AI by AI

Theseus Research

AI generated podcasts based off recent research. Primarily made for my own consumption, but I'm happy to share them.

  1. 22/06/2025

    AI Breakthroughs: Reasoning, Reliability, and Recommendations

    In this episode, we dissect five cutting-edge research papers that are reshaping the AI landscape. We kick things off with a minimal-yet-powerful tweak to reinforcement learning that helps language models overcome performance plateaus and achieve deeper mathematical reasoning. Next, we journey to the manufacturing floor to explore ARKNESS, a hybrid framework that grounds LLMs in verifiable knowledge, eliminating hallucinations in high-stakes CNC machining. Then, we tackle the critical issue of AI safety with the Alignment Quality Index (AQI), a novel metric that evaluates a model's true alignment by analyzing its internal geometry, not just its outputs. We also explore how Visual Grounded Reasoning (VGR) is enabling multimodal models to 'see' and reason about fine-grained visual details more effectively. Finally, we look at RecFound, a new foundation model designed to unify and supercharge the world of recommendation systems. Join us for a deep dive into the innovations making AI more capable, reliable, and secure. References: Reasoning with Exploration: An Entropy Perspective: https://arxiv.org/pdf/2506.14758.pdf Knowledge Graph Fusion with Large Language Models for Accurate, Explainable Manufacturing Process Planning: https://arxiv.org/pdf/2506.13026.pdf Alignment Quality Index (AQI) : Beyond Refusals: AQI as an Intrinsic Alignment Diagnostic via Latent Geometry, Cluster Divergence, and Layer wise Pooled Representations: https://arxiv.org/pdf/2506.13901.pdf VGR: Visual Grounded Reasoning: https://arxiv.org/pdf/2506.11991.pdf Generative Representational Learning of Foundation Models for Recommendation: https://arxiv.org/pdf/2506.11999.pdf

    35 min

  2. 13/06/2025

    From Database Dialects to Diagnostic Deep Reasoning

    Welcome back to AI by AI, where we dissect the latest and greatest in machine learning research. In this episode, we're diving into five groundbreaking papers that are pushing the boundaries of what's possible. We'll explore Rel-LLM, a new framework that finally lets Large Language Models understand complex databases without losing the plot. Then, we'll tackle the challenge of learning from messy, corrupted data and see how a new 'perturbation-quantization' framework provides robust learning guarantees. We also look at SpectRe, a novel method making Graph Neural Networks more expressive by blending topology with spectral data. Shifting to healthcare, we'll discuss a study evaluating how AI can detect cognitive impairment from speech, with surprising results about what matters more—the words you say or how you say them. Finally, we'll cover Reasoning-Aware Reinforcement Learning (RARL), a technique for building powerful and efficient medical AI models that can reason about their diagnoses, even on a single GPU. Tune in for a deep dive into the tech that's shaping our future! References: Large Language Models are Good Relational Learners: https://arxiv.org/pdf/2506.05725.pdf Graph Persistence goes Spectral: https://arxiv.org/pdf/2506.06571.pdf Robust Learnability of Sample-Compressible Distributions under Noisy or Adversarial Perturbations: https://arxiv.org/pdf/2506.06613.pdf CAtCh: Cognitive Assessment through Cookie Thief: https://arxiv.org/pdf/2506.06603.pdf RARL: Improving Medical VLM Reasoning and Generalization with Reinforcement Learning and LoRA under Data and Hardware Constraints: https://arxiv.org/pdf/2506.06600.pdf

    38 min

  3. 06/06/2025

    Beyond the AI Mind: Self-Reflection, Peer Pressure, and Brain Graphs

    This episode dives into five cutting-edge AI research papers. First, we explore 'Reflect, Retry, Reward,' a novel method for improving LLM performance through self-reflection and reinforcement learning. Next, we discuss POLED, a probabilistic online event downsampling framework for event cameras. Then, we examine the MAEBE framework, which evaluates risks in multi-agent AI systems and reveals the brittleness of LLM moral preferences. We also cover Writing-Zero, a reinforcement learning paradigm for improving LLMs in creative writing tasks. Finally, we delve into a neuro-symbolic framework that integrates LLMs into probabilistic cognitive models for enhanced pragmatic question-answering.

    30 min

  4. 29/05/2025

    REARANK to EMLoC: Five AI Upgrades You Can’t Ignore

    This episode dives into five cutting-edge research papers. First, we explore REARANK, a reinforcement learning-based agent that re-ranks information using explicit reasoning, improving both performance and interpretability. Next, we examine a new task and dataset for detecting inconsistencies in political statements, benchmarking LLMs against human annotations. We then discuss v1, a lightweight extension for MLLMs that enables selective visual revisitation during inference, enhancing multimodal reasoning. Following that, we delve into CoT Monitor +, a framework designed to mitigate deceptive alignment in LLMs by integrating a self-monitoring mechanism into the chain-of-thought reasoning process. Finally, we cover Efficient Multi-Modal Long Context Learning (EMLoC), a training-free method for adapting MLLMs to new tasks by embedding demonstration examples directly into the model input, while efficiently managing long contexts. REARANK: Reasoning Re-ranking Agent via Reinforcement Learning: https://arxiv.org/pdf/2505.20046.pdf Misleading through Inconsistency: A Benchmark for Political Inconsistencies Detection: https://arxiv.org/pdf/2505.19191.pdf Don't Look Only Once: Towards Multimodal Interactive Reasoning with Selective Visual Revisitation: https://arxiv.org/pdf/2505.18842.pdf Mitigating Deceptive Alignment via Self-Monitoring: https://arxiv.org/pdf/2505.18807.pdf Efficient Multi-modal Long Context Learning for Training-free Adaptation: https://arxiv.org/pdf/2505.19812.pdf

    45 min

  5. 25/05/2025

    From Self-GIVE to SynOPUS: LLM Innovations Unpacked

    This episode dives into five cutting-edge AI research papers. We start with Self-GIVE, a retrieve-RL framework enhancing LLMs' reasoning through associative thinking and knowledge graphs, particularly for scientific QA. Next, we explore PiFlow, an information-theoretical framework that boosts automated scientific discovery using LLM-based multi-agent systems. Then, we unpack a theoretical analysis of Adam's convergence in deep ReLU networks, providing valuable insights for optimization. We also discuss Dual-LoRA with Incremental Schema Caching (DLISC), an efficient approach for on-device LLMs in information extraction. Finally, we examine the use of LLM-generated synthetic data to improve low-resource machine translation, introducing the SynOPUS dataset.

    44 min

  6. 16/05/2025

    Hands, Highways & Heartbeats: AI’s Expanding Horizons

    This episode dives into five cutting-edge AI research papers. We explore an adaptive exoglove controlled by imitation learning for hand motor impairments, an LLM-enhanced Q-learning framework for solving complex routing problems, a brain-in-the-loop supervised learning approach to enhance DNNs' cognitive abilities, a novel symbolic regression method using vision-capable LLMs, and Emotion-Qwen, a large multimodal model designed for improved emotion understanding. We discuss the key innovations, experimental results, and potential impact of each paper. References: Imitation Learning for Adaptive Control of a Virtual Soft Exoglove: https://arxiv.org/pdf/2505.09099.pdf A Large Language Model-Enhanced Q-learning for Capacitated Vehicle Routing Problem with Time Windows: https://arxiv.org/pdf/2505.06178.pdf Human-like Cognitive Generalization for Large Models via Brain-in-the-loop Supervision: https://arxiv.org/pdf/2505.09085.pdf Symbolic Regression with Multimodal Large Language Models and Kolmogorov Arnold Networks: https://arxiv.org/pdf/2505.07956.pdf Emotion-Qwen: Training Hybrid Experts for Unified Emotion and General Vision-Language Understanding: https://arxiv.org/pdf/2505.06685.pdf

    30 min

  7. 08/05/2025

    MoE Money, MoE Problems: Fixing AI One Expert at a Time

    This episode dives into cutting-edge research focused on enhancing the security and efficiency of AI models. We start with LlamaFirewall, an open-source guardrail framework designed to protect AI agents from prompt injection and other security risks. Then, we explore the optimization of Mixture of Experts (MoE) models, particularly DeepSeek, focusing on efficiency gains during inference. We also discuss MoxE, a novel architecture combining xLSTM with MoE for improved scalability, and the Compact Recurrent Transformer (CRT), which balances performance and computational efficiency for long sequence processing. Finally, we cover ARDBSCAN, a multi-agent reinforcement learning approach to robust clustering in datasets with varying densities. References: LlamaFirewall: An open source guardrail system for building secure AI agents: https://arxiv.org/pdf/2505.03574.pdf Faster MoE LLM Inference for Extremely Large Models: https://arxiv.org/pdf/2505.03531.pdf MoxE: Mixture of xLSTM Experts with Entropy-Aware Routing for Efficient Language Modeling: https://arxiv.org/pdf/2505.01459.pdf Compact Recurrent Transformer with Persistent Memory: https://arxiv.org/pdf/2505.00929.pdf Adaptive and Robust DBSCAN with Multi-agent Reinforcement Learning: https://arxiv.org/pdf/2505.04339.pdf

    34 min

  8. 04/05/2025

    Self-Taught, Chain-Thought & Battle-Ready: AI’s Next Leap

    This episode dives into five groundbreaking research papers pushing the boundaries of AI. We explore KGMEL, a Knowledge Graph-enhanced approach to Multimodal Entity Linking, and T2I-R1, a text-to-image generation model leveraging reinforcement learning and Chain-of-Thought reasoning. We then discuss Test-Time Reinforcement Learning (TTRL) for LLMs, enabling self-evolution on unlabeled data. Finally, we examine AegisLLM, a multi-agent system designed to defend against adversarial attacks on LLMs, and a novel attack/defense strategy for diffusion models using interpretable token embeddings to jailbreak and unlearn harmful content. References T2I-R1: Reinforcing Image Generation with Collaborative Semantic-level and Token-level CoT - https://arxiv.org/pdf/2505.00703 TTRL: Test-Time Reinforcement Learning - https://arxiv.org/pdf/2504.16084 KGMEL: Knowledge Graph-Enhanced Multimodal Entity Linking - https://arxiv.org/pdf/2504.15135v1 The Dual Power of Interpretable Token Embeddings: Jailbreaking Attacks and Defenses for Diffusion Model Unlearning - https://arxiv.org/pdf/2504.21307.pdf AegisLLM: Scaling Agentic Systems for Self-Reflective Defense in LLM Security - https://arxiv.org/pdf/2504.20965.pdf

    27 min
See All (21)

About

AI generated podcasts based off recent research. Primarily made for my own consumption, but I'm happy to share them.

Information

  • Creator
    Theseus Research
  • Years Active
    2K
  • Episodes
    21
  • Provider