.

4.05

2. PCG-RAG: Progressive Compositional Grounding for Retrieval-Augmented Generation in Complex Web Question Answering

Complex web QA often requires first working out who or what several clues describe, then looking up an attribute of it, and a common mistake is to treat a name as settled after only one clue matches. The paper's example is a London power station converted into an art museum: a system that settles on Battersea from London and power station alone retrieves an architect who actually belongs to a different building.

4.05

3. HiddenPathQA: A Benchmark for Knowledge Graph Question Answering When Questions Hide Their Paths

Users ask questions without knowing the knowledge graph's structure. Asking which condition Andrew could have due to family history really means a multi-hop path through father, father's father, medical condition and hereditary disorder, yet the question names none of those relations. Existing KGQA test sets nearly spell the path out in the wording, so the three families of methods, picking relations hop by hop, planning then retrieving, and recalling by vector similarity, have rarely been tested on this case. The authors built over a thousand such questions and also replaced entity names with meaningless codes to rule out answering from memory: LLMs that do not query the graph drop about 40 points on the coded version, close to guessing. Their approach lists all candidate relation chains around the entity to the LLM at once for whole-chain comparison, with no pruning along the way. This approach beats six baselines by 8.8 to 17.9 points. The trouble with hop-by-hop scoring is that one wrong step midway cannot be undone.

4.05

4. SNAPRAG: Cost-Efficient Knowledge Extraction Attack on RAG Systems via Adversarial Perturbations

The private knowledge base behind a RAG system can be extracted one question at a time: the attacker appends an instruction to the query telling the model to repeat the retrieved documents verbatim, then builds the next batch of queries from what comes back. Existing attacks either have the attacker run a separate LLM to write queries or need many queries. SNAPRAG uses no LLM and instead swaps some tokens in already stolen text, so each new query lands far enough from seen content in the attacker's local embedding space to hit parts of the knowledge base not yet retrieved. To steal half a knowledge base, existing methods need up to 6.77 times as many queries. Within the same query budget it recovers up to 96.9% of the documents. When the targets add input and output moderation, it still leads on four of the five. Content moderation alone cannot stop an attack that extracts the data bit by bit.

3.95

5. MMGraph-Agent: Agentic Multimodal RAG via Cache-Inspired Multimodal Knowledge HyperGraphs

Multimodal RAG carries two real costs: building the knowledge graph means paying an LLM to extract entities, and multi-round retrieval keeps stretching the context. MMGraph-Agent builds its graph simply: a local named-entity tool links the entities in each sentence into one hyperedge, an edge that can connect several entities at once. Images are encoded straight into the index with CLIP, so the whole build calls no paid API and finishes in a few CPU hours. At retrieval time it mimics a CPU's two-level cache, condensing each round's evidence into a digest in a small cache that later rounds reuse first. Across 23 multimodal QA datasets it beats the second-best method on each dataset by 11.6 F1 points on average, and the cache cuts token usage by 67%. Switching graph building to cheap rule-based extraction did not hurt quality.

3.80

6. DeepResearch RewardBench: Evaluating Reward Models for Deep Research Report

Long reports written by deep research agents are too expensive for human review, so people score them with reward models or LLMs, yet few have systematically checked whether that works. The authors collected reports from several agents, ordered them by the original datasets' expert grading criteria, and grouped them into comparison items of two, three or four reports. Conventional reward models clearly struggle here, and LLMs used directly as judges do best. Having a judge compare several reports together is far more accurate than scoring each one alone, by 28.4 points on average when ranking four. When scoring alone, models keep giving 80 or 85, so good and bad reports blur together. Manual analysis of errors shows the most common mistake is the judge picking dimensions outside the grading criteria, such as caring more about how fresh the citations are or how many tables there are. Grading reports is hard because the judge's dimensions do not match the criteria.

3.80

7. Dhara: Evidence-Oriented Retrieval for Indian Knowledge Systems

Finding evidence in Indian classical texts is hard because answers are often spread across verses, translations and commentaries in different places, and askers often remember only half a line of a verse. Besides standard hybrid retrieval, Dhara adds two routes. One writes notes in advance on the people and concepts in each chapter and book, then follows links from a matched note back to source passages. The other pulls verse fragments out of the question for fuzzy matching. Fragment matching raises nDCG@10 on verse questions by about 20 percentage points. Single-vector retrieval performs close to multi-vector ColBERTv2, yet its mean retrieval time is 0.108 s against 0.962 s, with a smaller disk footprint. A stronger reranker does better on every question type, at the cost of several times more total computation time. For multi-passage questions, what matters is finding the full set at once, not only whether the first hit is right.

3.80

8. Cascading Circle Loss: Order-Preserving Contrastive Learning for Graded Legal Retrieval

Legal retrieval must find not only the clause that directly answers, but also rank related definitions, liability caps and force majeure clauses behind it by degree of relevance, while the common InfoNCE only separates relevant from irrelevant and ignores order among positives. CasCL splits these two jobs, with InfoNCE comparing only the most relevant passage against negatives to separate relevant from irrelevant. A second loss strings the graded passages into a chain by score, with the hardest negatives at the end, and requires each level to sit closer to the query than the next by a gap proportional to the score difference. Getting each adjacent pair right then orders the whole chain. On the company's contract clause data in six languages, its NDCG@10 beats Circle Loss by 2.94 points, and removing InfoNCE drops Recall@1 from 41.14 to 28.56. Each loss handles one job, and dropping either one hurts. The intermediate relevance levels were labeled by a vote of three LLMs and were not compared against human labels.

3.75

9. VisProve: Learning Typed Visual Provenance for Scientific Document Question Answering

Citations in scientific paper QA today mostly box where an answer comes from but not how that content is used: read off directly, condensed, or reached only by reasoning. The authors built nearly four thousand QA items over full papers, labeling each piece of evidence with its page, section, region box and one of these three uses. As the model generates each token, VisProve first predicts which region to look at and reads the visual features there, then estimates which use that evidence serves, adding both back into generation as a small bounded correction. Against plain supervised fine-tuning on the same backbone, section-level F1 rises from 45.99 to 53.24 while page-level localization barely moves. The F1 that requires both region and use to match is still only 5.48. Models mostly find the right page, but are far from explaining how each piece of evidence supports the answer.

3.65

10. ARMOR: Adaptive Curriculum Meta-Learning for Noise-Robust RAG Reasoning

Retrieved documents always carry irrelevant or even misleading content, and to withstand that noise a RAG model needs three skills at once: knowing whether it already knows the answer, picking out useful documents, and reasoning to an answer amid noise. Training all three together with reinforcement learning makes their rewards interfere. ARMOR trains only one skill per round, with a scheduler picking whichever skill's score changed most recently. It also changes how GRPO computes advantages: because the subtasks have small output spaces, repeated samples for the same question are often very similar and cannot be told apart within the group, so it normalizes across the whole batch instead. When the authors mix noise into five retrieved documents, even at 80% noise, its correctness stays at 69%. Training the three skills in a fixed order instead makes the model forget what it learned earlier. Letting training order follow the model's progress is steadier than fixing it in advance.