The Archive

Qwen 3.6 27B performance benchmarks across llama.cpp backends on RTX 3090: ik_llama.cpp achieved 1261 tok/s prefill, 72.9 tok/s decode with 156k context.

Integrates conformal prediction with neuro-symbolic concept models to quantify prediction confidence and improve trustworthiness.

PIPER uses LLM-generated pseudoqueries and profiling for content-based table search across data lakes.

RGB-only active 3D scene graph generation for mobile robots without depth sensors, with learned viewpoint selection.

MLLMs fail at grounded 3D spatial reasoning and multi-agent Theory of Mind due to text-based probability distributions lacking topological understanding.

PPR-GDE method combines pairwise preference rewards with group-based diversity enhancement to reduce RL diversity collapse in open-ended generation.

Dual-Rate Diffusion accelerates inference by interleaving sparse heavy context encoder with light denoising model for efficient feature reuse.

UTOPYA is a 15.2M-parameter multimodal framework for anomaly detection and time-series prediction in batch distillation using FiLM-based fusion.

Framework integrates fixed external RGB cameras as geometric priors for active 3D scene graph generation in robotic systems.

Position paper argues agent generalization requires scaling environment rule-sets and interaction interfaces, not just trajectories within fixed benchmarks.

Theoretical analysis of canonical regularisation in wide feature-learning neural networks, extending kernel regime NN-GP correspondence insights.

MARS system for CASTLE Challenge treats egocentric multimodal reasoning (4 days, 15 views, 8 modalities) as agentic evidence selection.

Ringmaster LMO enables asynchronous Linear Minimization Oracle training for heterogeneous distributed systems, improving on synchronous Muon.

Generative Visual Grounding uses EEG-to-image models to ground brain signals visually instead of via text, improving MLLM-based brain foundation models.

ML surrogate models for PCB signal integrity analysis that adapt to buffer parameter variations without retraining.

Elastic-dLLM reduces computational redundancy in diffusion language models via context compression and position-aware augmentation.

TRACE method corrects LLM hallucinations by analyzing cross-layer evidence to identify where factual information is suppressed.

SAGE framework improves VLM spatial reasoning via geometric logic consistency and duality operations in GRPO training.

Vision Inference Former addresses visual consistency in MLLMs by elevating visual features above text-token parity in connector-based architectures.

FOL2NS neurosymbolic framework generates natural language from first-order logic formulas with deep nesting and varying quantifier depths.

Reddit discussion on Claude deployment in enterprise, data security concerns, and comparison to Copilot for Enterprise.

Stratechery analyzes data center opposition and proposes financial compensation as the viable solution to community resistance.

OpenAI and Dell partner to deploy Codex in hybrid/on-premise environments for enterprise coding automation with security controls.

Multi-agent framework exposes concept erasure vulnerabilities in diffusion models, showing suppressed concepts can be reactivated via black-box awakening.

Benchmarking study on foundation models vs. gradient-boosting for credit risk prediction with interpretability via SHAP.

pArticleMap system uses embeddings and graph algorithms to map nanomedicine research frontiers and generate evidence-grounded hypotheses.

RCT shows LLM assistance improves learning task performance unevenly; gains correlate with domain knowledge, revealing productivity heterogeneity.

Empirical study of indirect prompt injection attacks enabling privacy leakage in black-box LLM chatbot agents with external tool access.

First source attribution benchmark for generative 3D models across multi-view, geometric, and frequency-domain fingerprints.

Novel GNN-based framework for multivariate time series anomaly detection using prior-observation adversarial learning.