We test whether large language models solve formally matched kinship problems equally well when relations are expressed in familiar vocabulary or by explicitly defined nonce predicates. Across 500 paired graphs, concrete accuracy exceeds arbitrary accuracy by 35.6 percentage points in local Qwen3.8-27B, 26.6 in Gemma 4 26B-A4B, 12.0 in Gemma 4 31B, and 5.4 in Qwen3.8-Max. All four paired gaps are statistically resolved. Reasoning budgets and prompt-language interventions can substantially reduce the difference, showing that it is modifiable rather than a fixed incapacity. The minimal conclusion is behavioral: on these tasks, the models' manifested relational competence is not indifferent to presentation. Explicit definitions provide the formal relations but do not make nonce predicates as usable as familiar vocabulary embedded in learned linguistic associations.
Figures & tables
Figure 1: Paired concrete-minus-arbitrary accuracy on the same 500 graphs per model. Diamonds denote strict whole-response reruns; circles denote original off cells whose saved outputs were observationally answer-only. Intervals resample graphs. Marker types flag distinct protocols, not model architectures.
Model
Concrete
Arbitrary
Gap
Gap 95% CI
C-only/A-only
Valid C/A
Qwen3.8-27B
64.6
29.0
+35.6
[30.6,40.6]
203/25
394/253
Gemma 4 26B
96.8
70.2
+26.6
[22.6,30.6]
135/2
499/496
Gemma 4 31B
98.2
86.2
+12.0
[9.0,15.2]
64/4
496/491
Qwen3.8-Max
98.4
93.0
+5.4
[3.2,7.8]
32/5
496/495
Table 1: Answer-only baseline percentages (gap in points). “Valid” gives canonical responses out of 500 in each presentation. Invalid outputs remain incorrect. Local Qwen and Gemma 26B use strict whole-response scoring; Gemma 31B and Qwen-Max use original final-line scoring, but every saved output is single-line.
Relation family ( n )
Q27
G26
G31
Q-Max
Grandchild (68)
+16.2
+2.9
+2.9
+1.5
Lateral (168)
+41.7
+19.6
+17.3
+0.0
Great-avuncular (66)
-4.5
+51.5
+31.8
+31.8
Marriage-linked (198)
+50.5
+32.3
+4.0
+2.5
Table 2: Answer-only C − A gap in points by gold relation family. Family counts sum to 500. These descriptive partitions show heterogeneity and are not four independent confirmatory tests per model.
Figure 2: Original common-protocol accuracy curves (500 graph pairs per point). Off* means the dedicated reasoning channel was disabled, but local Qwen and Gemma 26B sometimes elaborated in the answer channel. Unconnected diamonds to its left show their separate strict answer-only reruns; no line joins them to the budget curve. Horizontal positions are categorical.
Model
Output
C
A
Gap
Qwen3.8-27B
Free
61.2
49.6
+11.6
Qwen3.8-27B
Inventory
78.8
88.6
-9.8
Qwen3.8-27B
Coded
87.2
89.0
-1.8
Gemma 4 31B
Free
92.0
56.4
+35.6
Gemma 4 31B
Inventory
95.6
91.6
+4.0
Gemma 4 31B
Coded
99.4
99.0
+0.4
Table 3: Augmented 500-graph follow-up, percentages and C − A points. “Inventory” displays English relation labels; “coded” displays a randomized label-to-letter mapping. This is a different task set and protocol from Figure 2 .