The Archive

Analysis of ~100B language models' decodable night-sky representations in residual streams; mechanistic interpretability finding with 65-85% variance captured.

InferScale: GPU-native KV injection system for personalized LLM serving with persistent memory, reducing TTFT overhead from repeated prefill.

SciFigQual-Bench: benchmark for scientific figure quality assessment using full-manuscript context; evaluates caption alignment and visual misleadingness.

Cost-aware stopping mechanism (CAM-DF) for LLM agent tool acquisition balancing task coverage against cost, context load, and privacy.

On-Policy Distillation defense against fine-tuning poisoning attacks; routing-based approach to template-robust LLM safety realignment.

MemSecBench: benchmark tracing malicious content lifecycle in agent memory systems from persistence through recall to repair across backends.

Field codes for distributed optimal transport coupling sampling; mathematical framework for empirical Monge maps with certified marginals.

Google DeepMind releases Lyria 3.5 in Google Flow Music with improved musicality, lyrics, vocals, and creative control.

Parallel Trajectory Tempering improves Energy-Based Model training stability on multimodal scientific data via better MCMC mixing.

Single-beat cuffless blood pressure estimation via ear-PPG and ECG with lightweight hybrid learning for wearables.

HT-PAder: parameter-free online convex optimization algorithm for non-stationary environments with heavy-tailed noise.

Visual Credit Audit (VCA) measures whether multimodal models genuinely use image information vs. text-only reasoning in spatial benchmarks.

SciFigAlign evaluates scientific figures by alignment of visuals with manuscript claims, addressing peer-review assessment beyond generic image quality.

ScratchSim generates synthetic industrial defect data via procedural rendering (BlenderProc) with domain randomization for surface scratch detection.

PIKS develops kernel methods for physics-informed learning with closed-form solutions and analytical theory, contrasting with PINN complexity.

Microsoft is on a mad dash behind the scenes to patch exploits before hackers find them.

Setoka benchmark evaluates hierarchical user understanding in memory-augmented personalized agents beyond explicit fact retrieval.

CoCaRS uses correlation calibration to suppress redundancy in heterogeneous knowledge distillation across diverse model architectures.

AI home management startup Hint, co-founded by Martha Stewart, wants to become an “AI for your home,” combining property records, maintenance schedules, home documents, and an AI assistant into a single app.

GPTQ-2D extends adaptive matrix rounding (GPTQ) to two-sided case with nonsingular basis matrices for quantization optimization.

Video world models suffer dimensional collapse during long autoregressive generation; representation regularization stabilizes frame quality over 100+ steps.

Sparse lottery tickets matched to dense model accuracy fail to maintain behavioral equivalence in production deployment scenarios.

HoF-Bench evaluates 95 real CVEs discovered by LLM analyzers (OpenSSL, curl, GnuTLS) without frontier models, establishing reproducibility bar for AI security scanning.

TreeCCA applies gradient-boosted trees to canonical correlation analysis via Eckart-Young loss for tabular feature extraction.

BayesAME automatically determines coreset size for efficient LLM benchmark evaluation using sequential Bayesian inference.

Stereotypes-to-Decisions framework measures regional bias in six LLMs across China's 34 provinces, linking abstract stereotypes to concrete allocation decisions.

Certificate-gated protocol determines which physical parameters (mass, drag, stiffness) latent world models actually internalize from raw visual observations.

OpenAI reports 3x improvement on ARC-AGI-3 benchmark via two API settings enabling reasoning retention and compaction in GPT-5.6.

Overlap gap property extended to finite-temperature neural network optimization, characterizing algorithmic accessibility of noisy solutions.

AgentSnare uses adaptive deceptive observations to mislead LLM-based penetration testing agents, defeating static artifact recognition.