🔬 Research Pulse
Daily Digest
June 06, 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: Existing AI-text detection benchmarks evaluate only final outputs, leaving unclear how AI authorship signals emerge, accumulate, or fade across progressive human-AI co-editing at document, sentence, token, and span granularities.
Summary: OpAI-Bench is a benchmark that simulates progressive human-to-AI co-editing by constructing 9 sequential revisions per document under 5 edit operations across 4 domains, with provenance preserved at document/sentence/token/span levels. Evaluating 17 detectors reveals that detectability is non-monotonic in AI coverage, with intermediate mixed-authorship versions often the hardest to detect.
Key Results: Introduces OpAI-Bench: for each human document, 9 sequentially revised versions are generated across predefined AI coverage levels and 5 representative AI edit operations spanning 4 domains. Evaluated using 8 document-level, 7 sentence-level, and 2 token/span-level detectors. Shows detectability depends on edit operation, domain, and cumulative revision history — and that mixed-authorship intermediate versions are often harder to detect than fully human or heavily AI-edited endpoints (non-monotonic detection).
Key Findings:
- Detection difficulty is governed by edit operation type, domain, and cumulative revision history — not just AI content proportion
- Mixed-authorship intermediate versions are frequently harder to detect than both pure-human and heavily AI-edited endpoints
- Current detectors evaluated only on final outputs miss these non-monotonic patterns, indicating systematic blind spots in deployed systems
Technical Novelty: Operation-guided progressive transformation pipeline that preserves full authorship provenance across 4 granularities (document/sentence/token/span), rather than treating AI-text detection as binary classification on final outputs.
What's New: First benchmark to model AI-text detection as a trajectory problem with operation-level control and multi-granularity provenance, rather than a static binary classification of final outputs.
Extension Opportunities:
- Train detectors directly on intermediate mixed-authorship versions to close the non-monotonic detection gap exposed by the benchmark
- Extend the operation taxonomy to multi-turn agentic editing (e.g., tool-augmented revision, chain-of-thought rewrites) and code/multilingual domains
- Build a real-time IDE plugin that uses span-level detectors to attribute authorship provenance live during co-editing
Replicability: Code and benchmark publicly released at github.com/VILA-Lab/OpAI-Bench. Reproduction primarily requires inference compute over 17 detectors plus an LLM to regenerate the 9-step edit trajectories — moderate single-GPU workload, no training required for evaluation.
Research Gaps:
- No detectors specifically designed for mixed-authorship intermediate drafts where signals are weakest
- Limited understanding of how specific edit operations (paraphrase vs. expand vs. polish) leave distinct, separable forensic signatures
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 enable sustained self-evolution in long-horizon machine learning algorithm discovery?
Summary: MLEvolve is a self-evolving multi-agent LLM framework for automated ML algorithm discovery that combines Progressive Monte Carlo Graph Search, Retrospective Memory, and decoupled planning/coding to overcome the isolation and memory limitations of prior MLE agents. It achieves SOTA on MLE-Bench with half the standard runtime budget and beats AlphaEvolve on mathematical algorithm optimization, demonstrating strong cross-domain generalization.
Key Results: On MLE-Bench, MLEvolve achieves state-of-the-art results on average medal rate and valid submission rate using only a 12-hour budget (half the standard 24-hour runtime). It also outperforms specialized algorithm discovery methods including AlphaEvolve on mathematical algorithm optimization tasks, showing cross-domain generalization.
Key Findings:
- Progressive MCGS with cross-branch graph edges and entropy-scheduled exploration outperforms standard tree search for long-horizon MLE tasks
- Retrospective Memory (cold-start KB + dynamic global memory) enables agents to evolve via accumulated, retrievable experience rather than memoryless search
- Decoupling strategic planning from code generation with adaptive coding modes stabilizes long-horizon iteration and yields SOTA medal rate and valid submission rate on MLE-Bench at 12 hours
- The framework generalizes beyond MLE: it outperforms AlphaEvolve on mathematical algorithm optimization
Technical Novelty: Three combined innovations: (1) Progressive MCGS that extends tree search with graph-based reference edges enabling cross-branch information flow with an entropy-inspired exploration-to-exploitation schedule; (2) Retrospective Memory combining a cold-start domain knowledge base with dynamic global memory for task-specific retrieval; (3) decoupling strategic planning from code generation via adaptive coding modes for stable long-horizon iteration.
What's New: Unlike prior MLE agents that treat each search branch in isolation and lack persistent memory, MLEvolve introduces cross-branch information flow via graph references, a two-tier retrospective memory, and a hierarchical planner/coder split — together targeting the specific failure modes of long-horizon LLM-agent search rather than just scaling existing tree-search baselines.
Extension Opportunities:
- Apply the Progressive MCGS + Retrospective Memory framework to other long-horizon scientific discovery domains (e.g., chemistry, materials science, theorem proving) to test generalization beyond ML/math tasks
- Replace the entropy-inspired exploration-exploitation schedule with learned bandit/RL-based controllers that adapt the broad-to-focused transition dynamically per task difficulty
- Extend Retrospective Memory with cross-user/federated experience sharing so multiple MLEvolve instances collectively accumulate and retrieve algorithm-discovery experience
Replicability: Code is publicly available at github.com/InternScience/MLEvolve. Reproduction requires substantial compute — MLE-Bench evaluation under the reported 12-hour budget per task across multiple Kaggle-style competitions, plus LLM API costs for the multi-agent loop (likely tens to hundreds of GPU-hours plus significant LLM inference spend).
Research Gaps:
- No reported analysis of failure modes or per-competition variance on MLE-Bench — unclear where the framework breaks down
- Limited evidence on how memory quality/scale impacts performance as the global memory grows over many tasks (potential memory pollution or retrieval degradation)
3. Goedel-Architect: Streamlining Formal Theorem Proving with Blueprint Generation and Refinement
Authors: Jui-Hui Chung, Ziyang Cai, Zihao Li... Published: 2026-06-04 | Citations: 0 arXiv | PDF
Research Question: How can LLM-based formal theorem proving in Lean 4 avoid the inefficient dead-end loops of recursive lemma decomposition while scaling to hard competition problems?
Summary: Goedel-Architect is an agentic Lean 4 theorem-proving framework that generates a global blueprint of definitions and lemmas as a dependency graph, then closes each lemma in parallel with a tool-equipped prover and refines the blueprint on failure. It achieves state-of-the-art open-source results on MiniF2F, PutnamBench, IMO 2025, and Putnam 2025 at up to 500x lower cost than comparable pipelines.
Key Results: Using DeepSeek-V4-Flash (284B-A13B) backbone, Goedel-Architect achieves 99.2% pass@1 on MiniF2F-test and 75.6% pass@1 on PutnamBench. With natural-language proof seeding: 100% on MiniF2F-test, 88.8% (597/672) on PutnamBench, 4/6 on IMO 2025, 11/12 on Putnam 2025, and 3/6 on USAMO 2026 — at up to 500x lower cost than comparable open-source pipelines.
Key Findings:
- Blueprint-based global planning outperforms recursive lemma decomposition by avoiding dead-end loops
- Optional natural-language proof seeding meaningfully lifts performance on hard problems (PutnamBench 75.6% → 88.8%)
- Pipeline solves frontier competition math: 4/6 IMO 2025, 11/12 Putnam 2025, 3/6 USAMO 2026 with an open-weight backbone
Technical Novelty: Replaces recursive lemma decomposition with a global blueprint (dependency DAG of definitions/lemmas) that is generated up front, attacked in parallel by a tool-equipped prover, and refined globally when lemmas fail — avoiding the dead-end loops typical of top-down recursive approaches.
What's New: Reframes formal proving as blueprint generation + parallel lemma closing + global refinement, rather than the dominant recursive top-down or pure search-based paradigms; combines this with optional informal-proof guidance and a tool-equipped prover.
Extension Opportunities:
- Apply blueprint-then-refine paradigm to other proof assistants (Coq, Isabelle, Rocq) to test generality beyond Lean 4
- Integrate auto-formalization so blueprints can be seeded directly from informal textbook proofs at scale
- Use the blueprint dependency graph as training data for a smaller distilled model that learns lemma decomposition end-to-end
Replicability: Abstract claims open-source pipeline and uses open-weight DeepSeek-V4-Flash (284B-A13B MoE) — reproducible in principle but requires substantial compute for a 284B-param backbone plus Lean 4 tool infrastructure; code availability not explicitly stated in abstract.
Research Gaps:
- No clear mechanism described for verifying soundness of dependency declarations in generated blueprints
- Limited evidence on how the approach scales to research-level (non-competition) mathematics with deep theory hierarchies
🔥 GitHub Trending
1. anthony80188/medical-rag-chatbot
⭐ 17 stars | TypeScript
A medical FAQ chatbot powered by Retrieval-Augmented Generation (RAG) using FastAPI, LangChain, Qdrant, and React.
chatbot docker fastapi gemini langchain llm
2. Ericcc-Ma/learn-hermes-agent
⭐ 3 stars | Python
Build a Self-Evolving Agent Harness from Scratch | 12 lessons | Anthropic/DeepSeek/Qwen/GLM
agent ai anthropic artificial-intelligence claude deepseek
3. param087/agent-ml-skills
⭐ 2 stars | JavaScript
Production-grade Machine Learning, Data Science & MLOps skills for AI coding agents (Codex, Claude Code, Cursor, OpenCode). One npx command to install.
agent-skills ai-agents claude-code codex cursor data-science
4. wulongovo/llm-finetune-kit
⭐ 2 stars | Python
🔥 大模型LoRA微调工具包 | 针对消费级GPU优化 | 支持Qwen/DeepSeek/LLaMA | SFT + DPO
ai deepseek dpo fine-tuning gpu huggingface
5. quzhi-ai/deep-learning
⭐ 2 stars | Unknown
A Socratic tutor that refuses to hand you the answer — until you can prove you actually understand. 一个拒绝直接给答案的导师，直到你能证明自己真的懂了。
ai-agent claude-code claude-skill deep-learning learning socratic-method
6. br1ge/TelegramAI
⭐ 2 stars | Python
Telegram AI assistant powered by LLMs. Supports conversational AI, command handling, and seamless Telegram integration
ai ai-assistant artificial-intelligence assistant automation bot-framework
7. programmersd21/polaris_studio
⭐ 2 stars | Python
💫 Visual data pipelines with AI-reviewed transformations and live spreadsheet output, built offline-first with Polars and PySide6.
ai-assistant apache-arrow csv data-engineering data-ide data-pipeline
8. Shubham03-hub/Demand-Forecasting-Inventory-Optimization
⭐ 1 stars | Python
Machine Learning-powered Demand Forecasting and Inventory Optimization system using XGBoost, Streamlit, and MLflow. Forecasts future demand, calculates safety stock and reorder points, and provides in
datanalytics datascience demandforecasting inventory-management machine-learning mlflow
9. bsmensah-ctrl/nq-futures-ml-risk-audit
⭐ 1 stars | Python
Independent replication and overfitting audit for an ML-gated NQ futures research strategy
backtesting futures machine-learning monte-carlo quant-finance risk-management
10. m1n1v1rus/Semantic_Search_System___Flipkart_Reviews
⭐ 1 stars | Jupyter Notebook
An intelligent Semantic Search System built with Sentence Transformers and Python to perform context-aware semantic search on Flipkart product reviews.
flipkart-reviews jupyter-notebook machine-learning nlp python semantic-search
11. Noor-Sabah05/CodeRush-GDC-ITU-2026-Machine-Learning-Competition
⭐ 1 stars | Jupyter Notebook
Economic class classification (lower/middle/upper) from bag-structured demographic data. Multi-seed 5-fold ensemble of LightGBM & XGBoost with OOF threshold tuning. Top 5 · Coderush 2026
classification ensemble-learning--bagging kaggle-competition lightgbm-classifier machine-learning python
12. harshmukhi/Campus2Career
⭐ 1 stars | Python
Resume-based career recommendation platform built with Python, Streamlit, and Machine Learning.
ai machine-learning python resume-parser streamlit
13. Dong04-123/Dong-AI-Company
⭐ 1 stars | Python
Infinite Context AI Company — Graph Memory eliminates context windows. Corporate governance (red/blue debate, board review), dynamic worker pools, persistent cross-project memory. Ultra-large project
agent-framework ai-agent deepseek llm machine-learning mcp
14. siddemmohankrishna/Flappy-Bird-AI-using-Deep-Q-Network-DQN-
⭐ 1 stars | Python
Built a Deep Q-Network (DQN) agent using PyTorch and Gymnasium to learn and play Flappy Bird through Reinforcement Learning.
ai-agent artificial-intelligence data-science deep-learning dqn flappy-bird
15. m0k0ut/ai-product-manager-handbook
⭐ 1 stars | Python
A practical 2026 handbook for product managers and AI product managers — PM fundamentals, LLMs, RAG, evals, and responsible AI. 10 parts, 30 chapters, 2 glossaries.
ai ai-product-management artificial-intelligence generative-ai handbook llm
Generated by Research Pulse on 2026-06-06 06:02