Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
Modern machine learning relies on practical optimization methods that often differ from the algorithms studied in classical analyses: they train submodels, clip or normalize gradients, use layer-wise...
For safety of the intended functionality (SOTIF) analysis according to ISO 21448, a machine learning object detector for light detection and ranging (LiDAR) point clouds can meet a performance target,...
Connected and automated vehicles could allow reassigning lane directions and speed limits on every road frequently, but the benefit hinges on how traffic reroutes, and one-way roads can isolate juncti...
Text-steerable visual encoders condition a frozen Vision Transformer on a natural-language prompt and point at the matching image region, producing prompt-conditioned heatmaps and features. We audit t...
Automatic red teaming increasingly relies on large language models (LLMs) to iteratively diagnose failed attacks and propose improved adversarial prompts or injections. Although prior work strengthens...
Despite the remarkable capabilities of large language models (LLMs) across diverse applications, they remain vulnerable to generating content that violates safety regulations and policies. To mitigate...
An action seen once and the same action seen three times receive the same action-recognition label, so that supervision does not require a model to count. Answers shared across histories that differ i...
Grasping is a foundational capability in robotic manipulation. Scene-level 6-DoF grasp synthesis relies on point clouds for metric geometry, yet real-world point observations are often noisy, incomple...
WiMoT is a replay‑free continual Wi‑Fi sensing framework that stores per‑class feature means and covariances instead of raw samples, allowing it to adapt to new rooms or users while preserving earlier knowledge. It restricts representation changes to an affine map on a frozen feature basis, ensuring stored statistics remain valid without old data. Experiments on 3DO, WiMANS, and Widar3.0 show WiMoT achieves high final accuracies (up to 95.16%) with minimal forgetting (≤0.58 pp) and outperforms 21 baselines, including those that use replay, while keeping memory usage low by compressing covariances to rank‑32 factors.
The paper investigates how large language models (LLMs) can be distributionally aligned with human output distributions, particularly for applications like survey simulation. It examines both the alignment process and the models’ ability to generalize to unseen questions within the same domain and to related but different topics. The study finds that while LLMs can achieve alignment and generalize within a domain, cross‑topic generalization is variable and aligned models may unexpectedly lose creative capabilities, indicating unpredictable and unstable generalization properties.
PrefO2O introduces an offline-to-online latent preference transfer framework that enables zero‑shot preference‑based reinforcement learning on unseen tasks without human feedback. It learns a task‑invariant latent preference space from limited offline preference data, disentangling transferable preference structures from task‑specific variations. The method uses confidence‑weighted optimal transport to generate reliable pseudo‑preference supervision for online reward adaptation, achieving performance that matches or exceeds traditional PbRL methods trained with real human feedback across various robotic manipulation and locomotion benchmarks.
The paper "Rethinking Multi-View Fusion for Urban Representation Learning" introduces PRISM, a framework that addresses two key issues in urban data fusion: varying modality utility across regions and persistent gradient imbalance among heterogeneous objectives. PRISM employs sparse Mixture-of-Experts routing for region-specific view allocation and a PID-inspired regulation mechanism to reduce gradient conflicts during training. Experiments on three large-scale urban datasets demonstrate that PRISM achieves competitive performance while offering more stable optimization and balanced region representations.
ReMedI introduces an inference‑time refinement framework that combines a pre‑trained molecular language model with a Metropolis‑Hastings search to perform multi‑objective optimization of drug leads. The method proposes edits conditioned on the current molecule and refines them to improve properties, preserve lead similarity, and maintain scaffold integrity. Experiments on C‑MuMOInstruct and SMDD benchmarks show that ReMedI outperforms RL‑based baselines, achieving higher success rates and requiring far fewer training leads.
MG-REPA introduces Multi-Granularity Representation Alignment for generative target speaker extraction, aligning hidden states with a self‑supervised learning model at the frame level and a speaker verification model at the utterance level. Temporal normalization removes utterance‑level statistics from the frame‑level target, allowing each branch to focus on its specific granularity. On Libri2Mix, MG-REPA achieves comparable or better performance to a 2,000‑epoch baseline while requiring only 750 epochs, and ablation studies confirm its superiority over single‑granularity alignment.
ChoreoStudio is a music‑driven group choreography system that explicitly plans what formations dancers should adopt, when to transition between them, and which dancers execute each transition. It uses a music disentangler to extract rhythmic and structural features, a frozen omni‑modal model to generate a global semantic script, and a Director module that outputs centroid‑relative formation points, transition gates, and mobility allocations. These high‑level plans guide a diffusion generator that produces beat‑aligned body motion, resulting in state‑of‑the‑art motion quality, group organization, and formation dynamics on the AIOZ‑GDance benchmark.
The paper examines how large language models (LLMs) emulate voters in digital democracy and the demographic biases that arise. By analyzing three U.S. presidential elections, the authors compare a closed‑source and an open‑source LLM, finding that removing gender or education from voter profiles can significantly reduce group errors, while removing race increases errors. Common alignment tactics such as repeated queries or fairness instructions fail to mitigate these disparities and can even worsen them.
TensorCommitments (TCs) is a lightweight, tensor-native proof‑of‑inference scheme designed to enable verifiable inference for large language models (LLMs). It binds each LLM inference to a commitment using multivariate Terkle Trees, ensuring that any tampering breaks the commitment. For the LLaMA2 model, TC adds only 0.97% overhead for the prover and 0.12% for the verifier, while improving robustness to tailored attacks by up to 48% compared to prior work that requires a verifier GPU.
The paper introduces MAGIC, a method for improving test-time robustness of large vision‑language models by mitigating language bias. MAGIC evaluates decision margins to selectively activate textual and visual perturbation branches, integrates logits from multiple perturbed views, and restores the most informative corrupted views for ambiguous samples. It also uses a calibration dataset to prune ineffective branches, reducing inference cost while enhancing performance on benchmark datasets.
The paper "Forging Emotions" investigates how Generative Adversarial Networks (GANs) can be used to generate images that convey or trigger specific emotions. It highlights that current datasets for high‑level concepts like emotions are inadequate and suggests incorporating psychological emotional networks into dataset construction. The study also finds that simple color cues are insufficient for affect generation, advocating for deeper analysis of GAN structure and composition to create genuinely emotive imagery.
STROKE is a query‑efficient, score‑based black‑box adversarial attack that builds perturbations from simple oriented rectangular occlusions. It optimizes these occlusions using a lightweight greedy add‑and‑refine procedure, achieving high attack success with low query counts at nominal 1–2% area settings across diverse image classifiers. Compared to existing patch‑based attacks, STROKE shows favorable performance, reduced wall‑clock attack time, and the ablation studies confirm the effectiveness of its structural primitives.