The article examines IBM Research's Abstract‑CoT approach, which replaces verbose natural‑language chain‑of‑thought reasoning with a compact abstract token vocabulary, achieving up to an 11‑fold reduction in inference tokens while maintaining comparable accuracy across math, instruction‑following, and multi‑hop QA benchmarks.
This article breaks down three core prompt‑engineering techniques—Few‑Shot prompting for output format stability, Chain‑of‑Thought for multi‑step reasoning, and Self‑Consistency for answer robustness—showing when to use each, how to combine them in LangChain, and providing concrete code examples, performance data, and common pitfalls.
This tutorial walks you through the open‑source “AI Agents From Scratch” project, teaching how to build a fully local AI agent without any pre‑made framework by covering core modules, 14 step‑by‑step examples, advanced reasoning architectures, and minimal system requirements.
The ICLR‑2026 paper introduces LCPO, a lightweight preference‑optimization technique that uses only 800 curated examples and 50 training steps to cut large‑model chain‑of‑thought generation length by about 50% while maintaining or even improving answer accuracy, dramatically reducing training and inference costs.
The paper introduces Heima, a framework that compresses chain‑of‑thought reasoning into a small set of abstract “thinking tokens” for multimodal large models, dramatically reducing generated tokens while preserving inference capability, and provides an adaptive interpreter to reconstruct human‑readable reasoning for analysis.
This article breaks down the key algorithms that power large‑language models—Transformer, Mixture‑of‑Experts, Flash Attention, KV‑Cache, Multi‑Token Prediction, quantization, Chain‑of‑Thought and Retrieval‑Augmented Generation—explaining how each contributes to the performance of ChatGPT, GPT‑4 and DeepSeek.
The article analyzes why DeepSeek’s large‑model training struggles with Monte‑Carlo Tree Search, explains its use of Chain‑of‑Thought prompting, GRPO entropy‑boosting and rejection‑sampling fine‑tuning, compares these methods with Google’s OmegaPRM and PRM approaches, and proposes a concrete MCTS‑driven data‑generation pipeline to overcome the “chicken and liquor” trade‑off.
The article examines why DeepSeek’s large‑model training cannot yet leverage Monte‑Carlo Tree Search, detailing its reliance on SFT, GRPO‑driven CoT activation and rejection‑sampling, contrasting this with Google’s PRM‑based approaches, and proposing a MCTS‑powered data‑generation pipeline to overcome the “roast chicken and baijiu” training dilemma.
DeepSeek’s new paper "Thinking with Visual Primitives" tackles the reference gap in multimodal models by introducing points and boxes as reasoning units, achieving up to 8× token efficiency and leading benchmark scores in counting, spatial reasoning, and maze navigation compared with GPT‑5.4, Claude‑Sonnet‑4.6 and Gemini‑3‑Flash.
This article systematically reviews Anthropic and OpenAI’s public research on monitoring intelligent agent trajectories, covering infrastructure such as Clio, Petri, Bloom, chain‑of‑thought monitoring, the Confessions mechanism, internal coding‑agent audits, and the Docent tool, while highlighting mitigation strategies for reward hacking and hidden objectives.
The article explains how industrial‑grade AI agents require structured prompt engineering, chain‑of‑thought reasoning, task decomposition, and a three‑layer defense (sandbox, prompt isolation, and human approval) to prevent prompt‑injection attacks, while also covering context engineering, retrieval‑augmented generation, and tool design best practices.
This article explains how vague AI queries lead to generic answers and shows how structured prompt engineering—using clear roles, goals, constraints, and frameworks like RTF and BROKE—can turn ambiguous business needs into precise, high‑quality AI outputs, including advanced chain‑of‑thought and few‑shot techniques for agents.
Learn how to apply reverse debugging to AI prompts by letting the model act as an attacker, uncover hidden logical flaws, and use chain‑of‑thought logs to refine your instructions before they reach production, reducing costly errors and improving reliability.
A teenage prodigy engineered a "Thinking Claude" prompt that adds a human‑like chain‑of‑thought protocol to Claude 3.5, enabling free O1‑level reasoning and producing impressive outputs such as a functional calculator, sci‑fi story, and playable games, while the article details the prompt’s design process and usage.
From a simple black‑box view where an agent receives a user request and returns an answer, the article breaks down modern AI agent designs—detailing the pure Chain‑of‑Thought reasoning loop, the ReAct reasoning‑acting cycle, tool integration, iteration tuning, and how to choose the optimal architecture for production.
Vision‑R1 introduces a 7B multimodal large language model that leverages 200K unsupervised CoT data, Modality Bridging, and Progressive Thinking Suppression Training to overcome data scarcity and over‑thinking, achieving 73.5% accuracy on MathVista—within 0.4% of OpenAI’s O1.
The paper introduces TaYS (Think‑as‑You‑See), a streaming chain‑of‑thought framework that replaces the traditional “watch‑then‑think” video inference pipeline with a parallel, real‑time “watch‑and‑think” approach, dramatically reducing latency and improving accuracy on complex video reasoning tasks.
This article explains the fundamentals of prompt engineering for large language models, covering LLM output configuration, length and sampling controls, various prompt types, chain‑of‑thought and tree‑of‑thought reasoning methods, and practical best‑practice guidelines for creating high‑quality prompts.
The article introduces ViDA-UGC, a large‑scale UGC visual‑quality dataset and its companion benchmark ViDA‑Bench, explains the MILP‑driven sampling, expert annotation pipeline, and CoT‑based evaluation framework, and shows how fine‑tuning popular multimodal LLMs on this data markedly improves low‑level quality perception, grounding, and description capabilities.
This article introduces the Deep Thought Ratio (DTR) metric, explains how tracking token modifications across neural network layers quantifies genuine inference effort, and shows through extensive experiments that DTR predicts accuracy far better than token length while enabling a sampling strategy that halves computational cost.
This article introduces Render‑of‑Thought (RoT), a novel paradigm that compresses chain‑of‑thought reasoning into visual embeddings using frozen vision encoders, achieving 3‑4× token reduction, faster inference, and improved interpretability while requiring minimal pre‑training.
The paper introduces BLM‑Guard, an explainable multimodal ad‑moderation framework that combines interleaved‑modal chain‑of‑thought reasoning with a policy‑aligned reinforcement‑learning reward to detect hidden cross‑modal violations in short‑video ads, and presents a new benchmark that demonstrates state‑of‑the‑art performance across multiple risk scenarios.
OpenAI’s new GPT‑5.2‑Codex model achieves state‑of‑the‑art performance on SWE‑Bench Pro and Terminal‑Bench 2.0, and a 90‑page technical report introduces the concept of monitorability, defining metrics, benchmark suites, and key findings about chain‑of‑thought length, RL training, and model size.
SABER introduces a reinforcement‑learning framework that lets large language models dynamically switch among four token‑budgeted reasoning modes, dramatically cutting inference length while preserving or improving accuracy across math, code, and logic tasks.
This article explains how Chain‑of‑Thought prompting and its variants—zero‑shot CoT, Thread of Thought, Tabular CoT, Analogical Prompting, and Step‑back Prompting—enable large language models to perform multi‑step reasoning by breaking problems into intermediate steps, with practical prompts, examples, and implementation details.
This article explains how to converse effectively with large language models by mastering three core prompting techniques—Zero‑shot, Few‑shot, and Chain‑of‑Thought—illustrated with front‑end analogies, code snippets, and a step‑by‑step DeepSeek JSON‑generation exercise that shows common pitfalls and best practices.
The article details the development and deployment of LORE, a large generative model that reshapes e‑commerce search relevance by combining knowledge injection, chain‑of‑thought reasoning, and multimodal alignment, achieving simultaneous improvements in user experience and revenue metrics.
This article examines whether large language models truly think, explores the origins of deep reasoning through transformer architectures and scaling laws, reviews chain‑of‑thought and its variants, and analyzes how reinforcement learning from human feedback—including PPO, DPO, and GRPO—helps internalise step‑by‑step reasoning while pointing to future directions such as atomic thought, hierarchical models, and training‑free in‑context knowledge bases.
This guide outlines Claude team’s best practices for prompt engineering, covering core techniques such as clear instructions, background context, specificity, examples, and advanced methods like pre‑filled responses, chain‑of‑thought, output formatting, and prompt chaining, with concrete examples and code snippets.
Introducing Farsighted-LAM, a novel latent action model that integrates geometric perception and multi‑scale temporal modeling, and its end‑to‑end SSM‑VLA framework with a Chain‑of‑Thought reasoning module, the authors demonstrate markedly improved spatial‑temporal fidelity, interpretability, and state‑of‑the‑art performance on challenging VLA benchmarks.
Learn how to craft effective prompts for large language models by using a systematic structure—role and task, core principles, context handling, chain‑of‑thought, output specifications, and few‑shot examples—and discover techniques for generating and iteratively refining prompts with the model itself.
This article examines how large language model reasoning has progressed from linear Chain‑of‑Thought prompting to parallel Tree‑of‑Thought and flexible Graph‑of‑Thought approaches, highlighting each method’s mechanism, strengths, limitations, computational costs, and the broader shift toward cognitive‑centric AI research.
The article analyzes Denny Zhou's Stanford CS25 lecture on large language model reasoning, explaining how intermediate token generation, chain‑of‑thought prompting, self‑consistency, reinforcement‑learning fine‑tuning, and answer aggregation together unlock powerful reasoning capabilities beyond traditional greedy decoding.
The article analyzes OpenAI’s five‑layer AGI roadmap, compares it with DeepMind’s ECEVS framework, and examines the technical progress from L1 to L5—including RL‑enhanced chain‑of‑thought, ReAct agents, deep research, and upcoming innovations—while highlighting the commercial implications of each stage.
This article distills Denny Zhou’s Stanford CS25 lecture, explaining how large language models achieve reasoning through intermediate token generation, chain‑of‑thought prompting, self‑consistency, reinforcement‑learning fine‑tuning, and answer aggregation, while highlighting theoretical foundations and practical breakthroughs.
This guide systematically breaks down Anthropic's official prompt‑engineering recommendations—clear instructions, multishot examples, chain‑of‑thought prompting, XML structuring, response pre‑filling, prompt chaining, long‑context handling, extended thinking, and practical code snippets—showing how to unlock Claude's full potential across complex tasks.
This article explores the emerging field of prompt engineering, detailing its fundamentals, advanced strategies such as chain‑of‑thought, ReAct, and structured frameworks, and demonstrates practical applications in AI agents for data retrieval, SQL generation, and market insight, offering actionable guidance for developers and business users alike.
Chain‑of‑Thought (CoT) enables large language models to solve complex tasks by breaking problems into sequential reasoning steps, improving accuracy in mathematics, commonsense, code generation, business strategy, and medical diagnosis, while highlighting its principles, advantages, challenges, and future prospects.
This article presents a comprehensive approach that combines causal inference, large language models, and retrieval‑augmented generation to automate e‑commerce price recommendation, detailing the three‑step workflow, challenges across product categories, the RAG architecture, process‑reward‑guided tree search, reinforcement learning refinements, and experimental results showing significant accuracy and speed improvements.
The KuaiMM Conversation project introduces a multimodal large‑model‑driven dialogue system for Kuaishou, featuring the world‑first short‑video mixed‑dialogue dataset, a Chain‑of‑Thought interaction framework, and large‑scale industrial deployments that dramatically improve live‑stream comments and intelligent customer service.
This article introduces ConfuseBench, a benchmark that classifies LLM uncertainty into document‑missing, ability‑limited, and ambiguous types, and presents methods—including retrieval, chain‑of‑thought, and clarification—to detect and actively resolve uncertainty, improving answer quality across diverse tasks.
This article introduces a variable‑length chain‑of‑thought distillation technique built on Alibaba Cloud PAI’s EasyDistill toolkit, presents the high‑quality OmniThought‑0528 dataset, details the training of the DistillQwen‑ThoughtY 4B/8B/32B models, and provides code and usage examples for researchers and practitioners.
The article introduces the Think When You Need (TWYN) method, a reinforcement‑learning approach that dynamically adapts chain‑of‑thought length, dramatically cuts redundant token generation in large language models, and maintains or improves accuracy across diverse reasoning benchmarks.
The article analyzes how large language models can acquire true reasoning abilities for hard‑to‑score industry tasks by combining Chain‑of‑Thought prompting with reinforcement learning, addressing vague reward signals, reward hacking, and loyalty, and proposing a toolbox of reward engineering, synthetic data, hierarchical RL and multi‑agent collaboration.
By rethinking how we interact with large language models—using role‑play, task decomposition, chain‑of‑thought, ReAct, and other advanced prompting strategies—readers can transform generic ChatGPT answers into precise, context‑aware responses, leveraging pattern recognition and context windows for superior AI assistance.
Recent work from West Lake University's MAPLE Lab introduces a diffusion‑based “Divergent Thought Chain” that treats each intermediate denoising step of a diffusion language model as a reasoning step, using result‑based reinforcement learning to optimize non‑linear token generation and achieving state‑of‑the‑art performance on math and code tasks.
DeepSeek quietly launched the R1 0528 model, which early testers report matches OpenAI’s o3 in benchmarks and style, while adding deeper chain‑of‑thought reasoning, better writing output, and extended thinking windows, and the announcement is followed by a promotion for the GOPS Global Ops Conference.
This article compares two recent papers that investigate why large reasoning models such as Llama and Qwen show degraded instruction‑following performance when using chain‑of‑thought prompting, analyzing attention patterns, training effects, and proposed mitigation strategies.
The paper presents PENCIL, a novel alternating generation‑and‑erasure reasoning paradigm that achieves optimal space‑time complexity for chain‑of‑thought tasks, dramatically improves accuracy and efficiency on hard SAT, QBF, and Einstein puzzle benchmarks, and is provably Turing‑complete.
The paper introduces an explainable LLM framework (ELLM‑rele) that uses chain‑of‑thought reasoning and a multi‑dimensional knowledge distillation pipeline to compress large‑model relevance judgments into lightweight student models, achieving superior offline relevance scores and online click‑through and conversion improvements in Taobao’s search advertising.
This article introduces DistilQwen2.5-DS3-0324, a distilled language model series that balances rapid inference with strong reasoning by applying a fast‑thinking chain‑of‑thought strategy, details its two‑stage distillation framework, evaluation on diverse benchmarks, and provides code for downloading and using the models.
At ICASSP 2025, Gaode’s two accepted papers present Lenna, a language‑enhanced reasoning detection assistant that adds a DET token to multimodal LLMs and achieves state‑of‑the‑art accuracy on RefCOCO benchmarks, and a chain‑of‑thought image‑editing framework that converts complex prompts into segmented masks and repair prompts for diffusion‑based inpainting, surpassing existing methods.
The article analyzes how reinforcement learning, which transformed LLM reasoning in DeepSeek, can be applied to visual‑language models to overcome the limitations of traditional chain‑of‑thought prompting and supervised fine‑tuning, presenting concrete reward designs, training pipelines, and a critical assessment of their strengths and weaknesses.
This article surveys classic prompt‑engineering methods such as Chain‑of‑Thought, Self‑Consistency, Least‑to‑Most, Boosting of Thoughts, Tree of Thoughts, and AutoGPT, summarizing their core ideas, advantages, limitations, and experimental results to help readers understand how to enhance large language model reasoning without model fine‑tuning.
The article examines how modern reasoning models like Claude 3.7 Sonnet display chain‑of‑thought explanations, but often hide or distort their true reasoning, presenting challenges for AI safety and alignment, and evaluates methods to test and improve fidelity.
This article examines the fundamental reasoning limitations of large language models, illustrates real‑world failure cases, and outlines current research directions such as better datasets, chain‑of‑thought prompting, external verification, and specialized solvers to improve their logical capabilities.
The interview explores DeepSeek R1's open‑source weights, its multi‑stage training pipeline—including pre‑training, supervised fine‑tuning, and RLHF—alongside innovations such as self‑consistency, chain‑of‑thought prompting, distillation, MoE architectures, and cost considerations, highlighting its impact on the future of large language models.
The article examines how chain‑of‑thought (CoT) templates are evolving from short‑term heuristics to long‑range planning in large language models, highlighting recent advances such as OpenAI o1, DeepSeek R1, and Kimi 1.5, and explores template designs that boost reasoning performance, efficiency, and multimodal capabilities.
This article walks through the complete lifecycle of building a small large‑language model, covering token‑level inference, pre‑training, post‑training steps such as supervised fine‑tuning, reward‑model creation, and reinforcement‑learning methods like DPO, PPO and GRPO, culminating in a practical 0.5B model fine‑tuned for chain‑of‑thought reasoning.
DeepSeek4j 1.4 introduces a Java‑centric, Spring Boot‑compatible framework that fully preserves DeepSeek's chain‑of‑thought capabilities, adds reactive streaming, and provides simple one‑line API integration, addressing previous limitations in mainstream frameworks and offering ready‑to‑use configuration and code examples.
This tutorial explains how to enhance large language models' logical reasoning by using DeepSeek‑R1's deep‑thinking mode, few‑shot prompting, chain‑of‑thought, and zero‑shot chain‑of‑thought techniques, providing concrete examples, comparisons, and a step‑by‑step template for effective prompt design.
The article introduces DeepSeek4j 1.4, a Java Spring Boot library that overcomes existing framework limitations by preserving DeepSeek's chain‑of‑thought capabilities, adding full reactive streaming, and providing a simple one‑line API along with quick‑start instructions and code examples.
This article provides a comprehensive analysis of the Brain module in AI agents, covering its multi‑step workflow, knowledge integration, memory mechanisms, intent recognition, planning strategies, reasoning techniques, and the role of reflection and emotion in enhancing adaptability and robustness.
The article examines whether the latest AI models that claim to perform true reasoning—by breaking problems into steps and using chain‑of‑thought—actually reason like humans, presenting skeptical and supportive expert viewpoints, and offering practical guidance on how to use such models responsibly.
The article explains why existing Java AI frameworks struggle with DeepSeek R1’s chain‑of‑thought and streaming features, introduces DeepSeek4j 1.4 as a targeted solution, details its core capabilities, and provides a step‑by‑step guide to integrate it with Spring Boot and Project Reactor.
The article introduces DeepSeek4j 1.4, a Java‑based framework that overcomes Spring AI’s limitations by fully preserving DeepSeek’s chain‑of‑thought and billing features, adding reactive streaming, providing Spring Boot starter integration, and offering quick‑start code samples and configuration guidance.
The guide defines prompts as structured queries that unlock large‑language‑model abilities, outlines five core frameworks (RTF, Chain‑of‑Thought, RISEN, RODES, Density‑Chain), presents two key principles—clear, delimited instructions and explicit reasoning steps—to reduce hallucinations, and surveys advanced techniques such as zero‑shot, few‑shot, RAG, Tree‑of‑Thought and automatic prompt engineering.
This article compiles a curated reading list of foundational and recent research papers—from the original Transformer to chain‑of‑thought, mixture‑of‑experts, and reinforcement‑learning studies—that together explain the breakthroughs behind DeepSeek‑R1 and guide readers through the technical evolution of modern large language models.
OpenAI has publicly disclosed the chain‑of‑thought reasoning of its o3‑mini model, prompting a wave of community experiments, critiques about authenticity, and discussions on the model’s limitations, prompting insights into AI interpretability and the trade‑offs of revealing internal reasoning.
This article examines why mainstream AI frameworks like Spring AI and LangChain4j cannot fully support DeepSeek’s R1 model, explains its unique chain‑of‑thought response format and parameter constraints, and provides a complete Spring WebFlux‑based solution—including API calls, streaming handling, and response parsing—to preserve reasoning content.
This article reviews a comprehensive study that applies Chain‑of‑Thought reasoning to autoregressive text‑to‑image generation, introducing extended test‑time computation, direct preference optimization, and two custom reward models (PARM and PARM++) that together improve generation quality by up to 15% over Stable Diffusion 3.
The MktAI Assistant combines LLM‑powered memory, skill planning, and tool‑calling with real‑time API data to replace slow, manual SQL dashboards, delivering sub‑minute, fresh, explainable marketing queries and attribution insights that boost decision speed, accuracy, and collaboration between data scientists and business users.
The article introduces Marco‑o1, an open‑source LLM that enhances complex reasoning by fine‑tuning on Chain‑of‑Thought data, integrating Monte‑Carlo Tree Search, introducing mini‑step actions and a reflection mechanism, and evaluates its performance on multilingual math and translation benchmarks.
This article explains what Prompt Engineering is, traces its development from early NLP commands to modern adaptive and multimodal prompting techniques, describes various prompting strategies such as Zero‑shot, Few‑shot, Chain‑of‑Thought, Auto‑CoT, and showcases a JD Logistics case study using these methods to classify product types with code examples.
This article explains what Prompt Engineering is, traces its development from early command‑based interactions to modern adaptive and multimodal prompting, details various prompting techniques such as zero‑shot, few‑shot, Chain‑of‑Thought, hallucination‑reduction methods, and demonstrates their practical use in a JD Logistics SKU piece‑type classification case with code examples.
This article explains what Prompt Engineering is, traces its evolution from early NLP commands to modern adaptive and multimodal prompting, details core techniques such as Zero‑shot, Chain‑of‑Thought, Auto‑CoT, and reduction of hallucinations, and showcases a logistics case study using various prompting strategies.
The article explains how OpenAI’s o1 model leverages chain‑of‑thought prompting, dual‑system cognitive theory, and new scaling laws—pre‑training on code/math and post‑training reinforcement with step‑wise reward models—to achieve superior reasoning, safety, and performance over GPT‑4, heralding a shift toward models that learn to think.
This article provides a detailed technical analysis of OpenAI’s o1 model, exploring its enhanced logical reasoning, the likely use of reinforcement learning with hidden chain‑of‑thought generation, multi‑model architecture, training data pipelines, reward modeling, and how these innovations could reshape AI safety and scaling strategies.
iQIYI’s AI‑powered search expands beyond title‑only queries by handling fuzzy role, plot, star, award, and semantic searches, using Chain‑of‑Thought‑generated TIPS, Retrieval‑Augmented Generation with sophisticated indexing, chunking, embedding, reranking, and prompt‑engineering to deliver personalized, accurate video recommendations that boost user engagement.
The article examines OpenAI’s newly released o1 model, highlighting its superior performance in complex reasoning tasks such as math, programming, and science, and explains how model‑level chain‑of‑thought optimization and product‑level UI design give it an edge over competitors like Claude.
OpenAI’s newly released o1 series introduces a reinforcement‑learning‑trained LLM that generates long chain‑of‑thought reasoning, achieving top‑50% scores on IOI contests, high rankings on Codeforces and AIME, and dramatically outperforming GPT‑4o across scientific and mathematical tasks.
This guide walks readers from beginner to proficient Prompt Engineer by outlining the evolution of prompting, introducing a universal four‑component template, and detailing a five‑step workflow—including refinement, retrieval‑augmented generation, chain‑of‑thought reasoning, and advanced tuning techniques—plus evaluation metrics for LLM performance.
The paper introduces SymbCoT, a Symbolic Chain‑of‑Thought framework that translates natural‑language problems into symbolic form, plans, solves, and verifies reasoning steps, achieving significantly higher logical reasoning performance than traditional CoT methods across multiple benchmark datasets.
Early‑Stopping Self‑Consistency (ESC) dynamically halts sampling once a sliding‑window answer distribution reaches zero entropy, cutting the number of required LLM reasoning samples by 34‑84 % across arithmetic, commonsense, and symbolic benchmarks while preserving accuracy and offering a theoretically‑bounded, robust, budget‑adaptive alternative to traditional Self‑Consistency.
This paper presents a novel framework that leverages negative samples during large language model distillation through three stages—Negative Assistive Training, Negative Calibration Enhancement, and Adaptive Self‑Consistency—demonstrating significant accuracy gains on challenging mathematical reasoning benchmarks and improved generalization to out‑of‑distribution tasks.
The guide explains GPT’s core capabilities—natural language mastery, domain reasoning, and code generation—while detailing its limits such as prompt sensitivity, token caps, and lack of memory, then offers engineering workarounds like systematic prompting, chain‑of‑thought, external memory, tool integration, safety checks, and a six‑layer architecture for building robust commercial AI applications.
This comprehensive guide explores prompt engineering for large language models, covering its background, fundamental concepts, prompt formats, construction principles, advanced techniques like few‑shot, zero‑shot, and chain‑of‑thought prompting, as well as practical examples, evaluation metrics, and future directions.
Prompt engineering teaches how to craft effective instructions, context, input data, and output formats for large language models, using clear commands, iterative refinement, and advanced methods such as zero‑shot, few‑shot, chain‑of‑thought, Tree of Thoughts, retrieval‑augmented and progressive‑hint prompting to achieve precise, reliable results across diverse tasks.
This article explains how to craft efficient prompts by combining clear instructions and questions, discusses prompt injection risks and mitigation with delimiters, addresses hallucinations, and introduces zero‑shot, few‑shot, and chain‑of‑thought prompting techniques for large language models.
The article explains how ChatGPT’s surprising step‑by‑step reasoning, known as Chain of Thought, emerged as a technical breakthrough, links it to model scaling, cognitive System 1/2 theory, and the influence of code data in training large language models.
This article reviews the emergence phenomena observed in large language models, explains how model scale, in‑context learning and chain‑of‑thought prompting contribute to sudden performance gains, discusses small‑model alternatives, and explores the relationship between emergence and the training‑time Grokking effect.
This paper presents a large‑scale evaluation of ChatGPT on knowledge‑base complex question answering, introducing a feature‑driven multi‑label annotation framework and CheckList‑based functional, robustness, and controllability tests, and comparing its performance with other LLMs across multiple English and multilingual datasets.
Prof. Qiu Xipeng’s lecture provides a comprehensive overview of large language models—from their historical development and architectural foundations to key technologies such as in‑context learning, chain‑of‑thought, and natural‑instruction learning, as well as RLHF training, capability evaluation, and current limitations of ChatGPT.
The article explains how techniques such as Reinforcement Learning from Human Feedback, Instruction Fine‑Tuning, Supervised Fine‑tuning and Chain‑of‑Thought contribute to ChatGPT’s impressive general‑intelligence performance, as analyzed by DeepLearning.AI founder Andrew Ng.
