exp/active_cases_eval/model_comparison.md

VLM Comparison on 238 Hard Cases (False Negatives)

All 238 cases have gt_label=1. Acc / Recall@1 = fraction correctly identified as causally related.

Model	Acc	Correct	Rec@1	Prec@1	F1@1	Evaluated	ParseFail
qwen3vl_finetuned	53.4%	71/133	42.5%	68.0%	52.3%	133	0
qwen3vl_base	50.8%	63/124	32.4%	68.6%	44.0%	124	9
qwen3vl_instruct	48.7%	56/115	24.3%	73.9%	36.6%	115	18
qwen25vl_7b	42.9%	57/133	7.5%	75.0%	13.6%	133	0
moondream2	42.8%	56/131	32.9%	54.2%	40.9%	131	2
internvl35	42.1%	56/133	8.8%	63.6%	15.4%	133	0
phi35_vision	40.6%	54/133	2.5%	66.7%	4.8%	133	0
internvl3_8b	39.9%	53/133	0.0%	0.0%	0.0%	133	0
qwen2vl_7b	39.9%	53/133	1.2%	50.0%	2.4%	133	0
idefics3_8b	39.0%	37/95	5.0%	75.0%	9.4%	95	38
minicpm_v25	38.9%	7/18	33.3%	83.3%	47.6%	18	115
smolvlm	37.5%	6/16	40.0%	50.0%	44.4%	16	40
llama32_11b_vision	37.0%	27/73	32.6%	50.0%	39.5%	73	60
glm4v_9b	0.0%	0/0	N/A	N/A	N/A	0	0
molmo_7b	0.0%	0/0	N/A	N/A	N/A	0	0