QLoRA Fine-Tuning for Theory-of-Mind Reasoning in a 3B-Parameter Small Language Model: A ToMBench-Based Cross-Benchmark Generalization Study

We investigate whether quantized low-rank adaptation (QLoRA) fine-tuning improves theory-of-mind (ToM) reasoning in a small language model, and whether apparent cross-benchmark transfer survives control for response-position bias. We fine-tuned 4-bit Qwen2.5-3B-Instruct with low-rank (LoRA) adapters on ToMBench under a story-disjoint split with five seeds. Because fine-tuning shifted the model’s answer-letter preference on every benchmark, we evaluated all models under cyclic permutation of answer options, placing each gold answer in every position once. Held-out ToMBench accuracy rose by +8.15 points (between-seed 95% interval [+7.46, +8.85]), against +14.01 under an item-level split, consistent with story overlap inflating the item-level estimate. On external benchmarks, ToMi improved (+14.19 [+11.87, +16.50]), mostly on items where only the gold answer appears in the story, whereas SocialIQa, OpenToM and Hi-ToM showed no direction; a SocialIQa decline under the item-level split (−4.14) disappeared under the story-disjoint split. Training-free three-shot prompting under the same permutation raised ToMi accuracy by +7.23 points with content-free format exemplars and +12.96 with rationale exemplars but did not reproduce the in-domain gain. We therefore do not read the ToMi gain as transfer of ToM reasoning, and we conclude that cross-benchmark claims for ToM fine-tuning require a position-invariant evaluation and a prompting baseline under the same control.