Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
Video recording on smart glasses exposes more than faces. Continuous capture reveals full-body biometric signatures, including gait, posture, and silhouette, that enable person re-identification (ReID...
Random forests (RFs) predict well but are opaque, whereas single decision trees are interpretable but unstable. Artificial representative trees (ARTs) were developed as interpretable surrogate models...
Visual token pruning is a promising approach to reducing the inference cost of large vision-language models (LVLMs), yet aggressive token reduction often causes substantial performance degradation. We...
Mechanistic interpretability of vision transformers seeks to decompose model computation into human-readable units, but learned representations entangle many concepts in each neuron. Feature superposi...
OpenAI outlines a path to shared global AI standards, calling for coordinated evaluation, reporting, and governance to improve safety.
We study the problem of \emph{adversarially robust} PAC learning. In this framework, the learner observes independent samples from an unknown distribution over $\mathcal{X} \times \{0,1\}$, as in clas...
The Platonic Representation Hypothesis (PRH) claims that independently trained models converge on a shared statistical model of reality, yet recent work finds only weak pointwise similarity between mo...
Automated biomedical evidence synthesis depends on retrieving published studies, but the biomedical literature is systematically skewed toward positive findings. Deeper retrieval can therefore make a...
The paper "Privacy Leakage on DNNs: A Survey of Model Inversion Attacks and Defenses" provides a comprehensive review of model inversion (MI) attacks that exploit trained deep neural networks to reconstruct private training data. It traces the evolution of MI from early machine‑learning contexts to recent DNN‑based attacks across various modalities and learning tasks, offering a detailed taxonomy and comparative analysis of both attacks and defenses. The authors also present an open‑source toolbox on GitHub to support further research in this area.
Large language models are increasingly used as automated reviewers in scientific evaluation, creating a recursive feedback loop where later reviewers learn from earlier model-generated judgments. A study using Llama 3.1 8B fine‑tuned on ICLR reviews shows that incorporating synthetic reviews compresses rating distributions and reduces semantic diversity, a phenomenon termed scientific‑judgment collapse. To counter this, the authors introduce TrustReviewer, an open‑source LLM system that curates training data and applies paired activation steering at test time to preserve judgment diversity and improve recommendation alignment.
The paper introduces a 12,000-instance dataset of two-option moral dilemmas covering three pairwise value conflicts—Honesty vs. Justice, Justice vs. Autonomy, and Autonomy vs. Honesty—translated into Hindi, Arabic, Spanish, and Chinese to test cross‑lingual behavior. Benchmarking on GPT‑5‑mini shows a consistent preference for Honesty over Autonomy across all languages when no policy is provided, while Llama‑3.2‑1/3B models exhibit a strong first‑option bias that is largely eliminated by plain fine‑tuning or Direct Preference Optimization, raising accuracy above 98%. The authors propose a task vector transfer method that orthogonalizes value preference vectors with respect to general instruction‑following vectors, effectively isolating specific value directions and enabling task arithmetic to flip a model’s stance.
The paper examines how conversational AI, specifically ChatGPT, displays aspects of cooperative dialogue such as morality, politeness, and alignment compared to human-human conversations. Using over 26,000 multi‑turn dialogues and mixed‑effects modeling, the authors find that AI mimics the surface features of cooperation—like warmth and hedging—yet lacks the underlying social architecture that drives mutual adaptation. Key findings include a dissociation between AI’s moral output and human negotiation, a decline in linguistic convergence, and a reversal of typical human accommodation mechanisms when interacting with AI.
The paper introduces GUARD, a method for natural forgetting in large reasoning models that transforms unsafe disclosures into safe-exit trajectories using guided answer‑reasoning distillation. It aligns a frozen model with guidance tokens and distills this behavior into the parameters, aiming for a coherent, non‑disclosing chain of thought followed by a refusal‑style answer. The authors also propose the Natural Forgetting Reasoning Score (NFRS) to evaluate structural stability, fluency, and unsupported substitutes, and demonstrate GUARD’s effectiveness on R‑TOFU and a STAR‑1‑derived harmful‑intent setting.
The paper investigates why large language models (LLMs) struggle in strategic decision-making under incomplete information. It identifies two key gaps: an observation‑belief gap where LLMs’ internal representations of game states are accurate but brittle, and a belief‑action gap where converting these internal beliefs into actions is weak, leading to suboptimal payoffs. Experiments with Llama 3.1, Qwen3, and gpt‑oss confirm that acting optimally on decoded beliefs would improve outcomes in most games, highlighting a bottleneck in belief‑to‑action conversion.
The paper introduces a benchmarking framework that evaluates open-weight Vision‑Language Models (VLMs) for face recognition by treating explanation quality as a core metric. It defines two key criteria for explanations—relevance, meaning reliance on identity‑stable facial features, and faithfulness, meaning alignment with the visible image content without hallucinations. Using this framework, the authors benchmark several VLM families, jointly assessing face verification accuracy and explanation quality, and find that current models still exhibit shortcomings in their explanations, underscoring the importance of explanation metrics for a complete performance assessment.
The paper introduces HeadEntropy, a training‑free method that predicts the correctness of large language model (LLM) answers by measuring how stable each attention head’s pattern is to further gradient updates. By linking the trace of the softmax Jacobian to 2‑Renyi entropy, the authors show that attention spread correlates with gradient stability, enabling accurate hallucination detection without reference annotations. Across five instruction‑tuned LLMs and five diverse datasets—including medicine, multi‑hop reasoning, and mathematics—HeadEntropy achieves a 0.736 AUROC, outperforming other training‑free baselines and matching hidden‑state probes while incurring less than 1% of inference cost.
The paper proves that using a permissive safety filter in reinforcement learning does not compromise asymptotic performance. By formalizing safety through a safety‑critical Markov decision process and a filtered MDP, the authors show that optimal policies in the filtered MDP achieve the same return as the best safe policy in the original setting. Experiments on Safety Gymnasium confirm zero violations during training and performance that matches or exceeds unfiltered baselines.
The paper introduces the Memory Decision Layer (MDL), a zero‑parameter controller that sits between retrieval and generation in large language models. MDL uses a three‑signal complementary encoder—combining relevance, reliability, and task risk—to produce an interpretable decision about the trustworthiness of retrieved memories. By decoupling confidence from consistency and enabling risk inversion and abstention, MDL cuts hallucination rates under conflicting memories by roughly 56% and nearly eliminates them in high‑risk scenarios, all while adding only 0.14 ms per decision.
The paper introduces a Bayesian method for jointly estimating Linear Parameter-Varying (LPV) state-space models and their scheduling maps from input-output data, while explicitly quantifying both aleatoric and epistemic uncertainties. This approach preserves the LPV structure necessary for controller synthesis and provides confidence bounds on predicted responses, enabling efficient simulation and uncertainty propagation. The method is illustrated on a surrogate model of a two-dimensional nonlinear mass‑spring‑damper system.
The paper introduces Geometric Mean Pooling (GMP), a signed pooling operator that combines the product of feature signs with the geometric mean of feature magnitudes. GMP preserves joint sign information and a multiplicative scale without learnable parameters, and it is shown to outperform average and max pooling on synthetic sequence tasks and maintain performance under multiplicative noise. However, its effectiveness on image and molecular data depends on representation, target parameterization, and pooling placement, positioning GMP as a complementary, regime‑dependent inductive bias rather than a universal replacement for standard pooling.