Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
arXiv:2602.03753v2 Announce Type: replace Abstract: While representation alignment with self-supervised models has been shown to improve diffusion model training, its potential for enhancing inferenc...
arXiv:2606.10142v2 Announce Type: replace Abstract: Recent advances in 3D generation have led to substantial improvements in realism, controllability, and efficiency, yet the evaluation of 3D assets...
arXiv:2610.08460v1 Announce Type: new Abstract: We propose a simple framework for constructing differentially private confidence regions \textit{in one shot}, i.e., by adding noise only to the final...
The paper introduces QRAKEN, a training‑free, ontology‑agnostic neurosymbolic pipeline that grounds natural‑language queries to RDF knowledge graphs using empirical graph evidence instead of schema expectations. QRAKEN distills a compact TTQL representation offline, then uses it online to guide large language models, providing deterministic checks and an iterative refinement loop. On the CK25 benchmark, QRAKEN achieves a strict F1 of 0.643–0.652 with GPT‑4.1 mini and GPT‑5.4, outperforming state‑of‑the‑art systems and demonstrating the effectiveness of empirical pattern distillation over schema‑only approaches.
The paper proposes a framework for studying how alignment difficulty scales with model size, treating alignment as a family of measurable scaling relations for different risk categories. It introduces operational definitions of burden, distinguishes observed, audited, and true alignment, and proves that the largest exponent among corrected risks determines long‑run behavior. The authors preregistered and applied the protocol to public data, finding scaling exponents for attack success, truthfulness, dispositions, sycophancy, and a backdoor, and released interactive browser games to illustrate the laws.
The study investigates which image properties enable jailbreak attacks on multimodal models by analyzing four attack families across 313 prompts and five models, including an appendix evaluation of InternVL3.5-8B. It finds that attack images significantly increase success rates compared to benign images, while factors such as tile entropy, JPEG size, and tile count do not reliably predict attack tiles. Only the E4 manipulation, which removes query-specific relatedness, shows a measurable effect on attack success rates, though overall conclusions depend on the rubric judge used.
The paper introduces Guidance-Augmented GRPO (GA‑GRPO), a theoretical framework that unifies several external‑guidance methods for reinforcement learning with verifiable rewards (RLVR) used to improve large language model reasoning. GA‑GRPO treats external guidance as a stochastic operator that rewrites the question distribution, yielding a biased on‑policy policy‑gradient estimator whose bias is bounded by the total‑variation guidance divergence. The authors prove convergence rates, derive an optimal guidance‑weight formula, and validate their predictions experimentally on a math‑reasoning benchmark, showing that the optimal‑weight GA‑GRPO outperforms existing methods while saving GPU time.
Fed-BRDECS is a federated deep embedded clustering framework that protects client privacy by replacing globally normalized clustering objectives with a locally computable sample‑stability loss, eliminating the need to share sensitive soft‑assignment distributions. It addresses non‑IID client data through prediction‑balanced sampling, which oversamples rare predicted clusters, and centroid‑level restarting, which refreshes biased or inactive centroids. Experiments on image, text, and time‑series datasets demonstrate that Fed‑BRDECS outperforms existing federated and deep clustering methods while adding no inference‑time cost for anomaly detection.
The paper introduces an adversarial reinforcement learning framework to generate synthetic data for Argument Mining (AM). By jointly training a generator and a discriminator, the system produces structured AM instances that are both accurate and diverse. Experiments show consistent performance gains on three benchmark datasets in both full-data and low-resource scenarios.
The paper introduces STEP‑KD, a Sequential Timing Evaluation via Progressive Knowledge Distillation framework for predicting post‑routing timing from early‑stage design data. By training teacher models at post‑routing, post‑placement, and post‑floorplan stages and sequentially distilling their knowledge to a post‑synthesis student model through representation alignment, STEP‑KD bridges the abstraction gap that hampers existing machine learning approaches. Experiments on diverse circuits show that STEP‑KD reduces the weighted mean absolute percentage error of Total Negative Slack prediction to 19.78% versus 74.84% for the industry‑standard Static Timing Analysis tool.
The paper studies robust bandits, a framework for learning when the true environment lies outside the learner’s hypothesis class. It presents a polynomial‑time algorithm achieving ≠O(√T) regret for a specific tractable case, while showing that slight extensions of this case are NP‑hard. The work highlights the computational limits of robust bandits and suggests their relevance to AI alignment.
The paper audits five AffectNet‑pretrained facial affect models on child faces, revealing that all models perform worse on naturalistic child images and consistently confuse fear with surprise. The degradation is linked to specific expression morphologies—particularly open‑mouth faces—and to South‑Asian children, rather than skin tone. A linear probe shows that most of the gap lies in the classifier head, and fine‑tuning just this head on a small amount of child data improves performance by up to 28% on the largest child datasets.
Con-DSO introduces a consistency-aware RGB‑D direct sparse odometry framework that learns pixel‑level photometric and geometric uncertainty from adjacent RGB‑D frame pairs. The network predicts uncertainties that are converted into pairwise quality scores, guiding support‑pixel selection and forming a host‑side quality prior for keyframe tracking. Experiments on five public benchmarks show that this approach reduces absolute trajectory error by over 20% on ICL‑NUIM and by 50–80% on other datasets, improving robustness in challenging environments.
The paper introduces a method for discovering implicit negative candidates in recommender systems by extracting symbolic rules from observed customer behavior. These rules are scored on support, informativeness, and product relevance, then interpreted by a large language model to align with business objectives. Experiments in an industrial B2B setting and on five public datasets show that the approach improves precision and downstream PR-AUC compared to baseline negative sampling methods.
The paper demonstrates that a frozen pretrained diffusion model can be endowed with identifiable latent coordinates via a lightweight alignment map called Contrastive Diffusion Alignment (ConDA). By leveraging auxiliary-variable nonlinear ICA guarantees, ConDA aligns diffusion latents so that the resulting representation recovers latent dynamical states up to permutation and componentwise invertible transformations, preserves the underlying structural causal model, and simplifies lagged graph recovery. Experiments on physical and robotic video systems show that ConDA, especially under TCL and GCL assumptions, achieves near-perfect state recovery and accurate lagged graph reconstruction, even in simulated falling-body and bipedal robot scenarios.
The paper investigates what information must be preserved in replay buffers for class‑incremental learning. By treating cached predictions as temporally heterogeneous supervision, the authors separate classes known at storage time from those learned later, and evaluate the impact of deleting logit matching. Experiments on CIFAR‑100 with DER++ show that a simple task‑level offset can largely correct the cost of removing later‑class matching, while the cost of disrupting class correspondence remains.
The paper introduces HARP, a hard-budget allocation method for evaluating large language models (LLMs) in multi-turn interactions where the time-to-event is partially observed due to resource limits. HARP guarantees that the total computational budget is never exceeded, reallocates unused budget, and provides lower predictive bounds (LPBs) with finite-sample coverage and unbiased metric estimates. Experiments on tasks such as jailbreaks, toxic content, and hallucinations demonstrate that HARP achieves near-nominal coverage with low variance while respecting the fixed budget.
The paper introduces EmoNet‑Face‑HQ, a fine‑grained emotion recognition benchmark that uses generated portraits and a 40‑category taxonomy to evaluate vision‑language models (VLMs). It finds that VLMs perform poorly when asked to generate responses but can match or surpass a fine‑tuned model (Empathic‑Insight‑Face) when their logits are read directly as binary queries. The study shows that the benchmark’s difficulty lies in the readout process rather than in perception, and that graded probability outputs yield better performance than simple yes/no questions.
The study introduces a multitask conditional generative adversarial network (MT‑cGAN) that simultaneously synthesizes high‑resolution DESS‑like images and segments knee cartilage and menisci directly from quantitative MRI echo images. Evaluated on 508 knee MRIs from 361 subjects, MT‑cGAN achieved a mean Dice score of 0.84 for segmentation and the lowest coefficient of variation for T1ρ (1.84%) and T2 (1.81%) quantification, outperforming existing conditional GAN approaches. By eliminating the need for separate high‑resolution morphological scans, the method shortens scan times and supports clinical adoption of quantitative MRI.
The paper introduces SynEval, a six‑dimensional framework for evaluating multi‑table synthetic data. It assesses per‑column fidelity, multivariate structure (including a new conditional distribution check), cross‑table integrity, machine‑learning utility, privacy protection, and edge‑case robustness. SynEval outputs a unified weighted quality score and is generator‑agnostic, working on any real and synthetic CSV folders with automatic schema inference.