The Archive

Researchers test 'value transplant' to redirect language model search toward user-intended goals by shifting activations along learned value axes.

PReCache reduces KV cache redundancy in multi-LoRA agent systems via low-rank precomputation, improving memory efficiency for long-horizon tasks.

Information-theoretic lower bounds on sample complexity for learning cyclic causal structure in non-Gaussian linear models.

Empirical study of information retention in sliding-window KV inference across Qwen, Llama, Mistral models reveals cross-window state transmission.

ARCH-B benchmark evaluates 25 multimodal models on cross-representational building recognition across photographs, floor plans, and renderings.

Kafila distributes LLM inference across heterogeneous trusted commodity machines using device-aware partitioning to avoid bottlenecks.

Theoretical analysis of non-negative matrix factorization singularities and real log canonical threshold bounds for Bayesian inference.

VLM-Safe-RL framework integrates frozen vision-language models as dense safety signal in constrained RL for autonomous driving without collision anticipation.

CLEAR-Med dual-agent framework separates SQL generation from independent validation for clinical data analysis with deterministic safety checks.

OpenAI solicits user stories for Codex Originals program; community recruitment effort with no technical announcement.

GPT-6 Astra completes tax workbooks 2x faster than GPT-5.6 Sol with improved intent understanding; case study with Basis.

UOPD applies uncertainty-aware selective intervention during on-policy distillation to correct high-uncertainty agent steps and improve multi-turn task success.

State-space model for metric 3D point tracking from monocular video, enabling robot navigation and autonomous driving decisions in absolute meters rather than pixels.

ADPTNet proposes neuromorphic sequence model with adaptive prescriptive timescales and non-linear recurrence as energy-efficient Transformer alternative for long-range dependencies.

Simon Willison summarizes major LLM developments in 2026, including Claude Opus 4.5 and GPT-5.1 releases and industry trends.

AVPO framework reduces hallucinations in LLM activation verbalization via text reconstruction and QA-based validation, improving interpretability reliability.

Three-tier Hidden Markov Model progression for detecting fraud in financial transactions using variational Bayesian inference and learned latent behavioral regimes.

Pipeline to detect shortcut learning in audio classifiers via temporal explanations and Large Audio-Language Models, identifying spurious correlations from data artifacts.

Structure-Adaptive Tree Field Integrators (STAD-TFIs) enable near-linear tensor field integration on trees with distance-dependent kernels via path-backbone decomposition.

Jev 1.13, a non-generative System One model, evaluated on medical QA and diagnostic reasoning against GPT-6 Sol across MetaMedQA, PubMedQA, DiagnosisArena-MCQ, NEJM Case Challenges.

SR4-Fit unified rule-based ML framework balances interpretability-by-design with predictive performance for high-stakes decision-making via transparent, stable decision rules.

Reuse privileged state-based manipulation policies via learned perceptual reconstruction rather than action distillation, reducing sample complexity for visuomotor control.

Maat deterministic runtime governance layer validates LLM multi-agent handoffs against versioned workflow contracts, preventing error propagation without learned judges.

YOLOv8-based visual phishing detection treats webpages as images to identify spoofed brand layouts.

LTV-CTDNet forecasts turning movements at traffic intersections with physically constrained neural outputs.

Physics-informed residual constraints outperform generic regularizers in neural PDE surrogates at fixed capacity.

Simon Willison observes S3 pricing has stagnated for a decade at $0.023/GB-month, with no reductions since 2016.

FIRE stabilizes LLM self-distillation via Fisher-informed recalibration of feedback on correct/incorrect outputs.

EHRAdapt adapts frozen language models to clinical records via learned embeddings for rare event codes.