Tagged articles
beam search
24 articles · Page 1 of 1
This paper introduces GRC, a generation-reflection-correction framework for generative recommendation that enables models to self-diagnose and correct full draft semantic ID sequences via structured reflection labels, GRPO reinforcement learning, and entropy-guided scheduling, achieving state-of-the-art results on public and industrial datasets with minimal latency overhead.
CIKM 2026GRPOReflection
0 likes · 12 min read

Xiaohongshu built a custom inference engine GR-Inference for generative search retrieval, addressing unique load characteristics — long context, short decode, large dynamic beam, and constrained generation — that break general frameworks, achieving 1.5–3.6x throughput over SGLang and improving recall and click-through rates.
GR-InferenceInference OptimizationLLM serving
0 likes · 7 min read

A Tencent senior backend engineer with 15 years experience shares his transition to LLM inference engineering, detailing learning paths for model architecture, inference optimization, and self-assessment exercises, while arguing that traditional infrastructure skills remain crucial for AI Infra and Agent Infra development.
AI infrastructureBackend EngineeringKV Cache
0 likes · 38 min read

The article analyzes the limitations of traditional ID‑based and vector recall, proposes a generative Semantic‑ID‑based recall framework (ASGR), details a series of inference‑time optimizations—including a custom SDPA‑enabled optimizer, novel beam search, static‑cache compilation, tree‑aware attention, sliding‑window attention, quantization, ONNX/Optimum conversion, and TensorRT‑LLM integration—and reports concrete latency and throughput gains that enable large‑scale, low‑latency recommendation serving.
Inference OptimizationRecommendation Systembeam search
0 likes · 25 min read

This article walks through the mechanics of greedy search and beam search in large language models, demonstrates both methods with GPT‑2 on the prompt "I have a dream", visualizes the decoding trees, compares their scores, and discusses the trade‑offs between efficiency and output quality.
GPT-2Greedy SearchLLM
0 likes · 16 min read

The article explains how sampling parameters—Temperature, Top‑k, and Top‑p—shape the output of large language models, comparing greedy and beam search, illustrating probability changes with concrete examples, and offering practical guidance on adjusting these settings for different tasks.
Greedy SearchLLMbeam search
0 likes · 9 min read

This report details how large language models (LLMs) were integrated into Tencent's search advertising pipeline—from early extraction‑distillation experiments in 2023 to a 2024 end‑to‑end generative recall architecture—showing significant improvements in relevance, diversity, and revenue through knowledge injection, supervised fine‑tuning, constrained beam‑search decoding, and high‑performance inference services.
AILLMbeam search
0 likes · 11 min read

This review of the recent MindSpore NLP public class walks through the fundamentals of large language model text generation, detailing deterministic decoding such as greedy and beam search, stochastic sampling techniques like temperature, top‑k and top‑p, and advanced methods including constrained beam, contrastive, and assisted search, with illustrative examples.
Greedy SearchLLMMindSpore
0 likes · 5 min read

The article explains how generative models use deterministic methods like greedy and beam search and stochastic techniques such as top‑k, top‑p, contrastive search and sampling, describing their mechanisms, temperature control, repetition penalties, and practical trade‑offs for balancing fluency, diversity and coherence.
AIbeam searchcontrastive search
0 likes · 9 min read


ArchitectJul 1, 2023 · Artificial Intelligence
This tutorial explores various text generation decoding methods—including greedy search, beam search, top‑k/top‑p sampling, sample‑and‑rank, and group beam search—explaining their principles, providing detailed Python code examples, and comparing their use in modern large language models.
Greedy SearchHuggingFaceTransformers
0 likes · 59 min read

This guide thoroughly explains the major decoding strategies for neural text generation in HuggingFace Transformers—including greedy, beam, diverse beam, sampling, top‑k, top‑p, sample‑and‑rank, beam sampling, and group beam search—detailing their principles, Python implementations with LogitsProcessor components, workflow diagrams, comparative analysis, and references to original research.
HuggingFaceTransformersbeam search
0 likes · 60 min read

Taobao’s Daily Good Store tackles a two‑level recommendation challenge by jointly ranking shops and their items through a dual‑link system enhanced with a novel scatter‑score metric, personalized category scattering via Earth Mover’s Distance, beam‑search optimization, and UI upgrades, delivering higher efficiency, relevance, diversity, and ecosystem health.
Recommendation SystemUser Experiencebeam search
0 likes · 11 min read

The paper presents a lightweight on‑device re‑ranking system for short‑video recommendation that leverages real‑time user feedback and context‑aware generative ranking, detailing its architecture, feature engineering, beam‑search optimization, and both offline and online experimental results showing significant performance gains.
beam searchcontext-awarefeature engineering
0 likes · 12 min read

This article presents a method to improve keyword detection in large‑scale speech recognition by integrating a prefix automaton into the beam‑search decoding of seq2seq models, enabling real‑time addition of new terms while reducing computational overhead compared to traditional approaches.
Seq2SeqSpeech Recognitionbeam search
0 likes · 12 min read

Seq2Path treats each sentiment tuple as an independent tree path, training with average path loss and decoding via constrained beam search with a discriminative token, achieving state‑of‑the‑art results on five aspect‑based sentiment analysis datasets and deployment in Alibaba Entertainment AI Brain.
ACLSeq2SeqTree Structure
0 likes · 3 min read


Code DAODec 10, 2021 · Artificial Intelligence
This article explains the end‑to‑end deep‑learning pipeline for speech‑to‑text, covering audio digitization, preprocessing with librosa, conversion to Mel spectrograms and MFCCs, data augmentation, a CNN‑RNN architecture, CTC loss, decoding strategies and evaluation with word error rate.
ASRCTCMFCC
0 likes · 13 min read

This article introduces ByteDance's Deep Retrieval (DR) framework, describing its learnable index structure that aligns embedding training with retrieval objectives, detailing the core model, structure‑loss training via EM and online EM algorithms, beam‑search serving, multi‑task learning, and practical insights from Q&A.
EM algorithmbeam searchdeep retrieval
0 likes · 11 min read

Alibaba Mama’s advertising platform modernized its Tree‑based Deep Model by introducing a dual‑tower full‑library DNN with aggressive pre‑filtering and custom GPU TopK kernels, and a flattened‑tree model that retains beam search with multi‑head attention, while applying memory‑aware tricks such as attention swapping, softmax approximation, tiled‑matmul splitting, TensorCore batching, INT8 quantization and cache‑resident ad vectors, enabling multi‑fold latency reductions with minimal recall loss.
GPU accelerationbeam searchmodel optimization
0 likes · 15 min read

This article explains Alibaba's Deep Tree Matching (TDM) technology, covering the challenges of large‑scale match retrieval, the progression from classic two‑stage recall to tree‑based indexing, max‑heap tree modeling, beam‑search retrieval, and the joint model‑index learning across TDM 1.0, 2.0, and 3.0, highlighting significant offline and online performance gains and future research directions.
Alibababeam searchdeep learning
0 likes · 15 min read

This article explains key concepts of neural machine translation, covering Seq2Seq encoder‑decoder models, beam search strategies, BLEU evaluation, various attention mechanisms, and the enhancements introduced in Google's Neural Machine Translation system to improve speed, OOV handling, and translation quality.
AttentionBLEUGNMT
0 likes · 11 min read


Didi TechOct 10, 2020 · Artificial Intelligence
DiDi’s route engine, handling over 40 billion daily requests, replaces static graph algorithms with a deep‑reinforcement‑learning system that first learns intersection decisions via behavior‑cloning LSTM models and then refines them through self‑play Q‑learning, using beam‑search decoding to produce globally optimal, low‑deviation routes for ride‑hailing.
AIBehavior Cloningbeam search
0 likes · 12 min read

This article presents a comprehensive study of tree-based deep models for large-scale matching, introduces the theoretical framework of optimal tree models, proposes a Beam Search aware training algorithm (BSAT/OTM) to address training-test mismatch, and demonstrates significant recall improvements on Amazon Books and UserBehavior datasets.
beam searchdeep learninglarge-scale matching
0 likes · 23 min read

The article presents AutoCross, a system that automatically generates and selects high‑order feature crossings for tabular data using multi‑granularity discretization, beam search, field‑wise logistic regression and successive mini‑batch gradient descent, achieving superior accuracy and efficiency in large‑scale recommendation scenarios.
AutoCrossautomatic feature engineeringbeam search
0 likes · 10 min read

This article explains the Seq2Seq encoder‑decoder framework, its benefits for various sequence modeling tasks, and compares common decoding strategies such as greedy search and beam search, while also introducing attention and other enhancements for improved performance.
AttentionEncoder-Decoderbeam search
0 likes · 9 min read
