4.60
LEAD STORY ·
Perplexity /
TturbopufferPerplexityTturbopuffer When RAG splits a long document into chunks, answering a question often needs more than the chunk that states the answer: it also needs the passages saying which product or which year the answer is about. Current contextual embedding training accepts one gold answer chunk and treats every other chunk of the same document as a negative, so these supporting chunks get pushed down. This Perplexity paper changes the source of supervision. A compression model that sees the query scores every token of the document, and a chunk's score is the mean of its top-scoring tokens. The embedding model learns to match these graded scores, so no answer-chunk labels are needed. On a new diagnostic benchmark, the 9B model finds the answer chunk 14.4 points more often than the strongest baseline, voyage-context-4; without this distillation step its rate of finding answer chunks on real search queries drops by 15.2 points on average. How chunk-level supervision is given matters more than a bigger model.
4.35
A chain embedding retriever for multi-hop retrieval sees only correct passages in its retrieval history during training, but in use it sees the passages it found itself, which are often wrong. At each step it joins the question with the passages found so far to look for the next one, so one early wrong passage stays in every later step. On four-hop MuSiQue questions, 70% of the retrieval states already contain a wrong passage by the third step. The authors let the current model run on the training questions every epoch and train it on the states it actually reaches, wrong ones included, with the earliest still-missing correct passage as the target. With the same model and training recipe, the share of questions whose full evidence is in the top five rises from 44.6% to 50.5%. Retrievers also suffer from a gap between training and test histories, and training on their own mistakes closes part of it.
4.35
Multi-hop questions need evidence found link by link through intermediate entities that the question never names. Graph-based RAG starts from the entities the question mentions and mostly reaches only the first link. On three-hop questions it recovers the full evidence chain for less than a tenth of them. Iterative methods can search step by step, but the LLM must plan, order, rewrite and keep track at once, and each task fails in its own way. DeFT-RAG has the LLM write one sub-question at a time that names every entity it asks about, and adds each answered sub-question with its answer to a QA table by fixed rules, instead of the LLM keeping track. Yes-or-no decisions, such as whether a passage is relevant or whether the answers so far are enough, go to a small fine-tuned model. With a cache shared across 1,000 MuSiQue questions, 23.1% of sub-questions are answered from the cache and LLM calls drop by 12.9%, with no loss in accuracy.
4.35
Long-form RAG needs a set of documents that complement each other, but the ten documents picked by scoring relevance one by one often cover the same aspect again and again and leave others uncovered. The authors have an LLM split the question into sub-questions, use a cross-encoder to predict which sub-questions each document can answer, and then repeatedly pick the document that covers the most sub-questions not yet covered. On TREC RAG 2024, α-nDCG rises from 0.625 to 0.702, close to the ceiling reached when selection uses human-labeled answer nuggets. More interesting, documents chosen with the human labels cover more, yet put no more nuggets into the answer. Answer completeness depends not only on what is retrieved but also on whether the generator uses it.
4.35
In an unlabeled web page, which passages talk about related things is itself usable for training a retriever. xPRT works like this: it first picks passages whose vectors are closest to the average vector of the whole page as its representative passages. For each, it finds the four passages most related to it by vector similarity and word overlap. Most words of those passages are then masked, and a one-layer decoder must fill them back in with only the representative passage's vector as extra input, which forces that vector to hold their content. The authors continue pretraining E5, which already went through large-scale contrastive pretraining, and change only the objective. Five existing objectives gain at most 0.4 points on MTEB and most end below the starting point, while xPRT gains a steady 1.8 to 1.9 points. Replacing the selected related passages with random passages from the same document removes the gain, because web pages mix in navigation and link lists.
4.35
RAG systems usually serve a reranker and a generator side by side, and the candidate passages are read once by each model. Glance starts from the idea that the generator reads these passages anyway, so it may as well rank them. It prefills all candidate passages into the KV cache once and attaches a short scoring prompt for each passage on top of that shared cache. A small set of LoRA weights and a scoring head turn each prompt into a relevance score, and then the same model generates the answer. Each extra decision skill costs about 0.1 GiB of memory. After brief fine-tuning on MS MARCO, it reaches nDCG@10 of 0.741 on SciFact, against 0.721 for the dedicated reranker bge-reranker-v2-m3. The evaluation covers only two retrieval datasets and one model family, but serving one fewer resident model has a direct effect on serving cost.
4.00
When users ask a deep research agent a question, they often leave out the time range, use the wrong technical term, or skip their preferences. The authors remove or blur each kind of information in fully specified questions and compare the answers of three deep research agents. Without the scope or the terminology, the answers cover the user's real need much less well, yet citations stay accurate and there is no more hallucination. The search traces suggest why: with a vague question the agent searches more broadly, and fewer of its search terms come from the user's words while more come from what the model learned in training. Agents trained specifically for deep research, which score higher on complete questions, lose more. Deep research agents need to learn to ask first when a question is unclear, instead of guessing.
4.00
Projecting retrieval embeddings onto a few principal directions is a common way to cut storage and scoring cost, but how much weight each direction keeps follows only an overall trend from corpus statistics, not how much that direction helps separate relevant from irrelevant documents. RRC uses relevance labels to re-estimate each direction's weight: the larger and steadier a direction's contribution to the score gap between relevant and irrelevant documents, the higher its weight. It then fits a power-law curve to the overall trend of these weights and adds only each direction's departure from that curve to the existing weights, so the original trend of high weights on strong directions and low weights on weak ones stays. It only changes the linear transform shared by queries and documents, with no model training and no extra inference step. On five datasets of the reasoning-heavy BRIGHT benchmark, compressed to 768 dimensions, nDCG@10 is on average 2.34 points higher than the original compressor SpecTemp, and the gain holds after quantization.
4.00
When fusing results from several retrievers, unsupervised methods such as RRF are cheap but do not learn, while listwise LLM reranking works well but needs hundreds of calls per query. LLM-IRA takes a middle path. It picks only the document pairs on which the retrievers disagree most and asks an LLM which one is more relevant. These judgments are used to learn the fusion weights of the retrievers, so the final ranking is still a simple weighted sum of scores. With an open 8B instruction model as the judge, it matches or beats the best static fusion on most benchmarks; on two years of TREC DL it already passes the best unsupervised fusion with about 5 LLM calls per query. It needs no training data and no fine-tuning, which suits settings that want to spend only a small LLM budget at the fusion step.
3.75
RAG corpora often contain outdated, contradictory or unreliable documents, and labeling which ones are trustworthy by hand is expensive, so labels cover only a small part. This paper treats each document as a node in a graph: two documents from the same source, or that support or contradict each other, are joined by an edge, and these relations can be judged by an NLI model. Humans mark only a few single documents as reliable or not. It then solves an optimization problem in which one part asks the trust scores to agree with the relations on the edges and another asks them to fit the human feedback, so trust spreads from the few labeled documents to the whole corpus. Both retrieval and answer generation use the score. The authors report better retrieval quality and EM than the baselines, and stable results when feedback is sparse or noisy.