What if you could get the performance of a massive, 100-example prompt, but with 13 times fewer tokens?

That’s the breakthrough promise of "instruction induction" —teaching an AI to be the prompt engineer.

This week, we dive into PROMPT-MII , a new framework that essentially meta-learns how to write compact, high-performance instructions for LLMs. It’s a reinforcement learning approach that could make AI adaptation both cheaper and more effective.

This episode explores the original research by Emily Xiao, Yixiao Zeng, Ada Chen, Chin-Jou Li, Amanda Bertsch, and Graham Neubig from Carnegie Mellon University.

Read the full paper here for a deeperdive: https://arxiv.org/abs/2510.16932

The biggest bottleneck for AIs handling massive documents—the context window—just got a radical fix. DeepSeek AI's DeepSeek-GOCR uses a counterintuitive trick: it turns text into an image to compress it by up to 10 times without losing accuracy. That means your AI can suddenly read the equivalent of 20 million tokens (entire codebases or legal troves) efficiently! This episode dives into the elegant vision-based solution, the power of its Mixture of Experts architecture, and why some experts believe all AI input should become an image.

Original Research: DeepSeek-GOCR is a breakthrough by the DeepSeek AI team.

Content generated with the help of Google's NotebookLM.

Link to the Original Research Paper: https://deepseek.ai/blog/deepseek-ocr-context-compression

Training AI agents using Reinforcement Learning (RL) to handle complex, multi-turn tasks is notoriously difficult.Traditional methods face two major hurdles: high computational costs (generating numerous interaction scenarios, or "rollouts," is expensive) and sparse supervision (rewards are only given at the very end of a task, making it hard for the agent to learn which specific steps were useful).

In this episode, we explore "Tree Search for LLM Agent Reinforcement Learning," by researchers from Xiamen University, AMAP (Alibaba Group), and the Southern University of Science and Technology. They introduce a novel approach called Tree-GRPO (Tree-based Group Relative Policy Optimization) that fundamentally changes how agents explore possibilities.

Tree-GRPO replaces inefficient "chain-based" sampling with a tree-search strategy. By allowing different trajectories to share common prefixes (the initial steps of an interaction), the method significantly increases the number of scenarios explored within the same budget. Crucially, the tree structure allows the system to derive step-by-step "process supervision signals," even when only the final outcome reward is available. The results demonstrate superior performance over traditional methods, with some models achieving better results using only a quarter of the training budget.

📄 Paper: Tree Search for LLM Agent Reinforcement Learning https://arxiv.org/abs/2509.21240

Everyone's talking about AI agents that can automate complex tasks. But what happens when a cool demo meets the real world? We dive into hard-won, and often surprising, lessons from builders on the front lines. Discover why your first strategic choice isn't about a tool, but an entire ecosystem; why more agents can actually make things worse; and why the most critical skill is shifting from "prompt engineering" to "context engineering." This episode cuts through the noise to reveal what it really takes to build reliable AI agents that deliver value.

In this IA Odyssey episode, we unpack “overhearing agents”—AI systems that listen to human activity (audio, text, or video) and step in only when help is useful, like surfacing a diagram during a class discussion, prepping trail options while a family plans a hike, or pulling case notes in a medical consult.
While conversational AI (like chatbots) requires direct user engagement, overhearing agents continuously monitor ambient activities, such as human-to-human conversations, and intervene only to provide contextual assistance without interruption. Examples include silently providing data during a medical consultation or scheduling meetings as colleagues discuss availability.

The paper introduces a clear taxonomy for how these agents activate: always-on, user-initiated, post-hoc analysis, or rule-based triggers. This framework helps developers think about when and how an AI should “step in” without becoming intrusive.

Original paper: https://arxiv.org/pdf/2509.16325
Credits: Episode notes synthesized with Google’s NotebookLM to analyze and summarize the paper; all insights credit the original authors.

Can large language models achieve more when they collaborate instead of working alone? In this episode, we dive into “LLM Multi-Agent Systems: Challenges and Open Problems” by Shanshan Han, Qifan Zhang, Yuhang Yao, Weizhao Jin, and Zhaozhuo Xu.

We explore how multi-agent systems—where AI agents specialize, debate, and share knowledge—can tackle complex problems beyond the reach of a single model. The paper highlights open challenges such as:
• Optimizing task allocation across diverse agents
• Enhancing reasoning through debates and iterative loops
• Managing layered context and memory across multiple agents
• Ensuring security, privacy, and coordination in shared memory systems

We also discuss how these systems could reshape blockchain applications, from fraud detection to smarter contract negotiation.

This episode was generated with the help of Google’s NotebookLM.
Read the full paper here: https://arxiv.org/abs/2402.03578

Ever wondered why AI chatbots sometimes state things with complete confidence, only for you to find out it's completely wrong? This phenomenon, known as "hallucination," is a major roadblock to trusting AI. A recent paper from OpenAI explores why this happens, and the answer is surprisingly simple: we're training them to be good test-takers rather than honest partners.

This description is based on the paper "Why Language Models Hallucinate" by authors Adam Tauman Kalai, Ofir Nachum, Santosh S. Vempala, and Edwin Zhang. Content was generated using Google's NotebookLM.

Link to the original paper: https://openai.com/research/why-language-models-hallucinate

Ever feel like your AI assistants don't really get you? We're diving into how AI is moving beyond generic answers to offer truly personalized experiences. This episode explores the journey from Retrieval-Augmented Generation (RAG), a fancy term for AIs that look things up before they speak, to sophisticated AI Agents that can understand your unique needs, plan tasks, and act on your behalf. It's the next step in making AI a genuine partner in our digital lives.

This description was generated using Google's NotebookLM, based on the work of Xiaopeng Li, Pengyue Jia, and their co-authors.

Read the original paper here:

https://arxiv.org/abs/2504.10147

How can we get the best out of large language models without breaking the budget? This episode dives into Adaptive LLM Routing under Budget Constraints by Pranoy Panda, Raghav Magazine, Chaitanya Devaguptapu, Sho Takemori, and Vishal Sharma. The authors reimagine the problem of choosing the right LLM for each query as a contextual bandit task, learning from user feedback rather than costly full supervision. Their new method, PILOT, combines human preference data with online learning to route queries efficiently—achieving up to 93% of GPT-4’s performance at just 25% of its cost.

We also look at their budget-aware strategy, modeled as a multi-choice knapsack problem, that ensures smarter allocation of expensive queries to stronger models while keeping overall costs low.

Original paper: https://arxiv.org/abs/2508.21141
This podcast description was generated with the help of Google’s NotebookLM.

Google’s latest breakthrough, Gemini 2.5 Flash Image—nicknamed “Nano Banana”—is reshaping what’s possible in digital art and beyond. From keeping characters consistent across scenes to natural-language editing and even blending multiple images, this model is lowering the barrier to creation like never before. Imagine building entire fantasy worlds or accelerating scientific research without the traditional costs and time sinks.

But with this power comes profound questions: How do we handle the risks of fakes, hallucinations, and lost trust in what we see? What happens to human artists when machines can produce in seconds what once took weeks?

In this episode of IA Odyssey, we dive into the promises and perils of Gemini 2.5 Flash Image, exploring how it may democratize creativity, shift the role of artists, and force us all to rethink authenticity in the age of AI.

Original content generated with the help of Google’s NotebookLM.

Meet the Aime framework—ByteDance’s fresh take on multi-agent systems that lets AI teammates think on their feet instead of following brittle, pre-planned scripts. A dynamic planner keeps adjusting the big picture, an Actor Factory spins up just-right specialist agents on demand, and a shared progress board keeps everyone in sync. In tests ranging from general reasoning (GAIA) to software bug-fixing (SWE-Bench) and live web navigation (WebVoyager), Aime consistently out-performed hand-tuned rivals—showing that flexible, reactive collaboration beats static role-play every time.

This episode of IA Odyssey unpacks how Yexuan Shi and colleagues replace rigid “plan-and-execute” pipelines with fluid teamwork, why it matters for real-world tasks, and where adaptive agent swarms might head next.

Source paper: https://arxiv.org/abs/2507.11988

Content generated with help from Google’s NotebookLM.

In this episode of IA Odyssey, we explore a bold new approach in training intelligent AI agents: letting them invent their own problems.

We dive into “Self-Challenging Language Model Agents” by Yifei Zhou, Sergey Levine (UC Berkeley), Jason Weston, Xian Li, and Sainbayar Sukhbaatar (FAIR at Meta), which introduces a powerful framework called Self-Challenging Agents (SCA). Rather than relying on human-labeled tasks, this method enables AI agents to generate their own training tasks, assess their quality using executable code, and learn through reinforcement learning — all without external supervision.

Using the novel Code-as-Task format, agents first act as "challengers," designing high-quality, verifiable tasks, and then switch roles to "executors" to solve them. This process led to up to 2× performance improvements in multi-tool environments like web browsing, retail, and flight booking.

It’s a glimpse into a future where LLMs teach themselves to reason, plan, and act — autonomously.

Original research: https://arxiv.org/pdf/2506.01716
Generated with the help of Google’s NotebookLM.

We're witnessing one of the most profound shifts in the history of software—a rapid evolution from traditional coding (Software 1.0) to neural networks (Software 2.0) and now, the dawn of Software 3.0: large language models (LLMs) programmable with simple English. Inspired by insights from Andrej Karpathy, former AI Director at Tesla, we explore how this paradigm shift reshapes the very concept of programming and its profound implications for everyone engaging with technology.

From the "Iron Man" analogy, where AI augments human capabilities rather than replacing them, to the fascinating vision of LLMs as new operating systems, this episode dives deep into the practical challenges and enormous opportunities ahead. We discuss Karpathy’s real-world perspective versus the consultant-driven hype, emphasizing that the path forward lies in human-AI collaboration rather than immediate full automation.

Generated using Google's NotebookLM.

Inspired by Andrej Karpathy’s insights: https://youtu.be/LCEmiRjPEtQ?si=NulC7m-qN8FVvBhQ

Ever wondered how much information your favorite AI language models, like GPT, actually retain from their training data? In this episode of AI Odyssey, we delve into groundbreaking research by John X. Morris, Chawin Sitawarin, Chuan Guo, Narine Kokhlikyan, G. Edward Suh, Alexander M. Rush, Kamalika Chaudhuri, and Saeed Mahloujifar. The authors introduce a new method for quantifying memorization in AI, distinguishing between unintended memorization (dataset-specific information) and generalization (knowledge of underlying data patterns). With findings revealing that models like GPT have a surprising capacity of about 3.6 bits per parameter, this study explores how memorization plateaus and eventually gives way to true understanding, a phenomenon known as "grokking."

Created using Google's NotebookLM, this episode demystifies how language models balance memorization and generalization, offering fresh insights into model training and privacy implications.

Dive deeper into the full paper here: https://www.arxiv.org/abs/2505.24832

What if you could simulate a full-scale usability test—before involving a single human user? In this episode, we explore UXAgent, a groundbreaking system developed by researchers from Northeastern University, Amazon, and the University of Notre Dame. This tool leverages Large Language Models (LLMs) to create persona-driven agents that simulate real user interactions on web interfaces.

UXAgent's innovative architecture mimics both fast, intuitive decisions and deeper, reflective reasoning—bringing realistic and diverse user behavior into early-stage UX testing. The system enables rapid iteration of study designs, helps identify potential flaws, and even allows interviews with simulated users.

This episode is powered by insights generated using Google’s NotebookLM. Special thanks to the authors Yuxuan Lu, Bingsheng Yao, Hansu Gu, Jing Huang, Zheshen Wang, Yang Li, Jiri Gesi, Qi He, Toby Jia-Jun Li, and Dakuo Wang.

🔗 Read the full paper here: https://arxiv.org/abs/2504.09407

What if your AI didn't just follow instructions… but coordinated a whole team to solve complex problems on its own?

In this episode, we dive into the fascinating shift from traditional AI Agents to a bold new paradigm: Agentic AI. Based on the eye-opening paper “AI Agents vs. Agentic AI: A Conceptual Taxonomy, Applications and Challenges”, we unpack why single-task bots like AutoGPT are already being outpaced by swarms of intelligent agents that collaborate, strategize, and adapt—almost like digital organizations.

Discover how these systems are transforming research, medicine, robotics, and cybersecurity, and why Google’s new A2A protocol could be a game-changer. From hallucination traps to multi-agent breakthroughs, this is the frontier of AI you haven’t heard enough about.

Synthesized with help from Google’s NotebookLM.
Full paper here 👇
https://arxiv.org/abs/2505.10468

In this episode, we explore “The Illusion of Thinking”, a thought-provoking study from Apple researchers that dives into the true capabilities—and surprising limits—of Large Reasoning Models (LRMs). Despite being designed to "think harder," these advanced AI models often fall short when problem complexity increases, failing to generalize reasoning and even reducing effort just when it’s most needed.

Using controlled puzzle environments, the authors reveal a curious three-phase behavior: standard language models outperform LRMs on simple tasks, LRMs shine on moderately complex ones, but both collapse entirely under high complexity. Even with access to explicit algorithms, LRMs struggle to follow logical steps consistently.

This paper challenges our assumptions about AI reasoning and suggests we're still far from building models that trulythink. Generated using Google’s NotebookLM.

🎧 Listen in and learn why scaling up “thinking” might not be the answer we thought it was.

🔗 Read the full paper: https://ml-site.cdn-apple.com/papers/the-illusion-of-thinking.pdf
📚 Authors: Parshin Shojaee, Iman Mirzadeh, Keivan Alizadeh, Maxwell Horton, Samy Bengio, Mehrdad Farajtabar (Apple)

Prompting AI just got smarter. In this episode, we dive into Local Prompt Optimization (LPO) — a breakthrough approach that turbocharges prompt engineering by focusing edits on just the right words. Developed by Yash Jain and Vishal Chowdhary from Microsoft, LPO refines prompts with surgical precision, dramatically improving accuracy and speed across reasoning benchmarks like GSM8k, MultiArith, and BIG-bench Hard.

Forget rewriting entire prompts. LPO reduces the optimization space, speeding up convergence and enhancing performance — even in complex production environments. We explore how this technique integrates seamlessly into existing prompt optimization methods like APE, APO, and PE2, and how it delivers faster, smarter, and more controllable AI outputs.

This episode was generated using insights synthesized in Google’s NotebookLM.

Read the full paper here: https://arxiv.org/abs/2504.20355

AI is everywhere—but what is it, really? In this episode, we cut through the noise to explore the fundamentals of artificial intelligence, from narrow AI and reactive systems to generative models, AI agents, and the emerging frontier of agentic AI. Using insights from expert sources, articles, and research papers, we break down key concepts in simple, accessible terms.

You'll learn how tools like ChatGPT work under the hood, why generative AI felt like such a leap, and what it actually means for an AI to be an agent—or part of a multi-agent system. We explore the real capabilities and limits of today’s AI, as well as the ethical and societal questions shaping its future.

What if an AI could become smarter without being taught anything? In this episode, we dive into Absolute Zero, a groundbreaking framework where an AI model trains itself to reason—without any curated data, labeled examples, or human guidance. Developed by researchers from Tsinghua, BIGAI, and Penn State, this radical approach replaces traditional training with a bold form of self-play, where the model invents its own tasks and learns by solving them.

The result? Absolute Zero Reasoner (AZR) surpasses existing models that depend on tens of thousands of human-labeled examples, achieving state-of-the-art performance in math and code reasoning tasks. This paper doesn’t just raise the bar—it tears it down and rebuilds it.

Get ready to explore a future where models don’t just answer questions—they ask them too.

Original research by Andrew Zhao, Yiran Wu, Yang Yue, and colleagues. Content powered by Google’s NotebookLM.

Read the full paper: https://arxiv.org/abs/2505.03335