Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
arXiv:2610.27152v1 Announce Type: cross Abstract: Neural networks (NNs) are increasingly deployed on edge devices such as mobile and IoT platforms, offering low latency and improved privacy compared...
arXiv:2610.27041v1 Announce Type: cross Abstract: When given an image and a text prompt, vision-language models (VLMs) can produce seemingly sophisticated text responses. However, they are also prone...
arXiv:2610.28092v1 Announce Type: cross Abstract: Persuasive writing exploits cognitive biases, recruiting a reader's mental shortcuts to shape belief, often without the reader noticing. We call the...
arXiv:2609.32002v2 Announce Type: replace Abstract: The rank $r$ in LoRA is widely treated as a capacity control: a smaller rank is assumed to yield a simpler model that generalizes better. We show t...
WeGR is a unified generative retrieval framework that tackles two key challenges in industrial search: preserving collaborative signals in tokenization and mitigating reward sparsity in reinforcement learning. It introduces the Collaborative and Collapse-free Tokenizer (CoCo‑VAE) to embed query‑to‑item and item‑to‑item signals into continuous representations before quantization, while also refreshing underutilized codewords to maintain balanced codebook usage. For preference alignment, Dynamic Horizon Policy Optimization (DHPO) expands reward evaluation horizons to enable coarse‑to‑fine learning from SID prefixes to full items without auxiliary reward models. Extensive offline tests and a six‑week online A/B test on a large‑scale e‑commerce platform demonstrate that WeGR improves retrieval and ranking performance, yielding relative gains of 2.539% in orders, 1.508% in CTR, and 1.381% in CVR, and reduces CPU and GPU usage by 25.5% and 26.1% respectively.
arXiv:2609.34575v2 Announce Type: replace Abstract: Offline goal-conditioned reinforcement learning (GCRL) learns goal-directed policies from reward-free data, but in long-horizon tasks, goal-conditi...
CounterCount is a diagnostic dataset that pairs factual and counterfactual images to test whether Vision‑Language Models (VLMs) rely on visual evidence or on world knowledge when counting attributes. Humans achieve near‑perfect accuracy on both types of images, whereas VLMs drop from 91.6% to 49.5% on counterfactuals, often defaulting to canonical counts. The study shows that VLMs exhibit a System‑1‑like bias, underutilize visual evidence, and that targeted input manipulations can mitigate this bias.
Omni-Woodpecker (OWP) is a training‑free framework designed to reduce cross‑modal hallucinations in Omni LLMs by rebalancing fused representations. It identifies conflicts between modality‑specific views, separates grounding evidence from competing priors via orthogonal decomposition, and adaptively suppresses misleading priors while restoring ignored evidence before decoding. Experiments show OWP improves performance on CMM, AVHBench, and several general reasoning benchmarks, and generalizes across different Omni LLM architectures.
arXiv:2610.28467v1 Announce Type: new Abstract: Human mobility data are fundamental to Web-based location services and broader urban applications, and mobility generation provides synthetic trajector...
arXiv:2610.27735v1 Announce Type: new Abstract: Recursive self-improvement, in which frontier AI models work on training the next generation, is quickly becoming a reality. However, the safety of thi...
The paper introduces AuthorizationBench, a benchmark designed to evaluate whether large language model (LLM) agents act within the scope of user authorization when acting on their behalf. The benchmark comprises 286 executable tasks across ten application domains, covering situational triggers, presence or absence of direct harmful consequences, and varying reasoning effort. In experiments with four models and multiple agent harnesses, agents violated authorization in 19.2% of trajectories, with rates ranging from 10.5% to 26.6% across configurations, and increased reasoning effort did not substantially reduce these violations.
arXiv:2610.24993v2 Announce Type: replace Abstract: The throughput gains of speculative decoding depend on the trade-off between acceptance length and execution latency. We ask when additional accept...
arXiv:2610.27904v1 Announce Type: new Abstract: Staged post-training of multimodal foundation models chains vision-language alignment, reasoning distillation, and domain tuning, yet how each phase tr...
The paper investigates how language models fail to detect harmful intent after polarity-altering transformations, a vulnerability called polarity blindness. It examines grammatical negation, antonym replacement, and synonym replacement as controls, and tests the impact of adversarial wrappers. Across 260 prompts, 10 LLMs, and three languages, polarity blindness is widespread, with the most vulnerable model scoring an average harm level of 4.12 on a 0‑5 scale.
arXiv:2610.28579v1 Announce Type: cross Abstract: Cellular states are shaped by both intrinsic gene expression programs and the spatial microenvironment in which cells reside. Spatial transcriptomics...
arXiv:2610.28607v1 Announce Type: cross Abstract: Mathematical reasoning remains underexplored for low-resource languages, including Dialectal Arabic. To bridge this gap, we introduce the DAR-MGSM (D...
WaveCommitments is an open‑source framework that enables private and verifiable marketplaces for autonomous speech agents. It combines confidential asset valuation using Data Shapley within a trusted execution environment with multivariate polynomial commitments, allowing selective openings and tamper‑evident provenance. Experiments on three datasets and state‑of‑the‑art encoders, including Whisper, show up to 20× faster benchmarks and 39× smaller proofs, while also revealing how speaker value varies across tasks and encoders.
Diffusion language models (DLMs) are gaining traction as an alternative to autoregressive language models, but their fairness has not been well studied. This work investigates how sampling controls—such as block length and remasking strategy—affect social bias in three DLMs (LLaDA, LLaDA‑1.5, and Dream on Open‑BBQ). The authors find that both the magnitude of bias and the relative ranking of models vary significantly with sampling settings, and that even after applying mitigation strategies from autoregressive models, sampling dependence remains, indicating that fairness in DLMs cannot be assessed independently of sampling configuration.
HierDTA is a hierarchical structure‑aware framework for drug–target affinity prediction that represents drugs as atom–motif heterogeneous graphs and proteins as residue interaction graphs. It integrates atom and residue representations via cross‑attention, supplemented by molecule fingerprints and pretrained ESM‑2 embeddings, achieving superior performance on the Davis and KIBA datasets and strong generalization in cold‑start scenarios. Ablation studies and interpretability analyses confirm the contributions of its hierarchical augmentation and interaction components, and case studies demonstrate early enrichment of candidate bioactive compounds.
The paper addresses a flaw in differential privacy audits where selection bias in Membership Inference Attacks leads to overestimated lower bounds on privacy leakage. By applying statistical correction methods, the authors demonstrate that uncorrected audits can violate theoretical privacy guarantees up to four times more often than claimed, while corrected audits maintain valid bounds at the nominal confidence level.