The Archive

UEmbed is a decoder-only multimodal model producing sparse lexical and dense embeddings in one pass, extending learned sparse retrieval to multimodal RAG without auxiliary modules.

CoWAM adds selective intervention layer to bimanual robot policies using World Action Models, enforcing synchronization and collision-avoidance contracts only when beneficial.

Reparameterization technique for simplex-constrained optimization via smooth manifold embedding; applications to tensor decomposition and functional data registration.

Pseudorandom streams in diffusion models act as learnable inputs; predictable orbits from finite-precision hardware affect training and generation quality.

AtumAI uses agentic AI to generate datacenter control-plane policies via formal constraint-based search with learned transfer across tasks, outperforming off-shelf LLM approaches.

PRECOG collapses RAG prefill cost from O(L_context) to O(1) by injecting pre-computed context into SSM fixed-size hidden state.

First systematic benchmark of Sheaf Neural Networks on inductive tasks, extending prior transductive-only evaluations across three diffusion mechanisms.

Study of Arabizi (Arabic in Latin script) usage patterns and writing norms across five dialects with speaker interviews and released resources.

The EU made some AI labels that companies can use instead of designing their own. | Image: The European Commission / The Verge The European Union has ushered in some additional rules that aim to make it easier for people to identify chatbots and AI deepfakes online. The new transparency obligations under the bloc's landmark AI Act came into effect on August 2nd, requiring companies to disclose when people are interacting with AI models, and if content has been generated or altered by them. The transparency rules differ between providers (companies that develop and market AI systems) and deplo...

Taxonomy survey identifying cognitive capability gaps in generative and agentic AI: reasoning, adaptive behavior, memory, self-regulation.

Formalizes the missing-target problem in fairness audits: how to justify demographic distributions in open-ended generation when ground truth is undefined.

Gaussian approximation for finite-sample ridge regression distribution under nonstandard asymptotics with growing regularization parameter.

Proves one-bit mean estimation achieves order-optimal sample complexity without interaction, avoiding two-stage localization protocol.

Constructs unambiguous DNFs with width O(n) and 0-certificate complexity Ω(n²), refuting Alon-Saks-Seymour conjecture via lifting theorem.

VI-MoLE routes inputs through LoRA expert mixtures using value-of-information rather than uncertainty, avoiding spurious expert activation.

MedPRESS: 600-dialogue benchmark measuring patient-pressure-induced sycophancy in LLMs across medication, self-care, and triage scenarios.

Moment closure enables distribution-aware planning in stochastic RL without restrictive policy assumptions, addressing target variance and covariance gaps.

Magnet framework detects cross-session AI misuse where attackers decompose harmful goals into innocuous agentic tasks across isolated sessions.

LiveMem maintains persistent memory state in long-running LLM agents by decoupling working context from intrinsic memory lifecycle.

UMDP optimization balances performance and operational constraints by selecting minimal policy sets robust across uncertain environment models.

RoMeRL addresses memory-reward trap in self-evolving agents using reduced-order utility states to balance feedback coverage and irrelevant experience pollution.

Finite sample characterization of log-likelihood ratio statistics in logistic regression with nonasymptotic bounds independent of design regularity.

Identity abduction—inferring two structures as one object via representational grounding—occurs without continuous embodiment in scientific hypothesis generation.

CMuon optimizer stabilizes Diffusion Transformer training by chunking momentum orthogonalization on fused tensor weights, improving late-stage convergence.

SWE-Touch benchmark tests coding agents in shared workspaces with user code edits, exposing weaknesses in handling conflicting modifications during task execution.

DyFrDet detector improves small object detection via dynamic frequency suppression and label disambiguation to handle visual cues scarcity.

Study proposes longitudinal measurement framework to detect cognitive, developmental, and socio-affective behavioral changes from long-term LLM interaction rather than short-term evaluations.

Computational and statistical analysis of c-rectified flow framework used in FLUX.1 and Stable Diffusion 3, proving guarantees for cost-aware velocity field projection.

Analysis of 18 open-source LLMs showing cultural bias in mythology knowledge; models encode cross-cultural distinctions in residual streams but fail to decode non-Western traditions.

Private Bayesian bootstrap technique using group-level blocking weights to privatize both point estimates and uncertainty quantification for statistical reporting.