The Archive

MemT-1.0 internalizes persistent memory in foundation models via Gated DeltaNet-2 and Mixture-of-Storage for continuous information accumulation.

SpecMRL improves Matryoshka Representation Learning by reorganizing feature space with spectral basis ordered by importance for better compact encodings.

AfriBED introduces a speech recognition bias evaluation dataset for African languages to assess ASR performance disparity across demographic subgroups.

ASTOR benchmarks LLM autoformulation for soft-constraint optimization using formal grammar, addressing gaps in handling preference-based constraints.

Next Best Action frames customer intervention selection as constrained MDP with causal off-policy correction integrated into LP optimization.

Safety Mid-Training internalizes LLM safety during pre-training using shared reasoning structure with general tasks to prevent post-training degradation.

Parameter-efficient fine-tuning applied to brain encoding models achieves effective adaptation of large pretrained backbones with limited neural data.

“I think I have good news about the nature of neural networks.” With those simple words of optimism, Nick Cammarata, a researcher at the AI startup Goodfire, kicked off a 26,000-word blog post that has knocked the AI world on its head, sparking wild speculation as well as cautious (and not-so-cautious) celebration. Cammarata’s claim is…

Mise-en-Video enables fine-grained 3D control for multi-shot video generation with inter-shot consistency via spatial staging and neural rendering.

EdgeDRM enables decentralized federated learning inference-time collaboration via learned routing potentials allowing queries to reference neighboring clients.

Exact risk analysis of candidate-scoring models with teacher targets, frozen backbones, and temperature calibration in Gaussian surrogate settings.

Caterpillar Net constructs maximum of N inputs with two hidden ReLU layers of size O(N²) and 2^O(N log N), settling CPWL function expressiveness.

Study of training-data scaling in tabular ML with 1M+ samples shows GBDTs and MLPs scale similarly; tabular foundation models under-explored.

SOMA uses System One models as metacognitive monitors for reasoning tasks, enabling real-time intervention without fine-tuning.

Optimization framework for training models with mixed ground-truth and ML-labeled data via stochastic first/zeroth-order methods.

Multi-agent RL for market makers on thin prediction-market order books, trained on Polymarket data with novel simulation approaches.

LG-Bench introduces first graph-structured benchmark for life sciences with 10K+ questions in medicine, biology, chemistry.

RESIN uses entity-structured indexing to improve multi-chain reasoning by decomposing questions into focused retrieval tasks.

Game-theoretic framework enables LLMs to learn scam-baiting strategies through in-context learning without fine-tuning.

Sheaf-theoretic framework (TSI) for multi-agent task compatibility obstructions using topological methods.

Monsoon framework collects spontaneous-speech ASR data across 50 languages with multi-stage quality control, reducing WER.

HarnessPro recursively improves AI agent harnesses for performance-cost trade-offs via LLM-generated tests without ground truth.

Prior-Informed Adaptation (PIA) combines simulation pretraining with uncertainty-regularized planning for sample-efficient model-based RL in autonomous racing.

TokenGate enables learnable token pruning in LLMs for faster, memory-efficient long-context prefill using a differentiable top-K approximation.

Trajectory Replay adapts continuous-time generative models for lifelong learning by refining velocity fields instead of replaying samples.

Theoretical analysis showing gradient descent achieves ε-optimality on multiclass logistic regression in fixed iterations independent of ε.

recAMPLIFY applies recursive pretraining to protein language models, enabling test-time scaling and reducing memorization without massive scale increases.

Polar Switch achieves faster polynomial orthogonalization for gradient descent in language model optimization with lower FLOP cost than quintic methods.

CAVA uses optimal transport data valuation to identify harmful mislabeled samples in complementary-label learning from crowdsourced or VLM annotations.

NVIDIA CUDA Tile (cuTile) is a Python-based DSL and parallel programming model for NVIDIA GPUs that automatically leverages advanced hardware capabilities to... NVIDIA CUDA Tile (cuTile) is a Python-based DSL and parallel programming model for NVIDIA GPUs that automatically leverages advanced hardware capabilities to deliver peak performance for tile-level GPU programming. By abstracting thread- and warp-level scheduling behind high-level, tile-centric primitives, cuTile enables you to write GPU kernels in Python while compiler optimizations automatically… Source