Boundless DAS finds causal subspaces for group-binding, recency and stereotype pronoun mechanisms; recency is distinct, binding and stereotype overlap

Trinley et al. train Boundless DAS per layer on RUFF-based diagnostic datasets (14,400 examples each) in six instruction-tuned LLMs (Llama-3.1-8B, OLMo-2 1B/7B/13B, Gemma-2-9B, Qwen2.5-7B) and recover a causal subspace for each of group binding, recency and stereotype [trinley-etal-2026-pronoun-fidelity-mechanisms] Best-layer interchange-intervention accuracy is moderate to high (0.60-1.00 in the layer search), with the dominant mechanism varying by model and mechanisms often peaking at different layers within one model [trinley-etal-2026-pronoun-fidelity-mechanisms] Cross-mechanism transfer of frozen rotations shows the recency rotation is the most distinct, whereas group-binding and stereotype rotations score alike and are never clearly separable [trinley-etal-2026-pronoun-fidelity-mechanisms] A three-term mixture over inference-only patching distributions reaches Jensen-Shannon similarity 0.915-0.995 across ten models, with recency and stereotype carrying most of the signal and group binding contributing little [trinley-etal-2026-pronoun-fidelity-mechanisms] Head ablations show a localized concept-level copying route (few heads) versus a distributed token-level route; this is component-level evidence and not a subspace claim [trinley-etal-2026-pronoun-fidelity-mechanisms]