Safety & alignment

Moral sensitivity is the ability to identify the morally relevant features of a decision situation and use them as the basis for action. It is the foundation of broader moral competence: any other moral reasoning capabilities will be irrelevant if an agent lacks sensitivity to the relevant facts. In this paper, we offer a new evaluation of LLM moral sensitivity and in doing so, we address and resolve a central problem in AI alignment research: how to scale behavioural evaluations beyond expensiv

Large Vision-Language Models (LVLMs) exhibit remarkable vision-language capabilities and are increasingly deployed in real-world applications such as personal assistants, document analysis systems, and embodied agents. However, their dual-modal attack surfaces make them vulnerable to jailbreak attacks. Existing LVLM jailbreaks rely on simple designs, e.g., short text and out-of-distribution images. Nevertheless, recent advancements in both large language model backbones and multimodal mechanisms

In many scientific and engineering domains, maximizing discovery within a limited sampling budget demands strategic, observation-guided exploration. While generative models have enabled training-free reward alignment, current methods typically excel in local searches within narrow regions of the underlying distribution. These approaches struggle when preferences are unknown a priori and only revealed through sequential feedback-a scenario demanding broad exploration to uncover high-utility regio

Large language models (LLMs) have demonstrated remarkable capabilities across diverse applications, yet ensuring their simultaneous safety, helpfulness, and trustworthiness remains a persistent challenge. Conventional refusal-oriented alignment strategies mitigate harmful content generation but systematically fail to serve legitimate user needs, often withholding information that could safely and constructively address the underlying intent of sensitive queries. Building upon the constructive sa

Inference-time alignment steers a frozen language model during decoding using auxiliary reward signals, avoiding the cost of repeated weight updates. However, existing inference-time alignment methods typically optimize a single scalar score, so explicit safety constraints must either be ignored or encoded through manually tuned penalties. We propose Lagrangian Reward Augmentation (LARA), a general inference-time alignment framework under safety constraints. Starting from a KL-regularized constr

There is no doubt that safety alignment is an essential step in LLM training. However, conceptually it does not distinguish between various domains and the level of potential harm of a query, which creates significant complications in the fields like cyber security, where a model should not be constrained by its safety circuits to accomplish the goals of legitimate, authorized operations. In this work, we share our findings from a large scale abliteration experiment on 24 open-source LLMs and sh

We recently ran a strategy fellowship through Astra. As part of this, we ran a reading group for our fellows on some of the topics that we think are important for thinking about AI futurism (key dynamics in AI development, existential risk from AI, and approaches to mitigating risk). This post contains the reading list we used.

Local pairwise comparisons are a standard tool for learning how people want decision rules to work, e.g., in participatory design or alignment. However, their use builds in two strong assumptions: that local comparisons are sufficient evidence about how a person wants an automated decision rule to behave, and that people can always answer those comparisons decisively. We investigate how these assumptions may be compromised under internal pluralism: the idea that an individual evaluates decision

Despite alignment training, LLMs remain prone to generating unsafe outputs at deployment time. Monitoring outputs online and raising an alarm when safety can no longer be assumed is therefore critical. We study a simple real-time monitor that turns a verifier signal from an external model into an alarm decision by thresholding, with the threshold calibrated via risk control. In experiments on mathematical reasoning and red teaming datasets, we show that this simple design is competitive with mor

Vision-Language-Action (VLA) models are fundamentally bottlenecked by the scarcity of expert demonstrations -- triplets of observations, instructions, and actions that are costly to collect at scale. We argue that this bottleneck stems from conflating two distinct learning objectives: acquiring physical competence (how to move) and acquiring semantic alignment (what to do). Crucially, only the latter requires language supervision. Building on this Decomposition Hypothesis, we propose Task-Agnost

Deep learning has achieved remarkable performance in medical image segmentation, yet it suffers from critical limitations: mathematical intractability, substantial parameter requirements, and lack of clinical interpretability. We propose RadiomicNet, a novel two-stream hybrid architecture that enhances standard deep learning by integrating handcrafted radiomics features directly into the segmentation learning process. The key contribution is the Radiomics Attention Gate (RAG), which leverages Gr

As Large Language Models (LLMs) and agentic systems become integrated into real-world applications, ensuring their safety and security is critical. Guardrail systems that detect and block malicious instructions sent to and from an LLM are an essential component of AI security. However, researchers conducting black-box adversarial emulation against production AI systems often struggle to determine whether a guardrail block or an LLM rejection has occurred. This distinction is important because th

Using molecular large language models (LLMs) as a unified framework for understanding molecular structures and functions is emerging as a new trend in tasks such as molecular design and drug discovery. However, these models struggle to fully capture the visual representation of molecular structures, limiting their potential. While existing molecular vision-language models (VLMs) show promise, they still face challenges in structural alignment and lack the necessary topological modeling for accur

We study the challenging problem of novel view video synthesis from single images or monocular videos. Existing methods, which operate under the assumption that pre-trained video models lack native novel view synthesis capability and enforce view alignment via camera conditioning, task-specific fine-tuning, or stepwise hard denoising guidance, often suffer from artifacts and compromised global scene consistency. In this paper, we introduce NeoMap, a novel training-free framework designed to loca

Organizations increasingly make strategic decisions about AI systems whose behaviour, failure modes, and institutional effects cannot be fully known at design time. This technical report reframes strategic red teaming as a board-level governance discipline for testing the assumptions under which AI-enabled strategies are approved, funded, and supervised. The report proposes a six-component model for strategic red teaming in AI governance: an explicit assumption register, an adversarial mandate,

During a recent conference on the People’s Liberation Army, I heard the same question posed to attendees and paper writers: “How would China react to U.S. force posture change X, Y, or Z?” or “How would the Chinese military respond to U.S. strikes in certain locations?” Having participated in dozens of unclassified wargames at the RAND Corporation and elsewhere, I hear a similar refrain when playing the “red team.”This is a reasonable and legitimate question. Policymakers and war planners should

Existing robust preference optimization for language-model alignment mainly studies pairwise supervision and places robustness at the dataset, prompt, or preference-pair level. We instead study listwise preference optimization under ranking-label uncertainty: given a prompt and a candidate list, the observed ranking over that list may be ambiguous due to annotator inconsistency, near-ties, lossy rankwise feedback, or reward-model noise. We propose a pointwise total-variation robust Plackett--Luc

Vision-language-action models (VLAs) have recently advanced robotic manipulation, yet the effects of different robot-data pre-training paradigms remain difficult to compare because existing models often differ in architecture, data, action space, and evaluation protocol. We present VLAFlow (Vision-Language-Action Flow), a unified flow-matching framework for controlled comparison of VLA training objectives. Using a heterogeneous robot corpus, OXEMix, containing approximately 5,000 hours of data f

The emergence of the transformer architecture has ushered in a new era of possibilities, showcasing remarkable capabilities in generative tasks exemplified by models like GPT4o, Claude 3, and Llama 3. However, these advancements come with a caveat: predominantly trained on data gleaned from social media platforms, these systems inadvertently perpetuate societal biases and toxicity. Recognizing the paramount importance of AI Safety and Alignment, our study embarks on a thorough exploration throug

Safety evaluations for language models increasingly depend on judgments about ambiguous natural-language behaviour: whether a model has followed an instruction, refused appropriately, complied with a policy, resisted an embedded command, or misreported progress in an agentic task. Existing benchmarks often compress these distinctions into pass/fail labels, obscuring whether failures arise from capability limits, policy ambiguity, instruction conflict, scaffold failure, or unstable evaluator judg

While generative models have enabled training-free reward alignment, current methods typically excel in local exploration within narrow regions of the underlying distribution. These approaches struggle when preferences are unknown a priori and only revealed through sequential feedback-a scenario demanding broad exploration to uncover high-utility regions. To address this, we propose Sequentially-Controlled Interactive Multi-Particle Flow-Maps (IMPFM), a framework for sample-efficient online feed

Public First Action routed $2m to support Colorado House candidate Manny Rutinel through Latino Victory Fund — and didn’t publicly announce it ahead of Rutinel’s victory

Millimeter-wave (mmWave) beam alignment plays a critical role in next-generation wireless systems, yet its efficient implementation remains challenging. Meta-learning and transfer learning have been explored to enable deep learning-based beam prediction models to rapidly adapt to unseen environments; however, existing meta-learning approaches adapt the entire network and are trained from random initialization, leading to a large number of updated parameters and a high meta-training cost, while t

The bird's-eye view (BEV) representation enables multi-sensor features to be fused within a unified space, serving as the primary approach for achieving comprehensive 3D perception. However, the discrete grid representation of BEV leads to significant detail loss and limits feature alignment and cross-modal information interaction in multimodal fusion perception. In this work, we break from the conventional BEV paradigm and propose a new universal framework for multi-modal fusion based on 3D Gau

Congratulations to the Berkeley Artificial Intelligence Research (BAIR) Lab class of 2026! This year, BAIR celebrates another remarkable group of Ph.D. graduates whose curiosity, creativity, and perseverance have pushed the frontiers of artificial intelligence and machine learning. Their work spans the breadth of modern AI — robotics and embodied intelligence, large language models and reasoning, computer vision, generative modeling, AI safety, human-AI interaction, AI for science and healthcare

Understanding how aligned LLMs internally represent safety is critical for diagnosing alignment vulnerabilities, as it explains why jailbreaks succeed and informs the design of robust alignment strategies. Prior work shows that aligned LLMs encode harmfulness and refusal as separable directions in the residual stream at prompt-side token positions. We show that jailbreaks succeed at prompt encoding by suppressing either the refusal or harmfulness direction before any token is generated, with dis

The research proposes a multilayer Q-matrix-embedded neural network for cognitive diagnosis (M-QCDNet), which integrates the structural interpretability of cognitive diagnostic models (CDMs) with the deep learning neural network (NN). M-QCDNet structures the item-skill relationship using the Q-matrix as a structural prior, ensuring latent mastery profiles remain interpretable and consistent with cognitive theory, followed by the proposed loss function with an L2 penalty to penalize skills not al

Diffusion models are highly effective at modeling complex data distributions, including images and text. However, in applications like personalized recommender systems, the objective often shifts to modeling specific regions of the distribution that maximize user preferences-initially unknown but gradually uncovered through interactive feedback. This can naturally be framed as a reinforcement learning problem, where the goal is to fine-tune a diffusion model to maximize a reward function based o

Safety alignment of text-to-image (T2I) diffusion models aims to suppress harmful generations while preserving utility on benign prompts. Recent methods often appear to deliver high safety with high utility, but this conclusion rests largely on coarse global utility metrics (e.g., FID, CLIPScore) that are insensitive to fine-grained semantic correctness, creating an illusion of high utility. We show that when utility is measured with structured evaluation, this illusion breaks: on TIFA (Text-to-

Visual AutoRegressive modeling (VAR) has pioneered a coarse-to-fine multi-scale autoregressive generative paradigm, demonstrating strong capabilities in image generation. However, VAR still suffers from inherent deficiencies in multi-scale representation learning. Specifically, lower scales primarily capture global semantics, while higher scales focus on fine-grained details. Employing a shared architecture across scales induces optimization conflicts. Moreover, due to the causal autoregressive

Cyberspace has emerged as a fourth fundamental domain of human existence, yet philosophy lacks an integrated framework for addressing the ontological, epistemological, and ethical transformations it produces. This paper develops such a framework through the concept of cyber, defined here as the existential condition in which human life becomes constitutively entangled with digital technologies such that the virtual and the real cease to function as separable categories. Distinguished from Wiener

Zero-shot Object Goal Navigation (ZSON) with RGB-only perception poses a fundamental challenge for embodied agents, as the absence of explicit depth information introduces severe physical uncertainty and semantic-physical misalignment. Existing approaches either rely on high-level semantic reasoning without geometric grounding or learn end-to-end policies that lack explicit physical constraints, often resulting in semantically plausible but physically unsafe behaviors. In this paper, we propose

To improve safety in Large Language Models (LLMs) we can either perform post-training alignment or exploit refusal directions in the activation space. Both strategies are less feasible in Multimodal LLMs (MLLMs) as they require unsafe multimodal data, harder to collect than their unimodal counterpart. In this work, we relax this constraint and investigate whether textual refusal directions, extracted directly from the LLM backbone, generalize across modalities (i.e., image, video). Preliminary f

Explainable AI (XAI) methods have demonstrated significant success in recent years at identifying relevant features in input data that drive deep learning model decisions, enhancing interpretability for users. However, the potential of XAI beyond providing model transparency has remained largely unexplored in adjacent machine learning domains. In this paper, we show for the first time how XAI can be utilized in the context of federated learning. Specifically, while federated learning enables col

Engineering specifications such as interlocks, alarm rationalization tables, and cause-and-effect (C&E) matrices remain central to process control and safety, yet their creation is still predominantly manual, document-driven, and prone to inconsistency. This paper presents a semantic-AI framework that automates the generation of C&E logic by combining a knowledge graph (KG) with a constrained large language model (LLM) layer. The KG builds on an established modular alignment ontology to represen

Radar sensors provide reliable perception under adverse weather and lighting conditions, but their sparse, noisy, and weakly semantic measurements make dense semantic segmentation challenging. Most existing radar segmentation methods rely on grid-based encodings and pairwise interactions, which struggle to capture the higher-order relational structure formed by multiple radar returns from the same physical object. We introduce a unified higher-order structural alignment framework for multi-view

Emergent misalignment (EM) is a recently discovered phenomenon in LLMs where fine-tuning on a narrow misaligned task, such as writing insecure code, leads to broadly misaligned behaviour on unrelated prompts. Previous work has noted that the severity of EM is highly sensitive to training choices; however, we still lack a systematic characterisation of this sensitivity. We perform a sweep over several Qwen3 models, optimisers, datasets, and batch sizes, and find that the choice of optimiser has t

Flaw reporting for deployed AI systems is fundamental to identifying system failures and improving AI safety. Yet the AI reporting ecosystem is fragmented: researchers who identify flaws often do not know what or where to report, and groups who receive reports rarely share them with other relevant stakeholders. As a result, good-faith reporters duplicate effort by submitting many different forms, and recipients lack standardized, triage-ready information. We audit 12 reporting systems published

The Self-Improving Alignment (SAIL) algorithm addresses distribution shift by reducing a bilevel formulation of the problem to an efficient, single-level method. Empirically, SAIL has demonstrated strong performance on this task. However, a formal analysis of its convergence properties has been lacking. We identify a key theoretical challenge: the standard SAIL objective function is not guaranteed to be strongly concave due to unfavorable properties of its Hessian. To address this limitation, we

Artificial intelligence is transforming our capability to solve biological challenges. In dimensionality bottleneck regimes exacerbated by high-dimensional biological data, neural networks force distinct concepts into the lower dimensions known as superposition. Although this superposition is widely known to hinder interpretability, its impact on corrupting the geometry of latent spaces remains critically overlooked. Here, we utilized sparse autoencoders (SAEs) trained on over 100,000 multiplexe