Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
arXiv:2508.14313v4 Announce Type: replace-cross Abstract: Test-time scaling strategies for Large Language Models predominantly rely on either reinforcement learning with sparse outcome rewards or sea...
arXiv:2509.20377v2 Announce Type: replace-cross Abstract: Retrieval-Augmented Generation (RAG) has significantly improved the performance of large language models (LLMs) on knowledge-intensive tasks...
arXiv:2608.21023v1 Announce Type: new Abstract: Word alignment has traditionally been studied between sentences, but many cross-lingual tasks increasingly require correspondences across full document...
arXiv:2608.20394v1 Announce Type: cross Abstract: Industry pipelines that turn speech into supervised fine-tuning (SFT) data via multi-stage refinement are increasingly adopted but, to our knowledge,...
arXiv:2601.20476v2 Announce Type: replace Abstract: GenAI is widespread in educational applications; however, it is known to generate content with intrinsic and extrinsic hallucination. We introduce...
arXiv:2608.20430v1 Announce Type: new Abstract: World Action Models (WAMs) improve planning by incorporating future world evolution into action generation, yet existing methods allocate a fixed imagi...
The paper examines how large language models (LLMs) can assist in creating regulatory compliance artifacts for EU sustainability and privacy laws, specifically Digital Product Passports (DPPs) under the Ecodesign for Sustainable Products Regulation and Data Protection Impact Assessments (DPIAs) under the General Data Protection Regulation. It investigates the effects of data extraction instructions and regulatory ambiguity on the quality and consistency of LLM-generated artifacts, benchmarking various models against manually crafted ground‑truth schemas. Findings indicate that looser guidelines, like those for DPIAs, demand more extensive prompts to achieve consistency, whereas stricter formatting rules for DPPs yield consistent outputs but may introduce hallucinations.
The paper introduces STAR-OPD, a structured aspect‑cascade‑aware on‑policy reward distillation method for aspect‑based sentiment analysis (ABSA) quadruple extraction. It addresses a specific failure mode where distilled models produce structurally invalid target‑aspect bindings, leading to hallucinated targets and corrupted downstream predictions. By training on student rollouts and applying set‑structured rewards that enforce binding consistency, target grounding, and fine‑grained aspect disambiguation, STAR‑OPD outperforms both off‑policy and generic on‑policy baselines on E‑ABSA20K and SemEval‑2014, reducing target hallucination and improving performance on structurally hard cases.
The study introduces PV‑SST, a peer‑voted social‑platform testbed, and conducts a preregistered matched‑exposure experiment across four topics, four seeds, four model families, and three larger variants, totaling 448 trials. Results show that feeding agents a ranked list of prior peer posts increases lexical similarity in both the core panel and larger variants, but does not produce a reliable advantage in opinion alignment or survival rates. The only robust finding is lexical convergence driven by the peer‑ranked feed, with no consistent coordination benefit across models or topics.
The paper defends the 'Whole Hog Thesis', arguing that sophisticated large language models such as ChatGPT are full linguistic and cognitive agents, possessing understanding, beliefs, desires, knowledge, and intentions. It rejects low‑level computational starting points and instead builds its case from high‑level behavioral observations, using Holistic Network Assumptions to link actions to mental states. The authors systematically rebut common objections—such as hallucinations and planning errors—by showing these resemble human fallibility and by challenging the necessity of traditional conditions like embodiment or semantic grounding.
COEC (Calibrated Orthogonal-Equivalence Compensation) is a training‑free framework that improves structured pruning of large language models by applying alternating left and right orthogonal rotations to the retained weight matrix. The method optimizes the right rotation on a reduced Stiefel manifold, rescales singular values via generalized cross‑validation, tempers the calibration Gram matrix, and adds an alignment penalty to preserve geometric relations between attention projections. Experiments on Llama‑3, Llama‑3.1, and Qwen2.5 show that COEC consistently improves perplexity and zero‑shot accuracy across multiple sparsity levels, outperforming existing compensation techniques.
StruProKGR is a Structural and Probabilistic framework designed for sparse Knowledge Graph (KG) reasoning. It reduces computational costs by using a distance‑guided path collection mechanism and improves interpretability by aggregating paths probabilistically to prioritize reinforcing paths. Experiments on five sparse KG benchmarks show that StruProKGR outperforms existing path‑based methods in both effectiveness and efficiency.
The paper shows that large language models (LLMs) naturally organize their hidden state manifolds into small‑world networks, enabling efficient multi‑hop reasoning. By converting similarity matrices into unweighted graphs, the authors trace connectivity between distant semantic anchors and find a sharp topological phase transition: deep reasoning layers compress conceptual distances into paths bounded by six semantic hops, while early syntactic layers remain fragmented. The framework is applied to zero‑shot hallucination detection in Retrieval‑Augmented Generation, revealing that factual generations preserve a ~3‑hop structure, whereas hallucinations collapse the topology.
The paper introduces a new object detection approach that protects sensitive visual data in test images. It is the first to apply perceptual encryption to object detection, leveraging the embedding structure of Vision Transformers and a key-based domain adaptation to maintain accuracy comparable to unprotected models. Experiments on ViTdet confirm the method’s effectiveness in both accuracy and visual protection.
The paper introduces a new framework for identifying average dose-response functions in the presence of unmeasured confounding by using instrumental variables. It defines a uniform regular weighting function and partitions the treatment space into open sets where local identification is possible. For estimation, the authors propose an augmented inverse probability weighted score within a debiased machine learning setting, along with practical guidance for constructing weighting functions, falsification tests for the additive IV condition, and asymptotic theory for kernel regression or empirical risk minimization estimators.
arXiv:2608.20925v1 Announce Type: cross Abstract: Reference-based metrics remain the standard choice in machine translation evaluation, partly because quality estimation methods often correlate less...
arXiv:2608.21133v1 Announce Type: new Abstract: Medical image-text data can expose protected health information (PHI) through both visible image content as well as accompanying text, creating a barri...
arXiv:2606.13971v3 Announce Type: replace Abstract: While personalizing Image-to-Video (I2V) diffusion models with specific visual effects is increasingly demanded for high-end generation, current pr...
arXiv:2608.15698v2 Announce Type: replace Abstract: Visual document retrieval is a critical component of multimodal retrieval-augmented generation, aiming to identify query-relevant pages from docume...
arXiv:2411.02771v3 Announce Type: replace-cross Abstract: Doubly robust estimators are widely used for estimating average treatment effects and other linear summaries of regression functions. While c...