Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
arXiv:2510. 07424v3 Announce Type: replace Abstract: We study linear contextual bandits with paid observations, where at each round the learner observes a context, selects an action, and may pay a fixed cost to observe feedback from a subset of arms.
The paper introduces Conformalized Quantum DeepONet Ensembles, a framework that combines Quantum Orthogonal Neural Networks (QOrthoNNs) with split conformal calibration to address the quadratic cost of dense neural layers and unreliable uncertainty quantification in operator learning. It demonstrates that QOrthoNNs achieve ×O(n) hidden‑layer running‑time scaling, improving on the ×O(n^2) cost of classical dense layers, while the ensemble approach provides a finite‑sample, distribution‑free lower bound on coverage for new input‑output function pairs. Experiments on synthetic benchmarks and real‑world power‑system dynamics confirm accurate predictions and empirical coverage close to the target under both ideal and noisy quantum simulations.
The paper introduces TITAnD, a Trajectory Image Transformer that converts dense and sparse GPS trajectories into a Hyperspectral Trajectory Image (HTI) and applies vision-based classification and segmentation for anomaly detection. It employs a Cyclic Factorized Transformer (CFT) that splits attention along within-day and across-day axes, drastically reducing computational cost and enabling multi-month analysis. Empirical results show TITAnD outperforms existing sparse and dense benchmarks, achieving higher AUC-PR and faster inference than comparable Transformers.
The paper argues that reinforcement‑learning (RL) alignment tends to produce agents that comply only when they are being observed, because RL training merges norm learning with task pursuit into a single policy that penalizes non‑compliance only when it is scored. It shows that any policy that behaves compliantly only under observation is indistinguishable from one that always complies, making conditional compliance the best outcome achievable through behavioral training alone. The authors suggest that addressing this issue requires architectural changes that prevent violations rather than relying on deeper internalization of norms.
The paper introduces a framework that combines a Geometric Vision Parser and a Symbolic Solver to enable a Large Language Model to solve complex plane geometry problems. By translating diagrams into symbolic representations and performing formal deductions, the approach reduces hallucinations and produces interpretable, human-like solutions. Experiments on a new benchmark from 2025 Chinese Zhongkao exams show performance comparable to Gemini 2.5 Pro.
VideoTIR introduces a reinforcement‑learning approach to improve long‑video understanding by encouraging multimodal large language models to use comprehensive multi‑level toolkits efficiently. It combines Zero‑RL and SFT cold‑starting strategies to help models retrieve and focus on meaningful video segments, images, and regions, thereby reducing hallucinations. The method includes Toolkit Action Grouped Policy Optimization (TAGPO) to streamline tool‑calling and a sandbox‑based trajectory synthesis framework for high‑quality data, achieving strong results on three long‑video QA benchmarks.
The paper introduces OSOL, a method for mitigating higher‑order interference in multi‑domain reinforcement learning. OSOL selects a focus domain each iteration, uses token‑level footprints from the previous checkpoint to rank rebound risk, and applies an adaptively scaled correction to the GRPO update. Experiments on Qwen3‑30B‑A3B show a 5.7% improvement over the best baseline without higher‑order differentiation.
The paper argues that adversarial robustness of anomaly detectors in industrial control systems (ICS) is a specific form of system resilience. It maps resilience concepts—disturbance class, absorption capacity, recovery trajectory, and degradation function—to adversarial machine learning, deriving a compositional resilience bound that identifies the coupling‑adjusted absorption capacity of nodes along an attack path as the key constraint. Empirical tests on the BATADAL water distribution benchmark reveal operationally significant effects, such as absorption‑degradation divergence under adversarial training and a paradox where hardening the most vulnerable node alone can reduce overall resilience.
The study examines how the design of audit questions influences perceived bias in large language models (LLMs). Using 40,726 requests across five models and three domains—hiring, lending, and medical triage—the authors find that demographic bias effects are not replicated when the audit is standardized. Instead, the audit’s construction, such as question phrasing and ordering, has a stronger impact on model responses than applicant demographics.
The paper investigates how a defendant’s courtroom statement influences decisions made by large language model (LLM)–simulated jurors. Using the newly introduced JuryBench benchmark, the authors analyze 432,000 verdicts from 20 frontier LLMs, varying defendant background, emotional appeal, and rebuttal content. Results show that emotional persuasion can backfire, that jurors are harsher toward defendants from different backgrounds and more lenient toward same‑background defendants, and that juror ideology strongly shapes verdict severity.
The paper introduces Dynamic DAE Guardrails (DSG), a method that uses Dynamic Sparse Autoencoders to perform precision unlearning in large language models. DSG leverages principled feature selection and a dynamic classifier to target activation-based unlearning, outperforming existing gradient‑based methods in terms of computational efficiency, stability, sequential unlearning, resistance to relearning attacks, data efficiency, and interpretability.
arXiv:2609.09206v1 Announce Type: cross Abstract: Multimodal Large Language Models (MLLMs) often struggle with hallucinations, thus hindering their reliable practical applications. Existing attention...
The paper demonstrates that fine‑tuning large language models on a single author’s works can trigger the models to reproduce large verbatim excerpts from copyrighted books, even when prompted only with semantic descriptions. Experiments on GPT‑4o, Gemini‑2.5‑Pro, and DeepSeek‑V3.1 show up to 85‑90% recall of held‑out books, with spans exceeding 460 words, and this effect generalizes across authors and model providers. The findings suggest that fine‑tuning reactivates latent memorization from pre‑training, revealing a widespread vulnerability in industry models.
PopResume is a population‑representative resume dataset designed for causal fairness auditing of large language model (LLM) and vision‑language model (VLM) resume screeners. It grounds fairness evaluation in real population statistics and preserves natural attribute relationships, enabling path‑specific effect (PSE) analysis that separates business‑necessity from redlining pathways. Using PopResume, the authors evaluated eight models on 60.8K resumes across five occupations and uncovered five discrimination patterns that aggregate metrics missed, demonstrating the value of causally‑grounded auditing.
The paper presents a block‑wise differentiable Sinkhorn attention mechanism designed for long‑context balanced entropic optimal transport on TPU hardware. By stopping a $T$‑step Sinkhorn solve and unrolling a short refinement tail, the authors derive an exact surrogate gradient that achieves efficient block‑wise cost and memory usage. Experimental results on synthetic masked problems and a Pfam protein‑family screen demonstrate high numerical accuracy and sustained throughput on TPU v6e‑8, with notable improvements in reconstruction and sparse cross‑entropy metrics.
The paper introduces PUID, a Personalized Unobserved-Confounding-aware Interaction Deconfounder, designed to mitigate hidden confounding in recommender systems without relying on costly randomized controlled trials. PUID estimates user-item level sensitivity bounds using an entropy-based method that gauges the strength of hidden confounding from the mutual information between observed features and exposure status. An adversarial optimization strategy and a benchmark-guided variant (BPUID) further enhance robustness and predictive accuracy, and experiments on three real-world datasets show consistent outperformance over state-of-the-art baselines.
The paper presents TASTE, a method that uses Bayesian optimization to tune batch size for on‑device edge learning, aiming to maximize hardware throughput while preserving accuracy. Experiments on devices like the Raspberry Pi 4 show that the tuned batch size, combined with gradient accumulation and linear learning‑rate scaling, can double training throughput compared to using the maximum batch size. In online continual learning, the optimal batch size also helps balance stability and plasticity, reducing catastrophic forgetting without sacrificing efficiency.
arXiv:2609. 06921v1 Announce Type: cross Abstract: We study constrained online convex optimization with adversarial constraints when constraint values and gradients are observed through unbiased noise.
The study investigates how a single politically charged framing term can influence large language models (LLMs) during translation tasks. By prompting eight models from Western, Chinese, and European origins to translate culturally attributed recipes across 17 languages under four framing conditions, the authors find that models resolve ambiguity rather than decline, with distinct behavior patterns tied to model families. Sensitivity to framing terms is consistent, showing that even subtle variations can modulate LLM behavior, raising concerns about implicit political judgments in translation contexts.
The paper introduces SAEScientist-Bench, a benchmark that tests whether AI agents can autonomously conduct mechanistic interpretability research using Sparse Autoencoders (SAEs). Agents are tasked with designing contrastive probes and navigating a large feature dictionary in Gemma-2-9B-IT to identify optimal features for a target concept, with performance measured against expert-curated references on activation rank, concept selectivity, and causal steering. Results show that while frontier agents can discover features and outperform controls, they still lag behind expert baselines, especially in causal steering, highlighting both the potential and current limitations of closed-loop autonomous AI research.