A new research framework called LOM-action addresses a structural flaw in enterprise AI systems: large language models produce answers that sound correct but are untethered from the specific business context of a given decision — and leave no audit trail.
The paper, published on ArXiv in April 2025, argues that existing LLM-based agent systems draw from an unrestricted knowledge space rather than simulating how a specific business event reshapes that space. The result, according to the authors, is decisions that are fluent but ungrounded — a problem they label 'illusive accuracy'.
Why High Accuracy Scores Can Be Misleading
The illusive accuracy phenomenon sits at the heart of the paper's argument. Frontier models Doubao-1.8 and DeepSeek-V3.2 achieved roughly 80% accuracy on the benchmark tasks used in the study — a figure that appears acceptable in isolation. But their tool-chain F1 scores, which measure whether the model correctly executed the full sequence of actions needed to reach a decision, fell to just 24–36%.
Ontology-governed, event-driven simulation, not model scale, is the architectural prerequisite for trustworthy enterprise decision intelligence.
This gap matters because in enterprise environments, reaching the right answer through the wrong process can be just as damaging as a wrong answer. A decision that appears correct but was produced without following proper procedures, checking the right data sources, or logging its reasoning carries real compliance and liability risk.
How LOM-action Works
LOM-action is built around a three-stage pipeline: event → simulation → decision. When a business event occurs — say, a procurement request or a customer escalation — it triggers a set of scenario conditions encoded in what the authors call an Enterprise Ontology (EO). These conditions drive deterministic mutations in an isolated sandbox environment, producing what the paper calls a simulation graph (G-sim): a structured, scenario-specific version of the relevant business knowledge graph.
All decisions are then derived exclusively from this evolved graph, not from the model's general training knowledge. The architecture operates in two modes — skill mode for routine, rule-based decisions, and reasoning mode for more complex judgements requiring inference across the graph. Every decision produces a fully traceable audit log, which the authors position as a core requirement for regulated industries.
The use of a sandboxed simulation graph is a meaningful architectural choice. It means the AI cannot 'hallucinate' context that doesn't exist in the enterprise's actual data structures, and every step in the decision chain can be inspected after the fact.
Benchmark Results and What They Measure
According to the paper — and these benchmarks are self-reported by the research team — LOM-action achieved 93.82% accuracy and 98.74% tool-chain F1 against the frontier model baselines. The authors emphasize the fourfold F1 advantage over Doubao-1.8 and DeepSeek-V3.2, arguing it demonstrates that architectural design outweighs raw model scale for enterprise decision tasks.
The choice of tool-chain F1 as the primary metric is deliberate. Standard accuracy benchmarks ask whether the final answer is correct; F1 over tool chains asks whether the model chose and executed the right sequence of tools and actions to get there. In agentic enterprise systems — where an AI might need to query databases, check policy documents, and log a rationale before acting — process correctness is as important as outcome correctness.
The paper does not detail the full composition of the benchmark dataset used, which makes independent replication difficult to assess from the abstract alone.
The Compliance Case for Auditable AI
The audit trail component of LOM-action addresses a growing pressure on enterprise AI deployments. Regulators in the EU, UK, and increasingly the US are moving toward requirements that high-stakes automated decisions be explainable and logged. Industries including financial services, healthcare, and insurance face particular scrutiny over how AI-generated recommendations are documented.
Current LLM-based agents typically cannot produce a step-by-step record of why they reached a decision — they generate an output, but the internal reasoning is opaque and non-deterministic. LOM-action's graph-based approach produces a deterministic decision path that can, in principle, be reconstructed and reviewed.
The researchers frame this not as an optional feature but as a prerequisite for trustworthy enterprise AI — a distinction that positions the work against the dominant approach of deploying general-purpose frontier models with prompt engineering and hoping the outputs are sufficiently grounded.
What This Means
For enterprises evaluating AI for regulated or high-stakes decision processes, LOM-action's reported results suggest that raw model capability is a poor proxy for operational trustworthiness — and that purpose-built, ontology-constrained architectures may be a viable path to genuine auditability.