AI Search
Contextual embeddings are trained and evaluated as if a question had only one relevant chunk in a document, and this paper shows that assumption fails in RAG.
AI Search
Chain embedding retrievers for multi-hop retrieval see different histories in training and in use, and this paper fixes that without adding any LLM call.
AI Search
The work that multi-hop question answering piles onto one LLM can be split up.
AI Search
In long-form RAG, more complete retrieval does not necessarily mean a more complete answer.
AI Search
Which passages of the same web page talk about related things is, by itself, enough to train a retriever. xPRT first takes passages whose vectors are close enough to the page's mean vector as anchor passages, then, by vector similarity and word overlap, finds for each anchor the four most related passages of the same document.
AI Search
RAG systems usually serve a reranker and a generator side by side, and the candidate passages are read once by each model.
Personalization
The same cooking video is a pastime for a casual viewer and source material for a food creator.
Personalization
This Meta paper gives long-sequence ranking a new structure in which reading more history and stacking more layers become two separately tunable knobs: with 10K history events, its NE beats HSTU by 0.19% while using 69.6% less forward compute.
Personalization
Before an ad system launches a new ranking model, it must pick the best candidate offline from logs.
Infra
The table in vLLM's documentation that sets how many tokens speculative decoding guesses per step at each batch size can visibly cut throughput on a different GPU or model.
Infra
After analyzing eight production LLM request traces, the authors find that the common practice of treating host memory as a KV cache layer in front of the SSD barely filters any reads under these workloads.
Infra
In LLM serving, running prefill and decode together lets their resource needs complement each other, since prefill uses compute and decode uses memory bandwidth.
Infra
Open models built for agents often have to choose between capability and inference cost. NVIDIA's Nemotron 3.5 Super optimizes both.
Infra
The Kling team's TIDE builds the layout conversion that FSDP2 performs every step directly into communication.
Infra
Putting an MoE model's expert weights and its KV cache into one shared pool of GPU memory, and shifting how much each side gets as load changes, lets a memory-limited machine serve many more concurrent requests.
Generative
Kuaishou made three changes to GR4AD, its deployed generative ad retrieval system, and the result, GR4AD-X, raises online ad revenue by a cumulative 3.07% over GR4AD.
Generative
The authors at Alibaba found that when a generative recommender decodes with beam search, attention concentrates on a few past behaviors as decoding proceeds, and which behaviors matter stays largely the same from step to step.
Generative
Taobao's generative ranking framework UGR has one generative model directly produce the screen the user finally sees.
Generative
A generative recommender produces an item's SID one code at a time, and once the first two codes pick the wrong branch, no amount of later precision can recover the relevant items.
Embodied AI
Most of the training data came from handheld grippers, collected at a different place and under different lighting than the test scene, with only a small share of real-robot teleoperation.