AI-RAN distributes network control across multiple autonomous agents, namely xApps and rApps operating over shared radio resources. When an alert or degradation appears, operators currently lack a structured way to trace which decision caused it, what context triggered that decision, and how other agents interacted. Existing explainability tools focus on single-model feature attribution or linear event logging, neither of which captures cross-agent decision chains. We propose the Context Graph, a heterogeneous directed graph that organizes policies, measurements, decisions, and alerts into a queryable explanation layer. Its foundational primitive is the Decision-Event node acting as a causal junction, where incoming edges capture triggering context and outgoing edges track downstream consequences that are progressively upgraded through counterfactual verification. Instantiated over existing E2, A1, and O1 data flows, our proof-of-concept demonstrates that compliance and diagnosis queries are resolved at sub-millisecond latency over graphs of tens of thousands of nodes, while reliably distinguishing true causal impacts from spurious correlations.
Figures & tables
Fig. 1: Three-layer architecture of the Context Graph in AI-RAN. E2 streams enter through the Context-Logger xApp in the Near-RT RIC for sub-second timestamping. A1 policies and O1 VES events are ingested at the Non-RT RIC and SMO. All streams converge in the Context Graph Engine, which assembles nodes and edges, runs counterfactual inference, and serves queries.
Fig. 2: Graph structure of the Context Graph, centered on the Decision-Event node as a causal junction. Incoming edges (left) capture the triggering context; outgoing edges (right) capture verified consequences. For clarity, only the upgraded causes edge is shown. Before counterfactual verification, it exists as a candidate_effect edge.
Type
Description
Source
KPM-Snapshot
Measurement snapshot of a cell
E2SM-KPM report
Policy
Operator intent or SLA constraint
A1 interface
Slice
Network slice
A1 policy
Alert-Event
SLA violation or anomaly alert
VES over O1
Cell
Cell entity
Topology configuration
Agent
xApp or rApp instance
Agent identifier in messages, and RIC platform records
TABLE I: Context Graph Node Types
Type
Meaning
Relation Class
temporal
A occurs before B
Temporal
manages
Agent governs a cell or slice
Entity
executes
Agent performs a decision
Entity
observes
Cell or slice produces a KPM report
Entity
triggers
Policy or KPM context triggers a decision
Semantic
correlates
KPM changes are associated within the same analysis window
Semantic
TABLE II: Context Graph Edge Types
Fig. 3: Explanatory subgraph for the walkthrough chain. Incoming edges to each Decision-Event node capture the policy and KPM snapshot that triggered the decision, while the highlighted causal edges capture its counterfactually verified consequences. Policy identifiers use a version counter, while KPM, decision, and alert identifiers encode cell index and event time. Agent, cell, and slice entity nodes are omitted to focus on the event-level explanation path.
Config
Nodes
Edges
Assembly (s)
Candidate (s)
Annotation (s)
Compliance (ms)
Diagnosis (ms)
6 cells, 5 min
1,944
4,162
0.029
0.0062
0.018
0.0051
0.0053
12 cells, 5 min
3,820
8,216
0.053
0.010
0.080
0.0049
0.0055
24 cells, 5 min
7,666
16,702
0.11
0.023
0.50
0.0049
0.0054
12 cells, 30 min
22,883
48,723
0.33
0.053
0.40
0.0049
0.0052
24 cells, 30 min
45,267
96,002
0.76
0.11
2.0
0.0048
0.0052
TABLE III: Graph Size, Time of Each Stage, and Query Latency Across Configurations
Xijun Wang ([email protected]) received the B.S. degree from Xidian University, China, in 2005, and the Ph.D. degree from Tsinghua University, China, in 2012. Currently, he is an Associate Professor at Sun Yat-sen University, China.
Table 7
Zhaoyang Liu ([email protected]) received the B.E. degree from Sun Yat-sen University, China, in 2022, where he is currently pursuing the Ph.D. degree.
Table 8
Xiang Chen ([email protected]) received the B.E. and Ph.D. degrees from Tsinghua University in 2002 and 2008, respectively. Currently, he is a Full Professor at Sun Yat-sen University, China.
Table 9
Howard H. Yang ([email protected]) received the Ph.D. degree from Singapore University of Technology and Design in 2017. Currently, he is an assistant professor at Zhejiang University, China.
Table 10
Chenyuan Feng ([email protected]) received the Ph.D. degree from Singapore University of Technology and Design in 2021. Currently, she is a researcher at University of Exeter, U.K and a Marie Skłodowska-Curie Scholar.
Future AI-integrated Radio Access Networks (AI-RAN) will combine open programmability with learning-enabled xApps, rApps, and control functions that act on shared parameters and key performance indicators (KPIs). For conflict monitoring, it is not enough to know which applications are deployed; the system must also know whether the parameter--KPI dependencies assumed by runtime diagnosis remain valid under the current operating regime. This paper studies a lightweight monitoring primitive for that purpose: tracking an interpretable dependency representation from streaming telemetry events. We represent active dependencies by a Boolean matrix and use Boolean matrix multiplication to check whether recent parameter-activity and KPI-response events are consistent with the current estimate. We propose a sliding-window inference procedure that reuses the estimate when it remains consistent and recomputes it when recent observations indicate structural change. The tracker is intended as an explainable signal for conflict diagnosis and slow-loop model refresh, not as an autonomous mitigation mechanism. Experiments on controlled Boolean event streams show efficient and accurate tracking under dependency changes and Boolean observation noise.
Christie Djidjev, Nicholas Kaminski
Idaho National Laboratory Idaho Falls, United States
As multi-agent AI systems are increasingly deployed in real-world settings - from automated customer support to DevOps remediation - failures become harder to diagnose due to cascading effects, hidden dependencies, and long execution traces. We present AgentTrace, a lightweight causal tracing framework for post-hoc failure diagnosis in deployed multi-agent workflows. AgentTrace reconstructs causal graphs from execution logs, traces backward from error manifestations, and ranks candidate root causes using interpretable structural and positional signals - without requiring LLM inference at debugging time. Across a diverse benchmark of multi-agent failure scenarios designed to reflect common deployment patterns, AgentTrace localizes root causes with high accuracy and sub-second latency, significantly outperforming both heuristic and LLM-based baselines. Our results suggest that causal tracing provides a practical foundation for improving the reliability and trustworthiness of agentic systems in the wild.
Traces record what LLM agents did, leaving implicit what each step relied on. GRADE holds both in one graph; six published graph views are its projections. Execution edges come free; each costly dependency edge is graded as observed in the trace, declared by added logging, or inferred under an assumption, a schema we test one grade at a time. We recover the dependency layer from six tool-use, coding, and web corpora and measure its failure-prediction gain over run size. First, over a linear run-size baseline, it improves prediction on three corpora, but the gain depends on the learner and task sampling. On held-out corpora the layer stays above chance where run size does not, under one learner, with the same task-sampling caveat. Second, structure can pass for signal it does not carry: a layer inferred under full history restates run size. Within a corpus, a preregistered test does not separate the recovered layer from a degree-matched counterfeit that rewires each step's dependencies but keeps their number. The recovered layer's largest transfer lead over it reverses when the held-out corpus's task family is withheld. Before crediting structure, grade each edge, keep a run-size baseline, and score degree-matched attachment controls.