The Archive

Simon Willison's September sponsors-only newsletter covers Fable models, LLM pricing competition, and mathematical capabilities.

Visual force prediction from gripper deformation enables force-aware robot manipulation without specialized sensors.

Compact language models with tool augmentation perform early-stage drug discovery planning via LoRA fine-tuning on 1,263 workflow pairs.

PyINE framework addresses reasoning model oversight via instrumented Python execution traces to detect plausible but incomplete model reasoning.

Verb-ICL improves in-context learning for structured prediction via token-level coverage-based example selection.

Latent Lagrangian neural networks reduce order modeling of nonlinear dynamical systems using learned kinetic/potential energies.

Knossos and Ariadne benchmark diagram-to-graph topology extraction with VLMs on structural diagram understanding.

WNet replaces Transformer self-attention with discrete wavelet transform for linear-cost token mixing on long sequences.

RL framework for AI assistants to clarify underspecified user requests by value-of-information-driven clarification queries.

Study of coding agents for logging code removal identifies 10 patterns/11 reasons from Python and Java repositories.

GPU-accelerated Bregman Douglas-Rachford splitting for discrete optimal transport with three input format implementations.

Framework for combining diverse LLM agent assurance checks (rule checkers, monitors, gates) via temporal logic and assurance ledgers to identify unchecked risks.

HiRefPOU: hierarchical partition-of-unity mixture-of-experts architecture for operator learning on PDEs with sharp interfaces and multiscale structures.

The CEO of Amazon Web Services tried to push back against widespread suspicion of data centers.

VoCa: voice agent that coordinates speech with visual canvas interaction for multi-turn conversations, grounded in observational study of speech-boardwork coordination.

Conceptual framework for neurodiversity-aware multimodal affective computing in developmental assessment, challenging norm-referenced classification assumptions.

Drifting Models improve one-step generation via discriminative geometry in feature space; introduces persistent feature learning to explain pretrained space superiority.

LLMs cannot reliably self-determine verifiability boundaries for agent rules; proposes runtime escalation schemes for finance/healthcare/law compliance checking.

Score-Calibrated Flow: efficient training of diffusion/flow policies for online RL via score-based sampling from unnormalized critic densities.

Neural eigenvalue solvers accelerated via low-fidelity finite-difference method guidance, reducing training steps for high-dimensional operator problems.

Penumbra: sample-efficient adversarial search for finding minimal-edit compliance violations in finance/healthcare/law, targeting omissions and material disclosures.

RAGStress: benchmark for stress-testing RAG systems under realistic KB degradation (factual corruption, typos, poisoning, contradictions) across severity levels.

SepRQ: self-supervised speech framework using mask-free source separation achieves SOTA on Speaker Diarization and Speech Separation, surpassing WavLM.

Study on LLM error discovery in table question-answering: row order affects error detection; error marking alone insufficient for accuracy.

Multi-agent debate improves reasoning via proposal supply and verification readout; introduces Latent Verification Debate (LVD) to identify correct answers.

Contrastive Activation Addition (CAA): training-free steering vectors for SpeechLLMs enable task-specific specialization without retraining.

Nullspace projection method extracts persona subspaces in LLMs for fine-grained behavioral control beyond single-direction steering approaches.

MASBench: new benchmark evaluates LLM-based multi-agent systems under partial observability, measuring collaboration effectiveness.

Federated learning weights treated as wisdom-of-crowds mechanism; optimal weights inversely proportional to client error energies.