From Apparent Self-Reference to First-Person Framing: A Causal Ablation Study in Gemma 3 4B

This paper revisits an apparent self-reference signal in Gemma 3 4B through a same-dataset,post-hoc reinterpretation of a 32-triplet MLP-ablation study. Across 6,528 interventions spanningall 34 layers, zero-ablation self-vs-third-person comparisons initially highlighted Layers 9, 26, and27, but matched first-person non-AI controls shifted the interpretation toward first-person framingrather than AI-specific self-reference. Layer 26 showed a stable ∼3.2-nat differential KL acrossself-recognition, capability awareness, and metacognition prompts. At that layer, the AI-specificresidual under the self-vs-control diagnostic was small and statistically indistinguishable fromzero in all three domains (range: −0.027 to 0.261; all p > 0.38); Layers 9 and 27 showed thesame cross-domain pattern qualitatively, with smaller and non-replicating residuals than theself-vs-third-person effect. Training-knowledge prompts, the strongest signal in an earlier v1study, yielded no FDR-significant layers under the tighter v2 design, though that null remainsambiguous because prompt matching is weakest in that domain. A narrow v3 attention-headfollow-up found no replicated AI-specific alternative. Within this English prompt setting, theevidence is more consistent with first-person framing than with AI-specific self-reference inGemma 3 4B’s MLP layers.