AI Research Reproduction Rate Only 13.98%, as Geely Smart Charging Puts Large Models Into the Charging Network

AI is being tested on two fronts at once. One is in research: PaperBenchX uses 93 real paper reproduction tasks to measure whether models can rerun scientific workflows. The other is in industry: Geely is putting an energy large model into its charging network, seeking to turn fast charging, thermal control and battery management into verifiable system capabilities. What both threads share is that AI's value is no longer just about giving answers, but about whether it can reliably run a process end to end and take responsibility for the results.^[1][2]

Paper Reproduction Becomes a Touchstone for AI Research

UniPat AI has released PaperBenchX, building 93 reproduction tasks from 93 research papers, spanning 12 research directions including electromagnetics, photonics, chemistry, materials, biology and robotics, and 10 domain-native scientific environments, with 3168 expert-validated scoring items. It is the world's first benchmark for multidisciplinary end-to-end reproduction of paper results. An agent receives a real paper, a containerized environment with the corresponding scientific software installed, and a task sheet stating which part to reproduce. After submission, the system deletes all outputs, disconnects the network, and reruns the workflow from scratch in an isolated environment. The scorer trusts only the replayed artifacts. What it tests is not guessing answers, but rebuilding a scientifically valid process.^[1]

Full Reproduction Rate Only 13.98%, Verification Is the Weakest Link

Across the 93 paper reproduction tasks, the strongest-performing model, GPT-6 Astra, had a full reproduction rate of only 13.98%. Stage scores show that modeling and execution averaged 62.70% and 61.84% respectively, while verification scored only 42.80%. This means an agent can complete part of the modeling, call solvers and generate result files, but those results are not necessarily correct, nor can they necessarily prove that the results truly support the paper's conclusions. Trajectory analysis also shows that more interaction rounds do not necessarily mean a higher score; long runtimes often indicate repeated attempts, failure recovery, or continued computation under the wrong model setup. Fable 5 spent 37.1% of its time on paper reconstruction and code implementation, yet achieved near-optimal partial scores with fewer interactions. This suggests that whether early decisions on paper understanding, scientific modeling and parameter selection are reasonable largely determines whether later computation becomes a waste of resources.^[1]

Geely Smart Charging: AI Charging Manages Speed, Temperature and Lifespan at Once

On September 23, 2026, Geely Auto Group officially launched its next-generation AI charging technology, 'Geely Smart Charging', achieving for the first time a single-gun peak charging power of 2250 kW, keeping the maximum charging temperature within 65°C at all times, and increasing battery cycle life by 20%. Relying on the Xingrui PowerMind energy brain jointly developed by Geely and StepFun, Geely Smart Charging builds an integrated collaborative architecture of people, vehicles, chargers, cloud, network and storage. Its AI thermal-control technology can predict internal battery temperature changes 30 seconds in advance and intelligently allocate power dynamically once signs of overheating are detected. Geely also launched a new generation of its Aegis Golden Brick Battery Ultra Short Blade series. The 395 mm version used in the Lynk & Co 10 has a maximum charging rate of 12C, while the 370 mm version raises peak charging rate to 6C, and it will debut in the new Geely Galaxy E5.^[2]

The Charging Network Must Turn Peak Power Into a Stable Experience

Geely Haohan Energy has integrated the charging maps of Zeekr, Lynk & Co and Galaxy, and opened them to new-energy vehicle owners across society. To date, it has deployed more than 2500 charging stations and more than 12,000 charging guns, covering 232 cities nationwide, including more than 1500 ultra-fast charging stations and more than 7000 fast-charging guns. According to the plan, by the end of 2027, Geely aims to build more than 22,000 charging stations and more than 100,000 charging guns, including more than 15,000 Geely Smart Charging stations and more than 50,000 smart charging guns, taking the lead nationwide in achieving 'county-to-county coverage'. These figures show that AI charging is not a single-point technology, but depends on coordination among chargers, vehicles, cloud, network and storage. According to tests by CATARC (China Automotive Technology and Research Center), during the ultra-fast charging of the Lynk & Co 10 from 10% to 97% SOC, peak charging power reached 1093 kW, charging temperature stayed below 55°C for 57% of the time, and the maximum temperature throughout was only 64°C. Its AI ultra-fast charging function averages only 4 minutes 30 seconds to go from 10% to 70% SOC at room temperature, and only 8 minutes 40 seconds from 10% to 97%.^[2]

Common Trend: AI Enters a Verifiability Race

PaperBenchX uses reproduction as a ruler, and Geely uses the charging system as an examination room. Both are pushing AI from demonstration to verification. The research side needs evidence that can be regenerated, while the industry side needs hard metrics such as temperature, lifespan and network scheduling; only when processes can be rerun and results can be verified can AI capabilities expand from individual tasks to broader scientific and industrial scenarios. What is worth watching next is whether multidisciplinary end-to-end reproduction can form a comparable long-term evaluation, and whether AI charging solutions can maintain a balance among safety, efficiency and battery health on a larger-scale network.^[1][2]