95

Papers

10/10

Roadblocks Active

3

Connections

This paper benchmarks a full end-to-end ML pipeline (GUARDML) across three security tasks — network intrusion, malware, and phishing — finding transformer-based models beat gradient-boosted trees by 4–11 F1 points. Crucially, lightweight adversarial guardrails maintain a 0.91 true positive rate under evasion attacks (versus 0.71 without them) while adding only 7–9% latency, showing robustness need not cost much performance. This matters for AI reliability because it provides a concrete, quantified template for deploying ML models in adversarial environments where inputs are actively manipulated.

██████████ 0.8 hallucination-grounding Peer-reviewed

Read

A near-duplicate entry reporting the same GUARDML pipeline results: transformers achieving 0.91–0.95 macro-F1 across security tasks, with composable guardrails reducing adversarial attack success rates by 32–61%. While the duplication on Zenodo undermines originality, the underlying empirical data on adversarial robustness trade-offs remains the most grounded quantitative contribution in today's pipeline. The consistent replication of figures across both deposits increases confidence in the reported numbers.

██████████ 0.8 hallucination-grounding Peer-reviewed

Read

This dataset evaluates 14 LLMs on retrieving current SaaS product pricing, using human-verified ground truth collected in June 2026 with a ±15% tolerance threshold. It directly quantifies knowledge-currency failures — a concrete form of hallucination where models confidently state outdated facts — across a controlled, real-world domain. The methodology is commercially motivated and lacks rigor (no prompting protocol archived), but the framing of temporal knowledge decay as a measurable benchmark is practically useful for teams tracking LLM factuality over time.

██████████ 0.7 hallucination-grounding Peer-reviewed

Read

This paper decomposes a complex visual recognition task — classifying wood vessel anatomy — into sequential specialist stages: MobileNetV3 for coarse classification (90% accuracy), U-Net for segmentation (0.939 mIoU), and YOLOv11 for detection (0.845 mAP). The decomposed pipeline substantially outperforms what a monolithic approach would likely achieve on this fine-grained spatial reasoning task. This is relevant to the multimodal-understanding roadblock because it demonstrates that breaking spatial reasoning into hierarchical subtasks is a viable path when end-to-end models struggle with detail.

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

Read

The authors extracted a focused subset of 16 thoracic pathology labels from the MultiCaRe clinical case dataset by applying a negation-aware NLP pipeline to radiology reports — addressing the common problem that raw medical corpora contain many negated findings that flip the ground truth label. The curation is publicly deposited on Zenodo with checksums, making it a reproducible resource for chest X-ray classification research. Its main limitation is that the NLP extraction pipeline code is not directly linked in the deposit record, leaving the labeling process partially opaque.

██████████ 0.6 data-quality-curation Peer-reviewed

Read

This purely theoretical paper proposes encoding physical laws and logical axioms as high-priority parameters (Meta-Weights) within a neural network's weight space, arguing this bridges connectionist and symbolic AI without requiring consciousness. There are no experiments, no formal mathematics, and no falsifiable predictions — the framework exists entirely as prose analogy. It is included here because it addresses active questions about alignment and reasoning reliability, but readers should weight it accordingly: it is a conceptual sketch, not a research contribution.

██████████ 0.5 alignment-safety Peer-reviewed

Read

This Zenodo record is a software release (v2.0.0) for a deep learning system that grades liver fibrosis severity from ultrasound images, archived on GitHub and Software Heritage under Apache 2.0. The actual paper describing the model architecture, training data, and performance metrics is referenced but not included in the deposit. Its relevance to AI lies in demonstrating that clinical-grade image grading pipelines can be open-sourced, but the absence of the paper itself makes the contribution impossible to evaluate technically.

██████████ 0.5 interpretability Peer-reviewed

Read

This supplementary data package accompanies an in-person controlled study comparing participants who used ChatGPT versus those who did not when tasked with comprehending and modifying unfamiliar code in three-hour sessions. Qualitative coding used the Polya problem-solving framework, and the replication package includes raw ChatGPT logs, VS Code logs, and analysis scripts — a reasonable level of openness for a qualitative study. The key finding signal is that most participants ran out of time regardless of AI assistance, suggesting ChatGPT may not straightforwardly reduce cognitive load on unfamiliar codebases.

██████████ 0.5 agent-tool-use Peer-reviewed

Read

This document maps a proprietary AI governance framework ('Convergent Architecture') onto the 72 subcategories of NIST AI RMF 1.0, covering GOVERN, MAP, MEASURE, and MANAGE functions across five structural nodes. It is a conceptual alignment exercise with no empirical validation, and the primary framework being mapped references the author's own unpublished works, making independent verification difficult. Its practical value is limited to organizations already using Convergent Architecture who need to demonstrate NIST compliance on paper.

██████████ 0.5 alignment-safety Peer-reviewed

Read

A second supplementary deposit for the same ChatGPT-versus-no-AI code comprehension study, archiving raw participant data including screen recordings metadata, questionnaire responses, and sketched component diagrams from approximately 11 participants. The open data package is a positive contribution to reproducibility in human-AI interaction research. Like its companion deposit, the primary paper is absent, so the connection between raw data and reported conclusions cannot be verified from this record alone.

██████████ 0.4 agent-tool-use Peer-reviewed

Read