AI SEARCH · 2026-11-04
AI Search — Paper Daily, 2026-11-04
Today's AI Search track has two themes. One is making retrieval and reranking fit what users actually need. WeSearch at WeChat Search stops forcing multi-topic documents into a single Semantic ID path and now serves all traffic. Re3Rank has the model write a reasoning trace and uses it to read attention scores, and DMReranker splits the requirements attached to a query into small questions judged one by one. The other is agent-style search. DR Pixel retrieves and reasons directly over paper page images, Mind2Report writes commercial analysis reports, and another paper trains models to ask a follow-up question when a request is missing information.
4.70
LEAD STORY ·
TencentTencent Generative retrieval usually gives each document one fixed hierarchical Semantic ID, which assumes a document follows a single semantic path. Many articles cover several topics at once, so that assumption breaks. WeSearch at WeChat Search changes the order: it first finds the document clusters relevant to the query, then scores individual documents inside those clusters, separating coarse pruning from fine relevance judgment. The index updates incrementally, so the model can be refreshed weekly or even daily. It now serves all WeChat Search traffic, and a month-long full-traffic A/B test showed CTR up 0.68% and retrieval resource cost down 41%. Multi-topic documents no longer have to squeeze into one path.
4.45
Attention-based reranking looks at a few retrieval-focused attention heads in an LLM and measures how much attention they send from the question to each candidate document. But the question is often short and carries few clues. Re3Rank first has the model write a reasoning trace over the question and the candidates, then uses the trace tokens to query those heads, reading the attention each candidate receives as its relevance score. The trace is made of sampled discrete tokens, so gradients cannot flow back. During training, the forward pass scores with the sampled trace as usual, while the backward pass uses the token probabilities from the generator to weight the embeddings, so the ranking loss reaches the part that writes the trace. Across 9 benchmarks it beats the strongest baseline by 3.11 NDCG points on average.
4.45
Deep research agents that read papers usually OCR each page into text first, and much of the tables, figures and equations is lost at that step. DR Pixel skips the text conversion. It renders 1.06 million paper pages as images and does both retrieval and reasoning directly on the page images. Without any training, this beats the OCR-first pipeline on several of the strongest closed and open models, and it calls tools less often. The authors then fine-tune an open model on teacher trajectories, and give each step its own weight. They remove the step and see how much the final answer quality changes, and the smaller the change, the lower the weight. After training, retrieval recall rises by about 15 points and answer accuracy by more than 10 points. The OCR-then-retrieve step is not just unnecessary, it holds the agent back.
4.45
People often attach requirements to search queries, such as pages for beginners, documents in a certain language, or official sources only. Most rerankers were only trained to judge topical relevance, so they treat these requirements as a few extra words. The InfoSearch benchmark measures this with WISE, which gives credit when documents that meet the requirement move up under it and move down under the opposite requirement, with moves near the top counting more. Even GPT-4o used as a reranker averages only 33.5 on WISE. DMReranker splits an instructed query into a topic and a few concrete conditions, has a small model answer one narrow question per condition, and multiplies the probabilities that each condition is met into one score. Nothing is trained, and WISE reaches 50.9. Splitting one complex requirement into several small questions works better than asking a large model to judge the whole sentence at once.
4.15
Using Monte Carlo tree search to help an LLM find answers step by step over a knowledge base runs into one hard question: how to score the intermediate steps. One option is to simulate every step to the end and check the answer, which is expensive. Another is to train a separate reward model, which needs a lot of data. This paper instead gives the query history up to that step to an open instruction model and computes the perplexity of that history with and without the original question attached. The more the perplexity drops once the question is attached, the more relevant the findings are to the question, and the higher the reward for that step. It takes only a few forward passes, with no simulation to the end and no reward model to train. On 4 KBQA benchmarks it consistently beats step-by-step methods without search, and compared with classic tree search that simulates to the end it usually offers a better trade-off between accuracy and cost.
4.13
User questions often lack key information, yet LLMs rarely ask back. Testing several models of different sizes on real user questions, the authors found that fewer than 11% of underspecified questions get a follow-up question. They remove details from complete questions to create vague versions, so what was removed is known in advance. Their training method, DialogGRPO, first uses an external LLM to expand conversations offline into a graph, where nodes record which details have been drawn out and edges are candidate follow-up questions. The new details each question draws out can simply be counted. The method then samples a few follow-up paths on the graph from the state after that question and averages what they draw out, and the two together give the question its score. After training, the model draws out the missing information on more than 88% of vague questions and does not ask needlessly when the question is complete.
4.12
Judging whether a product matches a query in e-commerce search is hard when evidence conflicts: the attributes look right, the user intent does not, and business rules add their own requirements. A single reasoning pass struggles to sort these conflicts out. CADRE has three experts examine product attributes, user intent and business rules separately; they share one frozen LLM, and a trainable verifier model checks each expert's evidence. When the experts disagree, the verifier asks targeted follow-up questions so the experts clarify or revise before the final call. The full process is too heavy to serve online, so the authors use it as a teacher and distill it into a lightweight model for production. In a 7-day A/B test on a search advertising platform with 10% of traffic per arm, both business metrics and relevance satisfaction improved.
4.05
In multimodal retrieval, a video may lack audio and have noisy captions, so learning a good representation from that one sample's own modalities is hard. During training, this paper treats each modality of each sample as a node, and hypergraph edges link the modalities of the same sample and those of semantically related samples, with one edge able to connect several nodes at once. Features pass along these edges, so a sample's representation absorbs information from similar samples. A missing modality simply has no node and is not filled in. This information is learned into the encoder during training, so at retrieval time the encoder produces representations directly and the hypergraph is dropped, leaving the pipeline and its cost unchanged. Across 17 video settings it beats GRAM, an existing multimodal alignment method, in both retrieval directions, and mean text-to-video R@1 rises from 63.4 to 68.2. The gain holds even when nine in ten clips lack one modality.
4.05
CIKM 2026IIFlyTek (China)UUniversity of Science and Technology of ChinaDeep research agents that write commercial analysis reports tend to have three problems. Unclear intent makes the search drift off topic. Web content quickly fills up the context window. The report is written in one pass and misses things. Mind2Report follows how an analyst works: it first breaks down what the user really wants and drafts an outline. It then searches the web layer by layer and distills verified evidence into a research memory, instead of stuffing raw pages into the context. The outline keeps adjusting as the memory grows, and the report is finally written over several rounds following the outline. The authors also built an evaluation set of 200 real commercial tasks scored on quality, reliability and coverage, where Mind2Report beats leading closed and open agents of the same kind.
3.80
CIKM 2026
Beihang University
Chinese Academy of SciencesIInstitute of Computing Technology Documents retrieved by RAG may carry private information, and erasing names within a single document is not enough: combine several documents and the removed information can still be inferred. This paper first extracts the retrieved documents into one shared knowledge graph and lets the user specify which facts count as private, then trains Flan-T5 to rewrite the documents, removing private facts and keeping public ones. It then continues optimizing with PPO, giving a higher reward when less private information remains and more public knowledge is kept. On 4 question-answering datasets it erases private information better than rewriting with GPT-4o directly.
Fudan University
Peking University
EMNLP 2026