The Archive

RICE-Alpha uses event graphs and point-in-time modeling for LLM-based equity forecasting with temporal reliability weighting.

T-SNN models EEG brain states via simplicial complexes to capture higher-order neural interactions over time.

Digital hardware twin simulates megahertz plasma diagnostics and edge-AI integration for tokamak control commissioning.

ASTRA applies ADMM-accelerated online learning to dynamic topology reconfiguration in partial satellite constellations.

RewardExplainer generates interpretable explanations for LLM reward model decisions via counterfactual preference feedback.

Opera framework enables persistent verbal criticism for long-horizon coding agents, using typed operators and event-driven triggers to track and resolve diagnosed issues.

ICMAPE framework extends active sequential hypothesis testing to decentralized multi-agent systems using Bayesian learning for exploration driven by inference objectives.

Crowdsourced SFT pipelines vulnerable to data poisoning attacks enabling extraction of proprietary instructions via topic-based manipulation of fine-tuning data.

Hankel-Toeplitz forecaster achieves compact linear time series prediction using H+L-1 parameters via classical stationary prediction theory for long-term forecasting.

Framework addresses semantic diffusion in discursive text analysis by preprocessing documents that critique and contextualize positions rather than directly asserting them.

Future Information-Directed Sampling (FIDS) algorithm explores nonstationary bandits by gathering information about future optimal arms rather than current uncertainty reduction.

DynGraphAgentBench evaluates agentic controllers for dynamic graph anomaly detection under delayed feedback across seven temporal datasets with chronological deployment windows.

Neural operators adapt to changing operating conditions in mechanics simulations via reuse of high-fidelity solutions to improve command selection accuracy.

GroupMask enables layer-adaptive group-wise sparsity for semi-structured LLM pruning, improving compression by varying sparsity ratios across layers via grouped weight partitioning.

Conditional Progressive Pruning (CPP) outperforms consistency baselines in LLM multi-agent debate by exploiting multi-round communication under strict cost constraints.

HARMONIA: interpretable graph learning framework using mixtures of neural bases to improve scalability and feature specialization.

Study reveals probabilistic incoherence in LLMs (Jev, English Laya) when decomposing classification decisions across hierarchical category levels.

Token-level log probability signals characterize non-uniform value distribution in CoT reasoning, enabling inference optimization.

Theoretical analysis closes gap in sample complexity bounds for two-sample testing on inhomogeneous random graph models.

ThinkNet: subject-independent motor-imagery EEG decoder framework with validation-gated ensembles for brain-computer interfaces.

Greenpixie methodology quantifies per-token energy/water/CO₂ costs for LLM inference across open and closed-weight models.

EEG-Fusion: source-free expert routing framework for detecting and mitigating subject-level failures in motor-imagery decoding.

Integrated methodology for LLM-as-judge evaluation systems in multi-turn business agents with human-in-the-loop governance.

Sampler sharpening optimization shows masked diffusion language models achieve competitive few-step generation without retraining.

Jacobian centroids enable behavioral monitoring of JEPA world models for task alignment and failure detection in planning.

Reanalysis challenges evidence for language model homogeneity/Artificial Hivemind thesis via visualization, spectral analysis, and clustering contradicting prior collapse claims.

ViGeo framework extends visual in-context learning to video via spatiotemporal canvas completion, generalizing to unseen video manipulations without task-specific fine-tuning.

Diffusion-based rollouts suppress recursive error growth in environmental forecasting but do not guarantee forecast fidelity for long-horizon lead times.

QLoRA-optimized Phi-2 small language model for real-time chatbot applications reduces memory overhead in mobile and edge computing via 4-bit quantization and PEFT.

Single random Gaussian probe estimates layer quantization error sensitivity spectrally flat; enables data-free, budget-targeted mixed-precision quantization for 35B MoE and 9B models.