Papers

Listen to this article.

Only the latest audio is kept; older files are removed on each update.

Problem

Large language models (LLMs) are increasingly being used in complex, multi-step tasks that require them to manage lots of information across many turns. The standard approach – storing all previous interactions as context – quickly becomes unsustainable because the context window grows and grows. Existing “proactive context management” techniques try to solve this by allowing LLMs to edit their own working memory (e.g., searching, deleting, or summarizing). However, these methods are limited; they offer a basic toolset, don’t support complex planning or long-term memory, and struggle with how to best learn which edits are most effective during training.

Listen to this article.

Only the latest audio is kept; older files are removed on each update.

Problem

Current vision-language models excel at recognizing and describing physical events, but they struggle with the deeper task of reasoning about how those events unfold and how to predictably influence them. This boils down to a lack of explicit representation of the underlying physics – things like object states, how objects interact (dynamics), and measurable parameters governing their behavior. Essentially, these models understand what happens but not necessarily why.

Listen to this article.

Only the latest audio is kept; older files are removed on each update.

Problem

Scaling world models – AI systems that understand and predict environments – typically relies on feeding them vast amounts of video data alongside significant computational resources. This paper argues that this approach is inefficient because it lacks a crucial element: reliable, grounded reward signals to guide learning after initial training (often referred to as “post-training”). Current methods for assessing spatial generation quality often rely on fuzzy proxies like CLIP scores which are prone to bias and don’t effectively support Reinforcement Learning (RL).

Listen to this article.

Only the latest audio is kept; older files are removed on each update.

Problem

Vision-language-action (VLA) models are increasingly popular for creating generalist embodied agents – robots that can understand instructions, perceive their surroundings visually, and take actions in the real world. However, a key question remains: can we build even better VLA systems by rethinking their architecture and training them on more data? Current models may not be effectively leveraging architectural designs or scaling to truly massive datasets with varied types of data. Ultimately, existing systems might struggle to generalize well across different robots (“embodiments”) and tasks.

Listen to this article.

Only the latest audio is kept; older files are removed on each update.

Problem

Current large language models (LLMs) are impressive at reasoning and knowledge synthesis, but fall short when tackling complex real-world tasks that require sustained effort and interaction with external tools and data sources. This paper identifies a gap in the ability of LLMs to make sustained, verifiable progress towards long-term objectives—what they call “working capability”. Think about projects involving multiple files, searching online for information, running code, remembering what’s been done already, handling errors, and producing reliable results.

Listen to this article.

Only the latest audio is kept; older files are removed on each update.

Problem

Large Language Model (LLM) agents, enhanced with “skills” – essentially structured knowledge packages – have shown promise in improving task completion rates at runtime. However, there’s a gap in understanding why these skills work, and crucially, when they don’t. Existing evaluations largely focus on overall success, neglecting the underlying mechanisms that drive skill effectiveness. This paper aims to explore those nuances.

Listen to this article.

Only the latest audio is kept; older files are removed on each update.

Problem

Evaluating world models – AI systems that simulate and reason about environments – is currently a flawed process. Existing benchmarks typically provide only a single score, offering little insight into why the model performed well or poorly. Understanding the underlying reasoning behind these scores, especially concerning factors like physics, causality, and accurate state changes within the simulated world, is crucial for improving model reliability. Current methods lack transparency, essentially providing black-box evaluations with no explanation of how those scores were derived.

Listen to this article.

Only the latest audio is kept; older files are removed on each update.

Problem

Creating interactive 3D open worlds based on user requests is a significant goal, but existing methods often fall short when faced with realistic and complex prompts. Current evaluation techniques are limited to simplified scenarios, making it difficult to truly assess how well agents understand user intent, manipulate 3D objects, and reason about the combined textual and visual information within the world they’re building.

Listen to this article.

Only the latest audio is kept; older files are removed on each update.

Problem

Choosing the right Large Language Model (LLM) for a given query can be tricky. No single LLM excels at everything, and different models have varying costs associated with their use. This means efficiently routing incoming requests to the best-suited model is crucial for both performance and cost management in real-world deployments. However, existing routing solutions are often quite different from each other, making it hard to compare them or build upon them effectively.

Listen to this article.

Only the latest audio is kept; older files are removed on each update.

Problem

Generating dynamic 3D scenes (often referred to as “4D” because they involve space and time) from conditions like text or images is a challenging area in generative AI. Current methods for creating these 4D scenes have limitations: either they generate videos first and then reconstruct the 3D geometry with a separate model (leading to inconsistencies), or they directly predict the geometry, which ties their approach too closely to a specific video generator and makes it difficult to adapt as models evolve.

Paper: ContextPilot: Teaching Agents for Proactive Context Management via Fine-grained RL

Problem

Paper: Code as Worlds: Agentic Discovery of Executable World Representations for Physical Reasoning

Problem

Paper: Agentic Game Development as a Verifiable Trajectory Data Engine for Scaling World Models

Problem

Paper: GigaBrain-0.7: Scaling Embodied Foundation Models to Emergent Capabilities with a Three-System Ar...

Problem

Paper: Apodex 1.1: Scaling Agentic Intelligence for Complex Work

Problem

Paper: Demystifying Agent Skills: Why They Work-Until They Don't

Problem

Paper: HarnessEval-W: Agentifying the Evaluation of Visual Worlds

Problem

Paper: VibeWorlding: Can Multimodal Agents Construct 3D Open Worlds End-to-End?

Problem

Paper: LLMRouter: Unified Infrastructure for Developing, Evaluating, and Deploying LLM Routers

Problem

Paper: Beyond Pixels: From Video Priors to 4D Worlds

Problem