Live intelligence

Latest AI updates

14 updates
Research

PACT: Piecewise Adaptation via Chart Transport for Incremental Time-Series Foundation Models

This arXiv paper addresses incremental adaptation of pretrained time-series foundation models (TSFMs) under shifting operating conditions, where a model must learn new conditions without forgetting earlier ones and without knowing the condition at inference time. The authors observe that embedding-space shifts are non-uniform: some local regions change substantially while others stay nearly unchanged, so global adaptation can disturb already-correct embeddings. They propose PACT (Piecewise Adaptation via Chart Transport), which partitions the frozen source embedding space into local charts, learns lightweight residual transports for regions affected by each new condition, retains an identity path, and uses a learned gate to select among the identity and accumulated transports. With 5% labeled calibration data per arriving condition, PACT reports average macro-F1 above 0.95 across a four-condition bearing-fault stream, with similar retention-adaptation behavior on public benchmarks; code is available on GitHub.

Time-Series Foundation ModelsContinual LearningDomain Adaptation
Research

SPAR Uses Action-Free Egocentric Videos to Improve Robot Policy Learning

Researchers propose Shared Prediction on Action and Representation (SPAR), an MoT architecture built on a pretrained VLM for robot policy learning without action supervision. SPAR combines future visual state prediction with action prediction as independent and complementary learning signals. Representations learned from human egocentric videos improve downstream performance on out-of-distribution robot tasks, with further gains as video data scales. The work also examines objective choices and data curation for leveraging action-free videos in scalable robot learning.

robot policy learningegocentric videosvision-language models
Research

AURA Uses a Frozen General-Audio Codec for Spatial Audio-Language Tasks

AURA investigates whether a general-audio codec trained without spatial supervision can support both sound identification and localization in multi-source scenes. The authors find that spatial cues, including inter-channel differences, remain in the frozen codec but are not readily accessible to a language model due to audio-text information asymmetry. Their controlled experiments show that scene-description completion and inter-channel difference exposure each improve left/right accuracy, while binding directions to the queried source requires both. AURA keeps the codec and language model frozen and exceeds the strongest dedicated system on SpatialSoundQA by 10.31 points in dual-source direction accuracy using about 40% of its trainable parameters.

spatial audioaudio-language modelsrepresentation learning
Research

SemPair Relaxes Cross-Modal Correspondence for Remote Sensing Pretraining

The paper challenges the assumption that multimodal remote sensing pretraining requires exact instance-level cross-modal correspondence. It introduces Semantic Pairing (SemPair), which substitutes class-consistent cross-instance pairs for exact instance pairs, and a latent-factor model separating shared semantic structure from instance-specific correlation. A controlled interpolation, Mixed Pairing (MixPair), is used to study correspondence granularity, and experiments show SemPair reaches transfer performance comparable to exact pairing while MixPair reveals a non-monotonic relationship between downstream transfer and instance-pairing ratio. The authors conclude that exact instance correspondence is not a prerequisite for effective multimodal pretraining.

multimodal learningremote sensingpretraining
Research

Matched-Pair Study Finds Earth-Observation Pretraining Is Not Uniformly Better

A new arXiv study compares five matched pairs of Earth-observation and generic vision encoders across 14 frozen models and five RGB scene-classification datasets. EO pretraining significantly helped SatDINO, Scale-MAE and RemoteCLIP, but significantly hurt DINOv3-SAT and SatMAE on linear probing, with DINOv3-SAT dropping by up to 22.5 points. Two MAEs pretrained on the same fMoW-RGB imagery produced opposite outcomes, and gains or losses were largest with 1-5% labels. The authors conclude that satellite pretraining is not a reliable selection criterion and that candidate checkpoints should be validated on target data.

Earth observationfoundation modelsrepresentation learning
Research

Good Views Are Not Enough: Global Counterexamples for Gaussian-Regularized JEPAs

The paper presents globally certified counterexamples showing that good views alone do not guarantee scene invariance or uniqueness up to rotation at global optima of Gaussian-regularized JEPAs. In a four-state model, every global minimizer splits a scene and has an equal-value partner outside its rotation orbit; a 512-dimensional model shows pooled Epps-Pulley regularization can force retention of nuisance and positive cross-modal alignment error. The authors also report that a whitened prediction JEPA aligns better to the scene than next-token prediction but has 0.2872 nats/token higher linear-readout test NLL under matched readout budgets and 1% observation noise. A Bayes-risk decomposition separates information lost through the observation channel from readout difficulty.

JEPArepresentation learningself-supervised learning
Research

Shape-Teacher Alignment Boosts Image-Based Robotic Policies with 3D Shape Knowledge

Researchers propose shape-teacher alignment, a training-time method that transfers 3D shape information from a pretrained Point-MAE model to an image-based policy encoder without changing deployment inputs. This approach improves representation structure, accelerates convergence (up to 5.6x on LIBERO suites), and increases success rates. The method requires no depth input or shape teacher at deployment.

robotic manipulationrepresentation learning3D vision
Research

PLUS: A Detector-Agnostic Subspace Signal for Near-OOD Detection

The paper introduces PLUS, a post-hoc signal that complements existing out-of-distribution (OOD) detectors to improve near-OOD detection. PLUS builds per-class low-dimensional subspaces from sparse autoencoder decoder directions and scores inputs via whitened squared Mahalanobis distance. Fused additively with any calibrated base score at a fixed global weight, it improves near-OOD AUROC across all 126 detector-backbone combinations tested, while largely maintaining far-OOD performance.

out-of-distribution detectiondeep learninganomaly detection
Research

Task-Efficient Agent Evaluation Using Cached Traces

This paper introduces a method to estimate an agentic system's benchmark performance from a small number of task executions by leveraging rich execution traces. The approach extracts structured summaries with an LLM and embeds them to create representations that preserve task-related signal while suppressing identity-related noise. On SWE-bench Verified and Terminal-Bench 2.0, combining this geometry-based prediction with an item-response model reduces mean absolute error significantly compared to baselines.

AI evaluationagentic systemsefficient benchmarking
Research

TAPES: Task-Aligned Joint Predictive Pretraining for Transaction Event Sequences

TAPES is a pretraining framework for transaction event sequences that selects predictive target sets based on downstream utility and models joint likelihood via a shared discrete latent variable. It outperforms baselines on real-world and public financial benchmarks, with ablations confirming the value of target selection and joint modeling.

transaction event sequencespredictive pretrainingfinancial benchmarks
Research

EEG Foundation Models Show Convergent Representations with Decodability

This paper tests the Platonic Representation Hypothesis on EEG foundation models, finding that models with better EEG decodability show stronger cross-family alignment in local and global structure, though not consistently with parameter count. Cross-modal convergence is modality-dependent: stronger global alignment to language but not vision, with local alignment showing no consistent association. The results suggest a dissociation between local sensory features and global semantic structure.

EEG foundation modelsrepresentation learningcross-modal alignment
Research Editor selected

DINOFAS: Bonafide-Centric Representation Learning for Face Anti-Spoofing

DINOFAS is a DINOv3-based framework for face anti-spoofing that learns bonafide-centric representations rather than attack-specific cues. It uses LoRA for parameter-efficient adaptation, an attention style-content disentanglement module to reduce domain variations, and a dual-branch architecture separating prototype-based liveness learning from attack supervision. The model achieves strong cross-domain generalization, with an average HTER of 9.09% on the OCIM protocol and 21.41% on synthetic-to-real transfer. Evaluations on multiple benchmarks show robustness to unseen acquisition conditions and presentation attacks.

Why it matters: Presents a novel representation learning approach for face anti-spoofing that significantly improves generalization across domains and attack types.

face anti-spoofingrepresentation learningdomain generalization
Research

Aligning MLLMs with Human Cognition via Behavior-Derived Semantic Dimensions

Researchers propose a framework to align multimodal large language models with human cognitive structure by using a behavior-derived semantic embedding space of 66 dimensions from 4.7 million human Odd-One-Out judgments. The method infers salient decision dimensions for each triplet and generates dimension-grounded rationales, improving consistency with human similarity judgments and increasing representational alignment in fMRI data. The approach retains competitive performance on standard multimodal benchmarks.

multimodal large language modelshuman-AI alignmentcognitive science
Research Editor selected

Berkeley AI Research: Closed-Form Theory Explains What Word2vec Learns

Researchers at Berkeley AI Research have developed a quantitative theory showing that, under mild approximations, word2vec's learning problem reduces to unweighted least-squares matrix factorization. They prove that the learned embeddings are given by PCA on a specific matrix defined by corpus statistics and hyperparameters, and that training proceeds in discrete rank-incrementing steps. The theory predicts interpretable latent features, such as celebrity, government, and geography topics, and m

Why it matters: Readers gain a rigorous theoretical understanding of how word2vec learns interpretable embeddings, bridging classical NLP and modern representation learning.

word2vecrepresentation learningembedding theory