18 papers
Applied & Industry
Agents in the real world: supply chain, operations research, business process, discovery.
- Papers
- 18
- Read time
- 5.2h
- Avg length
- 17m
A Survey of Data Agents: Emerging Paradigm or Overstated Hype?
Borrowing the SAE self-driving levels, this survey defines a 6-level autonomy ladder (L0–L5) for "data agents," maps every published system onto it, and argues the honest answer to the hype question is "we're stuck at the L2→L3 jump — real autonomous orchestration over messy data doesn't exist yet."
- Read
- 22 min
- Source
- arXiv:2510.23587
America's AI Action Plan
The U.S. federal government's official roadmap to "win the AI race" by stripping regulation, building energy + data-center infrastructure at speed, and exporting the full American AI stack to allies while locking China out of compute.
- Read
- 18 min
An Agentic AI for a New Paradigm in Business Process Development
Instead of hard-wiring a business process as a fixed sequence of tasks, describe it as a set of *goals*, the *information objects* that prove each goal is met, and the autonomous *agents* that produce those objects — and let the runnable workflow emerge from which objects are ready.
- Read
- 14 min
Artificial Intelligence for Operations Research: Revolutionizing the OR Process
A field map of where machine learning plugs into the classic "predict → model → solve" optimization pipeline, so you can replace hand-tuned solver heuristics and human modeling with learned components that adapt to your data.
- Read
- 22 min
- Source
- arXiv:2401.03244
Clustering Market Regimes Using the Wasserstein Distance
Instead of describing each slice of market history by a handful of summary statistics, treat each slice as a full probability distribution and cluster those distributions directly using optimal-transport distance — a tiny, model-free tweak to k-means that detects bull/bear regimes more robustly than moment-matching or hidden Markov models.
- Read
- 16 min
- Source
- arXiv:2006.03487
Deep Reinforcement Learning for Machine Scheduling: Methodology, the State-of-the-Art, and Future Directions
A field map of how deep reinforcement learning is being used to schedule factory machines — sorting every approach into four "brains" (plain neural nets, encoder-decoders, graph neural nets, and metaheuristic hybrids) and telling you which one wins on speed, scale, and generalization.
- Read
- 22 min
- Source
- arXiv:2310.03195
Demand Forecasting at Wayfair
Wayfair forecasts 18 months of monthly demand for 4 million products by blending a stable statistical "top-down" forecast with a responsive ML "bottom-up" forecast, mixing them per-item and per-horizon — and it earned tens of millions of dollars a year.
- Read
- 14 min
Extract-0: A Specialized Language Model for Document Information Extraction
A 7B-parameter model, fine-tuned for $196 with synthetic data + LoRA + a semantic-similarity reward, beats GPT-4.1 and o3 at pulling structured JSON out of documents.
- Read
- 18 min
- Source
- arXiv:2509.22906
Fine-Tuning Vision-Language Model for Automated Engineering Drawing Information Extraction
A 0.23B-parameter open-source vision-language model, fully fine-tuned on just 400 expert-labeled engineering drawings, beats zero-shot GPT-4o and Claude-3.5-Sonnet by 30-52% on extracting tolerancing data — proving that for a narrow visual-extraction task, a tiny tuned model crushes a giant general one.
- Read
- 14 min
Future of Work with AI Agents: Auditing Automation and Augmentation Potential across the U.S. Workforce
Stanford asked 1,500 workers across 104 jobs which of their tasks they actually *want* AI agents to take over, scored those same tasks for technical feasibility with 52 AI experts, and built a public map (WORKBank) showing where agent builders are aiming at the wrong targets.
- Read
- 16 min
- Source
- arXiv:2506.06576
How People Use ChatGPT
OpenAI ran LLM classifiers over a privacy-scrubbed sample of ~1.1M real ChatGPT conversations to measure, for the first time from the inside, who uses the product and what they actually do with it — and the headline is that ~70% of usage is non-work, "writing" and "decision support" dominate, and coding/companionship are surprisingly tiny.
- Read
- 16 min
Kosmos: An AI Scientist for Autonomous Discovery
Kosmos is an AI scientist that runs for 12 hours straight — 200 agent rollouts, 42,000 lines of analysis code, 1,500 papers — staying coherent the whole time by routing everything through a shared, structured "world model" instead of one ballooning chat context.
- Read
- 22 min
- Source
- arXiv:2511.02824
Large Language Models and Operations Research: A Structured Survey
A map of how LLMs are eating operations research — turning plain-English problem descriptions into solvable optimization models, generating and evolving the heuristics that solve them, and in some cases solving the problem directly — organized into three clean pillars you can build a service line around.
- Read
- 22 min
- Source
- arXiv:2509.18180
Large Language Models for Supply Chain Decisions
Put an LLM in front of your optimization solver — not in place of it — so non-technical planners can interrogate, stress-test, and re-run supply-chain models in plain English, cutting decision time from days to minutes while your proprietary data never leaves your walls.
- Read
- 16 min
- Source
- arXiv:2307.03875
Optimizing Multi-Tier Supply Chain Ordering with LNN+XGBoost: Mitigating the Bullwhip Effect
Pair a tiny, fast "liquid" neural net (great at reading time-series dynamics) with XGBoost (great at squeezing accuracy out of tabular features), then drive a profit-maximizing order simulator with the combined forecast — and you get more profit and less of the demand-amplification ("bullwhip") chaos that wrecks multi-tier supply chains.
- Read
- 16 min
- Source
- arXiv:2507.21383
Optimizing Supply Chain Networks with the Power of Graph Neural Networks
This paper takes a real-world supply-chain dataset (41 FMCG products, their dependencies, and 221 days of demand), defines "predict next demand for one product" as a benchmark task, and shows that graph-aware neural networks can forecast it far more accurately than a plain neural net — establishing baselines others can build on.
- Read
- 14 min
- Source
- arXiv:2501.06221
Real-Time Health Supply Chain Optimization Using Digital Twin Technology
Wire live IoT sensor data into a continuously-updating virtual model of a health supply chain so AI can forecast disruptions (spoilage, stockouts, equipment failure) and recommend fixes *before* they happen — instead of after.
- Read
- 14 min
Tabular Data Understanding with LLMs: A Survey of Recent Advances and Challenges
Feeding a 2D table into a 1D language model is a lossy, format-sensitive mess, and this survey maps out how people represent tables, what tasks they test on, and the three structural gaps (shallow reasoning, fragility on complex inputs, no format generalization) that anyone building table-aware AI keeps tripping over.
- Read
- 16 min
- Source
- arXiv:2508.00217