Joined April 2010
- Tweets 10,121
- Following 506
- Followers 4,134
- Likes 1,887
1,682 Photos and videos
Pinned Tweet
8 Oct 2025
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.
blog.reachsumit.com/posts/20β¦
1
1
17
7,856
Jun 12
Reasoning with Memory: Adaptive Information Management for Retrieval-Augmented Generation
Amazon presents a RAG framework with an explicit working memory managed by a trainable extractor that filters and consolidates information across reasoning steps.
πamazon.science/publications/β¦
2
2
23
796
Jun 12
LoHoSearch: Benchmarking Long-Horizon Search Agents Beyond the Human Difficulty Ceiling
Meituan introduces a challenging search agent benchmark built from a knowledge graph of over 7 million Wikipedia entities, systematically maximizing search space.
π arxiv.org/abs/2606.12837
1
12
706
Jun 12
DailyReport: An Open-ended Benchmark for Evaluating Search Agents on Daily Search Tasks
Introduces an open-ended benchmark for evaluating search agents on everyday search tasks, using cascade rubrics.
π arxiv.org/abs/2606.12871
π¨π½βπ» github.com/AGI-Eval-Officialβ¦
2
13
423
Jun 12
Iterating Toward Better Search: A Two-Agent Simulation Framework for Evaluating Agentic Search Architectures in E-Commerce
eBay presents a modular two-agent simulation framework that pairs a configurable buyer agent with interchangeable responders.
π arxiv.org/abs/2606.12924
1
255
Jun 12
EvoBrowseComp: Benchmarking Search Agents on Evolving Knowledge
Tencent introduces an evolving benchmark of 400 English and 400 Chinese contamination-free questions, designed to prevent parametric memorization.
π arxiv.org/abs/2606.13120
π€ huggingface.co/datasets/Krysβ¦
2
275
Jun 12
How Fine-Grained Should a RAG Benchmark Be? A Hierarchical Framework for Synthetic Question Generation
@ChaseF1 et al. find optimal granularity to vary question characteristics when building synthetic RAG eval benchmarks.
π arxiv.org/abs/2606.12789
π¨π½βπ» github.com/fensorechase/rag-β¦
15
551
Jun 12
OneRetrieval: Unifying Multi-Branch E-commerce Retrieval with an Editable Generative Model
Kuaishou unifies multi-branch e-commerce retrieval into one generative model that keeps the inverted index's real-time editability.
π arxiv.org/abs/2606.13533
π¨π½βπ» github.com/xuxinzhang/oneretβ¦
5
352
Jun 12
When Does Mixing Help? Analyzing Query Embedding Interpolation in Multilingual Dense Retrieval
@tongyao_zhu et al. show that interpolating monolingual query embeddings often beats the best single-language query.
π arxiv.org/abs/2606.13537
π¨π½βπ» github.com/tongyao-zhu/queryβ¦
14
617
Jun 12
Learning to Reason by Analogy via Retrieval-Augmented Reinforcement Fine-Tuning
Meta presents a post-training framework that teaches LLMs to reason by analogy, retrieving examples by reasoning utility rather than semantic similarity.
πarxiv.org/abs/2606.13680
12
484
Jun 11
TreeSeeker: Tree-Structured Trial, Error, and Return in Deep Search
Introduces an inference-time framework that organizes deep search as branch-and-return search over tree-structured states.
π arxiv.org/abs/2606.11662
14
518
Jun 11
FORT-Searcher: Synthesizing Shortcut-Resistant Search Tasks for Training Deep Search Agents
Introduces a framework for synthesizing shortcut-resistant training data that forces genuine multi-step search.
π arxiv.org/abs/2606.12087
π¨π½βπ» github.com/RUCAIBox/FORT-Seaβ¦
1
14
797
Jun 11
The Structural Attention Tax: How Retrieval Format Hijacks In-Context Learning Independent of Content
Identifies a "structural attention tax" where knowledge graph triples capture 2β3Γ more attention than equivalent natural-language text.
π arxiv.org/abs/2606.11198
1
5
520
Jun 11
CompRank: Efficient LLM Reranking via Token-Level Compression and Decoding-Free Scoring
Introduces a token-efficient reranking framework that decouples document representations and uses attention-based scoring, retaining only ~10% of document tokens.
π arxiv.org/abs/2606.11700
5
345
Jun 11
What Limits Does Quantization Place on Dense Top-k Retrieval? A Theoretical Study
Presents a theoretical study showing that the corpus-independent O(k) embedding bound breaks under quantization.
π arxiv.org/abs/2606.11780
1
13
679
Jun 11
LLM-Based User Personas for Recommendations at Scale
Google generates real-time, natural-language user interest personas during serving, balancing interest summarization with novel-topic exploration via KD and asynchronous inference.
π arxiv.org/abs/2606.12198
3
18
854
Jun 11
Doc-to-Atom: Learning to Compile and Compose Memory Atoms
Introduces a compositional parametric memory framework that decomposes each document into semantically typed knowledge atoms.
π arxiv.org/abs/2606.12400
2
12
502
Jun 10
SIDInspector: A Mapping-First Diagnostic Resource for Semantic-ID Tokenizers
Huawei presents a tool to validate and profile semantic-ID tokenizer mappings before generator training.
π arxiv.org/abs/2606.10375
π¨π½βπ» github.com/jdding/sidinspectβ¦
1
7
591
Jun 10
STORM: Stepwise Token Optimization with Reward-Guided Beam Search
Trains LLM query rewriters with reward-guided beam search, turning retrieval rewards into a token-level signal so BM25-based lexical retrieval rivals dense retrievers at much lower cost.
πarxiv.org/abs/2606.10621
2
10
3,476
Jun 10
Effective Reinforcement Learning for Agentic Search by Recycling Zero-Variance Queries During Training
Proposes recycling zero-variance rollout groups back into the RL training pool, letting a 1.7B search agent match or surpass 7B systems.
πarxiv.org/abs/2606.10709
1
351
Jun 10
miniReranker: Efficient Multimodal Reranking through Visual Cache Reuse and Interaction Sparsity
Speeds up multimodal LLM reranking via vision-first prompting, early exit, and visual token pruning, cutting reranking runtime by up to 99%.
π arxiv.org/abs/2606.10759
1
1
18
648