Circuit Recovery on Qwen 0.5B -> 3B 2.5 benchmark leaderboard · SOTA2 Research

Circuit Recovery on Qwen 0.5B -> 3B 2.5

0.32Arithmetic (-) ScoreNAP-IG-Inputs Gold (Upper Bound)

Evaluation Results

Method
NAP-IG-Inputs Gold (Upper Bound)2026.04	0.32	0.91	1.13	—	1.06	0.94	0.87
NAP-IG-Inputs DFA (In-Distribution)2026.04	0.26	0.24	0.37	—	0.34	0.34	0.31
NAP-IG-Inputs DFA (Near-Distribution)2026.04	0.26	—	—	—	0.35	0.33	0.31
NAP DFA (In-Distribution)2026.04	0.26	0.25	0.37	—	0.37	0.38	0.33
NAP-IG-Activations DFA (In-Distribution)2026.04	0.26	0.25	0.36	—	0.29	0.3	0.29
Random Alignment2026.04	0.25	0.25	0.25	—	0.25	0.25	0.25
NAP-IG-Inputs DFA (Far/Zero-Shot)2026.04	0.25	0.3	0.32	—	0.28	0.28	0.29
NAP Gold (Upper Bound)2026.04	0.25	0.25	0.82	—	0.89	1.1	0.66
NAP DFA (Near-Distribution)2026.04	0.25	—	—	—	0.31	0.42	0.33
NAP-IG-Activations Gold (Upper Bound)2026.04	0.25	0.25	0.9	—	0.9	1.11	0.68
NAP-IG-Activations DFA (Near-Distribution)2026.04	0.25	—	—	—	0.34	0.39	0.33
NAP-IG-Activations DFA (Far/Zero-Shot)2026.04	0.25	0.25	0.34	—	0.34	0.48	0.33
NAP DFA (Far/Zero-Shot)2026.04	0.24	0.32	0.27	—	0.32	0.35	0.3