The Archive

AES AIMLA 2025 Challenge submission: contrastive and triplet learning for vocal sound imitation queries using frozen and fine-tuned encoders.

Modern vision-language models (VLMs) can support tasks such as visual question answering, captioning, and image-text reasoning. In practice, however, the data... Modern vision-language models (VLMs) can support tasks such as visual question answering, captioning, and image-text reasoning. In practice, however, the data needed to adapt these models may be distributed across institutions or organizations that cannot centralize their raw records. Federated learning provides a way to coordinate training across these data-local sites. For VLMs… Source

Lévy Attention: stochastic cross-attention for irregular time series that outputs predictive uncertainty in closed form with single-pass inference.

Measured counterfactual: 24 GPT-2 pretraining runs quantify a single training example's causal impact on model behavior via batch-level injection.

ChildSafeAds 2026 shared task: classify commercial content and legal risks in child-facing YouTube videos using SponsorBlock segments and transcripts.

Deep learning translates dynamical atmospheric circulation predictions into precipitation forecasts; predicts summer 2026 drought over central China.

Verifiable Latent Alignments (VLA): framework to detect and steer covert coordination in hidden multi-agent LLM communication via activation monitoring.

Continuous-time RL for non-Markovian control: Markovianization of multivariate Hawkes processes with convergence guarantees for jump-diffusion stochastic control.

Intel AI PCs with integrated GPUs/NPUs serve 70B LLMs via layer-wise pipeline parallelism and pre-compiled OpenVINO shards over local networks.

Geometric iterative retrieval improves neural audio codec resynthesis from RVQ tokens by treating layer hierarchy as natural continuous decomposition.

Frontier LLMs saturate on accuracy; precision (output consistency) rather than capability now determines practical advantage, reframing the metrics.

SCORE method recovers subject-invariant coordinate directions in EEG features to enable cross-subject brain-to-image retrieval without labeled calibration.

Dynamic topic modeling on 12.7B Reddit comments reveals semantic drift and discourse evolution patterns across 2006–2022.

The AI buildout shows no signs of slowing. And with hundreds of billions of dollars a year going into data centers and GPUs, compute has become the single biggest cost for anyone building AI products. But for all that spending, there still isn’t a straightforward way to put a price on compute — or for firms to hedge their exposure when the price changes. Silicon Data […]

Anchoring neural and visual representations reduces brain-to-image retrieval latency by decoding from one or few trials instead of 80-trial averages.

Gradient-boosted ensembles enable exact contrastive explanations by treating leaf values as coordinates in R^M, making feature attribution traceable to tree splits.

LLM correction-handling requires operations discipline: versioning, provenance, recurrence monitoring, and stale-rule retirement—applying systems engineering to AI stack governance.

PGFS++ improves molecular property optimization under synthesis constraints via direct reaction template and reactant representation instead of embedding prediction.

MDTIM applies masked diffusion training for time series imputation by separating missing/observed values and predicting original signal instead of noise.

SRGAN applied to EBSD battery material microstructure analysis for throughput improvement.

With a looming IPO, intense competition from Anthropic, and Chinese and open-weight rivals nipping at its heels, OpenAI has plenty of reasons to move fast. Instead, it hit the brakes. On Tuesday, the company said it had slowed the pace of some AI development while it tightened security and safeguards. That included a two-week pause in reinforcement learning training on its "latest models intended for deployment," and an ongoing delay to its "largest planned frontier RL run." The decision is a very public test of an idea AI safety advocates have pushed for for years: that companies should be w...

SIMPLE: task-conditioned multi-view in-context learner for reusable view fusion across downstream tasks.

Meta AI can create content and make suggestions based on what it “sees” on your screen. | Image: Meta Meta is launching a new Mac app dedicated to its AI chatbot. In an announcement on Wednesday, Meta says you can share your window with its AI chatbot, which can provide suggestions, answer questions, or create content based on what's on your screen. Meta AI on the Mac also supports dictation across all apps. The launch comes as Meta makes its AI chatbot more like a productivity-focused assistant to better compete with its AI rivals - many of which already have desktop apps. You can share your...

Open-MOPD benchmark diagnoses capability imbalance in multi-teacher RL distillation on SmolLM3-3B.

DistScan detects backdoors in object detection via prediction distribution shift, generalizes to scene-level attacks.

DA-WAM aligns world model latents with planning decisions for decision-informative autonomous driving prediction.

S-JEPA ablation: tests whether GMM component mapping affects encoder representations beyond probability values.

User study on AI translation readability vs. fidelity: evaluability gap when source shown but quality judgment misaligns.

Learning Random Geometric Graphs via probabilistic metric spaces for generic multivariate dataset representation.

SPK decodes latent priors from object detectors to detect OoD hallucinations in real-time detection.