AI/ML arXiv cs.AI

NormWorlds-CF: Solver-Verified Counterfactual Normative Reasoning with Metamorphic-Relation GRPO

Introduces NormWorlds-CF, a solver-verified environment for counterfactual normative reasoning, and MR-GRPO for improved reward conditioning in LLMs.

Cybersecurity arXiv cs.AI

Time-Frequency Consistency Learning for Robust Speech Deepfake Detection

Proposes the Time-Frequency Consistency Learning (TFCL) framework to improve the robustness of speech deepfake detection against real-world acoustic processing distortions.

AI/ML arXiv cs.AI

Filling Before Advancing: Capability-Gap-Driven Post-Training for Scenario-Specialized Remote Sensing MLLMs

Introduces 'Filling Before Advancing' (FBA), a post-training method for remote sensing MLLMs to bridge capability gaps before scenario specialization.

AI/ML arXiv cs.AI

scMIR: a vision-language foundation model for single-cell light microscopy image representation

Presents scMIR, a vision-language foundation model designed for high-throughput automated representation of single-cell light microscopy images.

Other arXiv cs.AI

An Explicit Counterexample to Stanley's Rankwise Lower-Bound Conjecture for Differential Posets

Provides a mathematical counterexample to disprove Stanley's Rankwise Lower-Bound Conjecture for differential posets for r >= 3.

AI/ML arXiv cs.AI

Directional Influence Function: Estimating Training Data Influence in Constrained Learning

Proposes the Directional Influence Function (DIF) to accurately estimate training data influence in models trained under explicit feasibility constraints.

AI/ML arXiv cs.AI

Moral Hazard in Multi-Agent Language Models

Explores 'moral hazard' in multi-agent LLMs through a dialogue game, analyzing how incentive structures affect cooperation and information sharing.

AI/ML arXiv cs.AI

AGMark: Attention-Guided Dynamic Watermarking for Large Vision-Language Models

Proposes AGMark, a dynamic watermarking framework for Large Vision-Language Models that uses attention weights to maintain visual-semantic fidelity while protecting IP.

AI/ML arXiv cs.AI

Breaking the Curse of Repulsion: Remoteness-Aware Control of Negative Off-Policy Updates

Introduces Dynamic Remoteness-Aware Policy Optimization (DRPO) to fix the 'curse of repulsion' in off-policy reinforcement learning by attenuating remote negative updates.

AI/ML arXiv cs.AI

Real-Time Driver Safety Scoring Through Inverse Crash Probability Modeling

Presents SafeDriver-IQ, a framework that converts binary crash classifiers into continuous 0-100 safety scores for real-time driver feedback.

AI/ML arXiv cs.AI

LLM-generated personalized nudges for improving pro-environmental behavior: Field evidence from resource conservation

A field study demonstrating that LLM-generated personalized nudges significantly improve pro-environmental behavior in resource conservation.

AI/ML arXiv cs.AI

RankFormer: A Propose-then-Select Transformer for Multi-Agent Multimodal Trajectory Prediction

Introduces RankFormer, a Transformer-based network for multi-agent multimodal trajectory prediction in autonomous driving without needing labeled intentions.

AI/ML arXiv cs.AI

Structured Scaling of AI Discovery Across Diverse Scientific Domains

Introduces SimpleTES, a framework for structured scaling of AI discovery loops that achieved new state-of-the-art results across 28 scientific domains.

AI/ML arXiv cs.AI

Short-Term-to-Long-Term Memory Transfer for Knowledge Graphs under Partial Observability

Proposes a neuro-symbolic value-based approach for short-term-to-long-term memory transfer in knowledge graphs under partial observability.

AI/ML arXiv cs.AI

GoQuant: Geometric Orthogonal Residual Projection for Multiplier-Free Power-of-Two Transformer Quantization

Presents GoQuant, a hardware-efficient Power-of-Two Transformer quantization framework that replaces multiplications with bit-shifts to optimize edge deployment.

AI/ML arXiv cs.AI

PatchWorld: Gradient-Free Optimization of Executable World Models for Agent Environments

Introduces PatchWorld, a gradient-free framework that converts offline trajectories into executable Python world models via counterexample-guided code repair.

AI/ML arXiv cs.AI

Detect Before You Leap: Mirage Detection in Vision-Language Models

Proposes TC-LIA, a model-agnostic method to detect 'mirages' (hallucinated visual evidence) in Vision-Language Models before answers are released.

AI/ML arXiv cs.AI

Building Large-Scale English-Romanian Literary Translation Resources with Open Models

Introduces the TinyFabulist Translation Framework (TF2) for high-quality English-to-Romanian literary translation using open-weight models.

AI/ML arXiv cs.AI

CIFNet: An Analytic Neural Learning Framework for Efficient and Calibrated Class-Incremental Learning

Presents CIFNet, a neural learning framework that uses closed-form classifier adaptations instead of backpropagation for efficient class-incremental learning.

AI/ML arXiv cs.AI

Comparing RAG and GraphRAG for Page-Level Retrieval Question Answering on a Math Textbook

Compares RAG and GraphRAG for textbook question answering, finding that standard embedding-based RAG is more accurate and efficient for page-level retrieval.