In briefMeasure Less, Know More: Self-Supervised Test-Time Feature Acquisition
SummaryThe paper introduces ECHO-k, a task-agnostic, self-supervised method for choosing which modalities to measure at test time when acquiring all features is costly. It uses a deep model's pretrained internal representations as proxy targets and provides theoretical guarantees in a linear setting, which motivates a reinforcement learning policy for sequential selection. Across task-agnostic and label-free baselines, ECHO-k consistently improves budgeted downstream performance across diverse foundation-model backends.
Eeshaan Jain, Linus Bleistein, Bart Deplancke et al.
In briefI2CD: Direct Image-to-Convex Decomposition for Physics-Ready Collision Shapes
SummaryPhysics simulators and motion planners need convex collision shapes, but image-to-3D generative models output dense, sometimes non-manifold meshes that require slow, brittle repair and decomposition. I2CD predicts a convex decomposition directly from a single RGB image by freezing a pretrained Hunyuan3D-2 model and training only a lightweight 38M-parameter cross-attention head. On 227 held-out objects it achieves the highest volumetric IoU among eight baselines while running 6 to 37 times faster, and its output loads directly into four simulators. A physical xArm7 experiment also shows faster planner-ready geometry for a cluttered scene.
Qian Wang, Liam Merz Hoffmeister, Brian Scassellati et al.
In briefCausal representation learning for time series with instant and lagged effects
SummaryThis paper studies causal representation learning for time-series data, aiming to recover latent states and their causal relations, including lagged effects across intervals and instantaneous effects within intervals. It gives sufficient conditions for identifiability using an observed auxiliary variable, such as time or a condition label, linked to changes in transition-noise distributions. The authors propose iCReN, a contrastive learning framework that estimates these structures. Experiments on synthetic data recover the latent states and both causal structures, and the learned representations are useful for forecasting on real-world data.
In briefCorrupted yet correct: why vision-language models resist attacks internally
SummaryThis paper studies how a vision-language model (Qwen2.5-VL-7B-Instruct) behaves under targeted adversarial perturbations. The attack drives training loss for a target caption near zero, yet free generation still outputs the original correct description. Using a two-stage PGD attack on 200 COCO images, the authors localise this train/inference gap to a single autoregressive step in the language decoder, using logit-lens and layer-wise analysis.
WhyNovel mechanistic account of a train/inference gap via logit lens and layer-wise analysis, with significance tests on 200 images, though sample size is modest and the probe has a flagged circularity c
Arun Josephraj Arokiaraj, Chenyi Huang, Zekun Wu et al.
In briefMachine-learned absorption spectra: electron density vs. molecular geometry as model inputs
SummaryThis study compares using the ground-state electron density versus molecular geometry as inputs to a machine-learning model that predicts molecular absorption spectra. The training set contains 6874 molecules from the QM7 dataset, with spectra computed by linear-response TDDFT and densities computed by DFT. The density-based convolutional neural network reached a validation correlation of 0.9926, versus 0.9795 for the best geometry-based graph model, reducing residual decorrelation by about 64%.
Siddharth Dhanpal, Peter Elliott, Paolo Emilio Trevisanutto et al.
In briefChromaGS: Text-Driven Recoloring of Animatable 3D Head Avatars
SummaryThis paper presents ChromaGS, a method that lets users recolor animatable 3D Gaussian head avatars in real time using natural language. Edits are applied at render time without retraining. The method splits colors into region-level base colors and per-Gaussian residuals, so changing a region's color keeps fine appearance details.
In briefEvent-image monocular depth estimation using iterative refinement guided by depth hypotheses
SummaryThis paper proposes HypoDepth, described as the first iterative refinement framework for monocular depth estimation that combines event camera and image data. It introduces a discrete Depth Hypothesis Volume, turning depth regression into a constrained search, and uses multi-scale correlation to guide residual optimization. The authors report state-of-the-art results on DSEC and MVSEC, strong zero-shot generalization, and a tiny model that runs in real time on resource-limited devices.
In briefThe Shape of Speech: A Geometric Measure of Coarticulation in Speech-Driven 3D Facial Animation
SummaryThe paper proposes a geometric measure of coarticulation, how neighboring sounds shape lip movement, for speech-driven 3D facial animation. It compares lip-path length with the shortest route through vowel and consonant positions. All four tested state-of-the-art methods produce flatter lip trajectories than real speech. In a pre-registered study with 97 viewers, real speech was preferred in 73.4% of sentence comparisons.
In briefTrust but Verify: Redesigning Noise for Discrete Diffusion Language Models
SummaryThis paper addresses discrete diffusion language models that fully trust corrupted context during generation. It combines absorbing and uniform-mixing corruption kernels with region-dependent rates, exposing the model to uncertain context during training. At inference, a heuristic sampler re-injects noise into suspicious context and freezes context updates late in generation. Finetuning a 170M-parameter masked diffusion model improves detection of uniform-random corruption and raises infill token-F1 at 10% and 20% corruption levels.
In briefOptiSelect: How the Optimizer Shapes Data Selection
SummaryThis paper studies how the choice of optimizer affects online data selection during LLM pretraining. The authors propose OptiSelect, a framework linking selection gains to how well optimizer-induced utility scores discriminate between candidates. They prove that sign-based and polar-tangential preconditioners (Lion, Muon) suffer discriminability collapse, while diagonal-adaptive optimizers (AdamW, Sophia) achieve better bounds. Pretraining experiments on 124M and 720M models are reported to be consistent with the theory.
Simin Fan, Alireza Abdollahpoorrostam, Martin Jaggi
In briefJumping the Line: Gaming Length Predictions in LLM Scheduling
SummaryThis paper presents JIL, an attack on LLM schedulers that rely on predicted output lengths to prioritize requests. By optimizing an adversarial suffix, the attack makes a lightweight length probe underestimate request size, gaining higher priority. Experiments across two datasets and four LLMs show predicted lengths drop by up to 83.4 percent and adversarial requests finish up to 1.53 times faster on average. The paper also notes a trade-off with response quality and finds that grouping predictions into coarse intervals reduces the attack's advantage.
In briefHow suspicion forms across borders as AI drives financial surveillance
SummaryThis paper examines how suspicion, a key concept in anti-money laundering and counter-terrorist financing, is changing as AI becomes part of financial surveillance. It argues that suspicion is increasingly produced by data-driven processes rather than only by human legal judgement within specific jurisdictions. The author introduces “algorithmic extraterritoriality,” a regulatory power exercised through data infrastructures rather than formal jurisdictional claims. The paper notes that treating individuals as datafied subjects of suspicion makes accountability and contestation harder.
In briefOuroReward: A Sequential Reward Scheduling Method for Reinforcement Learning in Text-to-3D Generation
SummaryThis paper addresses reinforcement learning for text-to-3D generation, where existing methods combine multiple quality rewards at once without modeling how they depend on one another, causing imbalanced optimization and interference. The authors propose OuroReward, which estimates pairwise dependencies, builds a cyclic path minimizing cumulative interference, and adaptively advances through rewards based on remaining optimization headroom. They also introduce AdaSelect to choose reliable prompts matched to the model's current capability for training stability. Experiments across different models and RL algorithms show consistent improvements in generation quality across multiple dimensions.
In briefCLIMB: Confidence-guided complementary evidence for multimodal retrieval-augmented generation
SummaryThis paper proposes CLIMB, a training-free inference-time framework for multimodal retrieval-augmented generation in knowledge-intensive visual question answering. It builds a compact complementary evidence pool that balances relevance and redundancy, then accepts answer updates only when an evidence-grounded confidence estimate increases. The authors report consistent improvements over retrieval-augmented multimodal baselines on Encyclopedic-VQA and InfoSeek.
In briefDeep Bayesian REFoCUS: Recovering Ultrasound Images from Incomplete Multistatic Data With a Learned Prior
SummaryThe paper frames ultrasound multistatic recovery from arbitrary transmit sequences as a Bayesian inference problem. It trains a deep generative prior on multistatic datasets to handle rank-deficient cases where classical linear REFoCUS decoders fail. The method outperforms linear baselines across rank-deficiency and noise levels, reduces to linear decoding when inversion is exact, and expresses uncertainty in the acquisition null space. It also shows strong generalization from simulation to in-vivo data without fine-tuning.
In briefChecking whether epistemic uncertainty in graph node classification shrinks as information grows
SummaryThis paper proposes a statistical framework to test whether epistemic uncertainty in node classification decreases as more information about the data-generating process becomes available. The authors argue that existing graph evidential deep learning methods regulate this uncertainty through model hyperparameters and therefore fail this consistency requirement. They propose graph bootstrap ensembles, which capture data and procedural uncertainty via graph resampling and randomized training, and report greater epistemic uncertainty reduction than standard deep ensembles under the same protocol.
Emma Meneghini, Francesco Ferrini, Bruno Lepri et al.
In briefIterated Consistency Models: Stability, Error Bounds, and Noise Schedules
SummaryThis paper analyzes the theoretical error of multistep sampling with consistency models (CMs), treating it as a composition of noising and approximate denoising operators. Under explicit, verifiable stability assumptions, it derives a non-asymptotic error bound that separates contraction of initialization error from accumulation of approximation error. The bound shows that large early noise levels drive contraction while small late noise levels control residual bias, with explicit constants for strongly log-concave and semi-log-concave targets. Experiments indicate that the contraction and approximation profiles can be reliably measured and closely match the predicted functional forms.
Alessio Spagnoletti, Abdul-Lateef Haji-Ali, Andr\'es Almansa et al.
In briefAIBL: Instance-based learning with structured memory and neural embeddings
SummaryThis paper introduces AIBL (Augmented Instance-Based Learning), which replaces the symbolic matching of Instance-Based Learning Theory with similarity in a learned vector space for high-dimensional sequential data such as text and images. Memory is organized into active, forgotten, and surprise stores, and an observation-driven graduation algorithm promotes recurring surprise instances to active memory. The authors evaluate the model on five machine learning tasks and three simulation tasks and report accuracy gains of 6 to 17 percentage points.
In briefContrastive EEG embeddings reveal individual traits beyond conversational role
SummaryThis paper applies the contrastive learning method CEBRA to EEG recorded from conversing dyads and analyzes the resulting embedding constrained to a 2D sphere. Labels such as the absolute difference in partners' autism-quotient scores decode above chance, but only the control that retrains the encoder under each label permutation tests the label itself, and it is not significant (p = 0.50). The embedding is organized by individual identity rather than autism trait differences, and speaker-versus-listener role analysis performs at chance. The authors suggest retraining-based nulls should be the default for grouped-data contrastive embeddings.
Hubert Huang, Michelle McCleod, Brendan Ames et al.
