Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
Vision-language-model-based embodied agents can complete instructed tasks but often violate safety constraints in the process, a problem recently framed as interactive safety. Training such agents to act safely is difficult, since safety and task success are distinct objectives, and safety arises only at a small number of safety-critical steps within a trajectory.
Reference-based diffusion stylization requires separating target geometry from transferable appearance. Existing tuning-based methods often rely on aligned content-style-target triplets or auxiliary visual encoders, which increases data cost and can transfer unintended scene structure from the style reference.
Long-term multi-agent systems continuously accumulate the memories produced by different agents. Existing memory methods typically treat retrieved memories as independent evidence and combine them through voting or weighting.
Large language models (LLMs) compression reduces deployment costs, but standard aggregate metrics like perplexity and accuracy often mask underlying behavioral shifts. In this work, we systematically evaluate 3 LLMs across 11 compression methods to investigate the effects of compression on knowledge retention, model confidence, and social bias.
The paper studies weak-to-strong generalization (W2SG), where a student model trained on a weaker teacher’s labels surpasses the teacher on the target task. Using a Bregman divergence bias‑variance decomposition, it shows that the student‑teacher risk gap depends on their expected misfit, without requiring convexity of the student hypothesis class. For squared loss, a sufficient condition is that the student converges to the teacher’s posterior mean, achievable by enlarging the student; for cross‑entropy loss, reducing the student’s predictive entropy and using reverse cross‑entropy can promote W2SG, which is empirically validated.
FedLNS is a server‑side framework that screens federated learning updates by representing each client’s contribution through changes in trainable normalization‑layer parameters, creating lightweight signatures that can be compared against a history‑aware cross‑client reference. The method requires no extra client‑to‑server communication, raw data, or labeled attack examples, and after screening, the remaining full‑model updates are aggregated with standard federated learning rules. Experiments on GPT‑style, BERT‑style, and LLaMA‑style models with 200 clients demonstrate that FedLNS achieves lower test perplexity than six baselines even when 40% of the population performs target manipulation under both IID and non‑IID data partitions.
The paper investigates how Vision Language Models (VLMs) can be fooled by small, human‑imperceptible changes to images. It introduces a gradient‑based attack that targets only the vision encoder, reducing computational cost while still effectively disrupting both untargeted and targeted multimodal alignment. Experiments on open‑source VLMs such as Qwen2.5‑VL, Granite‑Vision, FastVLM, and Phi‑3.5‑Vision demonstrate that these perturbations can significantly alter the models’ textual outputs.
The paper presents a super‑resolution generative adversarial network (SRGAN) that boosts electron backscatter diffraction (EBSD) throughput for lithium‑ion battery electrode materials. Trained on LiNixMnyCozO2 cathode data, the SRGAN outperforms classical interpolation across 2×–12× upscaling, especially preserving small grains and realistic boundaries. A 5× upscaling yields a 25× speed‑up or larger field of view with acceptable errors in grain size and shape metrics.
GrabVG is a visual grounding framework for UAV imagery that tackles the challenges of small, densely packed, and visually similar objects. It splits the task into preattentive hypothesis search and graph‑attentive feature binding, using distillation‑guided proposals and a sparse graph to capture intra‑ and inter‑instance relationships. Experiments on AerialVG and AerialSense show that GrabVG achieves higher accuracy and speed, outperforming baselines by significant margins.
The paper introduces SIMPLE, a prior‑fitted multi‑view in‑context learner that learns a reusable, task‑conditioned inference procedure instead of a fixed fusion function. By generating synthetic task priors in embedding space, SIMPLE can handle diverse view configurations, class structures, and missingness patterns. Experiments on multi‑view and multi‑omics benchmarks show that a frozen SIMPLE model performs competitively, and lightweight adapter calibration further improves performance across most datasets.
The OpenAI Privacy Filter is a 1.5‑billion‑parameter, bidirectional token‑classification model that detects and redacts personally identifiable information and secrets in unstructured text. It is built from an autoregressive checkpoint, converted into a banded‑attention classifier, and uses a constrained Viterbi decoder to produce coherent spans across eight privacy categories in a single forward pass. The model supports configurable precision‑recall tradeoffs, a 128,000‑token context window, and is designed for efficient local deployment and domain‑specific fine‑tuning as a data‑minimization component within layered privacy workflows.
The paper investigates how the dynamical regime of recurrent-depth reasoners—whether they settle, drift, or remain marginal—affects the reliability of test‑time depth. It establishes a depth‑safety condition based on per‑step displacement relative to the decoder margin, showing that operators in a settling regime can safely increase depth without degrading performance and can even improve accuracy on harder unseen tasks such as Sudoku. The authors provide empirical evidence from algorithmic tasks trained on limited data, demonstrate the impact of a terminal fixed‑point objective on depth behavior, and offer operational criteria to identify useful test‑time depth while cataloguing failure modes.
The paper introduces MIFR, a modality‑invariant and fair representation framework for skin disease classification that jointly processes clinical photographs and dermoscopic images using ViT‑based encoders. It employs a five‑component multi‑objective loss to balance classification accuracy, fairness across skin tones, class alignment, and modality invariance. Experiments on paired and external datasets demonstrate competitive predictive performance and fairness, with t‑SNE visualizations confirming alignment of embeddings from different modalities.
The paper investigates on‑policy self‑distillation (OPSD), where a student model learns from its own outputs using token‑level supervision conditioned on privileged reference information. Experiments with Qwen3 models on science and mathematics datasets show that the correct reference does not consistently improve performance; students can improve without it, and solutions from other problems sometimes outperform the correct reference. The study finds that student predictions align more closely with the base model’s reasoning than with the reference supervision, and that alignment alone does not reliably predict performance gains.
DeepTCM1.0 is a multi‑expert AI agent built on the DeepSeek V3.2 large language model, designed to integrate classical traditional Chinese medicine (TCM) theory with modern life sciences for mechanistic analysis of TCM compound formulas. The framework employs a three‑tier collaborative architecture and a three‑round iterative quality‑control workflow, simulating 11 interdisciplinary intelligent agents. It was validated on the Guizhi Decoction, with performance evaluated through double‑blind five‑dimensional scoring, ICC reliability testing, Mann‑Whitney U tests, and effect size analysis across 100 independent assessments by four large language models.
The study examines whether large language models (LLMs) discriminate against candidates based on institutional prestige and geographic location. Three factorial experiments (4,320 API calls across four LLMs and five domains) reveal a significant institution‑tier gradient (+0.297 points on a 10‑point scale), a stronger prestige effect than country‑of‑origin effect, and a dominant journal prestige effect (Nature vs. a peripheral journal). The authors also identify a ‘rescue effect’ where publishing in Nature mitigates low institutional prestige, and quantify bias using the Neutrosophic Bias Index, highlighting evaluation inconsistency for low‑prestige profiles.
The paper investigates the vulnerability of Gaussian‑noised text embeddings to inversion attacks. It identifies a "Double Noise Trap" that hampers standard generative methods and introduces DAEI, a denoising‑aware pipeline that significantly outperforms existing baselines in reconstructing original text. Experiments show DAEI improves BLEU by 154% and token‑level metrics by 32–60%.
The article reviews the growing use of Large Language Models (LLMs) for generating Verilog code, a key hardware description language in electronic design automation. It surveys 102 papers, covering conferences, journals, and preprints, and addresses four research questions about LLM selection, datasets, techniques, and alignment strategies. The review identifies current limitations and proposes a roadmap for future research in LLM-assisted hardware design.
SIGMA is a reinforcement‑learning framework for traffic signal control that incorporates a large language model to adaptively tune multiple objectives based on natural‑language emergency commands. It uses rotational data augmentation to learn orientation‑invariant policies and an offline‑to‑online training pipeline to ensure stable deployment. Experiments in SUMO on four Kolkata intersections show that SIGMA reduces waiting times, queue lengths, and improves throughput compared to fixed‑time, actuated, and DQN baselines, with ablation studies confirming robustness to component failures and geometric rotations.
The paper introduces Jacobian-Guided Anisotropic Noise Reshaping, a method that improves data utility under Local Differential Privacy by selectively reducing noise in task-relevant subspaces of data representations. It uses the Jacobian of a public downstream model to identify critical directions and reshapes isotropic LDP noise into an anisotropic distribution, preserving privacy while enhancing performance. Experiments on CIFAR-10-C show significant accuracy gains, especially for PrivUnit variants at ε=7.5.