GTA-Atomic: Atomic Evaluation in AI & Graphs

• GTA-Atomic is an evaluation protocol and formal concept defining indivisible, all-or-nothing operations in both AI tool benchmarking and graph transformation.
• In agent benchmarking, it isolates short-horizon tool precision, revealing that state-of-the-art models perform below 50% accuracy in precise tool invocation.
• In functional graph transformation, atomic primitives ensure complete, side-effect-free operations, mimicking transactional semantics without explicit mechanisms.

GTA-Atomic is an evaluation protocol for fine-grained, short-horizon, closed-ended tool use by AI agents and, in an unrelated context, designates a core formal notion of atomicity in functional graph transformation programs. In both settings, atomicity refers to the indivisible, all-or-nothing nature of operations—either in the invocation of application tools (benchmarking general agents), or in the pure functional rule-application over graphs. GTA-Atomic, as deployed in large-scale agent benchmarking and as a core concept in the Graph Transformation control Algebra (GTA), is central to the measurement and formalization of low-level, precise, composable actions, with strong semantics ensuring isolation, completeness, and minimal side effects (Wang et al., 17 Apr 2026, Weber, 2022).

1. GTA-Atomic in Agent Benchmarking

GTA-Atomic constitutes the closed-ended, single-step tool-use component of the hierarchical GTA-2 benchmark suite for General Tool Agents (Wang et al., 17 Apr 2026). This protocol focuses exclusively on short-horizon precision: rather than requiring agents to coordinate, iterate, or plan toward open-ended deliverables, GTA-Atomic tasks center on the correct invocation of exactly the right tool within a realistic, multimodal context. Each instance leverages real user queries and deployed, executable tools, diverging sharply from earlier benchmarks that rely on synthetic queries or dummy tool stubs.

Contrasting with GTA-Workflow—which encompasses long-horizon, open-ended, deliverable-oriented tasks—GTA-Atomic isolates the challenge of fine-grained tool selection and parameterization. Experimental results indicate that state-of-the-art frontier models perform below 50% accuracy on GTA-Atomic, establishing it as a stringent lower bound for compositional tool-use capability. Performance on GTA-Atomic is thus diagnostic of an agent's baseline competence in tool precision before more complex, multi-step workflow reasoning is assessed (Wang et al., 17 Apr 2026).

2. GTA-Atomic Principles in Functional Graph Transformation

Atomicity in functional graph transformation is formalized in the Graph Transformation control Algebra (GTA), which offers a stateless, purely functional foundation for control in graph rewriting systems (Weber, 2022). GTA discards in-place, stateful updates and instead interprets programs as total functions over grapes—finite sequences of graph sets—where each atomic program primitive produces a complete, side-effect-free new layer.

The full set of “atomic primitives” in GTA includes operations such as rule application ($\twoheadrightarrow(r)$), sequential composition ($\vertdiv(p_1,p_2)$), nondeterministic choice ($\div(p_1,p_2)$), and loops ($\circlearrowright(p)$), among others. Each primitive acts as an indivisible, transaction-like unit: an operation either completes entirely or not at all; no partial, intermediate, or interleaved state persists.

GTA Atomic Primitives	Description
$\trianglepa(c)$	Add graph-constraint $c$
$\trianglepacross(c)$	Remove graph-constraint $c$
$\twoheadrightarrow(r)$	Apply transformation rule $r$
$\vertdiv(p_1,p_2)$0	Select up to $\vertdiv(p_1,p_2)$1 by order $\vertdiv(p_1,p_2)$2
$\vertdiv(p_1,p_2)$3	Sequential composition
$\vertdiv(p_1,p_2)$4	Nondeterministic choice
$\vertdiv(p_1,p_2)$5	While-possible loop
$\vertdiv(p_1,p_2)$6	Search until constraint $\vertdiv(p_1,p_2)$7
$\vertdiv(p_1,p_2)$8	Drop isomorphic duplicates
$\vertdiv(p_1,p_2)$9	Cut history marker (“isolation”)

3. Formal Semantics: Atomicity, Isolation, and Non-determinism

Within GTA, every program $\div(p_1,p_2)$0 is interpreted as a (potentially partial) function

$\div(p_1,p_2)$1

over grapes. Application of a single transformation rule via $\div(p_1,p_2)$2 is inherently atomic: starting from a given layer, it deterministically generates a new set of all possible successor graphs without any observable partial state or side effect. Sequential composition $\div(p_1,p_2)$3 ensures that blocks of transformations are evaluated as indivisible units—either both stages complete, or neither leaves a partial trace.

Isolation is guaranteed by the pure functional model: all program executions produce new, immutable grapes, independent from external states or other computations. The $\div(p_1,p_2)$4 primitive optionally truncates the computation history, enforcing deliberate isolation of subcomputations.

Non-determinism, inherent in rule-matching, is encapsulated meta-deterministically: operators such as $\div(p_1,p_2)$5 explore all possible alternatives explicitly, and the power-set semantics of rule application guarantee that no alternative is silently discarded.

4. Transactional Composition and All-or-Nothing Behavior

Composite GTA programs built from atomic primitives inherit transactional semantics. For example, a composition

$\div(p_1,p_2)$6

sequentially applies $\div(p_1,p_2)$7 and $\div(p_1,p_2)$8, ensuring that if any rule is inapplicable (no matching in the current layer), subsequent stages are not executed and no partial results appear. There are no half-committed states at any point. This behavior mirrors traditional all-or-nothing (“atomic block”) semantics in database systems but is achieved without explicit transaction constructs due to the stateless, functional design.

A minimal example demonstrates the effect: if $\div(p_1,p_2)$9 fails to match, all subsequent layers are empty; if $\circlearrowright(p)$0 succeeds but $\circlearrowright(p)$1 fails, the output grape records the results of $\circlearrowright(p)$2 and the failure of $\circlearrowright(p)$3, with no partial update (Weber, 2022).

5. Computational Expressiveness and Minimality

GTA’s core atomic primitives underpin Turing completeness. The operators $\circlearrowright(p)$4 alone suffice to encode nondeterministic rule application, sequential composition, and iteration. The formal construction follows classical encoding of deterministic register machines into GT programs, with nondeterministic choice simulating rule selection, rule application effecting rewrites, sequencing implementing instruction ordering, and looping providing iteration.

Minimality of the GTA basis is formally established: omitting any of the core operators ($\circlearrowright(p)$5, $\circlearrowright(p)$6, $\circlearrowright(p)$7, or $\circlearrowright(p)$8) results in computational deficiency (e.g., loss of Turing completeness). Atomicity and isolation thus derive directly from the functional, three-valued “layer” structure of grapes, eliminating the need for explicit transaction mechanisms or locking protocols (Weber, 2022).

6. Comparative Perspective and Application in Evaluation

GTA-Atomic in benchmarking (as realized in GTA-2) and atomicity in GTA (graph transformation) both foreground indivisible, precise, and side-effect-free operations at the foundation of larger computational workflows. In GTA-2, atomic evaluation tasks establish a lower bound for agent tool-use precision before engagement with more complex, open-ended workflows; in GTA, atomic primitives guarantee reliable, compositional semantics for functional graph rewriting programs.

These atomic notions are critical for ensuring measurement integrity (in benchmarks) and semantic clarity (in functional programming for graph transformation), particularly as systems scale to handle real-world, multimodal, and nondeterministic contexts. The “real-world authenticity” principle in GTA-Atomic benchmarking and the transaction-like functional purity in GTA both reflect increasing demands for rigor, composability, and fidelity in formal and empirical evaluation (Wang et al., 17 Apr 2026, Weber, 2022).