Top AI Papers This Week

dVLA-RL introduces a novel reinforcement learning approach for discrete diffusion Vision-Language-Action models by optimizing denoising trajectories.

• Enables Reinforcement Learning (RL) for Discrete Diffusion VLAs (dVLAs) by optimizing denoising trajectories.
• Formulates the denoising process as an MDP, solving intractable marginal action probability for dVLAs.
• Introduces unified step scheduling, adapting denoising steps for efficient multi-task learning.

This paper presents a roofline-inspired scaling model to accurately predict the energy consumption of Transformer fine-tuning across multiple GPUs.

• Developed a framework to model energy consumption of Transformer training across multiple GPUs.
• Relates measured energy to proxies for compute, memory traffic, and hardware efficiency.
• Introduces a roofline-inspired hardware-efficiency factor to account for parallelism effects.

AutoSpec uses Inductive Logic Programming to automatically evolve and refine safety rules for LLM agents, reducing false positives while maintaining high recall.

• Introduces AutoSpec, a framework for automatically evolving safety rules for LLM agents.
• Leverages Inductive Logic Programming (ILP) to learn rule edits from user annotations and counterexamples.
• Reduces false positives by up to 94% and achieves high F1 scores (0.98, 0.93) in agent safety.

TriViewBench introduces a controlled benchmark to reveal MLLM limitations in multi-view structural reasoning, showing severe performance drops with complexity.

• Introduces TriViewBench, a novel 3-view benchmark with synthetic 3D scenes for controlled complexity scaling.
• Evaluates 18 MLLMs, revealing a consistent capability hierarchy and monotonic performance degradation with complexity.
• Identifies distinct failure modes in object counting: undercounting (occlusion) and overcounting (cross-view confusion).

This paper introduces the Unfireable Safety Kernel, an external, execution-time AI alignment mechanism that prevents AI agents from bypassing safety controls.

• Introduces "escapable AI systems" where agents can bypass internal safety controls.
• Proposes an "Unfireable Safety Kernel" for execution-time AI alignment with four architectural control properties.
• Implements the kernel in Rust, featuring machine-checked fail-closed invariants for robust security.

InvestPhilBench is a new multi-layer dynamic benchmark and scoring pipeline evaluating LLM procedural reasoning in expert investment philosophy.

• Introduces InvestPhilBench, a multi-layer dynamic benchmark with 8 cognitive tiers for LLM investment reasoning.
• Comprises 118 primary-source-verified principle cards, 25 decision frameworks, and 243 QA questions.
• Features the Benchmark Automated Scoring Pipeline (BASP) with 5 metrics and a Failure Mode Detection Protocol.

ForesightSafety-VLA is a new diagnostic benchmark for Vision-Language-Action (VLA) models, focusing on embodied safety evaluation.

• Introduces ForesightSafety-VLA, a diagnostic benchmark for evaluating embodied safety in VLA models.
• Defines a 13-category safety taxonomy across physical interaction, instruction, and perception.
• Evaluates policies under scene structure, language, and visual variations to diagnose failure sources.

LIBERO-Safety introduces a benchmark and dataset for evaluating physical and semantic safety in VLA models, revealing generalization-safety trade-offs.

• Introduces LIBERO-Safety, a parametric benchmark for procedurally generating safety-critical VLA scenarios.
• Develops a novel keypose-driven pipeline for scalable data generation, overcoming teleoperation bottlenecks.
• Curates a large-scale dataset of 19,664 strictly collision-free demonstrations with domain randomization.

VeriEvol scales multimodal mathematical reasoning by generating harder, image-grounded prompts and ensuring answer reliability through a novel verification framework.

• Introduces VeriEvol, an iterative framework for verifiable data construction in visual mathematical reasoning.
• Uses a type-aware evolution module to generate harder, image-grounded math prompts.
• Employs HTV-Agent, a verifier that ensures answer reliability through multi-source falsification.

A new pipeline uses LLM-annotated data and curriculum learning to scale dense retrieval for e-commerce search, significantly boosting performance over click-based methods.

• Multi-channel retrieval mining from three production systems for diverse training signals.
• LLM cascade for graded-relevance annotation, achieving 89.1% agreement with human labels.
• Three-stage progressive curriculum training with 240M+ examples across five difficulty levels.