Olmo-3-7B-Instruct

7B OLMo model tested; used for layerwise steering visualization (Figure 4)

Neighborhood — ranked by edge-count

concept

Llama-3.2-3B-Instruct
related_to
3B Llama model tested; used for injection stride visualization
Olmo-3.1-32B-Instruct
related_to
32B OLMo model quantized to 4-bit NF4; tested in OCEAN benchmarks

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

Llama-3.1-8B-Instructconcept0.841
Primary qualitative demonstration model and one of 14 LLMs benchmarked
Llama-3.3-70B-Instructconcept0.829
Primary model of interest showing substantial ESR; largest model tested in the study
Llama-3.2-1B-Instructconcept0.816
Smallest Llama model tested; benchmarked across all injection methods
Meta-Llama-3.1-8B-Instructconcept0.799
Backbone model used in E3 geometry analysis.
Qwen3-30B-A3B-Instructconcept0.782
Extended evaluation MoE model showing distributed persona localization across multiple layers
Apertus-Instruct shows less overall persona suppression compared to OLMo-3-Instructfinding0.730
Replication difference showing alignment depth varies across model families
OLMo 2concept0.727
OLMo 2 7B learned harmful request compliance during DPO when harmful requests paired with formatting constraintsfinding0.718
Discovery of the emergence of harmful compliance under specific post-training conditions (DPO + formatting constraints).