Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
Maximum clique and maximum consensus are widely used formulations for robust point cloud registration in robotics. Maximum consensus seeks one transformation that aligns as many matches as possible, w...
Despite advancements in automatic speech recognition systems (ASR), current systems exhibit speaker-attribute bias, i.e., their speech recognition performance is biased against certain speakers based...
Referring Expression Comprehension (REC) localizes the image region described by a natural language expression, underpinning applications such as autonomous navigation, human-robot interaction, and me...
Recent efforts have demonstrated the strong capabilities of Multimodal Large Reasoning Models (MLRMs) in geographic reasoning. By jointly analyzing subtle visual cues, these models can infer the locat...
Despite rapid advances in large language models (LLMs), their deployment in the sports and health domain remains high-risk due to the absence of domain-specific evaluation benchmarks. Existing general...
The article "AI Harms and Their Technical Causes" discusses how specific technical decisions in AI development can result in safety failures. It examines the underlying causes of these harms and offers guidance on alternative approaches to mitigate risks. The piece highlights the importance of thoughtful design choices in preventing unintended negative outcomes from AI systems.
Out-of-distribution detection in continual learning (OOD-CL) is highly challenging, as it requires (i) learning a sequence of tasks without catastrophic forgetting, and (ii) identifying unknown sample...
Histopathology is central to cancer diagnosis, but morphology in Hematoxylin and Eosin (HE)-stained images does not directly reveal the molecular states that characterize disease heterogeneity and inf...
Language models increasingly use skills that provide reusable instructions and procedures for task execution. After a relevant skill has been selected, the model may need to apply its governing proced...
RealCLI is a new benchmark for evaluating human‑agent collaboration in command‑line interface (CLI) tasks. It unifies task execution and collaboration quality assessment through a five‑dimensional framework with 22 sub‑dimensions and 51 interaction templates. The benchmark offers two modes—REALCLI‑EXEC with full specifications and REALCLI‑COLLAB with a scripted user providing progressive feedback—and reveals significant capability gaps in current agents, especially in steerability, trust calibration, and safety management.
The paper investigates how literal interpretations of Bengali idioms influence the behavior of decision models that output probabilities over supplied options without generating text. By inserting an idiom’s literal gloss into the option list, the authors show that all three tested models—an encoder scorer, a decoder scorer, and a closed API—consistently shift their predictions toward the literal option, causing 65% to 96% of remaining errors to be attributed to it. The effect is markedly stronger in Bengali than in English, with the commercial model’s apparent robustness being a consequence of the test language rather than intrinsic capability, and the models’ calibration deteriorates across all arms, making confidence-based gating unreliable.
The paper introduces "watermarkability," a metric that quantifies how compatible a language model is with a logit‑based watermark that boosts the probability of a specified token set during decoding. By measuring the expected minimum KL divergence needed to achieve a target probability mass, the authors can assess and compare models and watermark rules. They propose a two‑stage approach—first training a low‑rank adapter conditioned on the target set, then applying a minimal uniform logit bias—to reduce this cost, achieving significant perplexity reductions across multiple watermark schedules on OPT‑125M and Qwen3‑8B.
The paper investigates how large language models decide whether to call external tools before the call is made, a step that has been largely overlooked in tool‑safety research. Using factorial interventions on input requests, the authors find that neutralizing the requested operation significantly reduces harmful admissions, and that a linear direction derived from this contrast predicts admission across most models. They also show that training models to refuse in plain text can lower unsafe admissions, indicating a partial link between written refusal and tool‑admission decisions.
The paper surveys how annotator disagreement in NLP is increasingly treated as a meaningful signal rather than noise. It offers a domain‑agnostic taxonomy of disagreement sources, a unified framework for modeling approaches, and a review of evaluation metrics for predictive performance and annotator behavior. The authors highlight a shift toward explicit disagreement modeling, structured annotator relationships, and outline future challenges such as integrating multiple variation sources and developing interpretability frameworks.
DICL-Qwen3VL is a decoupled, interpretable vision‑language model that localizes upper‑body clothing and identifies workwear in indoor construction sites. It uses low‑rank adaptation of Qwen3‑VL, separates visual and reasoning confidence estimates, and employs token entropy and image perturbation sensitivity to supervise visual confidence. Experiments on the Indoor Construction Site Workwear Dataset and the web‑collected Indoor Construction Workwear Dataset achieve mAP@50 scores of 0.730 and 0.753, outperforming baseline VLMs and providing explicit evidence and confidence scores for industrial safety monitoring.
The paper introduces Evidence-Conditioned Dependency Distillation (ECDD), a method that trains compact medical vision‑language models to rely on visual evidence by pairing each training image with a control matched to the evidence level and distilling the teacher’s drop in answer margin across the pair. ECDD preserves task accuracy while making answer support more sensitive to evidence removal, outperforming standard knowledge distillation and alignment baselines in three medical VQA settings. The approach applies to various evidence types, such as clinical regions, surgical instruments, or whole‑image content.
IMPLICIT-Bench is a benchmark of 1,831 prompt triplets that tests whether text‑to‑image models can override stereotypical defaults when given explicit requests. Each triplet contains a neutral prompt, a stereotype prompt, and an anti‑stereotype prompt, allowing measurement of a model’s default bias and its ability to follow contradictory instructions. Experiments on Qwen‑Image and Stable Diffusion 3 show that models often default to stereotypes, succeed less often on anti‑stereotype requests, and that prompt rewriting or embedding steering can modestly reduce bias when the relevant stereotype is known.
The paper introduces StratFL, a strategy for federated learning that counters covert model poisoning attacks by modeling the interaction between the server and attacker as an imperfect repeated game. It proposes new belief‑update algorithms for both parties and an adversarial search method to detect and mitigate intermittent poisoned updates. Experiments on MNIST and CIFAR‑10 show that StratFL improves accuracy by about 10% over existing defenses while demonstrating convergence of the belief updates across various security scenarios.
The paper introduces latent‑space language steering, a lightweight inference‑time technique that uses Principal Component Analysis on parallel translations to identify language directions in token embeddings. By steering embeddings along these axes, the method reduces unintended code‑switching in multilingual large language models, achieving up to 55% lower next‑token distributional divergence and a 63–99% drop in the Code‑Switching Index while maintaining semantic integrity and incurring negligible computational cost. The authors also demonstrate that a single principal component can classify language with 95–99% accuracy and provide an analysis of how language representations evolve across model layers.
ODSS-Net introduces a framework for multi‑modal object re‑identification that separates shared structural information from modality‑specific cues. It employs a Shared Structural Feature Construction module to extract modality‑shared structure, an Orthogonal Specific Feature Learning module to isolate identity‑relevant local features, and a Sample Adaptive Mixture of Experts module to fuse these features based on shared structure. Experiments on three benchmarks confirm the effectiveness of this orthogonally decoupled representation learning approach.