The Archive

llama.cpp adds experimental DeepSeek v4 Flash support with aggressive 2-bit quantization, achieving 17 tokens/sec on M3 Max with 128GB RAM requirement.

Qualitative study of how smart-home AI teams (Amazon Alexa, Google Nest, Microsoft Azure) manage user expectations through design practices.

QACD framework treats causal discovery conditional-independence tests as defeasible arguments, mitigating cascading errors in finite-sample regimes.

Theoretical characterizations of admissible objective functions for hierarchical clustering, extending Dasgupta and Cohen-Addad frameworks.

First Romanian Grammatical Error Correction corpus (10k sentences) with adapted ERRANT toolkit and neural baselines for low-resource settings.

GraphPlanner: heterogeneous graph memory-augmented router for multi-agent LLM agentic systems with task planning and multi-round cooperation.

Tandem framework pairs large and small LLMs to reduce computational cost of reasoning-intensive inference while maintaining answer quality.

Transformer-based interpretable models for English reading comprehension with adversarial bias correction and attention visualization for educational use.

Llama.cpp benchmarks on Windows 11 vs Lubuntu 26.04 with RTX 5080 show significant OS-level performance variance in local inference.

Reddit user reports account suspension risk from OpenAI after attempting to automate YouTube downloads; anecdotal account of API guardrail enforcement.

Qwen3.6-27B-INT4 achieves 100+ tokens/sec with 256k context on RTX 5090 via vLLM 0.19, with KLD quantization validation.

HGIN jointly infers interaction graphs and predicts dynamics for lattice Hamiltonian systems from trajectory data without assuming homogeneity.

DxChain: clinical LLM agent using memory anchoring, navigation, and verification phases to reduce diagnostic tunnel vision and hallucinations in EHR analysis.

TimingLLM: two-stage retrieval-augmented LLM pipeline predicts post-synthesis timing (WNS/TNS) from Verilog without running EDA tools.

Paper argues companion chatbots must legally separate commercial from non-commercial contexts to protect user autonomy against undisclosed promotional content.

Empirical study shows persona conditioning in LLMs amplifies gender bias differently across English and Hindi in professional narrative generation.

Proposes Partition-of-Unity Gaussian Kolmogorov-Arnold Networks as spline activation alternative with partition-of-unity properties and kernel interpretation.

User demonstrates PaddleOCR-VL-1.5 multimodal inference via llama.cpp server for end-to-end document digitization with layout and table handling.

Paper documents LLM failure modes in peer review including prompt injection attacks and proposes safeguards for AI-assisted scientific evaluation.

XITE proposes embedding-based cross-lingual data augmentation via interpolation to improve transfer learning in low-resource multilingual settings.

FinGround detects computational and citation hallucinations in financial LLM systems via atomic claim verification before EU AI Act enforcement (Aug 2026).

Talker-T2AV decouples semantic and low-level modeling in autoregressive audio-video generation for improved talking head synthesis coherence.

ComplianceNLP integrates knowledge-graph-augmented RAG to automatically detect regulatory gaps across SEC, MiFID II, Basel III for financial institutions.

AgentEval introduces DAG-based step-level evaluation framework for agentic workflows with error propagation tracking and hierarchical failure taxonomy.

PhysCodeBench benchmarks physics-aware symbolic simulation across 700 samples to evaluate LLM semantic understanding of physical phenomena for robotics/embodied AI.

LLMs applied to model daily human behaviors for predictions and generation across personal assistants and recommendation systems.

RouteNLP framework routes queries across tiered LLM models to minimize inference costs while maintaining per-task quality thresholds.

CAPSULE framework provides hard safety constraints for RL exploration in high-dimensional systems using control-theoretic dynamics models.