The Archive

Pentesting agents evaluated on real-world targets show current benchmarks miss complexity and strategic decision-making required in practice.

MMVIAD introduces first multi-view video dataset for industrial anomaly detection with continuous 2-second inspection clips.

Framework enables visual-native multimodal search agents with on-policy data evolution and persistent visual evidence reuse.

SLIM uses sparse autoencoders to steer LLM hidden states for interpretable and controllable molecular property editing.

Combines NeRF and diffusion models for probabilistic 3D scene reconstruction via latent posterior sampling.

i knew it

Long-context LLM performance degrades nonlinearly with misleading information proportion, critical for RAG and agentic systems.

NoRIN applies nonlinear Johnson transform to time-series normalization, extending RevIN to reshape heavy-tailed distributions.

SensorFault-Bench stress-tests cyber-physical forecasting models under sensor noise, bias, and misalignment faults.

MaD Physics benchmarks agents on resource-constrained scientific discovery with real measurement trade-offs and planning.

ALAM extracts action priors from unlabeled video via algebraically consistent latent codes to improve vision-language-action robot models with limited action data.

Fourier embeddings represent periodic signals in high dimensions to improve angular encoding for ML models.

CLEF, a Transformer-based EEG foundation model using multitaper spectrograms, aligns clinical signals with neurologist reports via contrastive learning on 234-task benchmark.

Novel policy gradient framework jointly optimizes agent state dynamics and control policy for non-Markovian reinforcement learning without fixed state assumptions.

Mechanistic study probes cross-modal information flow and processing dynamics between audio and video in audio-visual LLMs.

NanoResearch co-evolves agent skills, memory, and policy to enable personalized research automation for heterogeneous user needs.

Joint spectral radius analysis provides convergence guarantees for deflated Q-value iteration in discounted Markov decision processes.

Label-free benchmark evaluates equation-suffix prediction via next-token likelihood scoring to test for shortcut vulnerabilities in technical language models.

Language generation task minimizes cumulative invalid outputs during learning via mistake-bounded generation framework.

Empirical evaluation compares domain-adapted and general-purpose LLMs/SLMs for structured threat modeling in cybersecurity.

Freelance programmer reflects on how AI coding assistants have made previously difficult tasks feel easy, raising questions about developer skill assessment.

Stockholm café deploys AI system to manage operations; human-interest feature.

For the first time, Google says it has spotted and stopped a zero-day exploit developed with AI. According to a report from Google Threat Intelligence Group (GTIG), "prominent cyber crime threat actors" were planning to use the vulnerability for a "mass exploitation event" that would have allowed them to bypass two-factor authentication on an unnamed "open-source, web-based system administration tool." Google's researchers found hints in the Python script used for the exploit that indicated help from AI, like a "hallucinated CVSS score" and "structured, textbook" formatting consistent with LL...

Anthropic's Claude Platform reaches general availability on AWS with managed agents, code execution, web search, batch processing, and same-day feature parity with native API.

Reddit user requests Anthropic maintain Claude Sonnet 4.5 for creative writing use.

Shopify's internal coding agent River enforces public Slack channels to enable collaborative code review and organizational learning at scale.

Solo creator built hurricane PSA entirely with AI over weekend; studio damaged during storm.

Reddit user requests Anthropic release Claude 4.5 Sonnet, context unclear without full thread.

Reddit discussion asking for recommendations on best 3B parameter open-weights model; no definitive answer or new information.

An interactive visualisation of Jensen–Shannon divergence - the symmetric, always-finite cousin of KL. Shape two distributions and watch JSD, its ceiling of one bit, and the per-point contribution respond in real time. https://robotchinwag.com/posts/jensen-shannon-divergence-visualisation/ Feedback welcome.