The Archive

AutoMat benchmark evaluates LLM agents on reproducing computational materials science findings, requiring domain knowledge and result interpretation beyond code quality.

Workflow decomposes statistical chart generation into screening, synthesis, rendering, and validation-driven refinement with aligned artifacts for LLM training.

More evidence of Grok CSAM seen as Minnesota passes nudifying app ban.

RunAgent multi-agent platform executes natural-language plans with constraint-guided execution and explicit control constructs (IF, GOTO, FORALL) for structured workflows.

Security assessment of patient-facing medical RAG chatbot reveals backend exposure risks, with governance lessons for safe clinical AI deployment.

Unsupervised low-dose CT denoising framework using Cycle-GAN-inspired deep learning to reduce noise while minimizing radiation exposure.

LightKV reduces LVLM KV cache memory overhead by exploiting vision-token embedding redundancy via cross-modality message passing during prefill.

SAVGO RL algorithm embeds state-action pairs with cosine similarity to shape policy updates, improving sample efficiency in continuous control.

Stratechery weekly commentary on Amazon, AI strategy, AR device futures, and Beijing tech policy (April 2026).

GeoContra framework verifies and repairs LLM-generated GIS Python code by enforcing geographic contracts including coordinate semantics, topology, and spatial predicates.

Biomechanical case study analyzes gait dynamics under occlusal constraint in Parkinson's patient, showing performance metrics don't fully reflect system organization.

LASE framework improves multilingual voice cloning speaker encoders for cross-script identity preservation in Indic languages using language-adversarial training.

Directed Social Regard (DSR) NLP method detects mixed pro-social and anti-social sentiments with fine-grained targets in online text, improving on binary sentiment tools.

Theoretical analysis of local attention in transformers characterizes expressivity trade-offs between computational cost and model quality versus global attention.

Reddit user complains about Claude Pro $20 tier rate limits and service degradation, considering upgrade to $100 plan.

Reddit discussion on ML conference peer review variability: strong papers consistently accept, weak papers reject, middle-tier papers vulnerable to reviewer mismatch and capacity constraints.

K-Shapley value extension for meritocratic fairness in budgeted combinatorial bandits with full-bandit feedback, applied to arm contribution attribution.

DeepONet-based neural operator learns solution to 2D Helmholtz equation on non-parametric domains with arbitrary scatterer geometries using signed distance encoding.

OpenAI removed AGI trigger clause from Microsoft deal, replacing it with 2032 date limit; enables multi-cloud licensing but signals shift away from founding governance principle.

Themis introduces multilingual code reward model framework and benchmark (Themis-CodeRewardBench) for multi-criteria code generation scoring beyond execution feedback.

NonZero scales cooperative multi-agent MCTS via interaction-guided exploration over low-dimensional representation instead of full joint-action space expansion.

The deals come as the DOD has doubled down on diversifying its exposure to AI vendors in the wake of its controversial dispute with Anthropic over usage terms of its AI models.

Quantum interval bound propagation (QIBP) method adapts certified adversarial training from classical ML to quantum neural networks using bound tracking.

I’ve noticed 2 things recently (even on 4.6). Time amnesia: 1. It used to be so good at understanding what the current day is and how far away a certain upcoming event is. Now even after a specific meeting it had in memory has passed it says it is upcoming and tries to get me prepared. And if I start a chat on a day of travel or at night or anything it has context on, it will forever think it is still that night or day. Pushing user to rest or not “spiral”: 2. The push to “rest”, refusing to give information more than once, or “it’s late, you’re spiraling” (even when it is incorrect and...

Cybersecurity was already under strain before AI entered the stack. Now, as AI expands the attack surface and adds new complexity, the limits of legacy approaches are becoming harder to ignore. This session from MIT Technology Review’s EmTech AI conference explores why security must be rethought with AI at its core, not layered on after…

Position paper argues agentic AI orchestration layers should use Bayesian decision theory for handling uncertainty in tool selection and resource allocation.

Technical paper on randomized-subspace Nesterov acceleration for first-order optimization with low-dimensional projected gradients.

Empirical study on EHR data windows for predicting hospital readmissions after joint arthroplasty using structured and unstructured clinical notes.

Framework to assess when LLMs should call external tools via decision theory, focusing on web search tool use and knowledge integration.

Federated multimodal unlearning approach (EASE) addressing cross-modal entanglement in decentralized image-text model training.