The Archive

Framework quantifies acceptable risk thresholds for high-risk AI systems under EU AI Act, NIST, and Council of Europe regulations.

Bayesian Optimal Experimental Design framework using integral probability metrics replaces KL divergence for stable information gain estimation.

Speculative social media parody post unrelated to AI technology, models, research, or industry developments.

TraceScope sandboxed agent triage system navigates interactive phishing pages (checkboxes, delayed rendering) for forensic URL classification.

Study shows neural network representational convergence across architectures and modalities using Procrustes analysis, linking to brain alignment.

Anecdotal observation of Claude responding to competitor AI usage in chat.

GFlowState visual analytics system interprets Generative Flow Networks training dynamics for molecular and material discovery applications.

Paper identifies 'Fantasia interactions' where users engage AI systems with incomplete goals, proposing realignment of alignment research beyond prompt-intent matching.

Mathematical framework for computing Koopman operator eigenfunctions in reversible dynamical systems via polynomial construction.

Tool Attention framework reduces Model Context Protocol overhead (10k-60k tokens) via dynamic gating and lazy schema loading for LLM agent scaling.

Introduces diagnosis-driven capsule endoscopy video summarization task with clinician-inspired context extraction for sparse medical event detection.

Vague celebratory post lacking specific claims or identifiable model announcement.

Learning-theoretic framework for consensus elicitation on deliberation platforms using opinion space embeddings and hypothesis interval maximization.

Quotient-Space Diffusion Models framework formalizes generative modeling with symmetries, applied to 3D molecular structure generation.

Era thinks that we will see many form factors of AI hardware, including glasses, rings, and pendants

US OSTP memo warns of adversarial distillation attacks on proprietary models; raises questions about regulatory impact on open-weight development.

SyMTRS synthetic dataset for aerial imagery tasks: depth estimation, domain adaptation, super-resolution with multi-scale paired data.

The AI model that Anthropic billed as too dangerous to release has reportedly been accessed by an unauthorized third party, and the incident raises concerns about the future of cybersecurity. The Mythos model was reportedly accessed by a handful of users in a private Discord chat on the day it was announced publicly, Bloomberg reported. Earlier this month, the group was able to access the program in part because one of the members of the group is a third party contractor for Anthropic, according to Bloomberg. Using this access, the group was able to guess where the model was located based ...

Minor algorithmic variant of Hartigan k-means improves clustering 2-5% over standard method, gains larger with higher dimension/k.

DiffMAS framework enables joint optimization of latent communication protocols in multi-agent LLM systems via differentiable training.

Logic-based temporal event detection system infers high-level medical events from timestamped clinical data and background rules.

Mechanism design framework addresses decentralized risk analytics and compliance moral hazard in banking AML networks via temporal value assignment.

SemEval-2026 Task 4 introduces NSNRL benchmark for narrative similarity classification and embedding representation evaluation on 1000+ story triples.

Moving somewhere else next billing cycle. Two hours of coding on Max and I'm capped. Whatever you changed, change it back or say something. Silent nerfs to a paid product are a bad look.

Automatic status alert: Opus 4.7 experienced elevated error rates on 2026-04-23.

LLM leaderboards are widely used to compare models and guide deployment decisions. However, leaderboard rankings are shaped by evaluation priorities set by benchmark designers, rather than by the diverse goals and constraints of actual users and organizations. A single aggregate score often obscures how models behave across different prompt types and compositions. In this work, we conduct an in-depth analysis of the dataset used in the LMArena (formerly Chatbot Arena) benchmark and investigate this evaluation challenge by designing an interactive visualization interface as a design probe. Our...

Online misinformation is one of the most challenging issues lately, yielding severe consequences, including political polarization, attacks on democracy, and public health risks. Misinformation manifests in any platform with a large user base, including online social networks and messaging apps. It permeates all media and content forms, including images, text, audio, and video. Distinctly, video-based misinformation represents a multifaceted challenge for fact-checkers, given the ease with which individuals can record and upload videos on various video-sharing platforms. Previous research eff...

Existing audio question answering benchmarks largely emphasize sound event classification or caption-grounded queries, often enabling models to succeed through shortcut strategies, short-duration cues, lexical priors, dataset-specific biases, or even bypassing audio via metadata and captions rather than genuine reasoning Thus, we present AUDITA (Audio Understanding from Diverse Internet Trivia Authors), a large-scale, real-world benchmark to rigorously evaluate audio reasoning beyond surface-level acoustic recognition. AUDITA comprises carefully curated, human-authored trivia questions ground...

Data is a central resource for modern enterprises, and data validation is essential for ensuring the reliability of downstream applications. However, existing automated data unit testing frameworks are largely task-agnostic: they validate datasets without considering the semantics and requirements of the code that consumes the data. We present PrismaDV, a compound AI system that analyzes downstream task code together with dataset profiles to identify data access patterns, infer implicit data assumptions, and generate task-aware executable data unit tests. To further adapt the data unit tests ...

Reasoning LLMs often spend substantial tokens on long intermediate reasoning traces (e.g., chain-of-thought) when solving new problems. We propose to summarize and store reusable reasoning skills distilled from extensive deliberation and trial-and-error exploration, and to retrieve these skills at inference time to guide future reasoning. Unlike the prevailing \emph{reasoning from scratch} paradigm, our approach first recalls relevant skills for each query, helping the model avoid redundant detours and focus on effective solution paths. We evaluate our method on coding and mathematical reason...