The Archive

RAPS-DA: regime-aware peer specialization framework handles heterogeneous knowledge conflicts in RAG by disentangling reliability levels of retrieved context.

Entropic Learnability Horizon (ELH) framework unifies information theory and topology to explain generalization in overparameterized deep networks.

DeepReinforce releases Ornith-1.0, MIT-licensed open-weights model (9B–397B variants) for agentic coding, built on Gemma 4 and Qwen 3.5, achieving SOTA on coding benchmarks.

Muon optimizer avoids slow saddle-to-saddle dynamics in matrix factorization; learns balanced solutions with uniform convergence rates across modes.

DR-ACI constructs prediction intervals for doubly robust pseudo-outcomes under temporal dependence via adaptive conformal inference.

Random Network Distillation enables lightweight clustering in federated learning without coupling to main training loop, reducing communication overhead.

Post-Hoc Concept Bottleneck Models often learn predictive artifacts rather than semantically meaningful concepts; proposes faithfulness evaluation beyond task accuracy.

CUDA optimization techniques (shared memory, pre-transposed weights, fused kernels) achieve 1.41x speedup on shallow networks on Tesla T4.

Multi-channel Multigrid preconditioner learned via neural networks to solve high-wavenumber Helmholtz equations with phase-space coarsening.

A new proposal would ban the sale of Americans' health and location information to data brokers - including information people reveal to an AI chatbot like ChatGPT or Claude. In the coming weeks, Senator Elizabeth Warren (D-MA) and Representative Mary Gay Scanlon (D-PA) are…

Google AI explains full-stack approach to AI development, covering integrated hardware, software, and model design.

SIMAX framework generates synthetic clinician-patient dialogues with human-coded annotations to evaluate AI communication coding at scale.

Factorizable Normalizing Flows model parameter-dependent density morphing by factorizing the flow, enabling tractable learning across exponentially many configurations.

AI agents are quickly moving beyond chat. They inspect code, run tests, read documents, search knowledge bases, query internal systems, and operate for hours on... AI agents are quickly moving beyond chat. They inspect code, run tests, read documents, search knowledge bases…

Argues LLMs lack situated perception of physical world causality and agency; identifies this primitive as necessary prerequisite for artificial superintelligence.

COHORT automates enterprise network threat mitigation via multi-agent LLM workflow proposing, implementing, and refining mitigations on emulated topologies.

Field order in serialized structured metadata silently impacts retrieval quality; proposes permutation-invariant fine-tuning to decouple field order from embeddings.

Non-parametric causal diffusion recovery from cross-sectional steady-state data; application to gene expression without requiring longitudinal observation.

MuonSSM stabilizes state space models for long-sequence modeling via momentum pathways and Newton-Schulz transformations on input injections.

HSAP addresses cross-contamination in causal attention for packed sequences by combining sequence parallelism paradigms for hybrid-context LLM training.

Curvature-Weighted Gradient Diversity (CWGD) refines SGD convergence analysis by weighting gradient noise by inverse Hessian to better model effective optimization dynamics.

Study compares nine open-weight LLMs against human participants in networked Prisoner's Dilemma, finding selected models reproduce cooperation macro-dynamics but lack individual fidelity.

Analysis of enterprise tabular data reveals markedly different distributions and performance profiles from public benchmarks for TabPFN, TabICL, and ConTextTab models.

Three-method study across Qwen2.5-Coder-32B, Llama-3.1-8B, and Gemma-3-27B shows internal probes read situation not pre-action intent, limiting misalignment monitoring efficacy.

Online imitation learning for LLM post-training succeeds when student policy class is realizable; error accumulation is not the primary advantage mechanism.

Theoretical analysis proves online minibatch SGD on two-layer ReLU networks prioritizes shortcut spurious features via optimization dynamics on Boolean hypercube XOR data.

FIL Hypothesis challenges Bitter Lesson by arguing inductive biases matter more when feedback loops are long; computation-scale approach fails in slow-feedback domains.

Framework translates natural language system requirements to ATL/A formal specifications for multi-agent systems verification using LLMs.

Enterprise investment in AI is booming. Gartner is calling 2026 an “inflection year” for organizations to align their AI projects with strategic business objectives. As the pressure to prove ROI mounts, executives and technology leaders are looking to agentic AI to drive the…

Hybrid active-online learning framework for optical network failure detection achieves 97% accuracy querying only 3.4% of streaming samples.