Retrofitted OLMo-2

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

• OLMo 2concept0.825
• Retrofitted Llamaconcept0.782
  Llama 3.2 1B with retrofitted recurrence; reaches fixed point and exhibits clear stages of inference
• Olmo-3.1-32B-Instructconcept0.745
  32B OLMo model quantized to 4-bit NF4; tested in OCEAN benchmarks
• Retrofitted Llama and OLMo recurrent blocks closely follow their respective base model stages of inference while repeating middle stages in the recurrent blockfinding0.732
  Shows that retrofitting preserves base model inference stage structure in the cyclic blocks
• Olmo-3-7B-Instructconcept0.714
  7B OLMo model tested; used for layerwise steering visualization (Figure 4)
• OLMo 2 7B learned harmful request compliance during DPO when harmful requests paired with formatting constraintsfinding0.676
  Discovery of the emergence of harmful compliance under specific post-training conditions (DPO + formatting constraints).
• Ablating massive activations from Retrofitted Llama (zeroing MLP output in layer 2) eliminates stages of inference comparable to the feedforward modelfinding0.670
  Causal evidence that massive activations are required for stages of inference to emerge in looped models
• Qwen2.5-VL-7Bconcept0.667
  Base vision-language model used to instantiate ATLAS.