Giving AI a toolkit to build better knowledge trees

Taxonomies are the organized hierarchies behind systems that group products, support search or help retrieve information. A model asked to build one must do more than produce plausible labels: it needs to place the supplied items in sensible relationships, include them all and avoid invalid structures. A fluent answer can still omit an item or put it in the wrong branch.

In a preprint, Jonas Limniatis and colleagues introduce TIGA, short for Taxonomy Induction with Tool-Grounded Agents. Their central question is whether an AI agent can make more reliable taxonomies by manipulating an explicit tree through constrained operations, instead of generating the whole hierarchy as text.

A tree, built through steps

The task in the study was specific: given a root concept and a fixed, flat list of candidate entities, construct a rooted tree connecting them. The system was not being asked to discover new concepts from documents or add items to an existing taxonomy. Each non-root item had to have one parent, and the tree had to be connected and free of cycles. Those structural requirements are distinct from whether the chosen parent-child relationships match the benchmark’s reference taxonomy.

TIGA keeps track of the current tree, items still to be placed, an evolving set of organizational guidelines, and a record of recent actions. The language model can use tools to add or remove an item, inspect a subtree, or revise the guidelines. Tool outcomes—including errors—are recorded alongside the model’s stated rationale. In this design, the model chooses an operation, while the tool layer applies it to the structured workspace.

Once the candidate set has been processed, an optional LLM reviewer checks the tree against the original items and the active guidelines. It can request revisions, but the experiments limit review to at most three rounds. This is a check by another model component, not an independent guarantee that the hierarchy is correct.

The researchers evaluated TIGA on four benchmarks: WordNet, DBLP, TExEval-2 and a sampled, depth-limited portion of Google’s product taxonomy. They compared it with two prompting-based approaches: HF, which generates a hierarchy in one pass, and Chain-of-Layer (CoL), which builds it layer by layer. Two language-model backbones were tested, with five independent trials for each configuration. Quality was measured by comparing predicted direct links and broader ancestor-descendant relationships with gold-standard trees. The team also tracked coverage, structural validity and computational measures such as tokens, model calls and inference time.

Better on many scores, not all

TIGA achieved the top edge-level or ancestor-level F1 score in seven of the eight main reported settings. Its strongest advantages were generally in recall and F1: it recovered more of the reference relationships in many configurations. But it did not lead on every precision measure. On some benchmarks, a prompting baseline made fewer parent-child predictions but had a higher proportion of correct ones. TIGA also showed greater run-to-run variability on some settings, including parts of the product-taxonomy evaluation.

Ablations—tests that remove one system component at a time—suggested that both the revisable policy and interaction memory helped on the two datasets tested. Removing either was associated with lower F1. The reviewer’s effect was less consistent: it improved some WordNet scores but slightly reduced scores on TExEval-2 under the revision cap. The study therefore does not show that adding a review step always helps.

The quality gains came with a clear cost. TIGA made many more model calls and used more tokens and inference time than the simpler approaches in the reported comparisons. For example, on TExEval-2 with one backbone, it used about 1,008 calls to construct 18 sub-taxonomies, versus about 161 for CoL in its five-shot setting. Whether the additional quality is worth that overhead depends on how costly taxonomy errors are in a particular application.

What remains uncertain

These results support a promising design, not a universal verdict on AI-generated knowledge structures. The evaluation covers four benchmarks and two main model backbones; some datasets were sampled or truncated to fit the systems’ limits. A small additional study with larger models examined selected cases, rather than providing a broad comparison. The benchmarks also define the task as a rooted tree, which may not suit domains where a concept can have multiple parents.

The authors note that the source of some increased variability is not established. The reviewer’s decisions were not independently validated by human experts, and its bounded revision budget can leave problems unresolved. More broadly, strong agreement with a benchmark tree does not prove that the tree is the only sensible way to organize a domain.

The manuscript identifies itself as a preprint submitted to Elsevier; the supplied source does not establish that it has undergone peer review. Its findings should therefore be read as preliminary. TIGA’s main contribution is a way to make taxonomy construction more explicit and inspectable—not a guarantee that a language model will make the right semantic choices.