The Archive

Covariance matching in Gaussian DDPMs improves path-KL divergence from Ω(1/T) to better scaling; improves classifier guidance stability.

Tests OpenAI, Anthropic, DeepSeek, xAI models for conflict-context failures: false atrocity equivalence, genocide denial, ethnic slur misrecognition.

Sparse autoencoder audit identifies 146 GPT-2 Small features correlating with IOI task failures; strongest predictor is 'cryptographic keys' feature.

Adapts audio diffusion models for interactive live music on consumer hardware via block-wise outpainting; avoids discrete-AR compute requirements.

Parametric modular logic programs extend ASP with subprogram parameters and intensionality; connects clingo collective control to traditional ASP.

AnyMo framework enables human motion recognition across heterogeneous wearable IMU sensors via geometry-aware setup-agnostic modeling.

Study of 75,898 API calls shows LLMs exhibit accumulated message effect bias when evaluating sequential items in single conversations.

Reddit speculation on Gorgon Halo vs Strix Halo memory bandwidth; claims 6.7% speedup insufficient for LLM inference, awaiting Medusa Halo.

Hierarchical offline GCRL framework using relativised options and representational abstraction to reuse experience across state-goal symmetries.

Qualitative study of 24 professionals at large tech firm shows AI adoption blurs role boundaries and transforms informal workplace culture.

ToaST tokenization method optimizes compression via recursive split trees and integer programming, decoupled from fixed vocabularies.

Analysis identifies hard clipping as bottleneck in RLVR training; proposes stochastic recovery of near-boundary signals to stabilize GRPO optimization.

Scout-Assisted Planning framework uses aerial drones to reduce backtracking for ground robot teams in partially known environments.

Theoretical analysis shows posterior collapse in β-VAEs implements automatic spectral pruning via Landau stability analysis.

ChronoVAE-HOPE time series foundation model replaces attention with VAE for specialized classification, addresses quadratic complexity.

CUSP benchmark evaluates AI's ability to forecast scientific progress across 4,760 events via feasibility, mechanistic reasoning, and temporal prediction.

CEDAR method disentangles vision-language model embeddings via invertible transformation without expanding dimensionality, enabling sparse interpretability.

Swift Sampling identifies high-information frames in long-form video via Taylor expansion of visual feature trajectories without training.

Self-policy distillation via subspace projection isolates task-relevant capabilities from style/formatting without external curation signals.

Just found out about this and had to share because almost nobody is talking about it yet. If you are tired of paying for AI courses or getting hit with paywalls just to get a certificate, Anthropic (the creators of Claude) quietly dropped a massive library of completely free, official training modules. Yes, they actually give you an official certificate of completion directly from Anthropic once you finish. Here is the breakdown of what is available and exactly how to get it without spending a dime. What is in the course catalog? They have split the training into a few different paths de...

Inverse scaling detected: more capable LLMs forecast worse on superlinear/regime-change time series; ForecastBench-Sim benchmark released.

Community joke post about lack of AGI announcements; no substantive technical content.

Holographic functions framework connects sampling, structural, and computational complexity bounds for Boolean functions and neural networks.

WorkstreamBench evaluates LLM agents on end-to-end spreadsheet construction in finance workflows, filling gap in agent evaluation.

Claw AI Lab platform enables multi-agent research teams with role customization, monitoring, and reproducibility dashboards.

LLM-based cross-lingual translation preserves moral semantics from English to Polish, enabling non-English moral classification datasets.

Reddit speculation about Anthropic missing a product deadline; unsubstantiated commentary.

SegCompass uses sparse autoencoders to create interpretable alignment between LLM reasoning and visual segmentation.

Image-semantic detection method enhances MLLM performance detecting AI-generated modern Chinese poetry.

Spotify and Universal Music Group (UMG) announced a licensing deal that will allow users to prompt the creation of AI-generated remixes and covers for streaming songs. The tool will be a paid add-on for Premium subscribers. Artists will be able to opt out of the program, but those who do participate will collect royalties on these AI remixes. In October of last year, Spotify announced that it was working with UMG, as well as other major labels, Sony Music Group, Warner Music Group, Merlin, and Believe, to create "responsible AI products." At the time, it was unclear exactly what that meant. B...