AI Model Compression, Embodied Perception, and Task-Oriented Scene Graphs

Latest research summaries in ML, Robotics, CV, NLP and AI

Welcome to today’s edition of State of AI 👋 And a warm welcome to our new subscribers since last edition!

This edition covers techniques to efficiently compress and deploy large language models on edge devices, to novel approaches for enhancing the reasoning and task planning capabilities of embodied AI agents in real-world environments. We also explore advancements in scene representation that integrate spatial, functional, and interactive elements to enable more effective task planning.

Here’s what caught our attention:

• TOGGLE: Temporal Logic-Guided Large Language Model Compression for Edge - A framework that leverages formal specifications to systematically compress LLMs while preserving critical linguistic properties, enabling efficient deployment on resource-constrained devices.

• Generative Adversarial Reasoner: Enhancing LLM Reasoning with Adversarial Reinforcement Learning - A method that co-evolves an LLM reasoner and a discriminator to provide dense, calibrated rewards that improve the model’s logical reasoning capabilities.

• MomaGraph: State-Aware Unified Scene Graphs with Vision-Language Model for Embodied Task Planning - A novel scene representation that integrates spatial, functional, and interactive elements to enable more effective task planning for embodied agents in household environments.

Let’s get into it 👇

Contents

1. Distributional AGI Safety

2. TOGGLE: Temporal Logic-Guided Large Language Model Compression for Edge

3. Towards Pervasive Distributed Agentic Generative AI -- A State of The Art

4. Alchemist: Unlocking Efficiency in Text-to-Image Model Training via Meta-Gradient Data Selection

5. GenEval 2: Addressing Benchmark Drift in Text-to-Image Evaluation

6. LinkedOut: Linking World Knowledge Representation Out of Video LLM for Next-Generation Video Recommendation

7. MEPIC: Memory Efficient Position Independent Caching for LLM Serving

8. Cartesian-nj: Extending e3nn to Irreducible Cartesian Tensor Product and Contracion

9. TACE: A unified Irreducible Cartesian Tensor Framework for Atomistic Machine Learning

10. Generative Adversarial Reasoner: Enhancing LLM Reasoning with Adversarial Reinforcement Learning

11. Constructive Circuit Amplification: Improving Math Reasoning in LLMs via Targeted Sub-Network Updates

12. AdaSearch: Balancing Parametric Knowledge and Search in Large Language Models via Reinforcement Learning

13. DVGT: Driving Visual Geometry Transformer

14. Posterior Behavioral Cloning: Pretraining BC Policies for Efficient RL Finetuning

15. MomaGraph: State-Aware Unified Scene Graphs with Vision-Language Model for Embodied Task Planning

Distributional AGI Safety

Authors: Nenad Tomašev, Matija Franklin, Julian Jacobs, Sébastien Krier, Simon Osindero

Source and references: https://arxiv.org/abs/2512.16856v1

---

Introduction

This paper proposes a framework for “Distributional AGI Safety” to address the potential risks of a spontaneous emergence of Artificial General Intelligence (AGI) through the interaction of a network of advanced AI agents with complementary skills, rather than a single monolithic AGI system.

Key Points

• The paper argues that the “patchwork AGI” hypothesis, where general capability levels emerge through coordination in groups of sub-AGI agents, needs to be taken seriously and should inform the development of corresponding safeguards and mitigations.

• The rapid deployment of advanced AI agents with tool-use capabilities and the ability to communicate and coordinate makes this an urgent safety consideration.

• The proposed framework centers on the design and implementation of virtual agentic sandbox economies, where agent-to-agent transactions are governed by robust market mechanisms, coupled with appropriate auditability, reputation management, and oversight to mitigate collective risks.

Methodology