The Archive

"Small weight-based replicators" could usher in a new generation of computer worms.

OpenAI's Amp platform enables AI agents to perform non-coding tasks using GPT-6 Astra and related models.

SentinelOne uses OpenAI Daybreak to reduce malware analysis time from weeks to hours, accelerating threat investigation capacity.

Attack method exposes flaws in OpenAI, Google safeguards against AI nudes.

Scimplify's annualized revenue run rate has tripled to $200 million as it eyes U.S. expansion.

Stratechery analyzes Apple's HomeView hardware and questions the market viability of premium devices reliant on limited AI capabilities.

Euler system enables multi-agent mathematical discovery with shared context and interpretable reasoning steps for collaborative proof development.

Theoretical analysis of optimization landscape for weighted MMD with Gaussian kernels shows local minima properties under gradient flow.

NLP-based data filtering leverages unstructured operational logs to improve industrial time-series model training quality.

Paired Fallback Robustness test benchmarks typed decision models' reliability in following fallback rules for deferral and model selection.

Comparative energy benchmark shows diffusion language models (LLaDA-8B, Dream-7B) achieve better Joules-per-token than autoregressive LLMs at scale.

DeltaStreamer combines change-aware memory and query-driven planning for streaming video understanding with evidence replay.

OpenTWBench releases 255K Taiwan exam questions and domain-specific benchmark suite; documents evaluation methodology impact on LLM scoring.

Residual correction framework reduces LLM training memory by maintaining optimizer states in low-rank subspace with orthogonal gradient recovery.

LAGO meta-agent framework treats specialized agents as reusable options; resolves 58.6% of SWE-bench tasks with frozen Qwen3.5-35B.

EGCForest multi-view hierarchical memory framework organizes enterprise chat logs for reliable cross-thread evidence retrieval by agents.

WebKD enables distillation of closed-model web agents into smaller open models by having teachers correct student failures in-context rather than offline.

Study reframes RLHF reward hacking as economic rent-seeking, quantifying proxy reward overoptimization as measurable capture and deadweight loss across model families.

CTIA-ODE applies continuous-time attention within ODE solvers for multivariate time series forecasting with exogenous variables.

MAPLE framework disaggregates knowledge-extraction evaluation beyond aggregate accuracy to measure diverse effects of steering and decoding techniques.

JointServe optimizes multi-model agent serving on memory-constrained GPUs via weight shard and KV block residency management without full model swapping.

ActionOPD distills agent behavior by supervising semantic actions rather than tokens, avoiding tokenizer incompatibility and token-level mismatch in on-policy settings.

AgenticRV formalizes runtime verification for agentic AI by decoupling monitoring boundaries from formalisms and enabling adapter-driven reaction policies.

Multi-turn evaluation of four steering methods shows inference-time value alignment techniques can handle conflicting user and model value orientations.

Analysis of transformer tabular foundation models reveals early layer predictive accessibility but quantifies per-input variation in depth benefits across layers.

VaryST generalizes spatio-temporal forecasting across shifted temporal distributions, spatial domains, and observation availability without node-specific coupling.

Risk-calibrated binary decision-making for LLM workflows using System One scoring with conservative fallbacks across moderation and reranking tasks.

ReproBAIT benchmark evaluates coding agents' ability to reconstruct biological AI tools from limited descriptions without internet access.

E2D.PMSide framework for contextual partial monitoring games using function approximation with efficient computation.

Negative guidance method for diffusion models targeting explicit density via positive-unlabeled learning to avoid undesired conditions.