$δ$-mem: Efficient Online Memory for Large Language Models
δ-mem adds compact online associative memory to LLM backbones via delta-rule updates, enabling efficient long-context reuse in agentic systems.
Search the full wire by company, model, lab, or keyword. Every story we have ever aggregated.
δ-mem adds compact online associative memory to LLM backbones via delta-rule updates, enabling efficient long-context reuse in agentic systems.
Transfer learning evaluation across 11 pre-trained models for image classification; standard practice study without novel methodology.
Random-Set GNNs introduce uncertainty quantification for graph neural networks via epistemic and aleatoric uncertainty modeling.
NOFE enables continuous dimensionality reduction via function-to-function mappings using Graph Kernel Operators.
Paper formalizes positional encoding requirements for Transformer-based neural combinatorial optimization on vehicle routing, accounting for spatial structure unlike NLP.
YFPO leverages neuron activation patterns for preference optimization in mathematical reasoning, using model internals to guide reward signals instead of external preference data.
Proposes Hierarchical-Cluster SOINN classifier for class-incremental learning that models classes as manifolds rather than point collapse, addressing Neural Collapse theory gaps.
Proteus red-teaming framework studies adaptive leakage of LLM agent skills via iterative adversarial revision, addressing real deployment risks beyond single-shot audits.
Spectral gradient clipping method preserving matrix structure and controlling singular value decay during neural network training.
Spiking neural network framework with gradient-trainable multi-timescale conductance dynamics for sparse temporal processing.
Shift-invariant transformer autoencoder for calibration-robust spectral unmixing in pharmaceutical and biomedical analysis.
Variational inference for Lévy-driven SDEs via neural tilting to model extreme events and heavy tails in safety-critical systems.
Proof that neural weight norms equal Kolmogorov complexity in fixed precision, explaining why weight decay induces Solomonoff's universal prior.
Neural1.5 method for clinical QA over EHRs using DSPy MIPROv2 optimizer for automated prompt tuning across four modular subtasks.
Framework enables visual-native multimodal search agents with on-policy data evolution and persistent visual evidence reuse.
CLEF, a Transformer-based EEG foundation model using multitaper spectrograms, aligns clinical signals with neurologist reports via contrastive learning on 234-task benchmark.
A while ago I posted my AI-made Battle of Vienna film, and the feedback from this community genuinely helped me improve. I’ve now finished my next one: a 15-minute cinematic vid about the Battle of the Teutoburg Forest, 9 AD. Arminius, Varus, and the day Rome lost three…
Import AI newsletter covers RSI economic impacts, AI regulation optionality frameworks, and neuromorphic computing advances; examines legal/governance demands of superintelligence.
Google Finance launches AI-powered redesign in Europe with local language support.
This was a really good talk, especially for anyone who's built things like the Karpathy wiki, Serena, or SQLite databases as memory for Claude. For any senior devs out there, are you spotting the solutions already implemented in distributed systems being reused? If many…
Been using Claude Code heavily and kept losing context between sessions. Built ProMa to fix it. It's a Claude Code plugin (8 skills + 1 hook) that maintains project state in markdown files. Every session, Claude reads the state and resumes exactly where it left off. Quick…
Cosine Gated Adam Decay optimizer improves asynchronous DiLoCo training by scaling stale pseudo-gradients using exponential decay.
Contextual Plackett-Luce model handles ambiguous sequence selection by matching multi-modal target distributions despite single-instance supervision.
Sub-network Laplace approximations optimized via formal parameter subset selection beyond heuristic layer-wise/diagonal approaches for neural network uncertainty.
Market-rule-informed neural network for electricity imbalance price forecasting embeds price formation rules into latent space.
AutoTTS framework uses agentic discovery to automatically design test-time scaling strategies for LLMs rather than hand-crafting heuristics.
ADD-PINN applies domain decomposition to physics-informed neural networks for traffic flow estimation with sparse sensors.
Extends convexification theory to train spiking neural networks with exact gradients, removing surrogate gradient approximation.
Susceptibilities technique extends neural network interpretability to deep RL agents, revealing parameter-space development undetectable from policy analysis alone.
Primer on susceptibilities and linear response theory for interpreting neural network posteriors via Bayesian influence functions.