Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
Despite recent advancements in 3D generation, models often struggle to produce assets with high visual fidelity. To bridge this gap, we propose OREO, an alignment framework that enhances the realism o...
Large Language Models (LLMs) consume and produce a single sequence of text; hence, if text can be added to the beginning of the LLM's response, i.e., an output prefix, then all subsequent tokens will...
The paper investigates how Large Language Models (LLMs) handle bug fixing compared to human-written patches by analyzing about 3,000 Codeforces submissions. It finds that LLMs often modify more lines than necessary and sometimes produce entirely new solutions, and that they solve more problems correctly when generating solutions from scratch rather than patching existing code. The study highlights implications for AI‑assisted programming tools, suggesting a shift toward incremental problem‑solving strategies.
The paper introduces Domain Recentering with Confidence Calibration (DRC), a training‑free technique that adapts CLIP to unlabeled target images by fitting a Gaussian mixture and applying posterior‑weighted mean subtraction, followed by a log‑prior correction based on confidence‑weighted predictions. DRC improves cross‑domain accuracy, surpassing zero‑shot CLIP by 4.13 and 5.07 points on ViT‑B/16 and ResNet‑50, and maintains gains under ImageNet distribution shifts.
The paper introduces Flow-Based Distribution Matching (FBDM), a non‑adversarial method that learns self‑supervised representations by aligning images to explicit geometric references through spherical conditional velocity regression. By using an ETF‑inspired reference, FBDM allows more reference components than the flow dimension while maintaining geometric separation, and it incorporates an alignment loss to bring augmented views closer together. Experiments on datasets from CIFAR to ImageNet show that FBDM performs nearly as well as adversarial distribution‑matching methods, achieves a 1.48‑ to 1.83‑fold speedup, and offers a theoretical bound on downstream misclassification rates.
The paper introduces a probe‑free method called the Neuron Separability Index (NSI) to assess how individual neurons in large language models distinguish grammatical from ungrammatical sentences using linguistic minimal pairs. Across 68 linguistic paradigms and seven model checkpoints, the study finds that while raw separability for morphology and syntax peaks early, single‑unit selectivity is sparse and weak, with rare strongly selective "grandmother neurons." Moreover, the research shows a dissociation between whole‑vector linear separability, single‑neuron selectivity, and behavioral competence, and demonstrates that targeted ablations can further separate activation selectivity from causal reliance.
The paper investigates whether large language models (LLMs) assess politeness in ways that match human judgments. Using two English datasets—one with continuous ratings and another with three‑way categorical labels—the authors find that LLMs agree more with each other than with humans. They observe that model–human alignment depends on explicit linguistic cues, while misaligned cases often involve rapport‑building strategies. Additionally, models tend to overproduce Neutral labels and underpredict Impolite labels, a pattern that persists even when expert consensus is used as a reference.
The paper introduces Median Temporal Ensembling, a training‑free aggregation method for action‑chunked visuomotor policies that replaces the standard exponential weighted mean with a coordinate‑wise median. This approach remains robust against adversarial corruption, maintaining a high recovery rate even as attack strength increases, and performs at least as well as the mean across numerous configurations while improving in many cases. It also handles non‑adversarial failures such as blank camera frames and shows limited impact on clean data, though it cannot counteract uniform shifts applied to all predictions.
FairTest is a search-based testing framework designed to uncover fairness failures in Multi-Agent Reinforcement Learning (MARL) systems. It guides candidate generation using three fitness functions—measuring observed fairness, predicting fairness from abstract states, and assessing policy decision uncertainty—and prioritizes tests based on predicted fairness and uncertainty. Evaluations on three environments and two MARL algorithms show FairTest detects significantly more fairness failures than three baselines, with a 221% increase in failure count and 23% better coverage on average.
This scoping review examines 421 studies (2015‑2026) on natural language processing applied to student evaluation of teaching comments. It maps the technical evolution from lexicons and classifiers to transformers and large language models, and evaluates four value dimensions. The review identifies a significant gap between actionable outputs (61.3%) and intended‑user evaluation (11.6%), highlighting limited progress in educational value and robustness.
The paper introduces LEAP-CBF, a safety filter that uses Least‑Effort Adversarial Potentials to quantify how much disturbance effort is needed to cause failure in nonlinear dynamical systems. LEAP serves as a control barrier function for the undisturbed system and can be combined with a robust safety filter that tolerates disturbances with bounded cumulative effort. The authors develop a deep reinforcement learning method to construct LEAPs and validate their effectiveness through simulations of multi‑agent systems and hardware experiments on a quadruped and quadrotors.
ChronoSteer is a decoupled agentic framework that bridges large language models and time series foundation models by learning cross‑modal alignment from synthetic paired supervision. It converts textual events into revision instructions that steer a frozen time‑series model, discretizes these instructions into a compact codebook to reduce semantic divergence, and then refines the predictions with a two‑stage training strategy. The authors also release a leakage‑controlled multimodal benchmark and report a 25.8% improvement in zero‑shot prediction accuracy over the unimodal backbone.
The paper presents artifact‑free B‑cos networks for chest X‑ray diagnosis, enhancing the original B‑cos architecture with anti‑aliasing techniques (ASAP and BlurPool) to eliminate aliasing artifacts in explanation maps. These modified networks, “B-cos_ASAP” and “B-cos_BP”, maintain competitive diagnostic performance while providing faithful, class‑specific contribution maps without post‑hoc explanations. Experiments on chest X‑ray datasets confirm that the new models deliver strong predictive accuracy and clear, artifact‑free explanations suitable for clinical use in multi‑class and multi‑label scenarios.
PHONOS is a real‑time streaming module for speaker anonymization that neutralizes accent cues by converting non‑native segmental realizations toward a target accent domain. It uses pre‑generated golden utterances that preserve timbre and rhythm, aligning them with silence‑aware DTW and applying zero‑shot voice conversion to supervise a causal accent translator. The system achieves an 81% reduction in non‑native accent confidence, improves accentedness ratings, reduces speaker linkability in embedding space, and operates with ≤241 ms end‑to‑end latency on a single GPU.
Mizar is a 159.3‑million‑parameter audio‑language model designed for devices with limited memory and computation. It couples a compact CED‑Small audio encoder with SmolLM2‑135M via a frequency‑merging mapper and is trained in three stages—audio‑language alignment, audio‑dependent fine‑tuning, and post‑training—to improve performance on audio‑question tasks. Across five random seeds, Mizar outperforms all other sub‑200M‑parameter ALMs on MMAU, MMAR, and ADQA‑clean, achieving mean accuracies of 52.92%, 42.42%, and 36.02% respectively, while enabling local inference on a single CPU with an average latency of 1.09 seconds for MMAU questions.
LiSeCo is a lightweight, gradient‑free method that controls language generation by directly intervening on the hidden activations of a token in embedding space. It uses control‑theoretic techniques to steer the generation trajectory away from undesired semantic regions and into a predefined allowed region, ensuring fine‑grained attribute control. The approach is computationally efficient, minimally impacts generation time, and is shown to be effective on tasks such as toxicity, sentiment, and bilingual language steering while preserving text quality.
The paper introduces HIVE, a method for detecting hallucinations in diffusion large language models by leveraging hidden trajectory evidence. HIVE compresses hidden states, selects informative step-layer evidence, and uses continuous prefix embeddings to condition a verifier that outputs a hallucination score and diagnostics. Experiments on two diffusion LLMs and three QA benchmarks show that HIVE outperforms eight baselines and improves AUROC and AUPRC by several points over text-only verification.
The study investigates whether a model trained on construction‑sector occupational accident narratives can accurately classify accident‑process roles in other sectors and reporting environments. Using 42,244 factual units from 6,040 construction narratives, the authors compared TF‑IDF, frozen pretrained representations, and task‑adapted pretrained models, achieving up to 85.7% balanced accuracy without retraining. The models performed consistently across metallurgy, chemistry‑plastics, and an independent company corpus, though performance varied more on the latter due to differing reporting practices.
BrainFedFM is a structural brain MRI foundation model that was federatively pretrained on 164,707 3‑D scans from 42 sites using a dual‑priority approach that emphasizes informative anatomical regions locally and prioritizes site contributions globally. The model outperformed seven baseline models—including four centralized foundation models—across 20 downstream tasks (classification, regression, segmentation), achieving a mean rank of 1.68 and a 50% performance gain, especially in classification and regression and among underrepresented populations. These results demonstrate the model’s generalizability and show that federated pretraining can effectively develop neuroimaging foundation models without pooling raw images.
CasCVS‑Net is a staged multi‑task cascade that jointly performs object detection, semantic segmentation, and Critical View of Safety (CVS) assessment for laparoscopic cholecystectomy. The model couples tasks through predicted anatomy—boxes guide segmentation and masks provide region‑level features for CVS classification—allowing CVS assessment to rely solely on model predictions. Trained on the Endoscapes dataset, CasCVS‑Net outperforms state‑of‑the‑art methods, achieving higher mAP and mIoU scores across detection, segmentation, and CVS tasks, especially for rare hepatocystic structures.