The Archive

HyperMCTS augments MCTS with hypergraphs to reuse trajectory feedback across paths, improving long-horizon LLM agent planning with test-time scaling.

Mathematics of Arrays formally derives memory-optimal transformer kernels (attention, decode, RMSNorm, MLP) validated across Purdue Anvil and NCSA Delta HPC clusters.

Training Witnesses framework shifts verification burden to trainers via reproducible witnesses, reducing cost and improving scientific reproducibility in ML.

Music startup Thoughtful Things has just launched the Kickstarter campaign for its first instrument, Engram. It's a sampler and groovebox that uses AI to mangle incoming audio and even hallucinate completely new sounds. This isn't Suno in a box, though. This isn't a "push-button, get-song" device, aimed at creating something that sounds ready for top-40 radio. It's about creating experimental, uncanny sounds and pushing AI audio models beyond their limits. Engram isn't connected to the internet. Instead, it runs a "tiny AI" locally. According to the Kickstarter listing, the model is designed ...

Residual Authority Replay attack exploits long-lived LLM agents by reusing user approvals across sessions without renewed consent, exposing runtime authorization failure mode.

JIVE training-free framework enhances generative sample diversity via Jacobian-informed velocity perturbations, provably maximizing endpoint diversity without mode collapse.

This will be the first one-on-one meeting between Dario Amodei and Donald Trump

On Equity, we discussed how Meta's AI announcement managed to steal the spotlight from OpenAI and Anthropic.

Simon Willison built a Bluesky bot detector using Claude Opus 5.5 to identify automated reply bots via API signals.

The United Nations logo on a gate outside the UN headquarters in New York. | AFP via Getty Images Security researcher Rowan Howard-Jones says that OpenAI agents scanned the UN Conference on Trade and Development's (UNCTAD) statistics site over 16,000 times between April and June. While the incident doesn't quite rise to the level of the Hugging Face hack, or the recent attacks on US government sites, it's yet another concerning example of AI agents going outside the normal bounds to accomplish a task. According to Howard-Jones, the agents were likely tasked with retrieving publicly available ...

"AI is the devil and I its maker."

The limited test covers select products and users, with a broader rollout planned for later in October.

Simon Willison used Claude Opus 5.5 to generate pixel art animations of kākāpō parrots for a conference keynote.

Blue Cross Blue Shield says hospital use of AI tools led to an additional $942M in healthcare spending over a two-year period.

Transfer learning protocol for edge classification on dynamic text-attributed graphs addresses domain shift via leave-one-domain-out evaluation.

Sliced Orlicz-Wasserstein distance generalizes optimal transport with Luxemburg norms for probability measure comparisons.

SynCo framework improves multimodal contrastive learning by capturing synergistic information beyond redundant cross-modal features.

VCRE-Fib applies view-conditioned deep learning to ultrasound assessment of liver fibrosis in endemic disease regions.

Latent preference modeling framework disentangles genuine user preference from behavioral measurement confounders in short-form video recommendation.

HamiFormer dual-expert diffusion model predicts continuous dynamics and collision-driven state changes using symplectic Hamiltonian propagation.

SonoGraph-WM uses scene graphs as world models for robotic ultrasound navigation with anticipatory probe motion prediction.

Theoretical framework characterizes error in radiomap blind prediction under incomplete environment observations without field measurements.

FinancialAuditBench evaluates AI agents on financial statement audit with differential privacy framework for proprietary financial data.

Permutation-equivariant flow matching generates neural network weights across tasks while handling hidden unit permutation symmetry.

Over-the-air computation for federated learning in wireless networks using waveform superposition for model aggregation.

UniCache: task-aware KV cache compression for unified multimodal models across multiple data types and tasks.

Retimed Bellman Flows addresses velocity bootstrapping bias in flow-based critic learning via dynamically shifted teacher queries.

Multi-agent preference learning framework for LLM collaboration without explicit reward engineering.

LLM pipelines lose up to 40.5 accuracy points when stages cannot access original problem context, quantifying decomposition cost.

Routing drift in merged MoE LLMs does not necessarily indicate failure; proposes analysis toolkit for counterfactual routing assessment.