The Archive

Multilingual whole-brain encoding study confirms LLM-brain alignment for language comprehension across Mandarin, English, French.

Qwen 3.6 35B MoE benchmark on RTX 5080: 56 tok/s at 128k context; Multi-Token Prediction offers no speed gain at scale.

OpenAI announces multi-year partnership in Singapore for AI deployment, talent development, and enterprise/public sector adoption.

Equivalence between Gaussian processes and linear diffusion models enabling likelihood-guided conditioning beyond conjugate settings.

Data valuation, the task of quantifying the contribution of individual data points to model performance, has emerged as a fundamental challenge in machine learning. Game-theoretic approaches, such as the Banzhaf value, offer principled frameworks for fair data valuation; however, they suffer from exponential computational complexity. We address this challenge by developing efficient algorithms specifically tailored for computing Banzhaf values in $k$-nearest neighbor ($k$NN) classifiers. We first establish the theoretical hardness of the problem by proving that it is \#P-hard. Despite this in...

Reddit anecdote about Claude failing a task; no substantive technical content.

Utilizing LLMs for automated taxonomy construction presents a clear opportunity for the comprehensive, yet efficient mapping of potentially complex domains. When contending with high volumes of rapidly growing corpora, however, it becomes unclear how to best leverage such data for optimal taxonomy construction. Taking the case of systematizing AI skills in the workplace, we use two large-scale job postings corpora to investigate key design decisions for the inclusion (or exclusion) of data points for taxonomy construction. We propose TaxonomyBuilder as a blueprint for our systematic study, wi...

DySink proposes dynamic frame caching for long-form video generation, replacing static early-frame anchors with adaptive context selection to reduce bias from outdated visual cues.

System extends Text-to-SQL LLMs with agentic capability for governed enterprise APIs, handling complex business logic, auditability, and non-technical user access to analytics.

Figure AI's 24/7 livestream showcases human soft spot for humanoid robots.

Off-the-shelf persona steering vectors reduce model sycophancy as effectively as targeted Contrastive Activation Addition, lowering agreement-bias to 9–68% without sycophancy-specific training.

Theoretical analysis of concentration bounds for stochastic approximation under heavy-tailed Markovian noise with heterogeneous step sizes and operator types.

DABS framework reduces redundant computation in aspect-term sentiment analysis via single-pass depth-selective reading of a shared Transformer representation.

Hybrid ML-physics model for forest height estimation from TanDEM-X interferometry, extending feature selection to resolve structural ambiguities in remote sensing data.

PG-DPO replaces Bellman recursion with Pontryagin Maximum Principle to enable RL under non-exponential discounting found in human preferences.

Context-invariant safety alignment framework enforces LLM refusal behavior independent of prompt surface form, using verifiable and noisy feedback selectively.

Generative model using latent Gaussian processes and optimal transport for temporal inference from scRNA-seq snapshot data.

PAC-Bayes analysis of Transformer generalization on boolean functions via Fourier spectra, showing sparse low-degree targets enable flat minima.

Reddit post about humorous courtroom moment from unspecified trial; no substantive AI/tech content.

DODOCO benchmarking tool diagnoses AlltoAll dispatch bottlenecks in MoE expert parallelism across five architectures, challenging assumptions about routing imbalance correction.

Point Cloud Sequence Encoding enables Graph Network Simulators to infer material parameters from observed scenes without explicit mesh access.

Proactive federated learning client selection reduces computational waste under non-IID data by filtering low-quality participants before aggregation.

Stratechery analysis of Google I/O's AI announcements and tension between DeepMind research priorities and Google's commercial strategy.

Comparative study of CNN architectures (VGG16, ResNet50, EfficientNetB0, XceptionNet) for GAN-generated image detection finds VGG16 highest accuracy.

ArPoMeme introduces 7,300 Arabic political memes dataset labeled by ideology (Leftist, Islamist, Pan-Arabist, Satirical) for multimodal polarization analysis.

Inter-layer visual attention discrepancy method reduces hallucinations in LVLMs by detecting insufficient attention to correct visual evidence during generation.

JobArabi corpus contains 20,528 Arabic job announcements from X (2024-2025) with gendered and dialectal recruitment language variants.

Leakage-aware deployment audit for conformal triage detects safety-critical release-side failures under prevalence shift in clinical settings.

SPpruner applies focus-then-context visual token reduction in VLMs, reducing inference cost while preserving subject-centric and contextual relationships.

Memory Grafting scales language model pre-training capacity via offline frozen hidden states from grafting models as conditional n-gram memory.