AI SEARCH · 2026-10-28
AI Search — Paper Daily, 2026-10-28
The retrieval track has 10 papers today, 5 with deep dives, centered on reranking and agentic search. ACARe finds that simply giving the reranker the original question clearly lifts a search agent's answer accuracy, with dedicated training adding a little more. A systems evaluation warns that before replacing a dedicated reranker with a large model, teams should tune the existing reranker's truncation policy. DecisionRAG drops the retriever and lets a decision model browse documents level by level. SkillQR mines rewriting skills from the corpus, and CA-HNM filters hard negatives with a knowledge graph. Teams working on search reranking or RAG can start with ACARe and the reranker-replacement study.
4.45
Dense retrievers usually take hard negatives from top-ranked documents that are not labeled positive, but ranking high only shows a document resembles the query, not why it should be excluded. CA-HNM checks each candidate against a course knowledge graph, wrong concept, wrong scope or wrong difficulty level, and keeps only those with a clear conflict as negatives. It beats a rank-matched control without the graph in most settings, yet randomly shuffling the graph on one dataset raises NDCG@10 from 0.0311 to 0.1174, so how much the graph helps is still open.
4.45
Most RAG systems still retrieve with BM25 or vector similarity, and evidence the retriever misses cannot be recovered by any generator. DecisionRAG drops the retriever and has a decision model browse like a person in a library, from topics to pages to passages, judging each level's candidates together in one context. Over 2.68 million Wikipedia passages its Hit@1 is 17.4 points above Qwen3-Embedding-8B retrieval plus a reranker, at about 1.7 seconds per query.
4.45
NVIDIA
University of WaterlooSearch agents often send sub-queries of just a few keywords, so a sub-query-only reranker promotes documents that match the words but do not help the real question. ACARe gives the reranker the original question and the agent's reasoning too. The key control: without changing Qwen3-Reranker-4B's weights, adding the question alone lifts a small agent's accuracy from 48.2% to 54.8%, and fine-tuning brings it to 57.6%. Most of the gain comes from the context itself, and all five agents need fewer searches.
4.35
LLM query rewriting usually applies one instruction to every query, yet the right direction for a rewrite depends on the corpus. SkillQR clusters the corpus offline and has an LLM write a few rewrite instructions for the document types, such as listing the required data structures and algorithms; at query time the same model picks one before rewriting, with no training data. Its mined skills match hand-written subject prompts, and the 14B version takes about 38 seconds per query versus 85 for iterative ThinkQE, showing what dropping multi-round iteration saves.
4.20
If a team wants to replace its RAG reranker with a large model, this paper says to tune the existing reranker first. Comparing the 27B decision model PPLX Decider with the 0.6B Jina v3.5 on identical candidates, default PPLX returns valid results for only 44.1% of requests; it looks better on completed requests but clearly loses once failures count. Splitting the input budget evenly and keeping each candidate's opening lets it match Jina, yet Jina with the same truncation drops to 245 ms and needs a fraction of the memory.
4.15
RAG poisoning defenses often report near-perfect detection, but this paper finds much of it comes from how poison sets are built: PoisonedRAG's code prepends the target question to every poison passage. Deleting only that prefix raises the published RAGuard filter's miss rate from 0.191 to 0.763, which means it mostly recognizes the prefix. The authors argue detection numbers mean little unless reported with the poison set's construction profile and the false-positive rate at the deployed threshold.
4.05
After each search, a search-augmented model must decide whether its evidence suffices: answer, search again, or abstain. Leaving this to holistic model judgement lets it answer when every entity is mentioned but one required fact is missing. Certify splits the question into fact requirements, stating which facts together yield the next, and searches for each; a passage satisfies a requirement only if it is about the right entity. It answers only when the requirements derive the answer, and otherwise names what is missing. Over 21 million Wikipedia passages, answers from its minimal evidence set are 80% correct versus 57% from all retrieved passages.
3.91
When RL teaches a model to abstain when unsure, typical rewards penalize every wrong answer equally and give abstention a fixed reward, ignoring confidence. CalibraRL adds a Brier penalty based on confidence and correctness, so confident errors cost more. For abstention, it estimates from other rollouts of the same question whether the model can answer reliably, rewarding abstention more when it cannot. It needs no extra labels, reduces hallucination on three QA datasets, and cuts mean calibration error by 30.8%.
3.91
Knowledge graphs in finance or industry rarely have QA annotations, so there is no data to improve LLM multi-hop graph reasoning with RL. EoG has one model alternate between proposer and solver: it samples reasoning structures on the graph and executes logic queries for exact answers, and the proposer writes multi-hop questions from them. The solver answers each question several times, and the proposer is rewarded most when results are mixed, not for questions that are too easy or too hard, keeping questions at the edge of the solver's ability. On six benchmarks over Freebase, Wikidata and an industrial graph, it beats prompting, retrieval and agent methods.
3.90
In formulation patents, relevant and distracting passages often name the same chemicals and differ only in the role they play, which generic embedding models struggle with. FormBench tests this with over 55,000 queries, with distractors chosen through a materials taxonomy to share most chemicals with the answer but differ in function. The results are striking: OpenAI's commercial embedding scores just 0.009 nDCG@10 on the hard corpus against 0.375 for BM25, which means the generic space collapses within the domain; only Qwen3-Embed fine-tuned on the benchmark's triplets beats BM25.