Back to newsletter
·Daily digest

🔬 Research Pulse

Daily Digest

June 05, 2026

🤖 AI

🧠 LLMs

1. Operation-Guided Progressive Human-to-AI Text Transformation Benchmark for Multi-Granularity AI-Text Detection

Authors: Sondos Mahmoud Bsharat, Jiacheng Liu, Xiaohan Zhao... Published: 2026-06-04 | Citations: 0 arXiv | PDF

Research Question: How do AI-authorship signals emerge, accumulate, or disappear during progressive human-AI co-editing, and can detectors identify AI involvement across multiple granularities (document, sentence, token, span) rather than only in final outputs?

Summary: OpAI-Bench introduces an operation-guided benchmark that simulates progressive human-to-AI text co-editing through 9 sequential revision versions per sample, enabling multi-granularity AI-text detection analysis at document, sentence, token, and span levels. Evaluation of 17 detectors reveals that AI detectability is non-monotonic — mixed-authorship intermediate drafts can be harder to detect than fully AI-edited ones — and depends on edit operation, domain, and revision history.

Key Results: Introduced OpAI-Bench with nine sequentially revised versions per sample across predefined AI coverage levels and 5 representative AI edit operations spanning 4 domains. Evaluated 8 document-level, 7 sentence-level, and 2 token/span-level detectors. Demonstrated that detectability depends on edit operation, domain, and cumulative revision history — not just AI proportion. Notably found non-monotonic detection patterns where mixed-authorship intermediate versions are harder to detect than fully human or heavily AI-edited endpoints.

Key Findings:

  • AI-text detectability is governed by edit operation type, domain, and cumulative revision history, not just AI content proportion
  • Non-monotonic detection patterns exist: mixed-authorship intermediate versions are often harder to detect than both fully human and heavily AI-edited endpoints
  • Existing detectors evaluated on final outputs miss critical authorship signals that emerge and dissipate during the revision process

Technical Novelty: First benchmark to preserve complete authorship provenance across progressive co-editing trajectories at multiple granularities simultaneously, using operation-guided sequential revisions with controlled AI coverage levels — rather than treating documents as static human-vs-AI binary artifacts.

What's New: Prior benchmarks treat AI-text detection as a binary classification on final documents. OpAI-Bench is the first to model the progressive revision trajectory itself, with controlled coverage levels, operation taxonomy, and granular provenance labels enabling token/span-level evaluation across editing stages.

Extension Opportunities:

  • Train a detector explicitly conditioned on edit-operation type and revision history to recover the non-monotonic signal that current detectors miss
  • Extend the benchmark to capture human-after-AI revision chains (AI-to-human direction) and multi-round collaborative loops rather than only human-to-AI progression
  • Build a browser/editor plugin that logs real provenance traces and uses OpAI-Bench-trained span detectors to highlight AI-edited regions in live documents

Replicability: Code and benchmark publicly available at github.com/VILA-Lab/OpAI-Bench. Reproducing evaluations requires running 17 detectors across ~9 revision versions per document over 4 domains — moderate GPU compute, feasible on a single workstation for inference-only detector evaluation.

Research Gaps:

  • No coverage of human-edits-on-AI-drafts direction or iterative back-and-forth collaborative editing
  • Limited to predefined synthetic edit operations rather than naturalistic logs of real human-AI editor sessions

2. MLEvolve: A Self-Evolving Framework for Automated Machine Learning Algorithm Discovery

Authors: Shangheng Du, Xiangchao Yan, Jinxin Shi... Published: 2026-06-04 | Citations: 0 arXiv | PDF

Research Question: How can LLM-based MLE agents overcome inter-branch information isolation, memoryless search, and lack of hierarchical control to sustain self-evolution across long-horizon machine learning algorithm discovery tasks?

Summary: MLEvolve is a self-evolving multi-agent LLM framework for automated ML algorithm discovery that addresses long-horizon optimization through Progressive MCGS (graph-based search with adaptive exploration scheduling), Retrospective Memory (cold-start knowledge + dynamic experience reuse), and decoupled planning/coding modes. It achieves SOTA on MLE-Bench at half the standard runtime and beats AlphaEvolve on mathematical algorithm optimization, showing strong cross-domain generalization.

Key Results: On MLE-Bench, MLEvolve achieves state-of-the-art results on average medal rate and valid submission rate under a 12-hour budget (half the standard runtime). It also outperforms specialized algorithm discovery methods including AlphaEvolve on mathematical algorithm optimization tasks, demonstrating cross-domain generalization.

Key Findings:

  • Progressive MCGS with cross-branch reference edges outperforms tree-search-based MLE agents by enabling information flow between exploration branches
  • Combining cold-start domain knowledge with dynamic global memory allows the agent to accumulate and reuse task-specific experience across iterations
  • MLEvolve reaches SOTA medal rate and valid submission rate on MLE-Bench using only 12 hours (half the standard budget) and generalizes beyond MLE to outperform AlphaEvolve on math algorithm tasks

Technical Novelty: Three coupled innovations: (1) Progressive Monte Carlo Graph Search (MCGS) that replaces tree search with graph reference edges enabling cross-branch information flow plus an entropy-inspired schedule shifting exploration to exploitation; (2) Retrospective Memory combining a cold-start domain knowledge base with dynamic global memory for experience retrieval; (3) decoupling of strategic planning from code generation via adaptive coding modes for stable long-horizon iteration.

What's New: Prior MLE agents use tree search with isolated branches, no persistent memory, and entangled planning+coding. MLEvolve is the first to combine graph-structured search with progressive exploration scheduling, persistent retrospective memory, and hierarchical planning/coding separation in a single self-evolving MLE framework — and to demonstrate that this generalizes beyond MLE benchmarks to algorithm discovery domains where AlphaEvolve is the prior SOTA.

Extension Opportunities:

  • Apply Progressive MCGS with graph-based reference edges to other long-horizon agentic domains (e.g., software engineering, theorem proving, hardware design) where branch isolation limits exploration
  • Extend Retrospective Memory with cross-task transfer learning so the dynamic global memory can be shared across users/projects, building a community-wide MLE experience corpus
  • Replace the entropy-inspired progressive exploration schedule with a learned controller (e.g., RL-trained meta-policy) that adapts exploration/exploitation tradeoffs per task domain

Replicability: Code is publicly released at https://github.com/InternScience/MLEvolve. Reproduction requires substantial compute — MLE-Bench evaluations under a 12-hour budget per task across multiple Kaggle-style competitions, plus LLM API calls for the multi-agent system (likely tens to hundreds of GPU-hours-equivalent in API spend per full benchmark run).

Research Gaps:

  • No discussion of how Retrospective Memory scales or degrades as the global memory grows large — retrieval quality and contamination risks are unaddressed
  • Cross-domain generalization is shown on math algorithm tasks but not on broader scientific discovery domains; limits of the approach outside MLE/math are unclear

🤖 Agents

1. The Post-GCN Decade Revisited: Curvature-Stratified Evaluation of Relational Learning

Authors: Shuo Wang, Xiangyu Wang, Quanxin Wang... Published: 2026-06-04 | Citations: 0 arXiv | PDF

Research Question: Do flat, averaged leaderboards in relational learning hide systematic, geometry-dependent performance differences across graph datasets with different intrinsic curvature?

Summary: The paper argues that averaged GNN leaderboards conflate datasets with fundamentally different intrinsic geometries, hiding real performance trade-offs. It introduces CurvBench, a curvature-stratified benchmark over 14 datasets and 18 models showing that model rankings are stable within a curvature regime but shift across regimes, and that Graph Foundation Models do not uniformly dominate geometry-aligned GNNs.

Key Results: Benchmarked 18 representative models (GCNs, Graph Foundation Models, and tabular methods) across 14 datasets stratified into positive, negative, and near-zero curvature regimes. Showed that model rankings are stable within a curvature regime but shift significantly across regimes, and that GFMs yield diminishing returns vs. geometry-aligned GNNs in certain regimes — evidence that aggregated metrics mask real trade-offs.

Key Findings:

  • Model rankings are highly stable within a curvature regime but shift significantly across positive/negative/near-zero regimes.
  • Graph Foundation Models show diminishing returns vs. geometry-aligned GNNs in specific curvature regimes, contradicting the 'universal model' narrative.
  • Standard aggregated metrics systematically mask geometry-dependent trade-offs and can lead to misleading generalization claims.

Technical Novelty: A curvature-stratified evaluation protocol (positive/negative/near-zero regimes) for relational learning benchmarks, rather than yet another model — reframing benchmarking itself around intrinsic graph geometry.

What's New: Prior benchmarks treat datasets as interchangeable and report averaged metrics; this work makes intrinsic curvature a first-class axis of evaluation and empirically shows it is a latent confounder behind reported GNN/GFM superiority.

Extension Opportunities:

  • Extend CurvBench to heterogeneous and temporal graphs, stratifying by curvature evolution over time or per edge-type subgraph
  • Use the curvature regime as a routing signal in a mixture-of-experts GNN that dispatches subgraphs to hyperbolic, Euclidean, or spherical encoders
  • Build an automated 'geometry-aware model selector' that estimates a new dataset's curvature distribution and recommends the top-ranked architecture for that regime

Replicability: Code, curvature-stratified dataset splits, and evaluation tools are publicly released at sirbabbage.github.io/CurvBench_HOME. Reproduction requires standard GNN/GFM training compute across 14 datasets × 18 models — feasible on a single multi-GPU workstation.

Research Gaps:

  • No principled, geometry-aware protocol existed for comparing GNNs, GFMs, and tabular methods on graph tasks.
  • Limited empirical understanding of when Graph Foundation Models actually generalize vs. when curvature-aligned specialist GNNs are preferable.

🔥 GitHub Trending

1. CHAITHANYAHEGDE/O-RAN-Fronthaul-Intelligence-Platform

⭐ 1 stars | TypeScript

AI-powered O-RAN fronthaul network intelligence platform for topology discovery, capacity estimation, traffic analytics, and explainable AI.

data-analytics fastapi machine-learning network-analysis networkx o-ran

2. J1mL1/arxiv2conf

⭐ 1 stars | Python

Resolve arXiv preprints to official conference proceedings pages — ACL, CVPR, ICLR, NeurIPS, ICML, AAAI, ECCV, and more.

arxiv conference machine-learning paper-search proceedings publication-venue

3. Hxh-yaoxing/skill-weave

⭐ 1 stars | Python

Adaptive skill routing for multi-agent systems — 4-dimension scoring engine

agent ai llm machine-learning multi-agent open-source

4. hitesh099/pg-smart-basket-analytics

⭐ 1 stars | Unknown

Smart Basket Retail Analytics & Demand Forecasting — Procter & Gamble

arima clustering demand-forecasting fmcg machine-learning power-bi

5. lfmen/world-cup-predictor

⭐ 1 stars | Python

Machine Learning model to predict the FIFA World Cup 2026 using XGBoost, Poisson regression, and EA FC 26 data.

machine-learning python world-cup

6. mukuldhattarwal/pitwall_intel

⭐ 1 stars | Python

🏎️ End-to-end Formula 1 race prediction platform using OpenF1 data, XGBoost, SHAP explainability, and Streamlit-powered what-if simulations.

data-science f1 formula1 machine-learning openf1 plotly

7. sumbal00001234-jpg/ML-HeartDisease-Clustering

⭐ 1 stars | Jupyter Notebook

Unsupervised Machine Learning: K-Means, Hierarchical Clustering, and DBSCAN on Heart Disease Dataset | Python | Scikit-learn

clustering data-science dbscan heart-disease-prediction hierarchical-clustering jupyter-notebook

8. shatakshisingh28/Portfolio_Analyzer

⭐ 1 stars | HTML

🚀 AI-powered GitHub Portfolio Analyzer that evaluates repositories, measures portfolio strength, and provides personalized recommendations for career growth.

artificial-intelligence developer-tools github-api github-insights machine-learning portfolio-scoring

9. AbhishekVerma3208/sales-forecasting-system

⭐ 1 stars | JavaScript

AI-powered Retail Sales Prediction System using Artificial Neural Networks (ANN) to forecast sales trends from 100,000+ records for data-driven business decisions.

ann deep-learning machine-learning retail-analytics

10. Manjeetkumarr/phishing-email-detection-model

⭐ 1 stars | Python

Machine Learning-based phishing email detection system using NLP, TF-IDF, and Logistic Regression with a Tkinter GUI for real-time email classification.

email-security machine-learning nlp phishing-detection python3 scikit-learn

11. Manshi08044M/facial-emotion-recognition

⭐ 1 stars | Python

Real-time facial emotion recognition using TensorFlow, Keras, OpenCV, and CNN/MobileNetV2.

cnn computer-vision deep-learning facial-emotion-recognition keras opencv

12. anuragN2107/Nexusnet-mlt-clv-Fraud-Engine

⭐ 0 stars | Jupyter Notebook

A Multi-Task Deep Learning Engine built in PyTorch that simultaneously predicts 90-day Customer Lifetime Value (CLV) and flags transaction fraud in a single forward pass (<50ms latency). Deployed on H

ann deep-learning entity-embedding fraud-detection gradio huggingface

13. mishra-khushboo/Tomato-Disease-Classification-and-Severity-Estimation

⭐ 0 stars | Jupyter Notebook

Dual-task EfficientNet for plant disease classification + severity estimation with Grad-CAM

agriculture computer-vision deep-learning efficientnet grad-cam pytorch

14. ekkkkki/GenerRNA

⭐ 0 stars | Python

GenerRNA: a generative pre-trained language model for de novo RNA design (PLOS ONE 2024). Code & docs — model weights on Hugging Face (pfnet/GenerRNA).

bioinformatics computational-biology de-novo-design deep-learning drug-discovery generative-model

15. stutikeerthana/AQA-GAN-Low-Light-Enhancement

⭐ 0 stars | Python

Adaptive Quality-Aware GAN for nighttime image enhancement using FiLM-conditioned U-Net and Quality-Aware PatchGAN.

computer-vision deep-learning gan image-enhancement image-processing pytorch

Generated by Research Pulse on 2026-06-05 06:03