The Archive

RoSHAP proposes distributional framework and robust metric for stable feature attribution rankings amid stochastic variation.

Pelican-Unified 1.0 is unified embodied foundation model using single VLM for understanding, reasoning, and action generation.

Quantization-conditioned attack exploits outlier injection to induce malicious behavior in LLMs via advanced quantization schemes.

Post-training quantization systematically reverses unlearning in LLMs; per-parameter forgetting updates fail under 4-bit compression.

Method to predict deployment-scale failure rates from limited evaluation sets using extreme-value extrapolation; quantifies inherent over-prediction bias.

First causal foundation model handling continuous treatments via meta-learning; extends beyond binary intervention setting.

APWA architecture parallelizes LLM-based multi-agent workflows to overcome reasoning and computational bottlenecks in complex tasks.

An interactive map tracking data center construction and AI policy, built by Isabelle Reksopuro. When Oregon resident Isabelle Reksopuro heard Google was gobbling up public land to fuel its data centers in her home state, she didn't initially know what to believe. "There's a lot of misinformation about data centers," she said. "Google has denied taking that land." Technically, she explains, The Dalles, a city near the Washington state border, sought to reclaim that land, "and Google is just a big, unnamed power user." The city had in fact asked for ownership of a 150-acre portion of Mount Hoo...

Neuro-symbolic approach couples large reasoning models with model checkers for reactive hardware synthesis via iterative Verilog repair.

MemEye benchmark evaluates multimodal agent memory preservation of visual evidence across granularity and temporal-change dimensions.

Survey of 60 international students on adoption of conversational AI chatbots for cross-cultural adaptation support.

CoCo-InEKF uses learned continuous contact covariances instead of binary states for robust legged-robot state estimation in dynamic contact scenarios.

Reddit post claiming major AI breakthrough in medicine/drug discovery; link-only with no substantive detail provided.

CLOVER closes training-evaluation mismatch in end-to-end autonomous driving by learning value estimates aligned with rule-based planning metrics.

Taxonomy and audit framework for LLM attack benchmarks using 4×6 Target×Technique matrix; reveals coverage gaps in HarmBench, InjecAgent, AgentDoor.

Made-up therapy referrals, incorrect prescriptions among the common mistakes.

User reports positive experience running local LLMs on NVIDIA RTX 5000 PRO 48GB GPU versus Mac Studio alternative.

Variational Policy Distillation addresses sparse rewards in RL by using adaptive language feedback to overcome teacher plateau and improve reasoning task performance.

Statistical string similarity features using co-occurrence and run-length matrices generalize across languages without linguistic information.

Agentic GraphRAG citations require trajectory-level faithfulness validation that accounts for graph traversal structure and uncited nodes influencing answers.

Off-policy evaluation logging policy design balances reward concentration and action coverage to minimize estimation error for deployment-free experimentation.

Dataset-agnostic audio framework converts text-based tool-calling benchmarks (Confetti, When2Call) to voice evaluations using TTS and noise without re-annotation.

Self-Recall Thinking framework improves multi-turn dialogue consistency by tracking non-adjacent turn dependencies without full-history context overhead.

Dual-Dimensional Consistency unifies sampling width and depth for inference-time LLM scaling, balancing reasoning quality against budget constraints.

Reddit appreciation post for Andrej Karpathy's influence on open-source AI development; no new technical content or announcements.

AI framework interprets petrochemical refinery LP optimization outputs using historical data analysis to validate and improve solver-generated decisions.

Dynamic Batch-Sensitive Adam optimizer improves convergence on imbalanced sequential datasets by scaling learning rates via batch difficulty scores.

Kernel ridge regression with Average Gradient Outer Product provably recovers low-dimensional central subspaces in multi-index models with sample complexity below prediction threshold.

ML-Embed introduces 3D Matryoshka Learning framework for efficient, multilingual text embeddings addressing computational cost and linguistic coverage gaps.