diff --git a/src/main/scala/eu/neverblink/jelly/cli/util/jena/riot/RiotParserUtil.scala b/src/main/scala/eu/neverblink/jelly/cli/util/jena/riot/RiotParserUtil.scala
index 99d43e2..1c757c1 100644
--- a/src/main/scala/eu/neverblink/jelly/cli/util/jena/riot/RiotParserUtil.scala
+++ b/src/main/scala/eu/neverblink/jelly/cli/util/jena/riot/RiotParserUtil.scala
@@ -1,6 +1,7 @@
 package eu.neverblink.jelly.cli.util.jena.riot
 
 import eu.neverblink.jelly.cli.command.rdf.util.RdfFormat
+import org.apache.jena.irix.IRIs
 import org.apache.jena.riot.lang.LabelToNode
 import org.apache.jena.riot.{RDFParser, RDFParserRegistry, RIOT}
 import org.apache.jena.riot.system.StreamRDF
@@ -18,9 +19,14 @@ object RiotParserUtil:
   ): Unit = {
     // Only really enable IRI resolution if the format supports it
     if resolveIris && format.supportsBaseIri then
-      // Parser with full IRI resolution
+      // Parser with full IRI resolution.
+      // We set the base explicitly to the system base (the current working directory), which is
+      // what Jena uses by default for stream-based Turtle/TriG parsing. Without this, readers that
+      // do their own base handling (notably RDF/XML via ARP) receive a null base and fail to
+      // resolve relative IRIs (e.g. "#foo"), throwing "Relative URI encountered".
       RDFParser.source(source)
         .lang(format.jenaLang)
+        .base(IRIs.getBaseStr)
         .labelToNode(LabelToNode.createUseLabelAsGiven())
         .checking(false)
         .strict(false)
diff --git a/src/test/scala/eu/neverblink/jelly/cli/command/rdf/RdfToJellySpec.scala b/src/test/scala/eu/neverblink/jelly/cli/command/rdf/RdfToJellySpec.scala
index c9aace6..9ab1d72 100644
--- a/src/test/scala/eu/neverblink/jelly/cli/command/rdf/RdfToJellySpec.scala
+++ b/src/test/scala/eu/neverblink/jelly/cli/command/rdf/RdfToJellySpec.scala
@@ -8,6 +8,7 @@ import eu.neverblink.jelly.core.proto.v1.{LogicalStreamType, PhysicalStreamType,
 import eu.neverblink.jelly.core.JellyOptions
 import eu.neverblink.jelly.core.proto.google.v1 as google
 import eu.neverblink.jelly.core.utils.IoUtils
+import org.apache.jena.irix.IRIs
 import org.apache.jena.rdf.model.{Model, ModelFactory}
 import org.apache.jena.riot.{RDFLanguages, RDFParser}
 import org.apache.jena.sparql.core.DatasetGraphFactory
@@ -934,5 +935,26 @@ class RdfToJellySpec extends AnyWordSpec with TestFixtureHelper with Matchers:
         stmts.head.getPredicate.getURI should be("http://example.org/p")
         stmts.head.getObject.asResource().getURI should be("b")
       }
+
+      "IRI resolution enabled (default), input RDF/XML stream with relative IRIs" in
+        withEmptyJellyFile { j =>
+          val input =
+            """<?xml version="1.0" encoding="utf-8"?>
+              |<rdf:RDF xmlns:rdf="http://www.w3.org/1999/02/22-rdf-syntax-ns#">
+              |  <rdf:Description rdf:about="#relative">
+              |    <__g0:test.org xmlns:__g0="http://">test</__g0:test.org>
+              |  </rdf:Description>
+              |</rdf:RDF>""".stripMargin
+          RdfToJelly.setStdIn(ByteArrayInputStream(input.getBytes))
+          RdfToJelly.runTestCommand(
+            List("rdf", "to-jelly", "--in-format", RdfFormat.RdfXml.cliOptions.head, "--to", j),
+          )
+          val content = translateJellyBack(new FileInputStream(j))
+          val stmts = content.listStatements().asScala.toSeq
+          stmts.size should be(1)
+          stmts.head.getSubject.getURI should be(s"${IRIs.getBaseStr}#relative")
+          stmts.head.getPredicate.getURI should be("http://test.org")
+          stmts.head.getObject.asLiteral().getString should be("test")
+        }
     }
   }