Fine-tuning LLMs improves capability but harms safety, study finds.
problem Balancing capability and safety in LLM fine-tuning.
method Theoretical framework and numerical experiments for two safety-aware fine-tuning strategies.
result Characterization of fundamental limits of safety-capability trade-off in LLM fine-tuning.
RLVR maintains safety while improving reasoning capabilities in LLMs.
problem Safety-capability tradeoff in fine-tuning LLMs.
method Reinforcement Learning with Verifiable Rewards (RLVR) and theoretical analysis.
result RLVR can enhance reasoning while maintaining safety guardrails.
A new framework ensures model safety by retaining old model capabilities while improving new tasks.
problem Catastrophic forgetting in continual learning systems.
method Retention-centric framework with data-dependent constraints and an efficient constrained optimization algorithm.
result The method ensures that new models retain important capabilities of old models while improving new tasks.
This paper tackles sandbagging in AI safety evaluations.
problem AI agents may hide dangerous capabilities to avoid being deactivated.
method Developed a simple model of strategic deception in sequential decision-making tasks.
result Demonstrated that optimal rational agents exhibit sandbagging behavior.
An active area of research is to increase the safety of self-driving vehicles. Although safety cannot be guarenteed completely, the capability of a vehicle to predict the future trajectories of its surrounding vehicles could help ensure this notion of safety to a greater deal. We cast the trajectory forecast problem in…
Synthesizes machine learning applications in reliability and safety.
problem Navigating the fragmented literature on ML for reliability and safety.
method Overview of ML categories, review of applications, discussion of Deep Learning.
result Machine learning can provide novel insights and improve accident prevention.
Paper explores anchoring for vision models, improving generalization and safety.
problem Anchoring can lead to undesirable shortcuts, limiting generalization.
method Introduced a new anchored training protocol with a regularizer to mitigate undesirable shortcuts.
result Significant performance gains in generalization and safety metrics.
Enhances safety of 3D object detection neural networks.
problem Ensuring robustness and safety of 3D object detection systems.
method Symbolic error propagation, specialized loss function, safety-aware non-max-inclusion algorithm.
result Improved safety and robustness of 3D object detection neural networks.
This paper formalizes AI safety using hypothesis testing in GenAI.
problem Ensuring safety of generative AI tools that create realistic content.
method Formalization of computational safety through hypothesis testing and signal processing.
result Demonstrates how AI safety can be assessed quantitatively using mathematical frameworks.
Meta-learning priors improves safe Bayesian optimization.
problem Optimizing robot controllers under safety constraints.
method Meta-learning priors from offline data using F-PACOH.
result Meta-learned priors accelerate safe BO convergence.
ContextBench benchmarks methods for generating linguistically fluent inputs that activate specific latent features in language models.
problem Identifying inputs that trigger specific behaviours or latent features in language models.
method Context modification and benchmarking methods like Evolutionary Prompt Optimisation (EPO) with LLM-assistance and diffusion model inpainting.
result Enhanced methods achieve state-of-the-art performance in balancing elicitation effectiveness and fluency.
CitySim dataset captures vehicle trajectories for safety research.
problem Lack of fine-grain vehicle trajectories for safety-oriented research.
method Five-step procedure: video stabilization, object filtering, stitching, detection, and error filtering.
result CitySim dataset improves safety evaluations and facilitates digital-twin research.
Generates multimodal safety-critical scenarios for robustness evaluation of decision-making algorithms.
problem Lack of comprehensive evaluation of neural network robustness under real-world scenarios.
method Proposes a flow-based multimodal scenario generator using weighted likelihood maximization and gradient-based sampling.
result Demonstrates improved testing efficiency and multimodal modeling capability compared to traditional methods.
Develops probabilistic safety regions for scalable classifiers.
problem Minimizing misclassification errors in supervised classification.
method Introduces probabilistic safety regions and scalable classifiers.
result Probabilistic certifications for classifier performance.
Reward hacking exploits misspecified rewards, affecting agent capabilities and true performance.
problem Reward hacking in RL models exploiting reward misspecifications.
method Constructed four RL environments with misspecified rewards; analyzed agent capabilities and behavior.
result More capable agents exploit reward misspecifications, achieving higher proxy reward but lower true reward.
Artificial Neural Networks (ANNs) have demonstrated remarkable utility in various challenging machine learning applications. While formally verified properties of their behaviors are highly desired, they have proven notoriously difficult to derive and enforce. Existing approaches typically formulate this problem as a p…
YRC-Bench benchmarks AI agents learning to collaborate with experts.
problem Learning to recognize when to consult an expert in novel situations.
method Validation strategy and proposer-validator decomposition.
result Insights for future AI safety research.
Practical reinforcement learning problems are often formulated as constrained Markov decision process (CMDP) problems, in which the agent has to maximize the expected return while satisfying a set of prescribed safety constraints. In this study, we propose a novel simulator-based method to approximately solve a CMDP pr…
Naturalistic driving trajectories are crucial for the performance of autonomous driving algorithms. However, most of the data is collected in safe scenarios leading to the duplication of trajectories which are easy to be handled by currently developed algorithms. When considering safety, testing algorithms in near-miss…
This paper uses MIL and MHCNN-RNN to predict precursors to aviation safety events.
problem Identifying events that precede aviation safety incidents.
method Multiple-instance learning (MIL) framework combined with a Multi-Head Convolutional Neural Network-Recurrent Neural Network (MHCNN-RNN) architecture.
result Multiple binary classifiers outperform in predicting high speed and high path angle events during the approach phase.
Semantic Embeddings are a popular way to represent knowledge in the field of zero-shot learning. We observe their interpretability and discuss their potential utility in a safety-critical context. Concretely, we propose to use them to add introspection and error detection capabilities to neural network classifiers. Fir…
Survey of alignment techniques for large language models.
problem Ensuring large language models align with human values.
method Analysis of diverse alignment methods and training paradigms.
result Preference-based methods offer more flexibility for nuanced alignment.
ML Compass helps organizations choose AI models that balance utility, cost, and compliance.
problem Selecting AI models that meet user utility, deployment costs, and compliance requirements.
method Develops ML Compass, a framework for constrained optimization over a capability-cost frontier, using internal measures and empirical data.
result ML Compass produces deployment-aware recommendations that differ from capability-only rankings, clarifying trade-offs between capability, cost, and safety.
Shai is a 10B model for asset management tasks, outperforming baselines.
problem Improving performance in asset management tasks.
method Continuous pre-training and fine-tuning on asset management-specific data.
result Shai outperforms baseline models in asset management tasks.
New approach extracts AI model representations for steering and monitoring.
problem Understanding and steering AI models' internal knowledge representation.
method Feature learning advances for scalable linear concept representations.
result Concept representations enable steering and monitoring of AI models.
The paper discusses safety assessment for AI systems, focusing on machine learning models.
problem Safety assessment of AI systems, especially machine learning models, in safety-related applications.
method Analyzed AI models as statistical models and proposed a new budget allocation for AI safety.
result Safety assessment of AI systems requires a new approach focusing on the model used, not just the system.
Autonomous vehicles rely on machine learning to solve challenging tasks in perception and motion planning. However, automotive software safety standards have not fully evolved to address the challenges of machine learning safety such as interpretability, verification, and performance limitations. In this paper, we revi…
New method predicts language model scaling risks with limited data.
problem Accurately predicting model capabilities and risks with limited data.
method Introduces a robust estimation framework using beta-binomial distribution and dynamic sampling strategy.
result More accurate predictions of rare risks and capabilities at lower computational cost.
RAGuard improves safety in LLMs for offshore wind maintenance.
problem Conventional LLMs fail with specialised or unexpected scenarios in offshore wind maintenance.
method Integrates safety-critical documents alongside technical manuals in RAG framework.
result RAGuard increases safety recall from almost 0% to over 50% while maintaining technical recall above 60%.
Survey of algorithms for testing AI-driven CPS safety.
problem Testing AI-driven CPS for safety in complex environments.
method Survey of applied algorithms for safety validation.
result Survey of existing tools and techniques for safety validation.
Deep neural network controllers for autonomous driving have recently benefited from significant performance improvements, and have begun deployment in the real world. Prior to their widespread adoption, safety guarantees are needed on the controller behaviour that properly take account of the uncertainty within the mod…
Safe imitation learning with a safety layer for flexible training.
problem Flexible yet safe imitation learning for complex tasks.
method Theory and modular method with a safety layer for continuous policy, adversarial training, and worst-case safety guarantees.
result Robustness advantage of safety layer during training compared to test time.
Deep learning methods are widely regarded as indispensable when it comes to designing perception pipelines for autonomous agents such as robots, drones or automated vehicles. The main reasons, however, for deep learning not being used for autonomous agents at large scale already are safety concerns. Deep learning appro…
A new algorithm for identifying the best arm in linear feedback with safety constraints.
problem Identifying the best arm in linear feedback with safety constraints.
method A gap-based algorithm that ensures safety while minimizing sample complexity.
result The algorithm achieves meaningful sample complexity while ensuring safety.
Safety filter for unknown discrete-time systems with learned models and noise covariance.
problem Ensuring safety for unknown discrete-time linear systems with Gaussian noise.
method Develops a learning-based safety filter using empirical model and noise covariance, optimizing control actions to stay within safety constraints.
result Minimally modifies nominal control actions to ensure safety with high probability, tightening constraints as more data is collected.
SafeML monitors ML systems for safety and security risks.
problem Ensuring safety and explainability of ML systems in safety-critical domains.
method Statistical difference measures of ECDF to detect distributional shifts.
result Approach can detect invalid application contexts of ML components.
Paper defines and quantifies safety risks in deep neural networks.
problem Safety concerns in deep neural networks applied to critical sectors.
method Defines safety property, computes maximum safe radius, identifies new risk class, develops algorithm.
result Method achieves competitive performance in safety quantification.
Study best arm identification with safety constraints in bandit problems.
problem Real-world decision-making with safety constraints.
method Analyzed linear and monotonic reward and safety constraints, proposed algorithms.
result Guaranteed safe learning in both linear and general reward/safety constraint settings.
Failures are challenging for learning to control physical systems since they risk damage, time-consuming resets, and often provide little gradient information. Adding safety constraints to exploration typically requires a lot of prior knowledge and domain expertise. We present a safety measure which implicitly captures…
Safe reinforcement learning tackles safety constraints with linear approximations.
problem Ensuring safety in reinforcement learning without violating constraints.
method Modeling safety as a linear cost function, developing SLUCB-QVI and RSLUCB-QVI algorithms for MDPs with linear function approximation.
result Achieved a nearly optimal regret bound for safe reinforcement learning, matching state-of-the-art unsafe algorithms.
Each year, around 6 million car accidents occur in the U.S. on average. Road safety features (e.g., concrete barriers, metal crash barriers, rumble strips) play an important role in preventing or mitigating vehicle crashes. Accurate maps of road safety features is an important component of safety management systems for…
Localized Multidirectional Correction improves non-refusal target-response behavior in foundation models.
problem Controlled post-training refusal suppression in routed MoE and hybrid-MoE foundation models.
method Introduce Localized Multidirectional Correction (LoMC), a support-gated intervention framework.
result Substantially improves non-refusal target-response behavior while maintaining general capability under a compact intervention footprint.
CARL safely adapts RL agents for safety-critical tasks.
problem Safety hazards in RL for safety-critical tasks.
method CARL combines model-based RL and cautious adaptation.
result CARL achieves higher rewards with fewer failures in safety-critical tasks.
A novel approach for safe offline RL using latent safety constraints.
problem Balancing safety constraints and reward maximization in offline RL.
method Conditional Variational Autoencoders for latent safety modeling, Constrained Reward-Return Maximization.
result Our approach maintains safety compliance while optimizing rewards, outperforming existing methods.
New method improves safety analytics by addressing imbalanced data issues.
problem Imbalanced safety datasets lead to inaccurate predictions and management problems.
method Extended accident triangle theory and three oversampling methods.
result Robust improvements in machine learning algorithms for safety analytics.
LLM safety alignment explained as divergence estimation.
problem Aligning large language models to avoid harmful outputs.
method Presented a theoretical framework showing alignment methods as divergence estimators.
result KLDO method improves safety alignment using compliance-refusal datasets.
Machine learning algorithms increasingly influence our decisions and interact with us in all parts of our daily lives. Therefore, just as we consider the safety of power plants, highways, and a variety of other engineered socio-technical systems, we must also take into account the safety of systems involving machine le…
Safe Bayesian Optimization algorithms are improved to ensure safety in real-world applications.
problem Ensuring safety in Bayesian Optimization algorithms for real-world applications.
method Investigated and improved three safety-related issues of SafeOpt-type algorithms: frequentist uncertainty bounds, RKHS norm assumptions, and discrete search spaces.
result Introduced Real-{eta}-SafeOpt, Lipschitz-only Safe Bayesian Optimization (LoSBO), and Lipschitz-only GP-UCB (LoS-GP-UCB) algorithms that retain safety guarantees and superior performance.