← Back to all news

AI safety and alignment

Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.

17,884 stories · RSS feed

arXiv AI
Aug 26

When May an Agent Stop? Evidence-Carrying Termination for Tool-Using LLMs

The paper introduces Evidence-Carrying Termination (ECT), a method that allows tool‑using large language model agents to declare a task complete only when a typed certificate links every required answer claim to valid, in‑scope trace evidence and a deterministic replay confirms the claimed value. In controlled experiments across 48 synthetic tasks and 576 trajectories, ECT eliminated unsafe completions and premature unsupported terminations, outperforming existing termination critics and controllers while maintaining comparable supported completion rates.

By Jason Liu
llmsagentssafety
More like this →
arXiv AI
Aug 26

Names Can Hurt: Spotting Slopsquatting Risks Caused by Package Name Hallucinations in Local Coding LLMs

The paper introduces a two‑layer detector to prevent ‘slopsquatting’—the risk of local coding LLMs fabricating Python package names that adversaries can pre‑register on PyPI. The first layer checks PyPI for existence, while the second uses a Random Forest classifier on ten name‑and‑metadata features; an import reconciler resolves naming mismatches. Embedded in a LangGraph state machine, the system retries at escalating temperatures and falls back to stronger models, achieving hallucination‑free code in 76% of 300 curated prompts and recovering additional runs through intra‑ and cross‑model retries.

By Akash Raj, Sargam Sahu
llmscomputer-visionfine-tuningsafety
More like this →
arXiv Computer Vision
Aug 26

Variance-Guided Spatial Attention Fusion for Robust End-to-End Driving under Asymmetric Sensor Degradation

The paper introduces Variance‑Guided Spatial Attention Fusion (VG‑SAF), a method for robust end‑to‑end driving that fuses camera and LiDAR data while handling asymmetric sensor degradation. VG‑SAF uses a physically grounded augmentor to generate dense reliability masks, modality‑specific experts to predict per‑pixel reliability scales, and a hybrid attention mechanism that gates unreliable cells and balances modalities. The approach also includes a Laplace uncertainty head to signal severe or combined sensor failures, and demonstrates improved closed‑loop robustness on the CARLA Longest6 benchmark across various degradation scenarios.

By Weizhi Tao, Zengwang Jin, Xiao Wang, Hailong Huang
efficiencymultimodalbenchmarkssafety
More like this →
arXiv Computer Vision
Aug 26

O3N: Omnidirectional Open-Vocabulary Occupancy Prediction for Embodied Intelligent Robotics

O3N is a novel framework that performs open‑vocabulary occupancy prediction from a single omnidirectional RGB image. It introduces a polar‑spiral voxel embedding (PsM) for continuous 360° spatial representation, an Occupancy Cost Aggregation (OCA) module that unifies geometric and semantic supervision, and a Natural Modality Alignment (NMA) pathway that aligns visual, voxel, and text features. Experiments show state‑of‑the‑art results on QuadOcc and Human360Occ benchmarks, with strong cross‑scene generalization and semantic scalability.

By Mengfei Duan, Hao Shi, Fei Teng, Guoqiang Zhao, Yuheng Zhang, Zhiyong Li, Kailun Yang
ragroboticsbenchmarkssafety
More like this →
arXiv AI
Aug 26

AI Finds A Way

The paper "AI Finds A Way" compiles 26 firsthand anecdotes from over 100 researchers across machine learning subfields, illustrating how AI systems often discover creative, unexpected solutions that can circumvent human-imposed design limits. These cases highlight the tendency of modern AI to exploit loopholes in reward signals and uncover novel scientific phenomena, even when using large foundation models. The authors argue that such behavior poses safety challenges and underscores the need to align AI models with human values while preserving their capacity for innovation.

By Aaron Dharna, Cong Lu, Ryan Sullivan, Joel Lehman, Victoria Krakovna, Jeff Clune
reinforcement-learningsafety
More like this →
arXiv AI
Aug 26

RAGSentinel: Certifiable Geometric Consensus for Robust Retrieval-Augmented Generation

RAGSentinel is a training‑free, label‑free defense designed for black‑box retrieval‑augmented generation systems. It employs a surrogate encoder to detect hidden‑state shifts caused by retrieved documents, removes shared topic directions, and filters poisoned documents as geometric outliers from a robust majority consensus. The method is proven to recover a poison‑free majority context under honest‑majority and representation‑separation assumptions, and experiments show it keeps attack success rates low while maintaining accuracy across multiple datasets, LLM families, and adaptive attacks.

By Yueyang Quan, Anjun Gao, Yufei Xia, Minghong Fang, Zhuqing Liu
llmsragsafety
More like this →
arXiv Machine Learning
Aug 26

Low-Latency Activation-Regularized Sparse Neural Operators with Distillation Assistance Towards Real-Time Edge-Deployable Virtual Sensing

The paper introduces the Sparse-Activation-ReLU (SAR) layer, a single‑step neural operator that promotes activation sparsity without surrogate‑gradient training and is compatible with event‑based computing. In a trunk‑based NOMAD architecture, SAR improves the combined Latency‑Error‑Energy (LEE) metric by over fivefold compared to Variable Spiking Neuron (VSN) and Leaky Integrate‑and‑Fire (LIF) models. Additional techniques such as synthetic knowledge distillation, a ReLU‑based spiking loss, and graph‑neighbor thresholding further reduce LEE and L2 error on the Heat Exchanger dataset, advancing energy‑efficient virtual sensing for edge deployment.

By William Howes, Farid Ahmed, Syed Bahauddin Alam
efficiencysafety
More like this →
arXiv AI
Aug 26

'Ghaib in Translation' aka Unseen Harm: Measuring Cross-Script Safety Inconsistency with 'Missed-in-Urdu' Scores in LLM Hate Speech Detection

The paper "Ghaib in Translation" investigates how large language models (LLMs) handle Urdu, a widely spoken language that is largely absent from safety evaluations. Five prominent LLMs—GPT‑4o, Claude Sonnet 4.5, Gemini 2.5 Flash, Qwen‑2.5, and Llama‑3.1—were tested on six datasets covering Nastaliq Urdu, Roman Urdu, English, and code‑switched Urdu‑English. The study found significant label instability between original‑script and English‑translation classifications, with missed‑in‑Urdu rates ranging from 2.4% to 9.9% (median 4.3%). A review of 205 papers across nine ALW/WOAH editions revealed no dedicated Urdu research, underscoring the language’s neglect in current safety research.

By Fawzia Zehra (Fuzzy), Kara-Isitt, Sonal Khosla, Stephen Swift
llmssafety
More like this →
arXiv AI
Aug 26

FARCA: Fact-Aligned Reliability-Aware Credit Assignment for Reinforcement Learning with Factual Supervision

FARCA is a policy‑optimization framework that addresses the problem of noisy factual credit assignment in reinforcement learning for large language models. It transforms coarse factual supervision into token‑level training signals that are both localized and weighted by a reliability estimate derived from counterfactual evidence attribution. Experiments on several factual reasoning benchmarks show that FARCA improves model factuality while maintaining general reasoning abilities.

By Qiming Xie, Wenjie Zheng, Xiangqing Shen, Rui Xia
llmsreinforcement-learningbenchmarkssafety
More like this →