The CoVar Zeitgeist: November, 2026¶

A curated list of the latest research in AI.

Featured¶

The Emergent Symbolic Structure of Artificial Neural Networks: This paper demonstrates that neural networks can implicitly represent and utilize symbolic structures similar to those in traditional logic and language, suggesting a bridge between connectionist models and symbolic AI.
Gaussian Splatting is an Effective Data Generator for 3D Object Detection: Reconstructing real driving scenes with Gaussian splatting and inserting 3D objects at physically plausible locations yields training data with exact 3D labels that boosts object detection and shows that random object orientations improved orientation error without changing mAP.

Infinite-Parameter LLMs: Generating and Adapting Weights from Live Data: The paper introduces Infinite-Parameter LLMs, which dynamically generate weights from live interaction data using a hypernetwork, enabling continuous learning and adaptation beyond static pretraining.
Kimi K2.5: Visual Agentic Intelligence: Kimi K2.5 is an open-source multimodal agentic model that combines text and vision for enhanced general intelligence, featuring Agent Swarm for efficient task execution, achieving state-of-the-art results across various domains.
EmbeddingGemma: Powerful and Lightweight Text Representations: EmbeddingGemma, a lightweight 300M parameter text embedding model, outperforms larger models across various benchmarks while maintaining low latency and high throughput, making it ideal for on-device applications.

Foundation-Model-Driven 3D Scene Graphs for Mobile Manipulation: This paper integrates FOUND-IT, a geometric and semantic foundation model, into a mobile manipulator for real-time 3D scene graph construction, enhancing object detection, path planning, and grasping tasks.
NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis: NeRF compresses a scene into a neural network, learned from photos with known camera positions, that can be queried at any 3D point for density and view-dependent color, so it can render realistic views from new angles.
Rendering Synthetic Objects into Legacy Photographs: A foundational graphics paper that turned a single real photo into a renderable 3D scene, so synthetic objects could be inserted with matching lighting, shadows and occlusion, the same recipe later used to generate realistic training data.

Joint Embeddings for Graph Instruction Tuning: This paper introduces joint embeddings to integrate graph data into large language models, enhancing their ability to understand and respond to graph-based instructions, outperforming traditional graph-to-text methods.

An Empirical Study of Example Forgetting during Deep Neural Network Learning: The paper tracks how often individual training examples flip from correctly to incorrectly classified during training, showing that some examples are forgotten repeatedly while others never are, that this pattern carries across architectures, and that a large share of the unforgettable examples can be dropped from the training set without hurting generalization.