EP287: Small models beat GPT-4o with Role-Agent

EP288: Test-Time Training shatters quadratic sampling limits

Title: The Power of Test-Time Training for Approximate Sampling Source: http://arxiv.org/abs/2606.11437v1 Summary: This work provides a foundational theoretical framework for Test-Time Training (TTT) by formalizing it as an approximate sampling problem and establishing query complexity lower bounds. It offers principled mathematical backing for the emerging paradigm of adapting model weights at inference time to solve complex reasoning tasks.

EP287: Small models beat GPT-4o with Role-Agent

Title: Role-Agent: Bootstrapping LLM Agents via Dual-Role Evolution Source: http://arxiv.org/abs/2606.10917v1 Summary: The Role-Agent framework introduces a dual-role evolution cycle where an LLM functions as both agent and environment to bootstrap its own reasoning capabilities. This novel agentic framework addresses the limitations of static training data by enabling self-correcting co-evolution through internal state prediction and failure mode analysis.

EP286: ReasonAlloc Solves the AI Memory Bottleneck

Title: ReasonAlloc: Hierarchical Decoding-Time KV Cache Budget Allocation for Reasoning Models Source: http://arxiv.org/abs/2606.11164v1 Summary: ReasonAlloc introduces a hierarchical KV cache allocation strategy that significantly optimizes memory usage during the long chain-of-thought trajectories characteristic of modern reasoning models. By identifying "Reasoning Wave" demand patterns, this training-free framework provides a foundational primitive for scaling inference efficiency in complex reasoning tasks.

EP285: How SkeMex builds medical AI intuition

Title: Experience Makes Skillful: Enabling Generalizable Medical Agent Reasoning via Self-Evolving Skill Memory Source: http://arxiv.org/abs/2606.09365v1 Summary: This paper proposes SkeMex, a foundational architecture for self-evolving agent memory that enables the distillation and governance of procedural skills from interaction trajectories. It introduces a novel "Read-Write-Assess-Govern" reasoning loop that allows agents to continuously improve their capabilities post-deployment without the need for model retraining.

EP284: Compressing massive context into soft tokens

Title: End-to-End Context Compression at Scale Source: http://arxiv.org/abs/2606.09659v1 Summary: This paper introduces Latent Context Language Models (LCLMs), a novel architectural primitive that utilizes encoder-decoder compression to efficiently handle long-context sequences at scale. It establishes a new Pareto frontier for accuracy and efficiency, providing a foundational backbone for next-generation agents that require massive context windows.