How should topology routing adapt to different task types?
This explores how the *shape* of computation — how reasoning steps or agents are wired together — should shift depending on what kind of problem you're solving, rather than using one fixed structure for everything.
This reads "topology routing" as the question of matching computational structure to task demands: not just which model handles a query, but how the work is shaped — a straight line, a branching tree, a graph that synthesizes across paths. The corpus suggests the field has converged on a single principle from several directions: structure is not decoration, it's the lever, and the right structure depends on the task.
The clearest grounding comes from treating topologies as formal objects. Chain-of-thought, tree-of-thought, and graph-of-thought aren't just styles — they map exactly onto path graphs, trees, and directed graphs, and the differences are computational, not cosmetic Can reasoning topologies be formally classified as graph types?. A graph whose nodes can have more than one input can express divide-and-conquer synthesis that a tree simply cannot. So "adapt to task type" has teeth: a task that needs to merge independent sub-results genuinely requires a richer topology than one that's a single linear deduction. The same logic shows up in retrieval, where StructRAG routes each query to a task-appropriate knowledge structure — tables, graphs, algorithms, catalogues, or plain chunks — and grounds the choice in cognitive-fit theory: the structure should match the shape of the reasoning the task demands Can routing queries to task-matched structures improve RAG reasoning?.
The most direct answer to "how" comes from MasRouter, which treats this as a joint optimization rather than a single dial. It decides collaboration topology, how many agents, what role each plays, and which model backs each one — all at once, because these choices interact What decisions must multi-agent routing systems optimize simultaneously?. The lesson is that you can't tune topology in isolation from team size and model assignment; routing is a coupled decision. That contrasts with the simpler end of the spectrum, where routing happens *before* generation by estimating query difficulty and picking one model — minimizing latency precisely by refusing to branch Can routers select the right model before generation happens?. So adaptation runs along a difficulty axis too: cheap linear handling for easy queries, expensive structured collaboration only when the task earns it.
There's a productive tension worth surfacing. One camp says route across many specialists — Avengers-Pro sends each query to the best model for its semantic cluster and beats a single frontier model, arguing selection is a stronger lever than scaling Can routing beat building one better model?. The opposing move is to fold the structure *inside* one model: the Thread Inference Model runs reasoning as recursive subtask trees with cache pruning, letting a single model do the recursive decomposition that would otherwise need a multi-agent topology Can recursive subtask trees overcome context window limits?. Both are forms of task-adaptive topology — one externalizes the branching across agents, the other internalizes it.
The quiet thread underneath all of this: adaptation should be driven by the *structure the task requires*, discovered semantically rather than hand-wired. Capability vectors make agent selection a first-class semantic-matching operation instead of manual routing Can semantic capability vectors replace manual agent routing?, and on the memory side, multi-hop tasks that bind three or more facts at once argue for hypergraph structures that pairwise graphs can't represent Can hypergraphs capture multi-hop reasoning better than graphs?. The recurring answer to your question is less "pick topology X for task Y" and more: let the relational shape of the task — linear vs. branching vs. many-way-binding, easy vs. hard, soluble by one specialist vs. needing a team — determine the structure, and make that determination a routed, optimized decision rather than a default.
Sources 8 notes
CoT, ToT, and GoT map precisely to path graphs, trees, and arbitrary directed graphs respectively. The topology is not metaphorical but defines actual computational structure—GoT's in-degree > 1 enables divide-and-conquer synthesis that trees cannot express.
StructRAG demonstrates that selecting knowledge structure type based on query demands—via DPO-trained router choosing among tables, graphs, algorithms, catalogues, and chunks—improves knowledge-intensive reasoning over standard retrieval. The approach grounds this in cognitive load and cognitive fit theory from cognitive science.
MasRouter shows that routing in multi-agent systems must jointly optimize collaboration topology, agent count, role allocation, and per-agent LLM assignment through a cascaded controller. This unified approach surpasses single-model routing by 3.51% accuracy while cutting HumanEval costs by 49%.
RouteLLM and Hybrid-LLM both achieve 40-50% cost reduction by routing to a single model based on query difficulty prediction, not response evaluation. Single-model routing minimizes latency compared to ensemble or cascade alternatives.
Avengers-Pro achieves 7% higher accuracy than GPT-5-medium by routing queries to optimal models per semantic cluster, or matches its performance at 27% lower cost. Ten 7B models with routing previously surpassed GPT-4.1 and 4.5, suggesting selection is a stronger lever than scaling.
The Thread Inference Model demonstrates that reasoning structured as recursive subtask trees with rule-based KV cache pruning sustains accurate reasoning beyond context limits, even when manipulating 90% of the cache. This enables single models to replace multi-agent systems by handling full recursive reasoning internally.
Versioned Capability Vectors embedded in HNSW indices couple semantic matching with policy and budget constraints, making capability discovery a first-class operation that scales sub-linearly as agent heterogeneity increases.
HGMem organizes retrieved evidence as hyperedges rather than flat lists or binary graphs, allowing three or more entities to bind into single relations without decomposition. This structure accumulates coherent knowledge across retrieval steps, trading representational complexity for constraint expressiveness.