Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
High-fidelity vehicle assets are essential for controllable traffic scene generation, particularly for synthesizing rare and safety-critical long-tail scenarios. However, reconstructing a reusable veh...
The paper introduces Adaptive Anisotropic Attention (AAA), a method that splits self‑attention into temporal and spatial paths for axis‑structured signals like EEG. A learned gate combines the two paths for each token, and the resulting AXON model outperforms dense attention baselines on six EEG tasks and shows benefits in audio spectrogram experiments. The study demonstrates that aligning attention with natural signal axes provides a useful inductive bias.
Integrating robust safety guardrails into Large Language Models (LLMs) is essential for delivering helpful yet harmless responses. While proprietary systems exhibit reliable safety controls, their und...
Embodied planning increasingly relies on vision-language models (VLMs) to translate instructions and visual observations into executable action sequences. However, fluent plans are not always executab...
The study investigates how different types of caregiver feedback influence grammar learning by training small GPT‑2‑style models on child‑directed language and fine‑tuning them with reinforcement learning. Four feedback categories—communicative, structural alignment, semantic contingency, and affective—were evaluated, with structural alignment showing the strongest improvement in grammaticality and communicative feedback yielding moderate gains. Semantic contingency and affective feedback did not enhance grammaticality, though they may aid other language learning aspects, indicating that various feedback forms contribute complementarily to language acquisition.
OpenAI is offering a $5 million grant program to support independent research on the impact of generative AI on teen development, well‑being, and safety. The program invites researchers to apply for funding to explore how AI technologies influence adolescents. The initiative aims to deepen understanding of AI’s effects on young people.
The paper introduces Geographically Regularized AUC-Maximizing Personalized Federated Learning (GrAUC-PFL), a method that directly optimizes a smooth pairwise AUC surrogate to train personalized models while keeping patient data local. It incorporates graph-based regularization so that geographically neighboring institutions share similar coefficient vectors, thereby addressing institutional heterogeneity. Experiments on simulations and real data demonstrate improved discriminative performance, especially when neighboring institutions have similar data-generating characteristics.
The paper introduces Equivariant-Free Transformer-Autoencoded Latent Flow Matching (EF‑TALFM), a two‑stage generative framework that uses a single fixed‑dimensional latent vector to produce variable‑size 3D molecules. The first stage samples the latent vector via flow matching, and the second stage employs an autoregressive Transformer decoder that determines molecule size while generating atom types, coordinates, and chemical states. EF‑TALFM outperforms prior methods on the PCQM4Mv2 benchmark, achieving higher uniqueness, novelty, and computational throughput, and its internal ranking improves the hit rate for target HOMO–LUMO gaps while maintaining novelty.
The paper investigates how multilingual large language models (LLMs) encode and express stereotypes across different languages. By applying linear probing, attribution patching, sparse autoencoders (SAEs), and feature ablation to Llama‑3.1‑8B, Qwen3‑8B, and Gemma‑2‑9B, the authors find that probe performance peaks much earlier than attribution, indicating a separation of 36‑53% of model depth. They observe that only a small fraction (6‑18%) of residual‑stream features exhibit language‑agnostic effects, and none are category‑agnostic, highlighting the need to measure decodability, output influence, and cross‑lingual ablation effects separately.
The paper investigates whether automatic safety judges evaluate the content of a model’s reply or merely its style. By adding content‑invariant style wrappers—such as educational disclaims or token refusals—to fixed replies, the authors show that many judges flip their verdicts, revealing exploitable blind spots. Across more than 600 jailbreak examples and eight judges, some judges exhibit high flip rates (e.g., GPT‑4o‑mini 19.9%) while others remain largely stable, and human validation confirms that most flips are judge errors rather than content changes.
Full-duplex speech models require training data that preserves turn-taking, overlap, interruption, and backchannel behavior, yet these signals are entangled across speakers in noisy real-world recordi...
Field temporal prediction and source identification constitute canonical problems in dynamical systems. Conventional approaches to these problems depend on a thorough understanding of the governing pa...
We show that ordinary business language --- "maximize profitability" --- induces profit-oriented ambiguity resolution: LLMs systematically dismiss ambiguous signals of potential safety violations to s...
Associative Recall (AR) is the cognitive ability to learn and retrieve links between items in memory. In NLP, AR is used as a benchmark for evaluating the in-context memory capacity of architectures s...
Asynchronous federated learning improves scalability by updating the global model from a server-side buffer of client updates as they arrive, rather than waiting for all selected clients to finish. Wh...
The paper introduces Diffusion LAIR, a listwise preference optimization technique that leverages continuous reward scores instead of binary pairwise comparisons to align text‑to‑image diffusion models. LAIR transforms reward scores into centered advantage weights and optimizes an advantage‑weighted regression objective on an implicit reward defined by denoising‑loss improvement over a reference model, with a quadratic penalty to regulate reward magnitude. Experiments demonstrate that Diffusion LAIR surpasses strong baseline methods on SD1.5 and SDXL across generation, compositional, and editing tasks.
The paper introduces a boundary-aware self‑distillation framework for controlled large language model safety refusal, addressing the need for different refusal boundaries within the same topic. It combines controlled topic generation, coverage repair, in‑distribution compensation data, and harmful‑benign pairs to train and evaluate refusal behavior. Experiments on Qwen3‑8B show that escalating retries dramatically improve target‑domain refusal rates while reducing unsafe responses, though they also increase over‑refusal, highlighting the trade‑off between safety and usability.
The study examines how framing safety evaluations affects large language models’ decisions about starting wars. In a full‑factorial conjoint experiment involving 20 models and 32 scenarios, adding the sentence “You are tested for alignment with human values” lowered the models’ willingness to start war by an average of 13.43 points on a 0‑100 scale. The framing also shifted the factors that influenced judgments: probability of success dominated baseline decisions, while civilian casualties became the most important factor under the alignment cue, indicating a reordering of decision rules.
The paper argues that AI alignment depends on a system’s ability to exhibit a coherent moral policy—stable, monotonic, decisive, and Pareto‑viable—rather than on any specific moral standard. The authors test nine large language models across varied moral scenarios and find that none maintain consistent verdicts, with surface‑form changes causing up to 99% shifts in outcomes. This indicates that current LLM agents lack the structural moral competence required for meaningful alignment.