Pruning Diffusion Models, Secure Code Generation, and Adaptive Reasoning for Embodied Navigation

Welcome to the first edition of State of AI for 2026👋 I hope you all enjoyed your holidays!

This edition covers a range of exciting technical advances, from improving the efficiency of large text-to-image diffusion models through novel pruning techniques, to ensuring the safety and reliability of LLM-generated code, and enhancing embodied navigation capabilities with adaptive reasoning and multi-modal memory.

Here’s what caught our attention:

• FastFLUX: Pruning FLUX with Block-wise Replacement and Sandwich Training - A framework for efficiently pruning large diffusion models while preserving high-quality text-to-image generation.

• STELP: Secure Transpilation and Execution of LLM-Generated Programs - A secure system for validating and executing LLM-generated code, addressing safety concerns.

• VLingNav: Embodied Navigation with Adaptive Reasoning and Visual-Assisted Linguistic Memory - A cognitive architecture that integrates adaptive reasoning and persistent multi-modal memory for improved embodied navigation.

• TerraFormer: Automated Infrastructure-as-Code with LLMs Fine-Tuned via Policy-Guided Verifier Feedback - A framework for automating Infrastructure-as-Code generation and mutation using LLMs fine-tuned with formal verification feedback.

• Multiplex Thinking: Reasoning via Token-wise Branch-and-Merge - A novel reasoning paradigm that combines the information density of continuous representations with the probabilistic structure of discrete sampling.

Let’s get into it 👇

Bi-Weekly AI Research Roundup

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

Contents

1. FastFLUX: Pruning FLUX with Block-wise Replacement and Sandwich Training

2. STELP: Secure Transpilation and Execution of LLM-Generated Programs

3. TerraFormer: Automated Infrastructure-as-Code with LLMs Fine-Tuned via Policy-Guided Verifier Feedback

4. A Vision for Multisensory Intelligence: Sensing, Science, and Synergy

5. DGAE: Diffusion-Guided Autoencoder for Efficient Latent Representation Learning

6. Aggregating Diverse Cue Experts for AI-Generated Image Detection

7. Efficient and Reproducible Biomedical Question Answering using Retrieval Augmented Generation

8. Precision over Diversity: High-Precision Reward Generalizes to Robust Instruction Following

9. Quantization Error Propagation: Revisiting Layer-Wise Post-Training Quantization

10. Multiplex Thinking: Reasoning via Token-wise Branch-and-Merge

11. Rewarding the Rare: Uniqueness-Aware RL for Creative Problem Solving in LLMs

12. TableCache: Primary Foreign Key Guided KV Cache Precomputation for Low Latency Text-to-SQL

13. Motion Attribution for Video Generation

14. VLingNav: Embodied Navigation with Adaptive Reasoning and Visual-Assisted Linguistic Memory

15. Real-Time Localization Framework for Autonomous Basketball Robots

FastFLUX: Pruning FLUX with Block-wise Replacement and Sandwich Training