86

Papers

10/10

Roadblocks Active

0

Connections

This case study documents three recurring failure patterns when deploying LLMs in real software development: performance degrades sharply (not gradually) as task interdependency rises, finite context windows cause silent errors across large codebases, and session resets erase architectural knowledge accumulated during a project. These are not benchmark failures — they occur in production, which makes the taxonomy practically useful for teams evaluating where LLM assistance is and is not safe to rely on.

██████████ 0.9 hallucination-grounding Peer-reviewed

Read

This thesis-level work advances three bottlenecks in robot manipulation simultaneously: reconstructing 3D object shapes from incomplete sensor data using a confidence-guided transformer architecture, learning to recognize novel objects from just a few examples, and filtering out poor-quality training trajectories when teaching robots via imitation learning. Addressing all three together matters because each bottleneck tends to block the others — better perception is wasted without reliable policy learning, and vice versa.

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

Read

Standard computer architectures waste most of their energy moving data between separate sensing, memory, and processing units — a bottleneck that worsens as AI models grow. This perspective surveys emerging hardware that collapses those three functions into one physical location, using memristive and ferroelectric devices plus spiking neural networks to process signals where they are captured. The approach points toward edge AI devices that could run continuous inference at a fraction of current energy cost, which is a prerequisite for real-world deployment of AI in sensors, wearables, and robotics.

██████████ 0.7 efficiency-scaling Peer-reviewed

Read

AI-GENIE uses large language models to automatically generate and iteratively refine psychological survey questions, then validates them using network-based psychometric methods across nearly 5,000 participants in nationally representative U.S. samples. The AI-generated scales matched the structural validity of expert-crafted ones, with measurable improvements in item pool quality (normalized mutual information gains of 8–20 points) across five different LLMs. This is a concrete, tested use case where LLM output quality was rigorously evaluated against human expert output — a methodological model other applied AI domains could adapt.

██████████ 0.7 data-quality-curation Peer-reviewed

Read

This narrative review maps how AI systems can combine MRI, PET, and blood biomarkers to improve diagnosis, tumor segmentation, and outcome prediction for brain cancer — tasks that currently require invasive biopsy. It is a review rather than an empirical study, so specific performance claims should be treated cautiously, but it usefully catalogues which imaging modalities (diffusion-weighted, perfusion, spectroscopic MRI; amino acid PET) carry the most biologically specific signal for multimodal AI fusion. The relevance for AI research is that medical imaging is one of the richest real-world proving grounds for multimodal understanding architectures.

██████████ 0.6 multimodal-understanding Peer-reviewed

Read

This paper applies ensemble classifiers (Random Forest, XGBoost, LightGBM, CatBoost) to detect security vulnerabilities in Ethereum smart contracts, achieving 87.67% accuracy on a curated benchmark of 143 annotated contracts and identifying four structural archetypes in nearly 48,000 real-world contracts via unsupervised clustering. SHAP values are used to explain predictions, and a blockchain oracle stores results on-chain to create an auditable record. The combination of explainability and immutable audit logging addresses a real deployment concern: automated security tools need to be accountable, not just accurate.

██████████ 0.6 interpretability Peer-reviewed

Read

This paper describes a robotic system that collects physiological data via non-contact sensors and uses embedded AI to perform emergency patient triage, reporting 92.5% agreement with expert clinical consensus and 98.1% sensitivity for high-acuity patients, with assessment time cut by roughly 70%. The concrete performance figures make this more evaluable than most applied-AI medical papers, though independent external validation is not reported. It represents a demanding real-world test of embodied AI: the system must sense, reason under time pressure, and produce decisions that clinicians can trust.

██████████ 0.6 embodied-ai Peer-reviewed

Read

Using critical discourse analysis and agricultural case studies, this preprint argues that generative AI like ChatGPT cannot capture knowledge that is relational, tacit, land-based, or transmitted across generations — the kind of knowledge that exists in practice rather than text. The empirical basis is thin (case study and discourse analysis), but the conceptual argument is relevant: it maps a category of knowledge that current training data pipelines cannot represent, which is a real constraint on where LLMs can be trusted. This complements the production failure taxonomy above by pointing to a structural rather than just a performance gap.

██████████ 0.5 hallucination-grounding Peer-reviewed

Read

This work develops machine learning interatomic potentials that can model how materials melt or reorganize at the atomic level — including ultrafast laser-induced melting in silicon and structural anomalies in liquid tellurium — with accuracy close to expensive quantum-mechanical calculations but at far lower computational cost. While the application is materials physics, the methodology directly advances a core AI challenge: building surrogate models that are accurate enough to replace high-fidelity simulators in scientific domains. The approach of combining constrained density functional perturbation theory with learned potentials is a transferable template for other simulation-heavy fields.

██████████ 0.5 efficiency-scaling Peer-reviewed

Read

This study applies machine learning to predict land surface temperature across space and time, using feature importance analysis to identify which environmental variables matter most and uncertainty quantification to bound prediction confidence. The environmental application is straightforward, but the methodological combination — spatiotemporal ML with explicit uncertainty reporting — is representative of a broader push to make applied ML predictions more trustworthy and interpretable in high-stakes domains. The paper adds to a growing body of work showing that interpretability tools are becoming standard practice outside pure ML research.

██████████ 0.4 interpretability Peer-reviewed

Read