Tag
IBM researchers introduce STAIR, a generative retriever that uses table of contents to preserve document structure, achieving 82.6% Recall@1 and reducing hallucination in RAG systems.
This paper explores using a model's semantic ID hierarchy for off-policy evaluation in generative recommenders, showing that coarsening to code-prefix clusters improves estimation accuracy under production logging constraints.
PRQ-KMeans is a post-hoc tokenization method that improves residual quantization for semantic identifiers by removing global-mean components, refining centroids, and using projection residuals, achieving significant performance gains in industrial search and public recommendation benchmarks.
Kuaishou researchers propose UniGD, a unified generative-discriminative framework for industrial retrieval that integrates retrieval and relevance scoring into a single model, with techniques like CAGE and CAM to improve effectiveness and reduce latency. Online A/B tests show a 5.78% ad revenue increase and 33.1% inference latency reduction.
FlashTrie presents a GPU-accelerated constrained beam search for generative retrieval, using a succinct trie layout and cooperative CUDA kernels to reduce decoding latency and enable real-time serving at scale, achieving up to 24× speedup and a 0.71% revenue lift in a commercial search engine.
ORBIT proposes a method to mitigate catastrophic forgetting in large language models fine-tuned for generative retrieval by tracking parameter distances and using weight averaging, outperforming common continual learning baselines.