In the final post of the Adaptive RAG series, we explore how to treat selective retrieval as a core, learned skill, moving from passive observation to active, intelligent decision-making.
https://t.co/MyjupeCBOS
Intelligent Embedding Pruning for Large-Scale Recommendation Models
Meta uses feature importance to decide how much to prune each embedding table in large recommendation models, cutting more memory than size-based pruning at equal or better accuracy.
๐ https://t.co/2YzKaw4OXB
LQ-Rec: Late Quantization for Sequential Recommendation
Proposes a recommender that stores items as semantic IDs but keeps the encoder in continuous space, so longer IDs improve ranking and one model supports anytime compression.
๐ https://t.co/62g1sIyIeB
AutoRAGBench: A Benchmark for LLM Agents Automating RAG System Development
Introduces a benchmark that tests how well LLM agents can build RAG systems from scratch and improve them through iteration across nine tasks.
๐ https://t.co/aUoMUTJtvZ
๐จ๐ฝโ๐ป https://t.co/D0ojGRSYFL
Cost-Aware Agentic Feature Optimization System for Fleet-Scale Ranking Models
Meta presents an LLM agent system that selects features for a whole fleet of ads ranking models while keeping serving cost and shared feature pipeline cost within budget.
๐ https://t.co/qEwAU4p2OB
Evidence-Based Credit Assignment for Search-Augmented Language Models
Introduces a training-only reward crediting each search by how much its retrieved documents raise the likelihood of the correct answer, so agents learn even when all rollouts fail.
๐ https://t.co/F0b9IUR8jT
Optimize the Cob, Then the Kernel: Optimizing Beam Search for Generative Recommendation Inference
Huawei cuts generative recommendation inference latency versus vLLM on Ascend NPUs by optimizing the whole pipeline rather than writing custom kernels.
๐ https://t.co/hSqXMiEiQx
MLE Agents: Autonomous ML Engineering in Large-scale Recommendation Systems
Meta uses LLM agents that propose, train, and evaluate changes to a large search ranker, guided by engineers.
๐ https://t.co/dNMlpYE0LR
CONCORD-Rec: Catalog-Conditioned Modality Utility for Semantic-ID Diffusion Recommendation
Tencent proposes a diffusion recommender that learns how much to trust each content modality, so missing or conflicting audio, video, image or text hurts less.
๐ https://t.co/RiAyKwqZjU
MARS Technical Report
Taobao presents an LLM agent that periodically sets user-specific retrieval quotas, model choices and score fusion weights across the whole recommendation pipeline, replacing one fixed configuration shared by all users.
๐ https://t.co/BLAvkXOhMX
X-split Argus: Learning Reusable History Tokens for Efficient CTR Prediction
Yandex presents an ad CTR ranker that compresses long user histories into reusable offline tokens, then scores ads with cross-attention to those tokens and recent events.
๐ https://t.co/vsI0KpPwnC
Dynamic Self-correcting Interaction Network
Proposes a dynamic interaction network that filters out unhelpful past interactions for both users and items, models the negative effect of new interactions, and handles cold-start users and items.
๐ https://t.co/8xW4CPR4Vm
Errpedia: Is Your Exact k-NN Vector Search Really Exact?
Introduces a dataset for testing numerical correctness of exact k-NN search; every tested library and database returns wrong top-10 results on some queries.
๐ https://t.co/Gj6GcoPbZJ
๐จ๐ฝโ๐ป https://t.co/v2onmdgScc
RAG in a Trenchcoat: When Minimal Memory Is Enough for Agentic Systems, and When It Isnโt
@Jingyu227 et al. show that simple RAG memory matches complex agentic memory on popular benchmarks, but no design wins everywhere.
๐ https://t.co/Xd0wUVwcfn
๐จ๐ฝโ๐ป https://t.co/FV39jO7n2B
KVUse: Reusable-KV User Memory for Efficient LLM Recommenders
Meta introduces a method that stores user profiles as reusable KV caches instead of text, so LLM ranking calls skip re-reading them, raising throughput without hurting ranking quality.
๐ https://t.co/SIwU2D3ZpM
SIMQ: A Single Graph-based Index for Multi-Vector Query
Introduces a single graph index for multi-vector retrieval. It approximates MaxSim via per-token graph search over data-aware clusters, with no interaction matrix.
๐ https://t.co/YiDhzGTqPr
Cluster-Guided Decoding with Trust-Region Constraints for Personalized Recommendation
Amazon adds a learned semantic cluster prior to item scores.
๐ https://t.co/xxsViwATlR
Efficient Wide-Beam Generative Retrieval with Model-Native Trie Decoding and Shared-Context Attention
LinkedIn replaces Python constraint callbacks with a GPU-native ID trie and shares prompt attention across beams, speeding up wide-beam retrieval.
๐ https://t.co/bp10mvq55M
Deal With It: On Better Schedules for Listwise Reranking
@bclavie et al. at Mixedbread show that dealing documents round-robin by retriever rank, like cards, is the best split for parallel LLM reranking calls.
๐ https://t.co/ek8BZbuBQh
๐จ๐ฝโ๐ป https://t.co/r9fm7q6NFq
More Than a Score: Reinforcement Learning with Verbalized Feedback for Generative Recommendation
Tencent introduces a system where a simulated user critiques each recommendation in words and the model refines its picks, going beyond scalar rewards.
๐ https://t.co/88G6tZB3qU
From Relevance to Utility: Rubric Alignment in Yuanbao Search
Tencent proposes rubric-aligned search, which trains retrieval, ranking, and document selection against one per-query rubric so LLMs get useful document sets.
๐ https://t.co/1jTWS7BJLP