88

Papers

10/10

Roadblocks Active

2

Connections

A stacking ensemble of Random Forest, XGBoost, and LightGBM models trained on 49,159 Indonesian tax records achieved 97% accuracy in predicting non-compliance, with SHAP and LIME used to identify which features drive predictions. The study matters because it tests a concrete workflow for deploying interpretable AI in high-stakes administrative decisions — not just building a black-box classifier. The main caveat is that the dataset is proprietary government data and no code is released, making independent replication impossible.

██████████ 0.8 interpretability Peer-reviewed

Read

This dataset deposit supports experiments on robot swarms (1–20 units) that physically entangle via Velcro links to produce coordinated collective locomotion on flat and inclined terrain. The finding that physical coupling alone can enable cohesive group motion without central coordination is relevant to embodied AI research on emergent behavior. Notably, all raw data (~20 GB) and code are publicly released under CC BY 4.0, representing above-average open science practice for robotics.

██████████ 0.8 embodied-ai Peer-reviewed

Read

A Multimodal Swin Transformer architecture combines street-view photographs with structured building attributes (age, value, wind speed exposure) to classify hurricane damage at 92.67% accuracy — outperforming image-only baselines. This illustrates a reusable design pattern: adding cheap tabular context to vision models meaningfully improves classification in domains where images alone are ambiguous. Model code and weights are open-sourced, though the StEER reconnaissance dataset is not freely accessible, limiting full reproducibility.

██████████ 0.8 multimodal-understanding Peer-reviewed

Read

A physics-informed Deep Dyna-Q framework (Dyna-PINN) embeds resistance-capacitance thermal constraints directly into neural surrogate models for heat pump control, improving data efficiency and generalization under limited training data. The work also introduces an operator-learning surrogate (ScaleONet) that transfers across different building cluster sizes without retraining. This is a practical demonstration that embedding domain physics into RL reduces the data requirements that make real-world deployment of AI control systems difficult.

██████████ 0.8 efficiency-scaling Peer-reviewed

Read

Markov Logic Networks (MLNs) fuse first-order logic with probabilistic graphical models, and this review maps the landscape of algorithms for automatically discovering the logical formulas and their weights from data. The relevance for AI interpretability is direct: unlike neural networks, MLNs output human-readable logical rules ('if A and B then C with weight w'), offering a path to explanations that are both probabilistically grounded and structurally transparent. Scalability — the exponential search space for formula discovery — remains the central unsolved obstacle to practical deployment.

██████████ 0.7 interpretability Peer-reviewed

Read

In a 40-patient pilot study, a commercial AI system (SONIX Health v2.0) analyzed cardiac ultrasound studies in a median of 94 seconds versus 490 seconds for a trained sonographer, while meeting a prespecified noninferiority criterion for left ventricular ejection fraction (mean difference 0.00 pp, upper 95% CI 1.41 pp). The study is notable for using a formal noninferiority statistical framework rather than just reporting raw accuracy, which is the appropriate test when the goal is replacing a human workflow step. Major limitations: 40-patient single-center sample, proprietary closed-source AI system, and a single cardiologist as the sole reference standard.

██████████ 0.7 multimodal-understanding Peer-reviewed

Read

The Hologram framework formalizes knowledge graph schema mapping as a Steiner Tree learning problem over relational graphs, using reinforcement learning guided by iterative human feedback to improve update quality and consistency. Combining RL with human-in-the-loop correction addresses a known failure mode of fully autonomous KG maintenance: accumulated errors from incorrect entity or relation additions that degrade downstream reasoning. Evaluation on TPC-H and a veterinary clinic dataset provides proof-of-concept, though no code is publicly released.

██████████ 0.7 agent-tool-use Peer-reviewed

Read

E2ER is an open-source software pipeline (v0.4.5, MIT license) that automates the full empirical research cycle from question formulation through data acquisition, analysis, and publication-ready output, with human approval checkpoints at data acquisition stages. The project is relevant as a concrete implementation of autonomous research agents, a topic that generates significant theoretical discussion but few working systems. However, the Zenodo deposit provides no benchmark results or example studies, so it is impossible to assess whether the outputs are scientifically reliable.

██████████ 0.6 agent-tool-use Peer-reviewed

Read

This working paper argues that LLMs cannot distinguish authoritative religious sources from lower-quality content because they optimize for linguistic fluency rather than truth, and proposes a ten-layer governance architecture (VERBUM-SPINE) embedding cryptographic provenance and domain-expert checkpoints for handling doctrinal queries. The underlying observation — that retrieval systems need source-authority metadata, not just text similarity — is a valid point about hallucination grounding in high-stakes domains. The paper is entirely conceptual with no experiments or implementation, so its claims remain untested.

██████████ 0.6 hallucination-grounding Peer-reviewed

Read

This theoretical report introduces 'significance fields' as a structural regulatory layer in a self-developed General Theory of Cognitive Structuring (GTCS), arguing that AI systems prioritize among admissible next actions via distributions over significance rather than flat selection. The framework attempts to provide a vocabulary for why AI agents weight certain continuations over others — a question relevant to alignment and interpretability. The work is purely definitional with no experiments, simulations, or empirical grounding, limiting its practical impact.

██████████ 0.6 alignment-safety Peer-reviewed

Read