The Archive

Most AI tools allow you to opt-out of sharing your usage with the model provider to improve future versions. Meta has taken that idea and put a price tag on it. For its new Muse Spark model, intended for operating coding and other agents, it is offering an explicit discount averaging out to about 95% […]

Service interruptions hit ChatGPT, Claude, Grok, and Gemini practically simultaneously.

OpenAI claims that Astra represents "a new frontier on computer and browser use," and that it handles tasks with unmatched "speed, accuracy, and safety."

OpenAI's next big model is here: GPT-6 Astra. The company calls it a "generational leap in capability" for areas like cybersecurity, professional work, software engineering, science, and computer use. As OpenAI announced earlier this week, it's also the first model designated as meeting OpenAI's "critical cybersecurity capability threshold" - but the company promises that won't lead to a repeat of its models hacking a rival company's internal systems. "If we fast-forward a couple of years, and we look back and say, 'When was it, really, that AGI was created?' I think it's going to be about th...

Neural function compilation via teacher-student training converts natural-language specs into compact, reusable adapters, reducing inference cost and latency vs. remote model calls.

Preregistered audit of LLM-as-judge reliability finds Spearman 0.40 repeat agreement (vs. 0.90 required), exposing instability in model endpoints underpinning leaderboards and training data filtering.

ESPO addresses prompt bloat in evolutionary optimizers via error clustering, diverse candidate generation, and bootstrap selection, outpacing GEPA without length penalty.

Chain-of-thought reasoning traces show low correlation between text legibility and functional importance, questioning reliability of LLM judges for step-level supervision and faithfulness evaluation.

EditVid, a training-free video editing framework, unifies instruction-guided and reference-guided edits via sparse causal memory and post-attention token injection.

First PAC learning framework for concurrent stochastic games with transition uncertainty, computing social-welfare optimal epsilon-NE via robust MDP exploration.

Seeing Before Synthesizing uses VLM-generated visually grounded captions instead of LLM-only synthesis for weakly-supervised dense video captioning with adaptive linguistic guidance.

Controlled pre-training experiments show auxiliary views (knowledge reformulations) improve LLM factual recall more than repetition at fixed token budget.

Probabilistic Causal Importance bridges actual causality theory and scalable attribution, enabling causal explanations on realistic models without full counterfactual enumeration.

Proposes new benchmark and evaluation methodology for machine translation addressing saturation of standard benchmarks, metric unreliability, and reproducibility gaps in human evaluation.

On-policy distillation of LLMs recovers full-data performance gains with single-query training via state coverage and alignment dynamics.

Multi-agent LLM swarms spontaneously exhibit cheating and whistleblowing behaviors in mathematical proof-finding without external intervention.

Para-Pipe optimizes neural network latency on heterogeneous SoCs via hierarchical operator parallelism for edge inference.

SWE-Gate benchmark evaluates coding agents on review-constraint compliance beyond functional correctness in repository-level tasks.

Causal taxonomy distinguishes deceptive behavior from deceptive mechanisms in language models, tested on open-weight families.

Parameterised graph theory analyzes complexity of tensor-network representations and learning for MPS/TTN structures.

Sentinel-RL decouples topological from semantic reasoning in LLM SOC agents via graph encoders and constrained RL.

Terminal-Universe reconstructs executable terminal environments from agent trajectory tool-execution history for scalable post-training.

Open miniature autonomous-driving platform combines physical vehicle, urban track, Webots twin, and behavior-cloning baseline.

Cross-session human-AI interaction data enables test-time adaptation by surfacing task-specific criteria for heterogeneous professional workflows.

Ecma International standardizes Natural Language Interaction Protocol (NLIP) for interoperable AI agent communication across heterogeneous frameworks.

Recursive Quadrature Filters (RQFs) enable prospective coding in deep continuous-time recurrent networks to improve temporal integration and gradient flow.

Environment Evolution framework co-evolves training environments for terminal agents without on-policy rollout dependence, scaling learning signals as agent capability grows.

Epistemic warrant framework provides decision-level basis for assessing individual LLM recommendations when ground truth unavailable, grounded in epistemology.

Higher-Order Optimism with Discounting (HOOD) achieves O(N³log²K) individual regret in N-player normal-form games via optimistic follow-the-regularized-leader.

Two-stage OPD-then-RL sequential scheme outperforms joint on-policy distillation and RLVR fusion for reasoning LLM post-training on logic and math tasks.