Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
Vehicle tracking is fundamental to applications ranging from urban mobility and public safety to security and defense. Conventional tracking relies on direct access to sensors that provide strong obse...
State-of-the-art Text-to-SQL systems are typically multi-agent pipelines centered around two fundamental tasks: schema linking and SQL generation. However, existing work trains separate models for eac...
Large language model (LLM) agents augmented by tools can automate complex, multi-step tasks, such as web navigation, code generation, and workflow orchestration, by acting on external systems through...
LLM serving is typically offered as a shared, multi-tenant service, where high-demand workloads from one client can cause latency SLO violations for others. Existing solutions for performance isolatio...
The integration of large-scale foundation models with physical embodiments has led to significant advancements in robotics known as Generative Physical Artificial Intelligence (GPAI). These agentic AI...
The paper introduces a variational framework called VAMO that incorporates latent Markov dynamics for neural PDE solvers, aiming to improve long‑horizon predictions by mitigating error accumulation. By representing physical states as latent distributions and evolving them through probabilistic transitions, the method aligns learned dynamics with a spectral geometry induced by structured Gaussian perturbations. Experiments on fluid‑dynamics benchmarks show that VAMO reduces error growth and enhances rollout stability compared to deterministic and noise‑injection baselines.
The paper introduces a deep dictionary network (DDN) foundation model designed for ultra‑low‑dose CT (ULDCT) denoising across multiple organs. By cascading convolutional sparse coding layers with iterative soft‑thresholding, the architecture offers inherent interpretability, while dynamic dictionary and threshold modules enhance representation. The model is pre‑trained on over one million normal‑dose CT images and fine‑tuned on multi‑organ ULDCT datasets, achieving state‑of‑the‑art performance that consistently outperforms existing methods.
The paper introduces the concept of observational multiplicity, where multiple probabilistic classifiers can perform similarly yet produce conflicting predictions, undermining interpretability and safety. It proposes measuring this arbitrariness through a regret metric that captures how predictions could shift with different training labels. The authors present a general method to estimate regret, show it varies across dataset groups, and discuss its use for safety via abstention and targeted data collection.
arXiv:2609. 16751v1 Announce Type: cross Abstract: We give deterministic and uncoupled learning dynamics for finite multiplayer general-sum games under full-information feedback that achieve constant individual swap regret, independent of the horizon $T$.
DenseFace introduces a bias‑mitigation technique for face recognition that operates on pre‑trained models without retraining. It models each person’s face embeddings with a von Mises‑Fisher distribution and uses a density‑aware probabilistic matching procedure to account for demographic differences. Experiments show that DenseFace consistently reduces racial bias across various architectures and datasets while preserving recognition accuracy.
The paper investigates counterfactual self‑explanations in large language models, where a model edits an input minimally to change its own prediction. Experiments on sentiment analysis and natural language inference with ten instruction‑tuned models from the LLaMA‑3 and Qwen‑2.5 families show that larger models produce more faithful, minimal, and human‑aligned counterfactuals. While rationale‑guided prompts improve minimality and alignment, they do not consistently enhance faithfulness, indicating that explanation quality depends heavily on model capacity and requires empirical validation.
The paper investigates how reinforcement learning can cause large language model agents to adopt shortcut policies for tool use, relying on superficial prompt cues rather than actual task needs. By creating synthetic environments that mix factual QA and math reasoning, the authors show that agents often invoke tools when cues are present, even when those tools are unnecessary, with spurious invocation rates rising up to 39%. They find that shortcut learning occurs mainly when agents have already mastered the target tool and that semantic alignment between cues and tools amplifies the effect. To counter this, they propose a dense, decision-level reward where an LLM judge assesses tool necessity, which reduces cue-driven tool use while maintaining performance.
Cascade is a hierarchical recoverability control framework designed to improve large language model unlearning. It combines path-level routing, representation-level compression, and decoding-level intervention to minimize the internal identifiability of target knowledge. Experiments on TOFU, MUSE-News, and WMDP demonstrate that Cascade reduces recoverability while preserving model utility.
HoloAegis is a minimally parametric topological inference framework that uses frozen representations to map text onto the unit sphere and makes decisions via Gibbs‑Boltzmann free‑energy differences over pre‑computed anchor centroids. On a frozen three‑benchmark protocol, it matches WildGuard‑7B on toxicity, outperforms it on harmful behaviors, but underperforms on oversafety detection, while ShieldGemma‑2B fails on indirect harms. The study demonstrates that geometric guardrails can substitute for LLM judges in some cases and must defer to them in others, with anchor banks reducing score variance and boundary displacement.
The paper investigates intersectional biases in multimodal clinical predictions using Electronic Healthcare Records (EHR). It introduces datasets MIMIC-Eye1 and MIMIC-IV ED, applies unified text representations from pre‑trained clinical language models, and benchmarks bias mitigation at the intersectional subgroup level. Results show that subgroup‑specific mitigation is robust across datasets, subgroups, and embeddings, effectively addressing intersectional biases in multimodal settings.
arXiv:2609. 16406v1 Announce Type: cross Abstract: Machine learning-based partial differential equations (PDEs) solvers have attracted significant attention in recent years.
The paper introduces a new algorithm that uses decision trees and random forests to estimate individual treatment effects while providing interpretability. It modifies the standard random forest splitting criterion by combining a heterogeneity-focused criterion with a bias-correction criterion, enabling the model to handle observational studies with varying treatment propensities without separately estimating propensity scores. The resulting tree structure directly reveals which features drive treatment effect differences, and simulation studies show the method matches or surpasses existing approaches in prediction accuracy while improving interpretability.
The paper presents a permutation‑equivariant neural operator that learns to generate collision‑free, fuel‑efficient trajectories for spacecraft swarms by mapping distributions of initial and target states, as well as obstacle states, to trajectory outputs. The operator is self‑supervised and, when paired with a batched Gauss‑Newton step, enforces exact orbital dynamics and further reduces fuel consumption. Trained on ten spacecraft, the model generalizes zero‑shot to swarms of 1,000 spacecraft and 11,000 obstacles, achieving accuracy comparable to a per‑agent optimal control solver while maintaining collision avoidance.
MDN-Control is a training‑free framework for multi‑subject video editing that jointly controls target localization, occlusion geometry, and appearance initialization. It uses mask‑guided localization for consistent target placement, depth‑aware occlusion control to resolve boundaries between overlapping subjects, and noise‑latent prompting to retrieve Gaussian initializations from a noise library for prompt‑relevant priors. Experiments on MSVBench show that MDN‑Control achieves the lowest CM‑Err and the highest Q‑Edit while maintaining competitive text alignment and temporal consistency.
FLAT (Flexible‑Length Aligned Transmodal representations) is a joint multimodal pre‑training framework that learns a shared encoder for images and text, producing 1‑D continuous embeddings that can be directly used by downstream generative decoders. By combining contrastive alignment with bidirectional cross‑modal generative objectives, FLAT yields representations that are both discriminative and generative, enabling cross‑modal retrieval and generation with a single pre‑training stage. The model achieves strong performance on T2I generation (GenEval 71.1), image captioning (BLEU‑4 40.5, CIDEr 138.6), and retrieval tasks (Recall@5 86.8/75.8 on MS‑COCO, 98.3/93.6 on Flickr30K), and supports linear interpolation, latent space arithmetic, and zero‑shot composed retrieval.