4.70
Search a vector index for what users dislike about a product, and the reviews you get back are usually on topic but a mix of praise and complaints. This paper measures why. Across 11 text embedding models, sentiment is present in the vectors and a simple linear classifier can read it, yet the cosine similarity that retrieval actually uses barely registers it. A small projection head on the frozen embedder, trained in seconds on a CPU, repairs this. It is trained contrastively on a few thousand sentiment-labeled sentences, pulling same-sentiment sentences together and pushing opposite ones apart. When a review sentence is used to find similar sentences, more of those with the same sentiment rank near the top, and the MAP measuring this rises by up to 26 points. The cost is a lower score on the general semantic similarity benchmark STS-B, but after concatenating the head's output with the original vector, it is only 2.4 to 12 points below the original vector.
4.45
When searching a classical Arabic poetry archive, which works better: keyword search with BM25 or an embedding model? This paper finds that the answer depends mainly on how much of the original wording the user still remembers. When users half-remember a few lines, BM25 beats all nine embedding models. When users can only describe what the poem is about, BM25 drops to 8th of 10 systems. By rewriting the same queries, the authors show that the ranking is driven by whether the query contains original wording. Adding half a verse of the original to a query raises BM25 recall by 37.8 points.
4.45
Two shoppers want the same product. One asks for strawberry, lactose-free Greek yogurt in a short phrase; the other writes a longer sentence and misspells a word along the way. Will search return the same results? This paper builds a benchmark, PQB, from real user reviews in three Amazon categories to test whether results hold up when the same need is worded differently. They often do not. With the core product terms identical and only sentence structure changed, among query pairs where one query puts the target product first, roughly 27% to 52% of the time the other does not. Misspellings that follow the user's own typing habits cut the top-20 hit rate by up to 12.3 percentage points. Models with better overall effectiveness are not necessarily more robust.
4.35
Cross-encoders rerank candidate documents most accurately, but every candidate must be fully processed for every query. This paper gives the faster MICE architecture a document storage scheme, MICE-P, which comes close to cross-encoder ranking quality at about a quarter of the per-query time. In MICE, the first layers encode the query and the document separately, so the document part can be precomputed; the later layers let query tokens attend to the stored document token representations. MICE-P compresses each document token vector to 128 dimensions, stores it in PLAID format, and decompresses it at query time. PLAID clusters token vectors into centroids and stores each vector as its nearest centroid's ID plus a quantized residual. On BEIR public datasets, its average nDCG@10 is only 0.9 points below a cross-encoder of similar size.
4.15
Retrieval-augmented question-answering systems often attach a collection date to web snippets, and this paper finds that this field, which only records the fetch day, can steer an LLM's answer the wrong way, even though it says nothing about when the content changed. Ask which team an ice hockey player played for on a given day, and put both an older and a newer Wikipedia revision in the context. The texts and revision timestamps stay fixed; only the collection dates are swapped, so the older revision's date equals the question date. The experiment covers 595 Wikipedia pages. The swap lowers five open-weight models' exact-match rate on the newer answer by 37.6 to 51.4 percentage points, and telling the model not to choose by collection time barely reduces the loss.
4.05
In retrieval-augmented few-shot decisions, such as making a call against a pile of policy documents, the key evidence is often buried in long, noisy retrieved passages, and the raw passages in the demonstrations add even more distraction. This paper replaces the raw retrieved passages in each demonstration with a compact evidence summary that an LLM writes for that decision, without seeing the expert answer, and then pairs it with the expert label and evidence. At test time a single call first writes the evidence summary, then gives the decision and its support. On a synthetic policy retrieval task, accuracy is 33.0 points higher than expert-only demonstrations of the same count, and 8.0 points higher on SciFact scientific claim verification.
3.95
When retrieval-augmented generation (RAG) fails on long-document questions, is retrieval missing evidence or is the model misreading it? This paper asks models to return line-level citations with each answer and checks both whether the answer is right and whether the cited lines support it. Across 8 frontier proprietary and 3 open-weight models, the bottleneck is mainly retrieval: similarity search easily finds one relevant passage but rarely finds every passage a multi-fact question needs. Chunking also discards document-level information, such as which paper a passage came from, so questions that depend on the source go wrong. Giving the whole material to a long-context model avoids both problems.
3.90
Most questions reaching a supercomputing center's help desk were already answered in the user guides or in someone else's old ticket, but the guides are hard to navigate and users cannot search other people's tickets. This Purdue paper builds an assistant: one tool runs retrieval-augmented generation over the live documentation, another is a Llama-3-8B fine-tuned on historical ticket questions and resolutions, and a final step combines the two answers into one, with source links. On 400 held-out tickets, fine-tuning on resolutions rewritten as self-service guides works best. The merged system, however, does not beat either tool alone, and its merged answer follows the documentation tool 78% of the time.
3.45
A city area can be described by satellite imagery, by text, or by the road and amenity structure in OpenStreetMap, yet most existing methods align only images and text. This paper proposes GeoLEON-CLIP, which aligns all three views of each grid cell in one vector space and trains both satellite images and text to align with the same cell's map-structure embedding, which anchors images to text. All three encoders stay frozen and only projection heads are trained. Retrieving the best-matching map cell from an image, then using that cell's map embedding to find text, beats direct image-to-text retrieval on top-1 hit rate by 4.8 points, and the reverse direction gains 4.4 points. In downstream prediction, Beijing housing prices and Airbnb data improve the most.
3.30
Finding a person in surveillance footage from one sentence means describing both appearance and abnormal behavior. The hard part is that training can use only synthetic image-text pairs, while testing uses real surveillance frames full of near-duplicate neighboring frames and look-alike distractors. This paper fine-tunes a Qwen3-VL embedding model on synthetic descriptions, retrieves by a weighted sum of several models' normalized similarity scores, and reranks the top candidates with a cross-encoder. Every setting is chosen on a validation set built only from synthetic data, without touching the official test images. The system reaches 94.78% mAP on the AI City Challenge text-based person anomaly search track, placing 8th on the public leaderboard.