Engineering
Keep-or-Drop? Adaptive Tokenizer for Compact Video Representation
KATok is a transformer-based VAE with an adaptive token selector for more efficient video compression, evaluating each token’s informativeness and…
Engineering
Chain-of-Thought Faithfulness of Reasoning Models Varies with Where and How Preference Cues Are Delivered
This paper introduces FACE-Eval, a 5,100-sample benchmark testing how chain-of-thought faithfulness varies based on whether preference cues originate from the…
Engineering
PaperBanana-Interact: Scientific Diagram Refinement with Multi-Turn Human Feedback
PaperBanana-Interact is a multi-agent system for refining scientific diagrams through iterative human feedback, accompanied by MTPaperBananaBench, a benchmark of 292…
Engineering
Scaffolding Foundation Models into Physical-World Agents Pushes the Frontier of Long-Horizon Navigation
NavMCP integrates vision-language models with navigation foundation models through three communication channels — intent, observation, and memory — to enable…
Engineering
MNIST-PRO: MNIST is Back as a Partially Observable World for AI Agents
MNIST-PRO transforms classic MNIST digit recognition into a sequential, glimpse-based search task under partial observability, evaluating how multimodal models construct…
Engineering
Weaving Visual Narratives: Agentic Image Bundle Composition Beyond Atomic Visual Matching
This paper introduces Image Bundle Composition (IBC), shifting image retrieval from scoring individual photos to dynamically assembling cohesive image collections…
Engineering
WebWorld: The Browser as a World Model for Self-Improving Web Code
WebWorld uses a web browser as a world model to improve vision-language models’ ability to generate and fix web code,…
Engineering
ContextBias: Controlled Evaluation of Bias Persistence Under Context Shift in Text-to-Image Models
ContextBias introduces ContextBench, a benchmark of 92 professional roles and 1,656 controlled prompts, to test whether stereotypical associations between professions…
Engineering
CoVA-SFT: A Large-Scale Dataset for Chain of Visual Abstractions
CoVA-SFT is a structured corpus of 51,900 samples containing over 222,000 multimodal reasoning steps, designed to train multimodal language models…
Engineering
EvoGenUI-Bench: Evaluating LLMs as Multi-Turn Generative UI Assistants
EvoGenUI-Bench evaluates LLMs’ ability to maintain and evolve interactive web interfaces across multiple turns, comprising 150 five-turn tasks across three…
Engineering
SafeAtlas-VL: Beyond Binary Multimodal Safety with Large-Scale Data and Guard Models
SafeAtlas-VL is a dataset of 1.5 million annotated instances for multimodal safety evaluation that replaces binary classification with a five-level…
Engineering
BLARM: Animating 3D Objects from Video via Blending Latent Rigid Motion Primitives
BLARM is a feed-forward method for generating animated 3D meshes from monocular video without explicit rigging or skeletal annotations, representing…