AI for science

• VenusRX-Bench standardises enzymatic reactions from several databases with leakage-controlled splits for reaction prediction, retrosynthesis and enzyme commission numbers, and the T5-style VenusRX model matches or beats existing systems on all three. arXiv
• A dual-rule contrastive model aligns mouse and human electrophysiology in one space, retrospectively tracking known clinical efficacy across ten model–drug combinations and linking Fmr1-knockout mice to human 16p11.2 copy-number carriers. arXiv
• Recasting neural decoding as Bayesian cognitive inference recovers stable cognitive states from variable recordings across five recording modalities and three cognitive domains, with representation geometry consistent across tasks. arXiv

Free offerings

Nothing notable today.

Building on frontier models

• TaReD organises an agent's tools by function and lets it discover them on demand while recursively decomposing the task, instead of injecting every tool description, lifting end-to-end success by up to 40 percentage points. arXiv
• OnTrack scores an agent's trajectory as it streams, about a millisecond per step, ranking runs 0.057 AUROC better than content-similarity baselines after eight steps and reclaiming roughly 18% of compute spent on failing runs. arXiv
• REMORY adds bounded soft memory tokens beside the text summary an agent compacts its history into, reaching near full-context SummHay performance from 5.2% of input positions and cutting repeated tool errors on Terminal-Bench 2.1. arXiv

Cost & hardware

• Nvidia-backed Firmus withdrew its Australian float on 9 October after weighing a cut from A$11 to A$8.25 a share, blaming market volatility, and will seek private money instead. ABC News (Australia)
• SpaceX is reportedly seeking a $40 billion debt package for Nvidia hardware, enough for roughly 360,000 Vera Rubin GPUs across 5,000 NVL72 racks. Tom's Hardware
• A Senate investigation says AI data centre job claims are misleading, with 100MW sites creating only about 100 permanent jobs and companies refusing to hand over employment data. Tom's Hardware
• A free Steam demo's AI inference ran to $1,000 a day, pushing its studio into a bank loan and toward local hardware models to escape cloud costs. Tom's Hardware

Evals & guardrails

• Typed decision models used as agent guardrails flip on one word: across seven open-weight models, log injection raised fail-open rates from 0% to 63% and misleading option names to 93–100%, defeating every tested defence. arXiv
• Agent safety counts forbidden actions but misses duties undone: 56.92% of GLM-5.3 trajectories leave obligations unfulfilled versus 30.00% with forbidden actions, and the best of 14 models reaches 48.97% recall on ObligationBench. arXiv
• An Identify–Solve–Escalate framework finds higher task accuracy brings no more epistemic humility, as agents spot knowledge conflicts early then drop them and model-level fixes trade accuracy for humility. arXiv

Algorithmic advances

• Spectral weight decay shrinks the spectrum additively after each step, giving 1.89x compression and 1.18x inference speedup on a 500M LLaMA at 4% distortion where standard weight decay manages 1.14x and 1.01x. arXiv
• Treating 4-bit weight-and-activation quantisation as coupled calibration lets a rank-4 correction beat rank-32 SVDQuant on FLUX.1-dev and SANA-1.6B with 6.25x faster quantisation, lifting Qwen3-8B MMLU from 61.50% to 68.17%. arXiv
• The Polytopal Neural Network forces each layer's latents onto a learned polytope via amortised simplex inference, giving a route to vector-quantised training and more compressed representations than standard vector quantisation. arXiv