Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
arXiv:2610.33078v1 Announce Type: cross Abstract: Visual grounding (VG) in visual question answering (VQA) is a fundamental capability of Vision-Language Models (VLMs), enabling the detection of the...
arXiv:2610.34365v1 Announce Type: cross Abstract: Science agents increasingly package domain expertise as reusable skills that plan analyses, select tools, and execute workflows. This modularity crea...
arXiv:2610.35532v1 Announce Type: cross Abstract: Unsafe deserialization turns untrusted data into attacker-controlled execution, and a single vulnerability can admit diverse exploit variants that ev...
arXiv:2610.33483v1 Announce Type: cross Abstract: Model merging combines task-specific expert models into a single model without joint training on their original datasets, but often incurs a performa...
arXiv:2610. 33882v1 Announce Type: cross Abstract: Hybrid post-training often mixes supervised fine-tuning (SFT) and reinforcement learning (RL), using training-side signals to set each sample's balance.
The paper introduces Steered-Teacher OPSD (ST-OPSD), a method that improves on-policy self-distillation for thinking models by steering the teacher’s hidden states to reduce reference hallucinations and encourage reconsideration. ST-OPSD extracts two directions from reasoning trajectories and applies them during distillation target computation, allowing the student to learn from a teacher that no longer invokes unavailable references and that promotes verification at uncertain steps. Experiments on AIME and HMMT demonstrate that ST-OPSD outperforms vanilla OPSD and other interventions, restoring reconsideration and reference-hallucination rates to base-model levels.
The paper presents a method for predicting the remaining useful life of lithium‑ion batteries using only the first few charge‑discharge cycles. It transforms voltage, current, and capacity data into time‑frequency representations via continuous wavelet transform, compresses the resulting scalograms to reduce redundancy, and feeds them into deep neural networks. Experiments on eight benchmark datasets show the approach outperforms previous methods across diverse chemistries and cycling protocols, and visual explanations confirm the model focuses on meaningful time‑frequency regions.
arXiv:2610.34208v1 Announce Type: cross Abstract: Offline safe reinforcement learning (RL) learns safety-aware policies from static datasets while satisfying safety constraints during deployment. Mot...
arXiv:2610.34385v1 Announce Type: cross Abstract: Agents that act on real systems should predict what their actions will cause before acting. We study when causal structure helps a learned forward mo...
arXiv:2610.34678v1 Announce Type: cross Abstract: Humans have been shown to use a "straightened" encoding to represent the natural visual world as it evolves in time (H\'enaff et al. 2019). In the co...
arXiv:2610.35480v1 Announce Type: cross Abstract: Multi-reward reinforcement learning offers a promising framework for aligning large language models (LLMs) with diverse objectives, but conflicting r...
arXiv:2610.35551v1 Announce Type: cross Abstract: In stochastic multi-objective bandits, Pareto optimal arms may represent different trade-offs among objectives. Balanced sampling across them is impo...
UMAS is a parameter‑free, cost‑neutral framework for system‑level uncertainty quantification in multi‑agent large language model systems. It estimates the credibility of inter‑agent influence using each generation’s intrinsic confidence, modulates uncertainty updates at each node, and aggregates these into a system‑level estimate. Evaluations on Debate and DyLAN show UMAS outperforms state‑of‑the‑art baselines by an average of 10.54 AUROC points and improves hallucination detection and answer selection, boosting MAS accuracy by up to 12.6 points on specific tasks.
arXiv:2610.33277v1 Announce Type: new Abstract: Engineering standards are challenging knowledge sources for large language models (LLMs) because they are hierarchical, multimodal, and densely connect...
arXiv:2610.33410v1 Announce Type: new Abstract: Time-series tasks require both accurate predictions and fast decisions. Advances in large models have introduced decision-oriented models that directly...
arXiv:2610.33442v1 Announce Type: new Abstract: Large language models (LLMs) are reported to be susceptible to reliability issues including hallucinations and reasoning errors. Existing quality detec...
arXiv:2610.33540v1 Announce Type: new Abstract: Image tokenizers convert pixels into compact latent sequences for efficient generative modeling. Most approaches arrange latents on a downsampled two-d...
The paper introduces T2D, a framework that prunes and adapts a target large language model into a draft model optimized for both inference latency and alignment with the target model’s outputs. By formulating draft model design as an optimization problem, T2D achieves up to 92% pruning while maintaining high speculative decoding performance across multiple target model families and domains. The resulting draft models can also be integrated with other speculative decoding techniques to yield additional gains.
arXiv:2610.34144v1 Announce Type: new Abstract: Fine-tuning adapts pretrained models to new tasks, but identifying which weight changes support the resulting behavior remains difficult. Weight-space...
arXiv:2610.34171v1 Announce Type: new Abstract: Time Series Foundation Models have demonstrated that pretraining on diverse datasets enables effective transfer to new forecasting tasks. Classical met...