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

91 episodes
daily
Avg 22 min
English

Episodes 91

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 · E482
Yesterday · 24 min
EP482: Training AI with Feedback Enriched Environments
Title: Environments as Scaffold: Enriching Feedback to Bootstrap Self-Evolving Agents in Long-Horizon TasksSource: http://arxiv.org/abs/2609.08404v1 Summary: This paper proposes a novel framework where environments act as 'scaffolds' to enrich feedback, enabling self-evolving agents to bootstrap their capabilities in long-horizon tasks. This introduces a foundational new paradigm for agent learning and adaptation, leading to more robust and capable AI agents by improving their ability to acquire knowledge and evolve effectively over time.
S1 · E481
Yesterday · 23 min
EP481: Why Expert Imitation Breaks Small AI Models
Title: Co-Evolving Harnesses and Models: On-Policy Correction Helps Weaker Models Catch Up Where Imitation Fails Source: http://arxiv.org/abs/2609.09134v1Summary: This work establishes a foundational co-evolution paradigm for model training where evaluation harnesses and model parameters continuously adapt through on-policy feedback. It solves key limitations of traditional imitation learning, allowing smaller or weaker models to efficiently close the performance gap with frontier systems.
S1 · E480
Friday · 20 min
EP480: Cloning Powerful AI Agents Just by Watching
Title: AgentLeak: Cloning Stronger LLM Agent Capabilities onto Weaker Agents Beyond Skill Stealing Source: http://arxiv.org/abs/2609.07131v1 Summary: This paper introduces a novel methodology for transferring complex agentic capabilities from frontier LLM agents to smaller, resource-efficient models beyond standard skill acquisition. It advances agent efficiency and distillation paradigms by enabling weaker models to replicate long-horizon reasoning and decision-making capabilities.
S1 · E479
Friday · 23 min
EP479: Multi-Agent AI Debates Are Pure Theater
Title: A Layered Analysis of Disagreement And Answer Quality in Multi-Agent LLM Debate Source: http://arxiv.org/abs/2609.08016v1 Summary: This paper provides a foundational analysis of how disagreement dynamics and consensus formation impact solution accuracy in multi-agent LLM debate frameworks. Its structural insights directly inform the design of more robust multi-agent reasoning loops and collaborative alignment strategies.
S1 · E478
Thursday · 24 min
EP478: How Geometry Fixes AI Memory Limits
Title: ECOKV: Geometry-Aware KV Cache Eviction via Complementary Diversity MetricsSource: http://arxiv.org/abs/2609.06663v1 Summary: This paper introduces a geometry-aware KV cache eviction strategy using complementary diversity metrics to optimize memory management during context processing. By improving memory retention efficiency without sacrificing model accuracy, it provides a crucial performance breakthrough for scaling context lengths in large language models.
S1 · E477
Thursday · 24 min
EP477: Why AI Decision Explanations Are Broken
Title: Causal Attribution for Agentic Decisions: Estimators, Coupling, and a Traceability Specification Source: http://arxiv.org/abs/2609.06445v1 Summary:This paper establishes a formal mathematical and causal framework for tracing, estimating, and attributing decisions within dynamic agentic systems. It provides essential foundational theoretical primitives required to formally audit, verify, and explain complex autonomous agent reasoning loops.
S1 · E476
Wednesday · 24 min
EP476: Hardcoding Values into AI Architecture
Title: Value-Preserving Architectures for Agentic AI Systems Source: http://arxiv.org/abs/2609.03920v1Summary: This paper likely proposes a fundamental architectural design paradigm that ensures the preservation of critical values or properties within AI agents. Such a design could introduce a new primitive for building robust, aligned, and safe agentic systems, addressing core challenges in agent deployment and reasoning.
S1 · E475
Wednesday · 22 min
EP475: Von Neumann Automata for Self Reproducing AI
Title: Dalek: A Constructive Agent MachineSource: http://arxiv.org/abs/2609.03546v1 Summary:This title strongly suggests the introduction of a novel architectural primitive or framework for agents, termed a 'Constructive Agent Machine.' This could represent a new foundational paradigm for agent design, enabling advanced agentic reasoning by allowing agents to build knowledge or plans in an inherently creative or self-organizing manner.
S1 · E474
Tuesday · 23 min
EP474: How SkillGLoW Fixes AI Memory Hoarding
Title: SkillGLoW: Procedural-Family Skill Consolidation for Self-Improving Agents on Long-Horizon Task Streams Source: http://arxiv.org/abs/2609.02217v1 Summary:This research presents 'procedural-family skill consolidation' as a novel framework for self-improving agents. It addresses fundamental challenges in agent autonomy and generalization by enabling agents to learn, adapt, and consolidate skills across diverse, long-horizon task streams efficiently.
S1 · E473
Tuesday · 24 min
EP473: How six templates make AI 120x faster
Title: Codebook Agent: Amortized Topology Design for LLM Multi-Agent SystemsSource: http://arxiv.org/abs/2609.02264v1 Summary: This paper introduces 'amortized topology design' as a novel architectural primitive for structuring interactions within LLM multi-agent systems. This approach promises significant efficiency and reasoning breakthroughs by optimizing how agents communicate and coordinate in complex agentic frameworks.
S1 · E472
Monday · 22 min
EP472: Why AI Reasons Better in Silence
Title: Latent Recurrent Thoughts: Recurrent Refinement of Proposed Latents for Reasoning with Frozen LLMs Source: http://arxiv.org/abs/2609.01117v1Summary: This work proposes a novel reasoning loop that employs recurrent refinement of latent representations to enhance reasoning capabilities, even when utilizing frozen large language models. This represents a significant breakthrough in improving the logical prowess and efficiency of existing LLMs for complex tasks, crucial for advanced agentic AI.
S1 · E471
Monday · 22 min
EP471: Self evolving autonomous agents with ARISE-RL
Title: ARISE-RL: Agentic Rubric-Grounded Iterative Self-Evolution with Reinforcement LearningSource: http://arxiv.org/abs/2609.01058v1 Summary: This paper introduces a novel agentic reasoning framework that enables iterative self-evolution in AI agents using reinforcement learning and rubric-grounded evaluation. This provides a foundational mechanism for agents to autonomously learn, improve, and refine their strategies based on explicit performance criteria.
S1 · E470
October 4 · 27 min
EP470: 1933 math fixes AI context limits
Title: Higher-Dimensional Rotary Position Embedding Source: http://arxiv.org/abs/2608.29715v1 Summary: This paper introduces an extension to Rotary Position Embedding (RoPE), a critical architectural primitive in Transformer models. This advancement can significantly enhance GenAI by improving how models process and understand complex, multi-dimensional data, leading to more capable and efficient models.
S1 · E469
October 4 · 23 min
EP469: How AutoCRAT stops AI from overthinking
Title: AutoCRAT: Within-trajectory Joint Control of Stochasticity and Compute for LLM ReasoningSource: http://arxiv.org/abs/2608.29988v1 Summary: This research proposes a novel framework for adaptively controlling stochasticity and computational resources during LLM reasoning. This represents a significant breakthrough in agentic reasoning, enabling more efficient, targeted, and adaptable decision-making for LLM-based agents and advanced GenAI systems.
S1 · E468
October 3 · 23 min
EP468: Steering AI logic with latent vectors
Title: Toward Latent Language Model Skills Steering and Optimization: An Empirical Study Source: http://arxiv.org/abs/2608.29459v1Summary: This research explores methods for steering and optimizing latent language model skills, offering a deeper level of control over LLM capabilities. This represents a significant reasoning breakthrough for GenAI by enabling more efficient and targeted utilization of generative models' intrinsic intelligence.
S1 · E467
October 3 · 23 min
EP467: LiteSearch-VL gives small AI search powers
Title: LiteSearch-VL: Small Multimodal Search Agents via Trajectory Distillation and Synthetic Step-DPO Source: http://arxiv.org/abs/2608.29357v1 Summary: This paper introduces novel techniques like trajectory distillation and synthetic Step-DPO to develop small, efficient multimodal search agents. Such advancements offer a significant efficiency and scaling breakthrough for Agentic AI, making sophisticated agents more practical and widely deployable.
S1 · E466
October 2 · 24 min
EP466: Sony OmniUE replaces typing with physical interaction
Title: Omni-Interactive Universal EmbedderSource: http://arxiv.org/abs/2608.27044v1 Summary: This paper proposes a "universal embedder" capable of processing and unifying diverse, interactive data modalities into a singular representation. Such a primitive could fundamentally change how GenAI and Agentic AI systems perceive and represent the complex, multimodal world, enabling more versatile and capable models.
S1 · E465
October 2 · 21 min
EP465: AI agents finally ditch computer screens
Title: ASIL: Replacing Screenshot-and-Click with Structured State and Semantic Actions Source: http://arxiv.org/abs/2608.26991v1Summary: This work introduces a novel framework for agent-environment interaction, moving beyond brittle pixel-based methods to structured state and semantic actions. This paradigm shift is foundational for creating robust and intelligent AI agents capable of reliably perceiving and operating within complex digital user interfaces and applications.
S1 · E464
October 1 · 22 min
EP464: PROGROUTER manages the cost of AI agents
Title: ProgRouter: Online Progress-Guided Orchestration for Multi-Agent LLM Workflows under Quality-Cost Tradeoffs Source: http://arxiv.org/abs/2608.25992v1 Summary: This work proposes a novel framework for orchestrating complex multi-agent LLM workflows. By intelligently guiding agent interactions based on progress and optimizing for quality-cost tradeoffs, it provides a foundational mechanism for building scalable and efficient AI agent systems.
S1 · E463
October 1 · 20 min
EP463: How PolyMemDB Solves AI Memory Hallucinations
Title: PolyMemDB: A Polyglot Database System for AI Memory Management Source: http://arxiv.org/abs/2608.25577v1 Summary: This paper introduces a polyglot database system specifically designed for comprehensive AI memory management, addressing a critical and pervasive bottleneck for scaling Generative AI and Agentic AI systems. It provides foundational infrastructure that enables agents to handle vastly larger contexts, persist long-term knowledge, and perform more complex, stateful reasoning.

Showing 1–20 of 91 episodes