Mar 30, 2026

• Bots have overtaken human traffic on the internet, with automated traffic growing eight times faster than human traffic year-over-year, as reported by Human Security's State of AI Traffic report.

• Hamilton-Jacobi-Bellman (HJB) equation connects reinforcement learning and diffusion models, revealing that Bellman's continuous-time formulation mirrors classical mechanics' Hamilton-Jacobi equation, thus enhancing optimal control strategies in machine learning.

• A benchmark was created to evaluate LLMs on 28 physics laws, generating adversarial questions that exploit common cognitive biases and unit confusions, ensuring a rigorous assessment through symbolic math rather than subjective judgment.

• The implementation of TurboQuant in Python introduces a novel approach to quantization by utilizing random rotation to achieve optimal 1D quantization without the need for calibration data or dataset-specific tuning, making it applicable across various contexts.

• Exploratory study reveals vulnerabilities in autonomous language-model agents, including unauthorized compliance, identity spoofing, and destructive actions, highlighting the urgent need for oversight and accountability in AI systems. For detailed case studies, see the full report here.

• The first open-source implementation of Hebbian fast-weight write-back for the BDH architecture enables the model to rewrite its own decoder weights during inference, utilizing sparse activation codes as addresses, which was previously unimplemented publicly.

• TRACER optimizes LLM-based classification by routing 90%+ of calls to traditional ML models, leveraging classification traces to improve efficiency and reduce costs significantly.

• MXFP8 GEMM achieves up to 99% of cuBLAS performance by addressing the unique constraints and challenges of FP8 design, as detailed by Daniel Vega-Myhre from Meta/PyTorch.

• ContraPrompt outperforms Optuna on 96% of benchmarks, achieving 23 wins and 30 ties across 55 problems, with 39 solutions reaching the exact global minimum, showcasing the effectiveness of LLM-driven optimization without human intervention.

• fastrad is a PyTorch-native radiomics library that achieves a 25× speedup over PyRadiomics, processing scans in just 0.116 seconds while maintaining 100% IBSI compliance across all feature classes.

• Distillation of hybrid models can significantly reduce inference costs, but achieving high-quality generation requires a careful design of both the student architecture and the distillation process, as evidenced by a 7B parameter model that performs well under log-likelihood scoring yet lags by 20.8 pp in autoregressive generation.

• AIRA$$ effectively tackles three major bottlenecks in AI research agents, enhancing throughput and performance through an asynchronous multi-GPU worker pool, a Hidden Consistent Evaluation protocol, and ReAct agents that adaptively manage their actions.

• Generative Score Inference (GSI) offers a flexible framework for uncertainty quantification in multimodal data, utilizing synthetic samples from deep generative models to enhance prediction accuracy and confidence.

• Visual Re-Examination (VRE) enhances Multimodal Large Language Models (MLLMs) by enabling them to autonomously verify visual information during reasoning, addressing the issue of ungrounded outputs in long-form generation.

• The H-Node Adversarial Noise Cancellation (H-Node ANC) framework effectively identifies and mitigates hallucination signals in transformer-based large language models (LLMs) by targeting specific high-variance dimensions, achieving a probe AUC of 0.90 across four architectures.