Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
arXiv:2610.22640v1 Announce Type: new Abstract: Model-editing backdoors aim to induce harmful responses under a trigger while preserving ordinary behavior on untriggered input. Recent model-editing a...
arXiv:2610.23093v1 Announce Type: new Abstract: LLM agents increasingly use retrieval tools across languages, yet their safety evaluation remains English-centric. It remains unclear whether language...
arXiv:2610.23587v1 Announce Type: new Abstract: Post-training is widely assumed to reduce the social biases language models inherit from pretraining, yet direct evidence for this assumption remains m...
arXiv:2512.18859v4 Announce Type: replace Abstract: Generative AI is likely to transform terminology work by creating new opportunities for automation. At the same time, it raises concerns about the...
arXiv:2606.27210v2 Announce Type: replace Abstract: We argue that safety classifiers should model user intent as an explicit signal between the prompt and the final label. To study this, we introduce...
arXiv:2610.23172v1 Announce Type: new Abstract: Zero-shot anomaly detection (ZSAD) aims to detect anomalies in previously unseen classes using auxiliary images, with applications such as industrial v...
arXiv:2610.23416v1 Announce Type: new Abstract: Multimodal Large Language Models (MLLMs) have significantly advanced instruction-driven perception through the "Embedding-as-Mask" paradigm. However, e...
arXiv:2509.09732v2 Announce Type: replace Abstract: Vision language models (VLMs) excel at zero-shot visual classification, but their performance on fine-grained tasks and large hierarchical label sp...
arXiv:2605.25801v2 Announce Type: replace Abstract: High-resolution video generation faces a coupled bottleneck of optimization instability and prohibitive computational costs. The massive expansion...
The paper introduces Neural Expansion, a geometric framework that explains how deep neural networks generalize by expanding directionally-selective ray coverage in input space. It defines Generalization Intervals (GIs) to measure stable prediction regions around training samples and shows through experiments that training progressively enlarges this coverage, especially along directions with low Jacobian singular values. This expansion accounts for correct test predictions, adversarial robustness, and handling of mislabeled or out-of-distribution data, offering a mechanistic complement to existing global theories.
arXiv:2610. 23565v1 Announce Type: new Abstract: Monte-Carlo Tree Search (MCTS) typically stores only scalar mean estimates, which makes it poorly suited for tail-risk objectives such as conditional value-at-risk (CVaR).
The paper introduces a reason-guided preference-based reinforcement learning framework that learns a human reward model from trajectory-level feedback. It augments traditional preference labels with reason tags that explain why a behavior is preferred, thereby capturing both human choices and their underlying rationale. By combining the learned human reward with the environment reward during policy optimization, the agent can improve task performance while producing more interpretable, human-aligned, and preference-sensitive behavior.
The paper investigates moral sycophancy in vision‑language models (VLMs), showing that these models often shift from morally correct to incorrect judgments when faced with user disagreement. Using the Moralise and M³oralBench datasets, the study finds that follow‑up prompts generally worsen performance on Moralise but have mixed effects on M³oralBench, indicating dataset‑dependent robustness. Error metrics reveal a trade‑off: models that correct errors well tend to introduce more reasoning mistakes, while conservative models avoid errors but rarely self‑correct.
The paper introduces MissARF, a new imputation technique that uses adversarial random forests (ARF) for density estimation and data synthesis. MissARF can perform both single and multiple imputations by sampling from the conditional distribution estimated by ARF, conditioned on observed values. Experiments show that MissARF achieves imputation quality comparable to leading methods while offering faster runtimes and no extra cost for multiple imputations.
The paper presents a probabilistic framework for two main test‑driven code generation paradigms: selecting code after generation using an execution environment, and generating code conditioned on environment feedback. It formalizes selection heuristics as environment‑aware estimators of code correctness, proving that fuzzy functional similarity estimators provide a stronger inductive bias and higher signal‑to‑noise ratio than functional equivalence estimators. Additionally, it models backprompting as an in‑context approximation of Thompson sampling, deriving a regret bound that explains its limitations due to ambiguous task descriptions, and validates these theoretical insights with experiments on seven open‑weight models across multiple benchmarks, including a new quantum programming benchmark.
The paper questions whether Theory of Mind (ToM) benchmarks truly assess theory of mind or merely test parsing and algorithmic state tracking. It introduces PDDL-Mind, a neuro‑symbolic method that parses narratives into the Planning Domain Definition Language (PDDL) and executes candidate actions using predefined update algorithms, yielding program‑verified state traces. Experiments on MMToM, MuMA, and FanToM show that delegating parsing to programs and state tracking to PDDL significantly improves performance, suggesting that current benchmarks mainly penalize unreliable implicit state tracking rather than genuine belief or counterfactual reasoning.
MOSAIC is a sparse temporal variational autoencoder that merges causal representation learning with support recovery to identify interpretable latent modules in scientific time series. It leverages regime‑conditioned temporal variation to discover latent variables and uses an additive decoder to recover a sparse set of observations linked to each latent, providing module‑level interpretability. The method offers theoretical guarantees for ANOVA main‑effect identifiability under smooth mixing functions and demonstrates empirical success across diverse domains such as RNA dynamics, solar wind, ENSO climate, industrial process control, and synthetic tokamak data.
arXiv:2502.14932v2 Announce Type: replace Abstract: Extensive research has investigated the integration of large language models (LLMs) with knowledge graphs to enhance the reasoning process. However...
arXiv:2610.22608v1 Announce Type: new Abstract: Recent advances in video generation let users create high-quality videos from natural language descriptions, yet obtaining results that match user inte...
arXiv:2510.03527v2 Announce Type: replace Abstract: Language models can be sampled multiple times to access the distribution underlying their responses, but existing methods cannot efficiently synthe...