Learning GenAI via SOTA Papers

Yun Wu

This podcast is focusing on sharing the papers on GenAI related topic, especially the SOTA (State of the Art) papers that are the foundations of GenAI work. It shows how these researches paved the way to the GenAI tools that we are using every day such as ChatGPT, Gemini, Claude Code etc.

Play

Open in

137 episodes
daily
Avg 21 min
English

Episodes 137

Newest Oldest Longest

All20 Unplayed20 In progress0 Played0

Counted on this page — what you have heard stays on this device, so it is not something the list can be paged by.

S1 · E528
Yesterday · 21 min
EP528: How Hindsight Makes Small AI Beat GPT
Title: Beyond Prediction: Steering VLM Agents with Retrospective World ModelingSource: http://arxiv.org/abs/2609.39101v1 Summary:This paper introduces a foundational framework that enhances Vision-Language Model agents by integrating retrospective world modeling to better steer autonomous actions. By improving how visual agents reason about and predict environmental consequences, it advances the state-of-the-art in agentic planning and decision-making.
S1 · E527
Yesterday · 22 min
EP527: How Confidence Filtering Fixes AI Reasoning
Title: Provable Test-Time Scaling for Beam Search in LLM Reasoning Source: http://arxiv.org/abs/2609.38672v1 Summary: This work establishes theoretical and algorithmic foundations for test-time scaling during beam search to enhance large language model reasoning. It provides provable guarantees that optimize computational inference budgets, marking a major breakthrough in test-time compute scaling laws.
S1 · E526
Sunday · 23 min
EP526: From Model Distillation to Agent Distillation
Title: Distilling Agentic Systems: A Roadmap across Models, Artifacts, and HarnessesSource: http://arxiv.org/abs/2609.36630v1 Summary:This work provides a comprehensive taxonomy and methodology for distilling complex agentic systems into efficient, deployable models. It represents a significant efficiency and reasoning breakthrough, bridging the gap between multi-component agent architectures and lightweight execution.
S1 · E525
Sunday · 21 min
EP525: How AnyAct Fixes Desktop AI Agents
Title: AnyAct: Universal Action for Self-Evolving AgentsSource: http://arxiv.org/abs/2609.37025v1 Summary: This paper introduces a foundational framework for universal action spaces that enables self-evolving AI agents to interact with diverse environments seamlessly. By generalizing agent capabilities across varied tasks, it establishes a core architectural primitive for autonomous agentic reasoning and adaptation.
S1 · E524
Saturday · 22 min
EP524: How SkillRubric Helps AI Grade Itself
Title: SkillRubric: Co-Evolving Actor Guidance and Evaluator Rubrics for Multimodal Agents Source: http://arxiv.org/abs/2609.34557v1Summary: This paper introduces a foundational framework for multimodal agents that co-evolves actor guidance and evaluator rubrics to improve task execution. By addressing the synergy between generation and evaluation, it provides a crucial mechanism for enhancing the reliability and autonomy of advanced AI agents.
S1 · E523
Saturday · 20 min
EP523: How Subgoals Stop AI Belief Drift
Title: SGG-ReflAct: Sub-Goal Guided ReflAct with Structured Planning for Reliable Long-Horizon Reasoning Source: http://arxiv.org/abs/2609.34548v1Summary: This work establishes a novel architectural approach for long-horizon agentic reasoning by integrating sub-goal guidance with structured reflection. It tackles core limitations in current AI agents regarding error accumulation and task decomposition over extended execution paths.
S1 · E522
Friday · 21 min
EP522: Why Limited Memory Makes AI Smarter
Title: Continuous Context Management Source: http://arxiv.org/abs/2609.35540v1Summary: This paper introduces a foundational mechanism for dynamically managing context streams in Generative AI systems. It provides critical architectural primitives for persistent memory retention and context efficiency, enabling LLMs and agents to operate seamlessly over infinite execution horizons.
S1 · E521
Friday · 21 min
EP521: HTN Planning Fixes AI Tool Orchestration
Title: HTN Planning as a Coordination Layer for Multi-Server MCP Tool Orchestration Source: http://arxiv.org/abs/2609.33731v1 Summary:This work presents Hierarchical Task Network (HTN) planning as a robust coordination layer for orchestrating tools across multiple servers via the Model Context Protocol (MCP). It establishes a foundational architectural approach for complex, multi-server agentic tool use and reasoning.
S1 · E520
Thursday · 20 min
EP520: How CodeSkill Solves Long Horizon Coding
Title: CodeSkill: Latent Skill Abstraction for Long-Horizon Code Agents Source: http://arxiv.org/abs/2609.33243v1Summary: This paper introduces a novel latent skill abstraction framework designed to enhance the capabilities of code agents over extended horizons. By efficiently abstracting and reusing skills, it addresses critical bottlenecks in agentic reasoning and long-term task execution.
S1 · E519
Thursday · 23 min
EP519: Multi Agent AI for Financial Valuation
Title: Self-Evolving Multi-Agent Symbolic Discovery for Financial Fundamental Analysis Source: http://arxiv.org/abs/2609.32746v1 Summary: This work presents a novel self-evolving multi-agent framework capable of autonomous symbolic discovery and reasoning for complex domains. By enabling agents to iteratively refine and generate symbolic knowledge without human intervention, it establishes a key primitive for advanced agentic reasoning loops.
S1 · E518
Wednesday · 22 min
EP518: Teaching Robots Without Human Puppeteers
Title: CAPEX: Efficiently Distilling Foundation Model Behavior into Deployable Robot Policies through Experience-Adaptive Reasoning Source: http://arxiv.org/abs/2609.33007v1 Summary:This paper introduces a foundational framework for distilling large foundation models into efficient, deployable robot policies via experience-adaptive reasoning. It bridges the gap between high-level generative reasoning and real-time physical execution, significantly advancing embodied agent architectures.
S1 · E517
Wednesday · 24 min
EP517: How DynBranch Eliminates AI Wait Times
Title: DynBranch: Speculative Subgraph Reuse for Dynamic Agentic LLM Serving Source: http://arxiv.org/abs/2609.31047v1 Summary: This work establishes a novel serving and execution paradigm tailored specifically for dynamic, branching agentic workflows that traditional static inference engines fail to handle efficiently. By introducing speculative subgraph reuse, it provides a critical infrastructural breakthrough for scaling complex, non-linear AI agent reasoning loops.
S1 · E516
October 27 · 21 min
EP516: ActKV Solves AI Agent Memory Overload
Title: ActKV: Efficient LLM Agents through Action-Guided KV Cache Management Source: http://arxiv.org/abs/2609.31395v1Summary: This paper introduces an architectural efficiency primitive that optimizes key-value cache management based on agent actions, directly tackling the high computational bottlenecks of LLM-based agents. By strategically retaining and pruning context during multi-step execution, it enables scalable and fast deployment of autonomous workflows.
S1 · E515
October 27 · 13 min
EP515: How IterSynth Solves AI Context Exhaustion
Title: IterSynth: Rethinking Deep Search Agents via Role-Decoupled Iterative Synthesis Source: http://arxiv.org/abs/2609.29444v1 Summary: This work introduces IterSynth, a novel agentic framework that rethinks deep search through role-decoupled iterative synthesis. This approach offers a foundational breakthrough in how AI agents can perform complex information discovery and synthesis, leading to more robust and comprehensive reasoning.
S1 · E514
October 26 · 21 min
EP514: How AI Agents Fix Their Own Mistakes
Title: A Wrong Turn Does Not Ruin the Journey: Deviation-Guided Skill Self-Evolution for LLM Agents Source: http://arxiv.org/abs/2609.29154v1 Summary: This paper proposes a novel reasoning loop for LLM agents, enabling them to autonomously evolve and improve their skills by learning from deviations or errors. This self-evolution mechanism represents a significant breakthrough in agentic AI's ability to adapt and refine its capabilities over time.
S1 · E513
October 26 · 19 min
EP513: How AI Drone Swarms Fly Without Text
Title: Controlling Collectives of AI Agents in Reasoning Space with Spatial Transformers Source: http://arxiv.org/abs/2609.28247v1 Summary: This paper establishes a foundational control mechanism for multi-agent systems by applying spatial transformers within reasoning space. It introduces architectural primitives for coordinating collectives of AI agents to collaboratively solve complex reasoning tasks.
S1 · E512
October 25 · 27 min
EP512: How Agent Editing Fixes AI Reasoning
Title: Agent-Editing World Model: Rethinking World Modeling for LLM AgentsSource: http://arxiv.org/abs/2609.28416v1 Summary:This paper introduces a novel architectural framework for LLM agents by reframing world modeling as an agent-editing process. It significantly advances agentic reasoning loops by enabling agents to dynamically update and reason over internal representations of their environment.
S1 · E511
October 25 · 22 min
EP511: How Human Explanations Teach AI Causality
Title: xWhyL: Causal Interactive Learning Source: http://arxiv.org/abs/2609.26037v1 Summary: This work pioneers a significant reasoning breakthrough for AI agents by focusing on causal interactive learning, moving beyond mere correlation to true understanding. It establishes novel agentic reasoning loops that enable agents to actively explore and learn causal relationships, enhancing their ability to explain, plan, and intervene effectively in complex environments.
S1 · E510
October 24 · 17 min
EP510: When AI Can Trust Its World Model
Title: Dual-Frontier: When Can an Agent Trust Its World Model? Source: http://arxiv.org/abs/2609.26293v1 Summary:This work establishes a foundational mechanism for autonomous agents to dynamically evaluate and quantify trust in their internal world models. By resolving when an agent should rely on world-model simulations versus real-world interactions, it advances fundamental agentic reasoning and planning loops.
S1 · E509
October 24 · 23 min
EP509: How AI Video Targets Cognitive Vulnerabilities
Title: CogenPVG: Cognitive-Enhanced Reflective Multi-Agent Framework for Persuasive Video Generation Source: http://arxiv.org/abs/2609.25821v1 Summary: This work advances agentic AI by introducing a novel multi-agent framework equipped with cognitive enhancement and reflective feedback loops specifically tailored for complex generative tasks. It establishes a sophisticated paradigm for how collaborative agent systems can iteratively plan, critique, and synthesize high-level generative content.

Showing 1–20 of 137 episodes