The Archive

Probability-Conserving Flow Guidance reformulates diffusion guidance through continuity equations to preserve learned manifold geometry under strong conditioning.

CopT reverses chain-of-thought order to draft answers before thinking, reducing token costs when LLMs solve problems without extended reasoning.

Theoretical analysis of when GNN distillation succeeds for combinatorial optimization under algorithmic alignment with problem structure.

Empirical study showing that higher-fidelity observations (RGB-D vs symbolic) degrade LLM performance on embodied reasoning tasks like mechanical puzzles.

Smooth Partial Lotteries stabilizes randomized selection processes by reducing selection probability volatility near score decision boundaries.

Tail Annealing applies soft-log transforms to heavy-tailed data before flow matching training, enabling power-law generation from Gaussian noise.

GRPO-based approach trains small Qwen3-1.7B model for zero-shot Text-to-SPARQL generation on DBLP using outcome-based RL rewards.

ReBel algorithm improves credit assignment for LLM agents in partially observable environments by explicitly modeling belief states rather than immediate rewards.

Reddit discussion seeking user opinions on Claude Mythos; unverified model name, no official source.

LLM-based agent system recovers hierarchical software architecture from ROS 2 robotic systems by analyzing distributed source artifacts.

PromptRad applies prompt-tuning with knowledge enhancement to multi-label radiology report classification in low-resource clinical settings.

Linguistic analysis shows conspiracy theories with greater semantic mutations persist longer on social media X over three-year dataset.

Minimal-pair evaluation protocol isolates code quality effects on autonomous coding agent performance independent of underlying capability.

Error analysis reveals sequence-to-sequence models struggle with Japanese morphological inflection when orthographic properties affect morphophonological distinctions.

Active context selection algorithm improves simple regret bounds in contextual bandits with instance-dependent guarantees over finite subpopulations.

Disagreement-Guided Reward Poisoning attack exploits critic disagreement to compromise Soft Actor-Critic agents in RIS-aided wireless control systems.

Reddit discussion of Microsoft economist's commentary on AI policy; lacks substantive detail or novel claims.

D³-Subsidy framework optimizes real-time driver subsidies for ride-hailing platforms balancing supply-demand while respecting rate caps and latency constraints.

CAMERA framework detects fraud in text-attributed graphs despite semantic camouflage where fraudsters mimic benign user language patterns.

Intent-controlled partial optimal transport extends OT theory with pointwise rejection mechanisms for structured mass-matching problems.

Diffusion-based emulators enable training-free, scalable particle filtering for nonlinear dynamical systems without classical solvers.

Glendale Community College president Tiffany Hernandez apologized for the mistakes and eventually offered many students a do-over. | Screenshot: YouTube The use of AI-powered tools to announce students as they walk on stage during graduation and commencement ceremonies has grown in popularity over the past few years, but it's not always succeeding at the one job it's there for. Many schools have switched to these systems as a way to ensure names are being pronounced correctly, but during a recent livestream of a Glendale Community College commencement ceremony in Phoenix, Arizona, the AI anno...

AutoResearchClaw introduces multi-agent debate, experiment cycling, and experience accumulation for iterative autonomous scientific discovery.

Empirical study shows agent skills hurt performance on 19% of CTF tasks, identifying conditions where procedural knowledge degrades autonomous cybersecurity agents.

FlexDraft enables parallel speculative decoding within single forward passes, reducing memory overhead and latency vs. sequential drafting.

Andrej Karpathy has joined Anthropic to work on pre-training. He previously co-founded and worked at OpenAI and led computer vision and AI at Tesla.

Probabilistic framework applies Bayesian optimization to learning rate selection in SGD, addressing hyperparameter tuning for neural networks.

GeoX uses self-play with verifiable program-based rewards to learn geospatial reasoning from satellite/aerial imagery without human annotations.

Loss-adaptive learning rates mitigate catastrophic forgetting during LLM fine-tuning without suppressing task-critical hard tokens.