Sumit

Senior ML Engineer @Meta | prev: @TikTok_us, @Amazon, @Samsung | UChicago Alum

[Sumit @_reachsumit]

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

[Sumit @_reachsumit]

Argus-Retriever: Vision-LLM Late-Interaction Retrieval with Region-Aware Query-Conditioned MoE for Visual Document Retrieval

Introduces a query-conditioned late-interaction visual document retriever.

📝 https://t.co/oB5oa4bgKc
👨🏽‍💻 https://t.co/SKzRERzgRh

[Sumit @_reachsumit]

DSIRM: Learning Query-Bridged Discrete Semantic Identifiers for E-commerce Relevance Modeling

Alibaba repositions discrete semantic identifiers as relevance features, using query-bridged contrastive quantization & LLM to predict item SIDs from queries.

📝https://t.co/iqTLbmvRZr

[Sumit @_reachsumit]

SAILRec: Steering LLM Attention to Dual-Side Semantically Aligned Collaborative Embeddings for Recommendation

Proposes an LLM-based recommender that improves the use of injected collaborative embeddings through dual-side semantic alignment.

📝 https://t.co/LLXCIqYmo5

[Sumit @_reachsumit]

ANN Search: Recall What Matters

Argues that Recall@k overstates the cost of approximation in nearest neighbor search and proposes 1/Ratio@k, a judge-free, hyperparameter-free quality measure that tracks downstream task quality more faithfully.

📝 https://t.co/zudyptX0hr

[Sumit @_reachsumit]

Cartridges at Scale: Training Modular KV Caches over Large Document Collections

@mhardalov et al. at Amazon present a training framework for scalable multi-cartridge learning that distills document collections into reusable KV caches.

📝 https://t.co/rCNEm9MaSZ

[Sumit @_reachsumit]

EviRank: Evidence-Based Confidence Estimation for LLM-Based Ranking

Estimates position-level confidence for LLM-based ranking by aggregating semantic, attention, and output evidence, with position-aware calibration.

📝 https://t.co/fCf9WXodCJ
👨🏽‍💻 https://t.co/noNieVWTQt

[Sumit @_reachsumit]

ARBOR: Online Process Rewards via a Reusable Rubric Buffer for Search Agents

Alibaba presents a reusable process-reward framework that maintains a shared rubric memory to supervise the search process.

📝 https://t.co/wXdiXGc6dj

[Sumit @_reachsumit]

Attention Calibration for Position-Fair Dense Information Retrieval

Introduces an inference-time attention calibration method with a tunable strength coefficient to reduce positional bias in dense retrieval.

📝 https://t.co/cNmqglKzuQ
👨🏽‍💻 https://t.co/2mdJThQC62

[Sumit @_reachsumit]

Do Neural Retrievers Prefer Certain Documents? Evidence of Learned Relevance Priors

Shows that supervised dense retrievers implicitly learn a query-independent relevance prior from annotation biases, making relevant but niche docs harder to retrieve.

📝 https://t.co/rZyuez4Far

[Sumit @_reachsumit]

LLM-Assisted Reranking to Operationalize Nuanced Objectives in Recommender Systems

Investigates how LLM-based reranking of news recommendations can amplify exposure to extreme or conspiratorial political content.

📝 https://t.co/vCUSHxveaj

[Sumit @_reachsumit]

Slipstream: Locality-Aware Graph Index Construction for Streaming Approximate Nearest Neighbor Search

Speeds up insertions in graph indexes for streaming nearest neighbor search by reusing candidates from previous insertions

📝 https://t.co/adYrwih5Wd
👨🏽‍💻 https://t.co/lJs69BgTeF

[Sumit @_reachsumit]

VirtualMLE: A Virtual ML Engineer that Optimizes Sequential Recommenders

Introduces an LLM-agent framework that tunes sequential recommenders through a closed loop of execution, reflection, and memory.

📝 https://t.co/qLfklrD9Ir
👨🏽‍💻 https://t.co/lHQdTEW9rw

[Sumit @_reachsumit]

Structures Facilitate Retrieve, Rerank, and Generate

Integrates document structural information across retrieval, reranking, and generation for document-grounded dialogue systems in both Chinese and English.

📝 https://t.co/yjcLth3HtO

[Sumit @_reachsumit]

Can LLM Rerankers Predict Their Own Ranking Performance?

@Shictyu et al. introduce reranker-internal query performance prediction, showing self-consistency is well-calibrated while verbalized confidence is overconfident & propose 2 methods to fix it.

📝 https://t.co/yWtA14EBbB

[Sumit @_reachsumit]

Skill Is Not Document: A Query-Conditional Benchmark and Two-Stage Retriever for LLM Agent Skill Routing

Tencent introduces a bilingual benchmark and two-stage retriever for agent skill routing.

📝 https://t.co/LS7cDMxuHk

[Sumit @_reachsumit]

Where Do Deep-Research Agents Go Wrong? Span-Level Error Localization in Agent Trajectories

Introduces a benchmark for span-level error localization in deep-research agent trajectories, and a claim-centric auditing framework that tracks agent claims

📝 https://t.co/WmSYtD6y9W

[Sumit @_reachsumit]

TVIR: Building Deep Research Agents Towards Text-Visual Interleaved Report Generation

Introduces a hierarchical multi-agent framework for deep research reports that interleave text with semantically grounded charts and images

📝https://t.co/JQV7RGEcoq
👨🏽‍💻https://t.co/jt7ywU8uJX

[Sumit @_reachsumit]

When Is 0.1% Enough? Analyzing the Combined Effects of Dimensionality Reduction and Quantization on Text Embedding Compression

Systematically studies combining dimensionality reduction and quantization for text embeddings.

📝 https://t.co/NhyGj2PxEy

[Sumit @_reachsumit]

FineVerify: Scaling Test-Time Compute with Fine-Grained Self-Verification for Agentic Search

Decomposes each question into checkable sub-questions, verifies sampled candidates against each & selects the highest-scoring answer.

📝https://t.co/JEgIuiquhF
👨🏽‍💻https://t.co/c8KU9HxL0N

[Sumit @_reachsumit]

OCC-RAG: Optimal Cognitive Core for Faithful Question Answering

Introduces a family of small language models for faithful context-grounded QA, mid-trained on a synthetic corpus of multi-hop examples.

📝https://t.co/3cLfGcFITg
👨🏽‍💻https://t.co/xkRnsd2HhA
🤗https://t.co/ErnNUU7oT1