AI SEARCH · 2026-10-27
AI Search — Paper Daily, 2026-10-27
Retrieval has 10 papers today, five with deep dives. A University of Passau study gives a direct answer: if you need a reranker, fine-tune a reranker, because the new decision models are not better. LightRetriever turns query encoding from a full LLM pass into a table lookup, which suits high-volume online search. Others cover the modality gap and reasoning-enhanced multimodal embeddings, plus retrieval over scanned PDFs. Teams building search or retrieval-augmented generation can start with the first two.
4.70
Using the new decision models to rerank search results does not beat a purpose-tuned reranker, concludes this University of Passau paper. Decision models, also called System One models, answer a question with a calibrated probability of yes instead of text, and within weeks of release they were used for reranking. The paper compares two fine-tunes of the same backbone: decision fine-tuning trains on all kinds of yes-or-no questions to give accurate probabilities, while reranker fine-tuning trains only on MS MARCO queries and passages to judge relevance. Across six Qwen backbones of different sizes and ten TREC DL and BEIR collections, reranker fine-tuning wins on every small and mid-sized backbone, by 5 to 29 points nDCG@10 up to 9B, and the two tie at 27B. Below that size, decision fine-tuning is no better than asking the untuned model. The authors' advice is blunt: if you need a reranker, fine-tune a reranker.
4.70
In LLM-based text retrieval, documents can be encoded offline, but queries arrive in real time, so every query runs the full LLM, which keeps throughput low and serving costly. LightRetriever breaks the symmetry: documents still go through the full LLM, while the query side is reduced to a table lookup. During training, a shared prompt plus a single token is fed through the LLM to get that token's vector, and a query vector is the average of its token vectors. Every vocabulary token's vector can be precomputed into one table, so at serving time a query is just lookups and an average. Versus running a full LLM on an A800 GPU, query encoding is over 1000 times faster and end-to-end retrieval throughput rises more than tenfold, while keeping about 95% of the full LLM retriever's retrieval scores on average across benchmarks.
4.45
In contrastive models like CLIP and CLAP, the two modalities are trained to align yet their embeddings often sit in offset regions, a phenomenon called the modality gap. This paper leaves the backbone untouched and changes only the space the embeddings live in: the unit sphere is replaced by the probability simplex, where every dimension is nonnegative and all dimensions sum to one. In practice a light softmax adapter is attached to each output, and similarity is computed as total variation distance between the two distributions. Both modalities are forced into the same coordinate system, and the offset nearly vanishes. On five cross-modal retrieval benchmarks, the distance between the modality centroids shrinks by 97% to 99%, with retrieval matching or beating standard models.
4.35
Some multimodal embedding models first have the model write reasoning about the input and then embed it along with the input, producing richer vectors. But training data only says which pairs are relevant, never what reasoning to write, so exploring from the input alone rarely finds reasoning that actually helps retrieval. PILOT has one multimodal LLM play two roles: one writes reasoning from the input only, while the other also sees the matching target during training and so finds useful reasoning more easily. Some of the second role's reasoning is transferred to the first, weighted by two things: how well the embedding built with that reasoning retrieves, and how likely the first role would be to write it from the input alone. Only the first is used at inference. On the 78-task MMEB-V2, overall scores rise by 3.2 and 3.3 points at 2B and 7B, with the largest gains in visual-document retrieval.
4.15
Persian dense retrieval has lagged for lack of dedicated training data and compact embedding models. This paper continues training Qwen3-Embedding-0.6B on over a million Persian query-document pairs, producing the roughly 600-million-parameter Yalda Embedding, with training and data released openly. On eight Persian retrieval tasks from FaMTEB, it has the best average among models its size, beats the original model and the same-sized F2LLM-v2 on every task, and approaches the several-times-larger Qwen3-Embedding-4B. Ablations show contrastive fine-tuning alone helps little and even hurts some tasks; raising negatives per query from 63 to 511, adding gradient clipping and learning-rate warmup, and having it learn the much larger F2LLM-v2-8B's relative ranking of each batch's documents are what lift the scores.
4.05
When retrieving images with long text, models tend to fixate on the few sentences describing prominent objects and barely use the surrounding context about setting and details. This paper calls it saliency collapse: removing part of the query's salient text cuts retrieval accuracy much faster than removing the same amount at random. RoSA splits each image with object-level visual saliency estimation and each text with an LLM, yielding a salient view and a contextual view of every pair. Training adds objectives that align each view on its own, so the model learns both kinds of information. Long-text retrieval improves clearly, holds up better when salient content is removed, and the extra objectives plug into existing methods.
3.80
Public-service questions often mention several places, but only one of them decides which service applies, say someone living in one emirate licensing a company in another. This paper builds a controlled English-Arabic retrieval diagnostic for the UAE and compares twelve methods. Deciding which place controls the service, keeping only that place's service entries, then fusing keyword and vector rankings works best: on answerable bilingual questions its first result is right about 96% of the time, versus 65% without the filter. But the routing module that decides which action and location the user needs is fragile: under paraphrases, irrelevant facts or negations, it gets only 2 of 168 variants right.
3.80
Vision-language models answering questions about tables in document images often misread rows and columns or miscalculate, and fine-tuning needs costly expert labels. Self-RefineVQA needs no training and no labels: one model takes three turns, first answering, then checking like a reviewer whether the answer matches the table, then revising according to the critique, for several rounds. Small models gain the most: Qwen2.5-VL 3B improves by up to 13% on WikiTableQuestions, rivaling far larger models. Inference cost grows linearly with rounds, but human labeling is eliminated.
ICLR 2026