NCA-GENM - Core Machine Learning and AI Knowledge (20% of the exam) - Section 2.4
Understand model fusion and the difference between modality and agent orchestration.
Distinguish model fusion - combining representations from multiple networks - from agent orchestration, which coordinates discrete model calls through a control layer. Choose the appropriate pattern based on whether integration must happen at the feature level or the workflow level.
Model fusionOrchestration
Practice question for this objective
lock_openFree sampleCore Machine Learning and AI Knowledgehard
An engineer is asked to explain the conceptual difference between modality orchestration and agent orchestration to a new team member. Which statement most accurately captures how the two differ in what they coordinate?
AModality orchestration coordinates how different data modalities such as image, audio, and text are encoded, aligned, and combined within a pipeline, whereas agent orchestration coordinates how autonomous agents or tools are selected, sequenced, and chained to accomplish a task.check_circle Correct
BModality orchestration coordinates which large language model answers a query and which prompt template is applied, whereas agent orchestration coordinates which GPU node executes the chosen model and how much memory it is allocated at inference time.
CModality orchestration decides the order in which model weights are loaded into device memory during start-up, whereas agent orchestration decides the order in which training batches are sampled and shuffled across each training epoch.
DModality orchestration and agent orchestration are interchangeable names for routing inference requests across replicas behind a load balancer, differing only in whether the incoming workload happens to be text based or image based.
Modality orchestration coordinates encoding and combining of data types, while agent orchestration coordinates selecting and sequencing autonomous agents or tools. The two orchestration concepts operate at different abstraction levels: one governs how heterogeneous modalities are processed and fused inside a model or pipeline, and the other governs how independent agents or tools are chosen and chained to complete a multi-step task, which is why they are not interchangeable.
Why A is correct: This correctly separates the two: modality orchestration concerns handling and combining heterogeneous data types, while agent orchestration concerns sequencing autonomous agents or tools toward a goal, which is the intended distinction.
Why B is wrong: This conflates orchestration with model selection and hardware scheduling; neither modality nor agent orchestration is defined by choosing prompts or GPUs, so the description misstates both concepts.
Why C is wrong: These describe memory management and data loading rather than coordination of reasoning or modalities, so the statement attaches the two terms to unrelated runtime and training concerns.
Why D is wrong: Treating the terms as synonyms for load balancing is a common oversimplification, but they describe distinct coordination problems, so calling them interchangeable is wrong.
Exam traps in Core Machine Learning and AI Knowledge
Answers that look right on this material and are not. Each one is a distractor from a different question in the NCA-GENM bank for this domain.
Modality orchestration coordinates which large language model answers a query, while agent orchestration coordinates how image and audio tensors are normalised before fusion.
errorWhy it is wrong: It is tempting because both terms involve coordination, but the two roles are reversed: handling input data types is the modality concern and routing decisions among reasoning components is the agent concern.
The upload pipeline is agent orchestration because it routes each input type to a different encoder, and the assistant feature is modality orchestration because it works with text requests.
errorWhy it is wrong: Routing fixed input types to encoders is deterministic modality handling, not agentic behaviour, and the assistant's runtime tool decisions are agentic, so both labels here are assigned to the wrong system.
Both systems perform agent orchestration, because each one routes inputs to more than one specialised component and combines the results into a final output for the user.
errorWhy it is wrong: Tempting because both do route to multiple components, but agent orchestration specifically requires runtime, data-dependent tool selection; the listing pipeline's path is fixed and predetermined, so it is not agent orchestration.