AI Briefing#neural-networks
The latest AI and developer news about #neural-networks, with the original source and a short summary.
The latest AI and developer news about #neural-networks, with the original source and a short summary.
The attack requires no retraining and passes standard safety evaluations, with triggers as simple as a date change.
A steering vector derived from the difference between loose and strict reward checkpoints allows extrapolation to reduce resource-limit failures from 14.5% to 8.0%.
The study demonstrates that gating mechanisms in SSMs cause models to learn in-weights memorization solutions, delaying or preventing convergence to correct in-context learning even without architectural memory limits.
The technique trains small neural networks offline to route and rewrite queries, achieving up to 8x faster search and higher recall than centroid-based indexing on the Natural Questions dataset.
GPT-6 Astra improved the MATH-500 score of Qwen3.5-0.8B to 48.1% by modifying its weights without training.
A recent paper's claim that GPT-6 Astra can improve a 0.8B Qwen model by editing weights is dismissed as a task any competent human researcher could perform using steering vectors.
The guide outlines the three-step process of voice synthesis, distinguishes between instant and professional cloning, and details detection methods for AI-generated audio.
A study on Gemma 3 270M shows that training just 1,000 random weights (0.001% of the model) enables hidden skills like sorting and JSON output that prompts alone could not unlock.
Randomly interconnected units governed by shared plasticity rules achieved pattern recognition performance similar to backpropagation-trained networks.
Handwired converts JavaScript code into fully differentiable neural networks that run in the browser without training or a GPU.
The study presents the first tri-modal Masked Diffusion Models pretrained on text, image-text, and audio-text data, achieving strong results at 3B scale.
Researchers constructed backdoor attacks using digital lockers that remain provably hard to elicit even with full access to model weights, challenging the reliability of current capability audits.
Apple researchers introduced Normalizing Trajectory Models (NTM), a framework that matches strong baselines in four sampling steps while retaining exact likelihood over the generative trajectory.
A new paper accepted at NeurIPS introduces adaptive representations for Functional Gradient Descent, ensuring convergence to the global minimizer and often outperforming neural networks by an order of magnitude.
Sakana AI proposed PC-ALM, which resolves the signal attenuation issue in existing Predictive Coding (PC), successfully training 1000-layer neural networks.
A new study demonstrates that neural network vector representations can be approximated by symbolic structures.
The paper and code for 'Spectral Neuron', a new ML model combining scalability and interpretability, have been released.
Sparse data is optimally reused 15–20 times, leading to the proposal of a scaling law that reflects this.
The new DD method ensures asymptotic alignment between training and test errors at each iteration step.
It proposes MLVC, a learned video codec capable of real-world deployment by solving the numerical inconsistency problem between NPUs.
The University of Tennessee has filed a lawsuit against Anthropic, alleging infringement of a patent related to neural networks.
A new neural network learning mechanism has been proposed that enables training while following Dale's principle, without the Weight Transport problem.
It has been revealed that LLM safety alignment is not distributed across the entire model but is concentrated in specific neurons, meaning that manipulating just a single neuron can disable safety guardrails.
This proposes Amortized MIPS, which dramatically reduces search computation cost by directly predicting MIPS solutions via neural networks.