The Archive

The brand trip is a right of passage for influencers. It's a mark of legitimacy that a sponsor wants to invite them on an all-expenses-paid vacation, often with luxurious freebies and activities. Trips can also spur hard feelings from uninvited influencers, trigger criticism from the public, and project a certain frivolousness. Usually it is fast fashion companies or beauty brands flying out influencers with the goal of flooding social media with sponsored content. This time, it was a major AI company. Over the weekend, several influencers began posting about OpenAI's "first ever brand trip,"...

Video-DeepResearch extends multimodal agents to video streams; identifies modality bias and knowledge leakage bottlenecks in current models.

ReflectRL improves LLM reasoning by learning from failed expert trajectories as reflective signals during on-policy post-training.

HIVE benchmark reveals voice transcription and keyboard perturbations reduce LLM accuracy across instruction tasks.

Information-geometric analysis of GFlowNet forward-policy training using Fisher-Rao metric and natural gradient optimization.

Hank Green, a popular YouTuber and science communicator, said he is stepping back from production amid intense criticism over his use of AI. Green described his AI usage as "not healthy," but stressed that he used it for finding research sources and not to write scripts. Much of the ensuing firestorm in this corner of the internet has centered on how a creator can square a brand built on authenticity and credibility with a technology trained on the (often uncompensated) works of others, and which has a well-known tendency to generate plausible-sounding falsehoods. Some attention has fallen on...

HalluTruthQA-4K: 4,000 expert-annotated Arabic QA instances for fine-grained hallucination detection across four knowledge domains.

Unconditional separations prove low-depth quantum circuits compute tasks intractable for constant-round diffusion and transformer LMs.

Adaptive test-time scaling via fuzzy controller allocates per-query compute based on prompt complexity and model confidence.

Game-theoretic analysis shows foundation model agents achieve rational cooperation through similarity inference, challenging decoupled-agency assumptions.

TACT: taxonomy-aligned post-training framework for pedagogically adaptive LLM-based ESL tutoring with human-grounded evaluation.

Release: llm 0.32 See my detailed blog post about this release . Tags: llm

Muon optimizer shows token-efficiency gains on Mamba-2 when applied selectively to output projections, with benefits localized to specific weight groups.

Logic pre-pretraining on formal derivations accelerates LM skill acquisition and improves compressibility compared to narrow symbolic tasks.

Mariano-Florentino Cuéllar joins Anthropic as first Chief Global Affairs Officer, signaling focus on policy and governance.

Latent Reward Registers enable preference alignment in Diffusion Transformers via learnable tokens that extract reward signals from intermediate noisy latents.

R-ItCUR algorithm recovers low-tubal-rank tensors from partial cross-concentrated sampling observations with sparse outlier corruption.

Physics-Flavored Neural Network automates kinetic phenotyping of engineered skeletal muscle tissues, replacing simplistic peak-force metrics.

PRISM meta-workflow converts multivariate time series to multi-channel images for anomaly detection, evaluating vision backbone alternatives.

Transformers execute Sequence-level Interactive Dynamic Parallel Processing (SIDPP), constructing prompt-dependent parameter transformations at inference rather than reproducing training statistics.

Standard music transformers lose equivariance to pitch transposition and time shifts as scale increases, allocating capacity to memorizing absolute patterns.

EcoFrame enables training-free adaptive frame scheduling for long video understanding in VLMs, using inference feedback to adjust sampling budgets.

First implementation of causal perception framework for competing Structural Causal Models, operationalizing fairness in agent reasoning systems.

Acceleration Matching algorithm for smooth trajectory inference from discrete snapshots.

Sparse Weight Decomposition enables efficient mechanistic interpretability of pretrained transformers without sparse retraining.

Social theory framework for pluralistic AI alignment recognizing multiple legitimate perspectives in diverse deployment contexts.

Lipschitz singularities in dynamic routing transformers create adversarial vulnerabilities in efficient UAV tracking systems.

ANNOTARES dataset for extracting logical structures (conditions and consequences) from German statutory texts.

KV cache transfer enables prefill reuse across different-sized LLM family models via closed-form linear mapping.

Contrastive activation addition steers temporal horizon preferences in Qwen3-32B via learned linear representations.