The Archive

Mechanistic analysis reveals transformer trained on navigation task maintains coherent spatial world model; behavioral failures traced to feature interference.

Single-loop stochastic optimization methods for nonconvex-concave minimax problems with complexity guarantees.

Test-time adaptation method for imbalanced binary segmentation prevents entropy-driven collapse via balanced anchor prompts.

RL method combining model-predictive path-integral control with expert-guided planning balances exploration-exploitation in continuous control.

Google customized Flow tools with designers Jane Wade and Sergio Hudson for New York Fashion Week preparation.

Benchmark evaluates LLM cross-lingual understanding of Chinese internet buzzwords with cultural semantics and safety implications.

TERMon detects runtime anomalies in edge AI accelerators via hardware-native ternary monitoring without model re-execution.

SpecQuant combines speculative decoding with multi-parent quantization for training-free efficient LLM inference on consumer hardware.

Bayesian methods for multi-way astronomical spectrum classification with uncertainty quantification for 4MOST survey.

Theoretical analysis of optimization geometry across equivalent Brownian RKHS coordinate representations.

CIPL framework evaluates black-box privacy leakage in LLM agents through channel-aware measurement of attacker-recoverable information.

ReACT-TTS uses listener facial reactions to plan emotion and prosody in conversational speech synthesis.

GUARD enables natural forgetting in large reasoning models via guided answer-reasoning distillation without hallucinated substitutes.

Spoken Wikipedia Presentation Corpus extends ASR datasets with LLM-generated slides for multimodal speech recognition evaluation.

PRISM-BN introduces 5054-instance benchmark for text-to-parameterized Bayesian Network extraction in neurosymbolic AI.

L0-MoE accelerates dense LLMs via L0-regularized Mixture-of-Experts, achieving 2.5x speedup with minimal performance loss.

Pipeline linking academic papers to source code repositories via Software Heritage and Wikidata for semantic discovery.

Samsone family of small audio language models (99M–356M params) achieving SOTA on edge inference benchmarks.

Measure quantization framework for multi-domain clustering using Sinkhorn divergence and optimal transport.

HATS-en benchmark shows WER poorly correlates with human judgment; BERTScore variants better track semantic fidelity in ASR.

OpenAI publishes six-pillar safety framework for youth in Australia, focused on protective guardrails rather than technical AI advances.

Activation steering less effective on latent chain-of-thought vs. explicit reasoning due to latent-to-language transition gap.

Cross-policy analysis of 15k LIBERO rollouts reveals vision-language-action policies achieve similar end-effector geometry when succeeding.

Theoretical extension of JEPAs beyond Gaussian latent spaces to Riemannian manifolds with conditions for representation identifiability.

Framework analyzing linear encoding of linguistic relations in GloVe, RoBERTa, ModernBERT; inflectional/derivational relations near-perfect, lexicographic/encyclopedic weaker.

Configurable multi-stage vision pipeline for crop disease/pest diagnosis in Farmer.Chat, enabling threshold tuning and new disease/crop addition.

SynthDemo-RL framework uses LLM-guided synthetic demos to overcome sparse-reward exploration in VLA fine-tuning via teacher-student distillation.

Riemannian Neural Hamiltonian Flows combine geodesic geometry with invertible generative modeling for improved interpretability in normalizing flows.

On Wednesday, MIT Technology Review hosted a live Roundtables event for subscribers that asked the question everyone’s asking right now: Could AI really kill us all? But attendees had so many more questions than we had time to answer in the 30 minute session. So we asked our senior AI editor Will Douglas Heaven and…

Chinese Competitive Debating Dataset: 182 professional debate matches with 360 judge annotations benchmarking LLM argument-tracking and adjudication.