Senior ML Engineer @Meta | prev: @TikTok_us, @Amazon, @Samsung | UChicago Alum blog.reachsumit.com/ ๐ฎ๐ณโ๐ฐ๐ทโ๐ฆ๐บโ๐จ๐ฆโ๐บ๐ฒ
Seattle, WA
- Tweets 10,713
- Following 512
- Followers 4,387
- Likes 1,923
Pinned Tweet
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/20โฆ
1
2
18
10,847
Is Reasoning Always Useful? Rethinking Reasoning Utility in Universal Multimodal Embeddings
Finds reasoning in multimodal embeddings can pull hard negatives closer too, then proposes a training-free router closing part of the gap.
๐ arxiv.org/abs/2609.29560
2
3
4
868
Learned Cross-Task Relationships in Multi-Task Models
Google adds auxiliary heads estimating cross-task covariance in multi-task models, deployed across YouTube surfaces.
๐ arxiv.org/abs/2609.28776
294
OneTrans-V2: Unifying Retrieval, Pre-rank, and Fine-rank with One Transformer in Industrial Recommender
ByteDance unifies retrieval, pre-rank, and fine-rank into one jointly trained Transformer, lifting GMV by 9.74% online.
๐ arxiv.org/abs/2609.28589
1
2
5
356
Seek: Self-Evaluative Exploration for Knowledge Retrieval
@amin_bigdelii et al. present a training-free iterative retrieval framework where an LLM generates pseudo-passages, retrieves candidates, and grades their relevance.
๐ arxiv.org/abs/2609.28980
๐จ๐ฝโ๐ป github.com/aminbigdeli/SEEK
4
38
1,562
X-Rec Technical Report
ByteDance presents a generative retrieval method that models recommendation as flow matching in item embedding space, generating multiple retrieval triggers in parallel instead of quantizing items into discrete tokens.
๐ arxiv.org/abs/2609.29180
2
8
519
ScalarLens: Numerical Embeddings with Stable Coordinates and Contextual Responses for CTR Prediction
Introduces a numerical embedding for CTR prediction that separates a fixed coordinate from a context-dependent response, skipping normalization.
๐ arxiv.org/abs/2609.29182
3
319
A Systematic Multi-Domain Evaluation of Document Retrievers
Benchmarks sparse, dense, and expansion-based retrievers across 7 IR datasets under a uniform setup, comparing effectiveness, latency, and failure points.
๐ arxiv.org/abs/2609.29455
๐จ๐ฝโ๐ป github.com/valentinsvelev/reโฆ
1
4
33
1,170
OBLIQ-IR: Training a Dense Retriever for Oblique Queries
@perdactor et al. train a dense retriever for oblique queries, where relevance depends on latent traits like stance or style rather than topic overlap.
๐ arxiv.org/abs/2609.29649
๐จ๐ฝโ๐ป github.com/DataScienceUIBK/oโฆ
1
5
33
2,038
SmallReason-ColBERT: An Ultra-Small Late-Interaction Retriever for Reasoning Intensive Retrieval
@perdactor et al. present a 32M retriever for reasoning-heavy queries, adding a learned token-weighting head.
๐arxiv.org/abs/2609.29652
๐จ๐ฝโ๐ปgithub.com/DataScienceUIBK/Sโฆ
18
121
5,113
Learning Better Reasoning for Generative Recommendation with Semantic IDs
Microsoft introduces a framework that selects reasoning traces by predictive value for fine-tuning, then refines them with ranking-aware RL feedback.
๐ arxiv.org/abs/2609.29973
๐จ๐ฝโ๐ป github.com/mengdanzhu/evo-reโฆ
1
16
757
From Interests to Semantic IDs: Retrieval-Grounded Credit Assignment for Generative Recommendation
Microsoft scores each generated interest by whether it retrieves the target, giving finer training credit than the final match.
๐arxiv.org/abs/2609.29983
๐จ๐ฝโ๐ปgithub.com/YuFan-Microsoft/Rโฆ
2
15
727
Advancing Model Research in AgentX: Long-Horizon Autonomy for Industrial Recommender Systems
Kuaishou presents their dual-agent framework where a Research Agent forms proposals and a Model Agent investigates them.
๐ arxiv.org/abs/2609.30001
1
7
310
BoundaryMORPH: Budgeted Reranking via Active Set Selection for Diffuse Retrieval
Amazon uses a Gaussian Process with the dual-encoder ranking as a prior to spend a small cross-encoder budget near the top-k cutoff, filling RAG context for broad queries
๐ arxiv.org/abs/2609.27213
399
The Recall Ceiling of LLM Recommendation Reranking
Shows that low retrieval recall caps any reranker, so oracle-style evaluation overstates LLM reranking, and under realistic retrieval, none beat a simple CF baseline.
๐ arxiv.org/abs/2609.27953
๐จ๐ฝโ๐ป github.com/GeoffreyWang1117/โฆ
5
15
909
Beyond a Scalar: Distributional Serving Interfaces for Watch-Time Prediction
Serves a compact summary of a learned watch-outcome distribution instead of one watch-time estimate, so lightweight task heads can reuse it.
๐ arxiv.org/abs/2609.28383
๐จ๐ฝโ๐ป github.com/Xuanxuana1/DSI
298
Ovis-Embedding: Pushing the Frontiers of Universal Omni-Modal Embeddings
Alibaba introduces an omni-modal embedding model that encodes text, images, video, and audio in one shared space, topping several multimodal benchmarks.
๐arxiv.org/abs/2609.25165
๐จ๐ฝโ๐ปgithub.com/ATH-MaaS/Ovis-Omnโฆ
1
6
39
1,794
Lightweight Ranking Heads: Accelerating Multi-Task Experimentation in Production Recommender Systems
Google injects new prediction heads into live ranking models without retraining the backbone, avoiding task conflicts and speeding up experimentation.
๐ arxiv.org/abs/2609.25433
4
27
1,013
Efficient Iterative Retrieval with Heterogeneous Batching
Introduces a serving system that runs embedding and generation in one GPU batch instead of separate GPUs, boosting throughput and latency for iterative RAG.
๐ arxiv.org/abs/2609.25405
๐จ๐ฝโ๐ป github.com/illinoisdata/Orthโฆ
1
10
603
Robust Fusion of Semantic and Behavioural Signals for LLM Reranking in Personalised Search
@asash et al. at Spotify show that behavioral stats boost an LLM reranker but become a shortcut on rare queries, and propose paired training to stay robust.
๐ arxiv.org/abs/2609.25825
4
21
839
CoVeR: Coverage-Based Routing of Verifier Calls in Agentic Retrieval
Skips an LLM verifier call in agentic search when a cheap embedding signal shows the evidence is still incomplete, cutting most calls while keeping accuracy.
๐ arxiv.org/abs/2609.26086
11
700