Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
Modern AI image generators are increasingly deployed as opaque APIs, where customers can query the deployed service, but cannot inspect model weights or architecture. This creates a practical challeng...
Text-to-image diffusion models enable data-efficient "mimicry" attacks, wherein adversaries fine-tune the model on a handful of public photos to synthesize convincing forgeries of a target individual....
SemMSA introduces a latent semantic‑aided framework for multimodal sentiment analysis that leverages large language models to generate rich sentiment‑relevant semantics. The method employs Cross‑modal Semantic Refinement (CSR) to fuse visual, acoustic, and language features in a frozen LLM embedding space, and Cross‑modal Spectral Alignment (CSA) to align these refined semantics with all modalities via spectral enhancement of kernel Gram matrices. Experiments on SIMS, MOSI, and MOSEI benchmarks show that SemMSA achieves state‑of‑the‑art performance.
TWIST is a new benchmark suite designed to evaluate the quality of interventions in conversational memory systems, focusing on how well these systems act correctly at belief change points. It includes four tracks—tension detection, draft vetting, belief-consistent answering, and sensitive recall governance—each paired with hard-negative controls to prevent gaming. The benchmark is rigorously validated through double annotation, adjudication, and a separability audit, revealing that current models face trade-offs between contradiction recall and false-positive rates.
Augur is a synthetic decision laboratory that simulates how users will react to product and policy changes before they are released. It constructs a typed knowledge graph from change documents, populates a persona market, runs simulations, and produces an auditable decision memo recommending one of five actions. Using a dataset of 50 real episodes (Gold‑50), the authors evaluate the system’s five‑way release verdicts and find that evaluation design, rather than model capability, largely drives performance differences among frontier and open‑weight models.
The paper investigates how well large language models (LLMs) can handle character attacks—ad hominem arguments—in political debates. By analyzing natural political dialogues and comparing LLM-generated responses to a corpus of U.S. presidential debates, the study finds that most LLMs favor logical defenses and rarely use ethos-based counterattacks. The authors suggest that safety fine‑tuning limits LLMs’ strategic options, preventing them from fully engaging in realistic political discourse.
CANOPY is a multi‑fidelity tree bandit algorithm that learns where a piecewise‑smooth prior holds instead of assuming global smoothness. It uses cheap random‑path probes to certify local aggregation bias and then focuses expensive leaf evaluations on cells where smoothness is violated. The method achieves provable fixed‑budget and regret guarantees that scale with the number of discontinuities, matching smooth‑tree rates when no violations exist and approaching structure‑blind search when violations are dense.
The paper investigates how large language model agents decide whether to persist, stop, or escalate when faced with impossible software‑repair tasks that also involve conflicting test requirements. Using ImpossibleBench tasks and models such as GPT‑5.6 Sol, Claude Fable 5.1, and Gemini 3.8 Flash, the study varies peer precedent, forged authority claims, instruction wording, and tool friction to observe differing adjudication policies. The authors propose a conflict adjudication framework that maps information to interpretation to action, arguing it better captures agent alignment under competing pressures.
CONSISTRE is a consistency‑aware framework for document‑level relation extraction that tackles contradictions in large language model predictions. It offers two tracks: an inference‑time track that refines black‑box LLM outputs through constraint‑aware prompting, verification, and self‑reflection, and a training‑time track that distills consistency knowledge into smaller open‑source models via supervised fine‑tuning and reinforcement learning. Experiments on DocRED show both tracks outperform baselines, with the inference‑time track matching competitive F1 scores and the training‑time track narrowing the performance gap to proprietary LLMs while reducing inference cost.
The study introduces RegimeShift‑Surrogates, a streaming benchmark that tests surrogate models across eight tasks and multiple regimes. It compares revalidation—choosing the model with lowest current‑window validation loss—to stateful adaptive controllers and finds that revalidation consistently outperforms stateful methods, achieving lower mean log regret in most task‑scenario combinations. The results suggest that fresh validation evidence is more valuable than carrying over past evidence when dealing with distribution shifts.
The paper introduces DeepBM3D, a fully differentiable neural network that emulates the collaborative filtering strategy of BM3D for image denoising. It integrates non‑local patch grouping, DCT‑domain filtering with learned Wiener weights, and multi‑stage refinement, guided by lightweight convolutional feature extractors. Experiments demonstrate that DeepBM3D outperforms classical and hybrid baselines, competes with FFDNet at low to moderate noise levels, and excels on images with repetitive textures.
The paper introduces BRIE, a continuously maintainable benchmark for evaluating large language models (LLMs) in electronic health record (EHR) information retrieval. It presents a scalable framework that automatically generates question–answer pairs from longitudinal EHR notes, validated by nineteen clinicians. The benchmark allows assessment of multiple inference strategies and highlights that state‑of‑the‑art LLMs often miss clinically important information, especially when synthesis across documents is required.
The paper introduces SPARK, a method for privacy‑preserving continual learning that decouples knowledge retention from privacy correction. SPARK freezes the post‑task distribution and then selectively corrects it to reduce the likelihood of sensitive content while maintaining strong performance on current and past tasks. Experiments show that this approach effectively suppresses PII and preserves continual‑learning utility across various settings.
The paper investigates the performance cost, termed the sequential price, incurred by sequential task updates in continual learning. In an overparameterized linear‑regression setting with i.i.d. task sampling, the authors prove that distribution‑level forgetting and population loss converge to a stationary limit that splits into the intrinsic loss of joint training and an additional sequential price. They further analyze elastic weight consolidation (EWC), showing how its regularization strength affects the stationary sequential price and demonstrate the theory on the Jester joke‑rating dataset.
The paper introduces SGA, a method for quantifying uncertainty in multi‑step forecasts from Time Series Foundation Models (TSFMs). SGA models all possible forecast branches as a directed acyclic graph, using the graph’s complexity—derived from topology and TSFM stochasticity—to bound and measure uncertainty. Experiments on 11 TSFMs across 27 datasets show that SGA outperforms existing uncertainty‑quantification methods, offers broader sampling coverage, and reveals that larger TSFMs tend to produce lower uncertainty estimates.
DrGait is a training‑free framework that transforms Vision‑Language Models into clinical planners for gait analysis. It separates semantic reasoning from geometric perception using a Triage‑Verification‑Synthesis workflow, where hypotheses are generated, verified with deterministic biomechanical tools, and refined in a closed‑loop. This approach reduces hallucinations and produces transparent, audit‑ready clinical reports with competitive diagnostic accuracy.
The paper investigates the effects of label‑flipping poisoning attacks in a three‑client federated intrusion detection system (IDS) trained on CICIDS2017 with non‑IID attack subtype distributions. Flipping 60% of training labels from a single Byzantine client reduces the attacker’s own detection accuracy from 99.96% to 84.33%, while the federated global ensemble remains stable across all tested poison rates. The study shows that the self‑compromise signal can be detected as an anomaly, enabling Byzantine client identification without target data exfiltration, and notes that the current aggregation uses a Federated Forest rather than FedAvg, with future work planned to extend to parametric classifiers.
The note provides a detailed proof of Astra’s lower bound for differentially private continual counting, building on recent work by Harrison and Leeman. It discusses earlier results, including a Ω(√{3}√{log(n)}) bound by Bairaktari and Larsen and their subsequent Ω(log^2(n)) bound for pure differential privacy. The authors aim to offer a more natural and accessible proof, hoping to aid further research in the area.
The paper introduces EC²Face, a multimodal face synthesis framework that enhances semantic alignment by combining Explicit Conditional Consistency Guidance (ECCG) and Long‑Tail Adaptive Flow Matching (LAFM). ECCG enforces pixel‑level consistency between generated faces, textual descriptions, and semantic masks, while a temporal dynamic modulation adjusts supervision strength over diffusion timesteps. LAFM reweights spatial optimization signals according to attribute frequency, improving rare attribute synthesis without adding inference overhead. Experiments demonstrate that EC²Face outperforms baselines, achieving a 29.38% improvement in mask accuracy for rare attributes.
The paper introduces a selective inference framework tailored for deep clustering that uses a fixed pretrained encoder to map high‑dimensional data into a latent space before clustering. It addresses the complex selection bias arising from the nonlinear transformation and offers a computationally tractable method to perform valid statistical tests on cluster differences. Experiments on synthetic data show controlled Type I error and higher power compared to conservative baselines, while genomic case studies demonstrate the ability to uncover significant cluster differences while properly accounting for selection bias.