inferwire
/
AI·5 min read

MAKA: LLM Agents With a Critic for Machining Titanium Blades

Researchers split an LLM assistant for CNC machining into four agents, with a critic that traces every number to a tool, and test it on 16 titanium rotor blades.

TL;DR

  • MAKA splits an LLM assistant for CNC machining into four agents; a critic checks numbers, sources and safety bounds before recommendations reach a human [^1].
  • On multi-step tool tasks for titanium rotor blades, the structure beat a single model with the same tools by up to 87.5 percentage points [^1].

Background

CNC machines cut along tool paths planned in CAM software, yet the part still deviates from its design: the machine tracks the path imperfectly, the tool wears and thin walls bend under cutting forces [^1]. Engineers correct this by adjusting offsets, paths and parameters, often by trial and error [^1]. Simulation, inspection and shop experience mostly meet through manual interpretation, and off-the-shelf LLM assistants, the authors say, can neither run multi-step numerical work reliably nor show where their numbers come from [^1].

What happened

Danny Hoang and Farhad Imani of the University of Connecticut, with colleagues from the Connecticut Center for Advanced Technology (CCAT), Quinnipiac University and the U.S. Army DEVCOM Ground Vehicle Systems Center, posted the paper to arXiv on May 5 [^1]. Their system, multi-agent knowledge analysis (MAKA), splits the assistant into four roles. A Central agent interprets a request and routes it. An Analysis agent may report a number only if a deterministic tool computed it, and records which tool call did. A Knowledge Graph agent answers from a graph built from machining literature and cites the entries it used. A Critic agent checks intent, that provenance, the cited evidence and safety bounds such as maximum offset sizes, then accepts the answer, sends it back for revision or escalates to a human [^1].

The team tested MAKA on rotor blades. At CCAT, 16 blades 3 inches (76.2 mm) long were cut from Ti-6Al-4V titanium on a 4-axis Trak VMC10si and scanned with a GOM ATOS ScanBox at 30 inspection points, paired across the two sides of the blade [^1]. MAKA combines those scans with deflection simulations from ThirdWave and path-tracking predictions from Siemens Run MyVirtual Machine. It splits each deviation into a path component, a drift across parts that stands in for tool wear, and a residual it treats as compliance; a separate variability measure flags unstable spots for caution or escalation [^1]. The knowledge graph, built automatically from four papers on machining Ti-6Al-4V with a locally run gpt-oss-120b model, holds 2,701 triples [^1].

The benchmark had 75 questions, split evenly into tasks needing one, two, or three or more dependent tool calls [^1]. Six open models, among them GPT-OSS-20b and Qwen3-VL-30b, reached high pass rates on single calls even without MAKA but fell behind on longer chains, where MAKA's gain reached 87.5 percentage points [^1]. When the researchers removed tool hints from the first routing step on purpose, the critic raised mean tool-selection F1, a score for calling the right tools, from 0.2919 to 0.6697 and fully repaired 61.19 percent of the damaged runs [^1]. The knowledge graph lifted the mean score on 75 open-ended machining questions from 0.3462 to 0.4736, while time per question rose from 0.7305 to 2.6496 seconds [^1].

Why it matters

The design takes the numbers away from the language model. Machining-critical quantities come from deterministic tools and traceable retrieval rather than free text generation, which the authors see as the basis for accountable operation [^1]. Their tests back the claim that the extra agents are reliability mechanisms rather than optional add-ons: on longer chains, single models failed through wrong call order, invalid arguments and lost intermediate results, the failure modes MAKA's structure targets [^1]. The full reasoning loop runs in 4.3 seconds, which the authors call negligible next to blade machining cycles of minutes to hours, so it could run between machining stages [^1].

The results stop at simulation, though. In a digital twin, MAKA's staged compensations cut predicted surface deviation from the order of 10⁻² inch to about ±0.001 inch over most of the blade, and the authors present this as a baseline before deployment, explicitly not as a physical closed-loop correction [^1]. Simulations and scans agree on where errors concentrate, at the blade tip and leading edge; matching them point by point was not the goal [^1]. The tests cover one blade design, one machine and a graph built from four papers; for another part, the authors say, only the tool layer would need replacing, without retraining the agents [^1].

Practical example

Say Dana is a process engineer at a small aerospace supplier that mills titanium rotor blades. After the latest batch, the scans show deviations creeping up from blade to blade, and Dana types a request like the one in the paper: load the aggregated inspection file and give compensation for parts 4 to 16 [^1].

The Central agent turns that into a structured instruction for the Analysis agent [^1]. That agent chains tools: it pairs the measurements from both sides of the blade, subtracts the path-tracking error predicted by the virtual machine, fits a drift across the parts, and converts the chosen correction into length and radius offsets for a tool tilted at 25 degrees [^1]. Each number carries the ID of the tool output that produced it [^1].

This is where a single model tends to slip. Given the same tools, it more often calls them in the wrong order, passes an invalid argument or loses an intermediate result [^1]. In MAKA, the Critic checks that every number traces to a tool output and that the offsets stay inside bounds from the knowledge graph and shop-floor limits [^1]. At one inspection pair the scans spread widely from blade to blade, so the variability measure marks it as low confidence, and the Critic asks for a conservative offset there or hands that decision to Dana [^1].

Dana gets a table of offsets with an audit trail, checks the predicted result in the digital twin and decides whether the change goes into the CAM program [^1].

Related gear

We recommend this book because it explains how CNC programs apply tool length offsets and cutter radius compensation, the kinds of correction MAKA calculates.

AdvertisementAmazon

CNC Programming Handbook

Sources

  1. [1]arXiv — Physics-Grounded Multi-Agent Architecture for Traceable, Risk-Aware Human–AI Decision Support in Manufacturing