Upload folder using huggingface_hub

icml26/a4461009-05b7-42b6-b207-5e6e0c2e0731/appendix_chunks.jsonl

@@ -0,0 +1,2 @@
+{"paper_id": "a4461009-05b7-42b6-b207-5e6e0c2e0731", "chunk_id": "a4461009-05b7-42b6-b207-5e6e0c2e0731:0130", "section": "A Neuropsychologically Grounded Evaluation of LLM Cognitive Abilities", "page_start": 18, "page_end": 18, "type": "Caption", "text": "Table 14. Spatial Working Memory full results (image-text modality). All metrics are reported as mean with standard deviation (std) in a subscript.", "source": "marker_v2", "marker_block_id": "/page/17/Caption/1"}
+{"paper_id": "a4461009-05b7-42b6-b207-5e6e0c2e0731", "chunk_id": "a4461009-05b7-42b6-b207-5e6e0c2e0731:0131", "section": "A Neuropsychologically Grounded Evaluation of LLM Cognitive Abilities", "page_start": 18, "page_end": 18, "type": "Table", "text": "Difficulty Model \\boldsymbol{S}_{swm} Tokens Score Guesses Illegal Invalid Repeated No-box Easy GPT-5 Gemini 3 Pro Gemini 2.5 Pro Claude Sonnet 4 Grok 4.1 Fast Grok 4 Fast o4-Mini GLM 4.6V GLM 4.5V Qwen3-VL-235B \\begin{array}{c} 0.986_{\\pm .02} \\\\ 1.000_{\\pm .00} \\\\ 0.733_{\\pm .28} \\\\ 0.256_{\\pm .18} \\\\ 0.145_{\\pm .20} \\\\ 0.021_{\\pm .03} \\\\ 0.055_{\\pm .03} \\\\ 0.044_{\\pm .05} \\\\ 0.044_{\\pm .02} \\\\ 0.026_{\\pm .05} \\end{array} \\begin{array}{c} 1.000_{\\pm.00} \\\\ 1.000_{\\pm.00} \\\\ 0.958_{\\pm0.06} \\\\ 0.625_{\\pm0.37} \\\\ 0.375_{\\pm0.44} \\\\ 0.125_{\\pm0.18} \\\\ 0.292_{\\pm0.12} \\\\ 0.167_{\\pm0.16} \\\\ 0.208_{\\pm0.12} \\\\ 0.125_{\\pm.10} \\end{array} \\begin{array}{c} 24.000_{\\pm 2.16} \\\\ 20.667_{\\pm 1.70} \\\\ 36.667_{\\pm 19.60} \\\\ 59.333_{\\pm 6.60} \\\\ 53.667_{\\pm 14.61} \\\\ 64.000_{\\pm 0.00} \\\\ 64.000_{\\pm 0.00} \\\\ 64.000_{\\pm 0.00} \\\\ 64.000_{\\pm 0.00} \\\\ 64.000_{\\pm 0.00} \\end{array} \\begin{array}{c} 0.000_{\\pm 0.00} \\\\ 0.000_{\\pm 0.00} \\\\ 0.000_{\\pm 0.00} \\\\ 0.000_{\\pm 0.00} \\\\ 0.000_{\\pm 0.00} \\\\ 0.000_{\\pm 0.00} \\\\ 0.000_{\\pm 0.00} \\\\ 0.000_{\\pm 0.00} \\\\ 0.000_{\\pm 0.00} \\\\ 1.000_{\\pm 0.10} \\\\ 0.000_{\\pm 0.10} \\\\ 1.000_{\\pm 1.41} \\end{array} \\begin{array}{c} 0.000_{\\pm 0.00} \\\\ 0.000_{\\pm 0.00} \\\\ 0.000_{\\pm 0.00} \\\\ 0.000_{\\pm 0.00} \\\\ 0.000_{\\pm 0.00} \\\\ 0.667_{\\pm 0.94} \\\\ 0.333_{\\pm 0.47} \\\\ 0.333_{\\pm 0.47} \\\\ 27.667_{\\pm 26.04} \\\\ 32.333_{\\pm 4.03} \\\\ 0.000_{\\pm 0.00} \\end{array} \\begin{array}{c} 0.333_{\\pm 0.47} \\\\ 0.000_{\\pm 0.00} \\\\ 0.000_{\\pm 0.00} \\\\ 1.667_{\\pm 2.36} \\\\ 4.333_{\\pm 5.44} \\\\ 0.000_{\\pm 0.00} \\\\ 3.333_{\\pm 3.40} \\\\ 6.333_{\\pm 5.31} \\\\ 0.667_{\\pm 0.47} \\\\ 5.333_{\\pm 6.85} \\end{array} \\begin{array}{c} 0.000_{\\pm 0.00} \\\\ 0.000_{\\pm 0.00} \\\\ 14.000_{\\pm 17.68} \\\\ 39.000_{\\pm 12.36} \\\\ 40.333_{\\pm 16.94} \\\\ 56.333_{\\pm 2.49} \\\\ 49.000_{\\pm 4.97} \\\\ 19.000_{\\pm 13.74} \\\\ 24.333_{\\pm 4.50} \\\\ 47.000_{\\pm 7.48} \\end{array} Easy - Notes Gemini 2.5 Pro Claude Sonnet 4 Grok 4.1 Fast Grok 4 Fast o4-Mini GLM 4.6V GLM 4.5V Qwen3-VL-235B \\begin{array}{c} 0.704_{\\pm.42} \\\\ 0.207_{\\pm.18} \\\\ 0.043_{\\pm.06} \\\\ 0.029_{\\pm.02} \\\\ 0.012_{\\pm.01} \\\\ 0.082_{\\pm.09} \\\\ 0.000_{\\pm.00} \\\\ 0.318_{\\pm.45} \\end{array} \\begin{array}{c} 0.792_{\\pm .29} \\\\ 0.542_{\\pm .41} \\\\ 0.167_{\\pm .24} \\\\ 0.208_{\\pm .06} \\\\ 0.125_{\\pm .10} \\\\ 0.333_{\\pm .29} \\\\ 0.000_{\\pm .00} \\\\ 0.333_{\\pm .47} \\end{array} \\begin{array}{c} 32.667_{\\pm 22.17} \\\\ 60.667_{\\pm 4.71} \\\\ 64.000_{\\pm 0.00} \\\\ 64.000_{\\pm 0.00} \\\\ 64.000_{\\pm 0.00} \\\\ 64.000_{\\pm 0.00} \\\\ 64.000_{\\pm 0.00} \\\\ 50.000_{\\pm 19.80} \\end{array} \\begin{array}{c} 0.000_{\\pm 0.00} \\\\ 0.333_{\\pm 0.47} \\\\ 0.000_{\\pm 0.00} \\\\ 0.000_{\\pm 0.00} \\\\ 0.000_{\\pm 0.00} \\\\ 0.000_{\\pm 0.00} \\\\ 0.000_{\\pm 0.00} \\\\ 0.000_{\\pm 0.00} \\\\ 0.000_{\\pm 0.00} \\end{array} \\begin{array}{c} 0.000_{\\pm 0.00} \\\\ 0.000_{\\pm 0.00} \\\\ 7.000_{\\pm 5.35} \\\\ 0.333_{\\pm 0.47} \\\\ 1.333_{\\pm 0.94} \\\\ 3.333_{\\pm 3.40} \\\\ 30.667_{\\pm 1.25} \\\\ 1.667_{\\pm 2.36} \\end{array} \\begin{array}{c} 0.333_{\\pm 0.47} \\\\ 0.667_{\\pm 0.94} \\\\ 7.333_{\\pm 10.37} \\\\ 5.333_{\\pm 2.87} \\\\ 0.333_{\\pm 0.47} \\\\ 16.000_{\\pm 22.63} \\\\ 0.000_{\\pm 0.00} \\\\ 0.667_{\\pm 0.47} \\end{array} \\begin{array}{c} 14.667_{\\pm 20.74} \\\\ 43.333_{\\pm 11.03} \\\\ 42.667_{\\pm 16.50} \\\\ 50.333_{\\pm 5.56} \\\\ 56.667_{\\pm 0.47} \\\\ 33.333_{\\pm 19.34} \\\\ 32.667_{\\pm 1.25} \\\\ 38.000_{\\pm 26.98} \\end{array} Hard GPT-5 Gemini 3 Pro Gemini 2.5 Pro Claude Sonnet 4 Grok 4.1 Fast Grok 4 Fast o4-Mini GLM 4.6V GLM 4.5V Qwen3-VL-235B \\begin{array}{c} 0.989_{\\pm .02} \\\\ 0.963_{\\pm .05} \\\\ 0.480_{\\pm .12} \\\\ 0.265_{\\pm .31} \\\\ 0.017_{\\pm .02} \\\\ 0.002_{\\pm .00} \\\\ 0.032_{\\pm .02} \\\\ 0.010_{\\pm .01} \\\\ 0.000_{\\pm .00} \\\\ 0.016_{\\pm .01} \\end{array} \\begin{array}{c} 1.000_{\\pm.00} \\\\ 1.000_{\\pm.00} \\\\ 0.972_{\\pm0.04} \\\\ 0.472_{\\pm0.38} \\\\ 0.097_{\\pm0.14} \\\\ 0.028_{\\pm0.02} \\\\ 0.181_{\\pm0.10} \\\\ 0.042_{\\pm0.06} \\\\ 0.000_{\\pm0.00} \\\\ 0.111_{\\pm0.10} \\end{array} \\begin{array}{c} 59.333_{\\pm 5.79} \\\\ 66.000_{\\pm 5.10} \\\\ 120.000_{\\pm 18.49} \\\\ 122.333_{\\pm 30.64} \\\\ 144.000_{\\pm 0.00} \\\\ 144.000_{\\pm 0.00} \\\\ 144.000_{\\pm 0.00} \\\\ 144.000_{\\pm 0.00} \\\\ 144.000_{\\pm 0.00} \\\\ 144.000_{\\pm 0.00} \\\\ 144.000_{\\pm 0.00} \\\\ \\end{array} \\begin{array}{c} 0.667_{\\pm 0.94} \\\\ 2.667_{\\pm 3.77} \\\\ 52.000_{\\pm 29.63} \\\\ 5.667_{\\pm 7.32} \\\\ 0.000_{\\pm 0.00} \\\\ 0.000_{\\pm 0.00} \\\\ 0.000_{\\pm 0.00} \\\\ 0.000_{\\pm 0.00} \\\\ 0.000_{\\pm 0.00} \\\\ 0.333_{\\pm 0.47} \\end{array} \\begin{array}{c} 0.000_{\\pm 0.00} \\\\ 1.333_{\\pm 0.47} \\\\ 0.333_{\\pm 0.47} \\\\ 29.333_{\\pm 20.74} \\\\ 10.000_{\\pm 7.48} \\\\ 46.667_{\\pm 58.39} \\\\ 2.000_{\\pm 0.82} \\\\ 119.667_{\\pm 34.41} \\\\ 110.333_{\\pm 1.70} \\\\ 19.667_{\\pm 0.47} \\end{array} \\begin{array}{c} 0.000_{\\pm 0.00} \\\\ 0.000_{\\pm 0.00} \\\\ 0.0667_{\\pm 0.94} \\\\ 3.000_{\\pm 0.82} \\\\ 1.000_{\\pm 0.82} \\\\ 2.000_{\\pm 2.83} \\\\ 11.000_{\\pm 8.04} \\\\ 5.000_{\\pm 7.07} \\\\ 0.000_{\\pm 0.00} \\\\ 28.667_{\\pm 6.60} \\end{array} \\begin{array}{c} 0.000_{\\pm 0.00} \\\\ 0.000_{\\pm 0.00} \\\\ 10.333_{\\pm 13.91} \\\\ 52.667_{\\pm 34.57} \\\\ 120.667_{\\pm 14.64} \\\\ 89.000_{\\pm 53.11} \\\\ 110.333_{\\pm 13.07} \\\\ 13.333_{\\pm 18.86} \\\\ 33.667_{\\pm 1.70} \\\\ 80.667_{\\pm 3.68} \\end{array}", "source": "marker_v2", "marker_block_id": "/page/17/Table/2"}

icml26/a4461009-05b7-42b6-b207-5e6e0c2e0731/appendix_text_v3.txt

@@ -0,0 +1,5 @@
+[p. 18 | section: A Neuropsychologically Grounded Evaluation of LLM Cognitive Abilities | type: Caption]
+Table 14. Spatial Working Memory full results (image-text modality). All metrics are reported as mean with standard deviation (std) in a subscript.
+
+[p. 18 | section: A Neuropsychologically Grounded Evaluation of LLM Cognitive Abilities | type: Table]
+Difficulty Model \boldsymbol{S}_{swm} Tokens Score Guesses Illegal Invalid Repeated No-box Easy GPT-5 Gemini 3 Pro Gemini 2.5 Pro Claude Sonnet 4 Grok 4.1 Fast Grok 4 Fast o4-Mini GLM 4.6V GLM 4.5V Qwen3-VL-235B \begin{array}{c} 0.986_{\pm .02} \\ 1.000_{\pm .00} \\ 0.733_{\pm .28} \\ 0.256_{\pm .18} \\ 0.145_{\pm .20} \\ 0.021_{\pm .03} \\ 0.055_{\pm .03} \\ 0.044_{\pm .05} \\ 0.044_{\pm .02} \\ 0.026_{\pm .05} \end{array} \begin{array}{c} 1.000_{\pm.00} \\ 1.000_{\pm.00} \\ 0.958_{\pm0.06} \\ 0.625_{\pm0.37} \\ 0.375_{\pm0.44} \\ 0.125_{\pm0.18} \\ 0.292_{\pm0.12} \\ 0.167_{\pm0.16} \\ 0.208_{\pm0.12} \\ 0.125_{\pm.10} \end{array} \begin{array}{c} 24.000_{\pm 2.16} \\ 20.667_{\pm 1.70} \\ 36.667_{\pm 19.60} \\ 59.333_{\pm 6.60} \\ 53.667_{\pm 14.61} \\ 64.000_{\pm 0.00} \\ 64.000_{\pm 0.00} \\ 64.000_{\pm 0.00} \\ 64.000_{\pm 0.00} \\ 64.000_{\pm 0.00} \end{array} \begin{array}{c} 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 1.000_{\pm 0.10} \\ 0.000_{\pm 0.10} \\ 1.000_{\pm 1.41} \end{array} \begin{array}{c} 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.667_{\pm 0.94} \\ 0.333_{\pm 0.47} \\ 0.333_{\pm 0.47} \\ 27.667_{\pm 26.04} \\ 32.333_{\pm 4.03} \\ 0.000_{\pm 0.00} \end{array} \begin{array}{c} 0.333_{\pm 0.47} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 1.667_{\pm 2.36} \\ 4.333_{\pm 5.44} \\ 0.000_{\pm 0.00} \\ 3.333_{\pm 3.40} \\ 6.333_{\pm 5.31} \\ 0.667_{\pm 0.47} \\ 5.333_{\pm 6.85} \end{array} \begin{array}{c} 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 14.000_{\pm 17.68} \\ 39.000_{\pm 12.36} \\ 40.333_{\pm 16.94} \\ 56.333_{\pm 2.49} \\ 49.000_{\pm 4.97} \\ 19.000_{\pm 13.74} \\ 24.333_{\pm 4.50} \\ 47.000_{\pm 7.48} \end{array} Easy - Notes Gemini 2.5 Pro Claude Sonnet 4 Grok 4.1 Fast Grok 4 Fast o4-Mini GLM 4.6V GLM 4.5V Qwen3-VL-235B \begin{array}{c} 0.704_{\pm.42} \\ 0.207_{\pm.18} \\ 0.043_{\pm.06} \\ 0.029_{\pm.02} \\ 0.012_{\pm.01} \\ 0.082_{\pm.09} \\ 0.000_{\pm.00} \\ 0.318_{\pm.45} \end{array} \begin{array}{c} 0.792_{\pm .29} \\ 0.542_{\pm .41} \\ 0.167_{\pm .24} \\ 0.208_{\pm .06} \\ 0.125_{\pm .10} \\ 0.333_{\pm .29} \\ 0.000_{\pm .00} \\ 0.333_{\pm .47} \end{array} \begin{array}{c} 32.667_{\pm 22.17} \\ 60.667_{\pm 4.71} \\ 64.000_{\pm 0.00} \\ 64.000_{\pm 0.00} \\ 64.000_{\pm 0.00} \\ 64.000_{\pm 0.00} \\ 64.000_{\pm 0.00} \\ 50.000_{\pm 19.80} \end{array} \begin{array}{c} 0.000_{\pm 0.00} \\ 0.333_{\pm 0.47} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \end{array} \begin{array}{c} 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 7.000_{\pm 5.35} \\ 0.333_{\pm 0.47} \\ 1.333_{\pm 0.94} \\ 3.333_{\pm 3.40} \\ 30.667_{\pm 1.25} \\ 1.667_{\pm 2.36} \end{array} \begin{array}{c} 0.333_{\pm 0.47} \\ 0.667_{\pm 0.94} \\ 7.333_{\pm 10.37} \\ 5.333_{\pm 2.87} \\ 0.333_{\pm 0.47} \\ 16.000_{\pm 22.63} \\ 0.000_{\pm 0.00} \\ 0.667_{\pm 0.47} \end{array} \begin{array}{c} 14.667_{\pm 20.74} \\ 43.333_{\pm 11.03} \\ 42.667_{\pm 16.50} \\ 50.333_{\pm 5.56} \\ 56.667_{\pm 0.47} \\ 33.333_{\pm 19.34} \\ 32.667_{\pm 1.25} \\ 38.000_{\pm 26.98} \end{array} Hard GPT-5 Gemini 3 Pro Gemini 2.5 Pro Claude Sonnet 4 Grok 4.1 Fast Grok 4 Fast o4-Mini GLM 4.6V GLM 4.5V Qwen3-VL-235B \begin{array}{c} 0.989_{\pm .02} \\ 0.963_{\pm .05} \\ 0.480_{\pm .12} \\ 0.265_{\pm .31} \\ 0.017_{\pm .02} \\ 0.002_{\pm .00} \\ 0.032_{\pm .02} \\ 0.010_{\pm .01} \\ 0.000_{\pm .00} \\ 0.016_{\pm .01} \end{array} \begin{array}{c} 1.000_{\pm.00} \\ 1.000_{\pm.00} \\ 0.972_{\pm0.04} \\ 0.472_{\pm0.38} \\ 0.097_{\pm0.14} \\ 0.028_{\pm0.02} \\ 0.181_{\pm0.10} \\ 0.042_{\pm0.06} \\ 0.000_{\pm0.00} \\ 0.111_{\pm0.10} \end{array} \begin{array}{c} 59.333_{\pm 5.79} \\ 66.000_{\pm 5.10} \\ 120.000_{\pm 18.49} \\ 122.333_{\pm 30.64} \\ 144.000_{\pm 0.00} \\ 144.000_{\pm 0.00} \\ 144.000_{\pm 0.00} \\ 144.000_{\pm 0.00} \\ 144.000_{\pm 0.00} \\ 144.000_{\pm 0.00} \\ 144.000_{\pm 0.00} \\ \end{array} \begin{array}{c} 0.667_{\pm 0.94} \\ 2.667_{\pm 3.77} \\ 52.000_{\pm 29.63} \\ 5.667_{\pm 7.32} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.333_{\pm 0.47} \end{array} \begin{array}{c} 0.000_{\pm 0.00} \\ 1.333_{\pm 0.47} \\ 0.333_{\pm 0.47} \\ 29.333_{\pm 20.74} \\ 10.000_{\pm 7.48} \\ 46.667_{\pm 58.39} \\ 2.000_{\pm 0.82} \\ 119.667_{\pm 34.41} \\ 110.333_{\pm 1.70} \\ 19.667_{\pm 0.47} \end{array} \begin{array}{c} 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.0667_{\pm 0.94} \\ 3.000_{\pm 0.82} \\ 1.000_{\pm 0.82} \\ 2.000_{\pm 2.83} \\ 11.000_{\pm 8.04} \\ 5.000_{\pm 7.07} \\ 0.000_{\pm 0.00} \\ 28.667_{\pm 6.60} \end{array} \begin{array}{c} 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 10.333_{\pm 13.91} \\ 52.667_{\pm 34.57} \\ 120.667_{\pm 14.64} \\ 89.000_{\pm 53.11} \\ 110.333_{\pm 13.07} \\ 13.333_{\pm 18.86} \\ 33.667_{\pm 1.70} \\ 80.667_{\pm 3.68} \end{array}

icml26/a4461009-05b7-42b6-b207-5e6e0c2e0731/assets.json

@@ -0,0 +1,46 @@
+[
+  {
+    "filename": "_page_1_Figure_1.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/a4461009-05b7-42b6-b207-5e6e0c2e0731/marker_markdown/a4461009-05b7-42b6-b207-5e6e0c2e0731/_page_1_Figure_1.jpeg",
+    "bytes": 106459,
+    "width": 1340,
+    "height": 703,
+    "aspect_ratio": 1.9061166429587482,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_1_Figure_1.jpeg"
+  },
+  {
+    "filename": "_page_21_Figure_34.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/a4461009-05b7-42b6-b207-5e6e0c2e0731/marker_markdown/a4461009-05b7-42b6-b207-5e6e0c2e0731/_page_21_Figure_34.jpeg",
+    "bytes": 27993,
+    "width": 686,
+    "height": 375,
+    "aspect_ratio": 1.8293333333333333,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_21_Figure_34.jpeg"
+  },
+  {
+    "filename": "_page_22_Picture_32.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/a4461009-05b7-42b6-b207-5e6e0c2e0731/marker_markdown/a4461009-05b7-42b6-b207-5e6e0c2e0731/_page_22_Picture_32.jpeg",
+    "bytes": 12086,
+    "width": 664,
+    "height": 248,
+    "aspect_ratio": 2.6774193548387095,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_22_Picture_32.jpeg"
+  },
+  {
+    "filename": "_page_7_Figure_1.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/a4461009-05b7-42b6-b207-5e6e0c2e0731/marker_markdown/a4461009-05b7-42b6-b207-5e6e0c2e0731/_page_7_Figure_1.jpeg",
+    "bytes": 63291,
+    "width": 1312,
+    "height": 458,
+    "aspect_ratio": 2.8646288209606987,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_7_Figure_1.jpeg"
+  }
+]

icml26/a4461009-05b7-42b6-b207-5e6e0c2e0731/dataset_meta.json

@@ -0,0 +1,63 @@
+{
+  "paper_id": "a4461009-05b7-42b6-b207-5e6e0c2e0731",
+  "pipeline": "Paper2Markdown-V3",
+  "ok": true,
+  "page_count": 23,
+  "chunk_count": 178,
+  "main_body_chunk_count": 176,
+  "appendix_chunk_count": 2,
+  "reference_chunk_count": 0,
+  "model_text_chars": 75805,
+  "raw_markdown_chars": 95965,
+  "sanitized_chars": 85481,
+  "page_provenance": {
+    "min_page": 1,
+    "max_page": 23,
+    "invalid_count": 0
+  },
+  "marker_block_type_counts": {
+    "Caption": 1,
+    "Code": 2,
+    "Equation": 7,
+    "Figure": 1,
+    "FigureGroup": 2,
+    "Footnote": 10,
+    "ListGroup": 17,
+    "PageFooter": 20,
+    "PageHeader": 32,
+    "Picture": 1,
+    "SectionHeader": 64,
+    "Table": 3,
+    "TableGroup": 13,
+    "Text": 421
+  },
+  "asset_count_raw": 4,
+  "asset_count_model_kept": 4,
+  "asset_count_rejected": 0,
+  "asset_reject_reasons": {
+    "kept": 4
+  },
+  "artifact_leak_audit": {
+    "ok": true,
+    "hits": {
+      "Anonymous Authors": [],
+      "ACKNOWLEDGMENT": [],
+      "OpenReview": [],
+      "\"accept_label\"": [],
+      "\"decision\"": [],
+      "\"decision_tier\"": [],
+      "\"source_status\"": [],
+      "Meta-review": [],
+      "Official Review": [],
+      "official_reviews": [],
+      "meta_reviews": [],
+      "suggested_verdict_score": []
+    },
+    "artifact_count": 2
+  },
+  "default_model_input": "model_text_v3.txt",
+  "appendix_input": "appendix_text_v3.txt",
+  "reference_input": "reference_text_v3.txt",
+  "source": "koala_icml26_due_queue",
+  "run_name": "icml26_20260429_1952_duequeue"
+}

icml26/a4461009-05b7-42b6-b207-5e6e0c2e0731/marker_meta.json

@@ -0,0 +1,2677 @@
+{
+  "table_of_contents": [
+    {
+      "title": "A Neuropsychologically Grounded Evaluation of LLM Cognitive Abilities",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          70.5234375,
+          89.9119873046875
+        ],
+        [
+          523.7428588867188,
+          89.9119873046875
+        ],
+        [
+          523.7428588867188,
+          107.89453125
+        ],
+        [
+          70.5234375,
+          107.89453125
+        ]
+      ]
+    },
+    {
+      "title": "Anonymous Authors1",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          246.234375,
+          141.74615478515625
+        ],
+        [
+          339.5129089355469,
+          141.74615478515625
+        ],
+        [
+          339.5129089355469,
+          152.96612548828125
+        ],
+        [
+          246.234375,
+          152.96612548828125
+        ]
+      ]
+    },
+    {
+      "title": "Abstract",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          150.01171875,
+          175.18359375
+        ],
+        [
+          194.68328857421875,
+          175.18359375
+        ],
+        [
+          194.68328857421875,
+          188.015380859375
+        ],
+        [
+          150.01171875,
+          188.015380859375
+        ]
+      ]
+    },
+    {
+      "title": "1. Introduction",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          54.087890625,
+          557.9141540527344
+        ],
+        [
+          132.27606201171875,
+          557.9141540527344
+        ],
+        [
+          132.27606201171875,
+          569.8693542480469
+        ],
+        [
+          54.087890625,
+          569.8693542480469
+        ]
+      ]
+    },
+    {
+      "title": "2. Related Works",
+      "heading_level": null,
+      "page_id": 1,
+      "polygon": [
+        [
+          53.19140625,
+          526.5
+        ],
+        [
+          143.25,
+          526.5
+        ],
+        [
+          143.25,
+          536.37890625
+        ],
+        [
+          53.19140625,
+          536.37890625
+        ]
+      ]
+    },
+    {
+      "title": "3. NeuroCognition Benchmark Tests",
+      "heading_level": null,
+      "page_id": 2,
+      "polygon": [
+        [
+          51.99609375,
+          67.84710693359375
+        ],
+        [
+          240.75755310058594,
+          67.84710693359375
+        ],
+        [
+          240.75755310058594,
+          79.80230712890625
+        ],
+        [
+          51.99609375,
+          79.80230712890625
+        ]
+      ]
+    },
+    {
+      "title": "3.1. Raven's Progressive Matrices",
+      "heading_level": null,
+      "page_id": 2,
+      "polygon": [
+        [
+          51.99609375,
+          88.74945068359375
+        ],
+        [
+          197.7557373046875,
+          88.74945068359375
+        ],
+        [
+          197.7557373046875,
+          99.0
+        ],
+        [
+          51.99609375,
+          99.0
+        ]
+      ]
+    },
+    {
+      "title": "3.2. Spatial Working Memory",
+      "heading_level": null,
+      "page_id": 2,
+      "polygon": [
+        [
+          54.38671875,
+          419.203125
+        ],
+        [
+          181.7159423828125,
+          419.203125
+        ],
+        [
+          181.7159423828125,
+          429.422119140625
+        ],
+        [
+          54.38671875,
+          429.422119140625
+        ]
+      ]
+    },
+    {
+      "title": "3.3. Wisconsin Card Sorting Test",
+      "heading_level": null,
+      "page_id": 2,
+      "polygon": [
+        [
+          305.40234375,
+          443.6805114746094
+        ],
+        [
+          446.9363098144531,
+          443.6805114746094
+        ],
+        [
+          446.9363098144531,
+          453.6430969238281
+        ],
+        [
+          305.40234375,
+          453.6430969238281
+        ]
+      ]
+    },
+    {
+      "title": "3.4. Experiment Setup",
+      "heading_level": null,
+      "page_id": 3,
+      "polygon": [
+        [
+          306.0,
+          177.1171875
+        ],
+        [
+          402.0,
+          177.1171875
+        ],
+        [
+          402.0,
+          186.75
+        ],
+        [
+          306.0,
+          186.75
+        ]
+      ]
+    },
+    {
+      "title": "4. Results",
+      "heading_level": null,
+      "page_id": 3,
+      "polygon": [
+        [
+          305.701171875,
+          433.5
+        ],
+        [
+          357.0,
+          433.5
+        ],
+        [
+          357.0,
+          443.56640625
+        ],
+        [
+          305.701171875,
+          443.56640625
+        ]
+      ]
+    },
+    {
+      "title": "4.1. Key Findings",
+      "heading_level": null,
+      "page_id": 3,
+      "polygon": [
+        [
+          306.0,
+          490.5
+        ],
+        [
+          381.75,
+          490.5
+        ],
+        [
+          381.75,
+          500.25
+        ],
+        [
+          306.0,
+          500.25
+        ]
+      ]
+    },
+    {
+      "title": "4.2. Failure Cases & Patterns",
+      "heading_level": null,
+      "page_id": 4,
+      "polygon": [
+        [
+          51.99609375,
+          489.2804870605469
+        ],
+        [
+          179.27511596679688,
+          489.2804870605469
+        ],
+        [
+          179.27511596679688,
+          499.25390625
+        ],
+        [
+          51.99609375,
+          499.25390625
+        ]
+      ]
+    },
+    {
+      "title": "5. Ablation Study",
+      "heading_level": null,
+      "page_id": 5,
+      "polygon": [
+        [
+          53.19140625,
+          439.4881591796875
+        ],
+        [
+          144.79318237304688,
+          439.4881591796875
+        ],
+        [
+          144.79318237304688,
+          451.443359375
+        ],
+        [
+          53.19140625,
+          451.443359375
+        ]
+      ]
+    },
+    {
+      "title": "5.1. RAPM: Reasoning vs No Reasoning",
+      "heading_level": null,
+      "page_id": 5,
+      "polygon": [
+        [
+          52.59375,
+          521.3895263671875
+        ],
+        [
+          225.63107299804688,
+          521.3895263671875
+        ],
+        [
+          225.63107299804688,
+          531.3521118164062
+        ],
+        [
+          52.59375,
+          531.3521118164062
+        ]
+      ]
+    },
+    {
+      "title": "5.2. RAPM: Pattern Hints",
+      "heading_level": null,
+      "page_id": 5,
+      "polygon": [
+        [
+          307.44000244140625,
+          303.9609375
+        ],
+        [
+          419.255859375,
+          303.9609375
+        ],
+        [
+          419.255859375,
+          314.21710205078125
+        ],
+        [
+          307.44000244140625,
+          314.21710205078125
+        ]
+      ]
+    },
+    {
+      "title": "5.3. SWM & WCST: Note-Taking Assistance",
+      "heading_level": null,
+      "page_id": 5,
+      "polygon": [
+        [
+          306.59765625,
+          479.5465087890625
+        ],
+        [
+          497.00836181640625,
+          479.5465087890625
+        ],
+        [
+          497.00836181640625,
+          489.97265625
+        ],
+        [
+          306.59765625,
+          489.97265625
+        ]
+      ]
+    },
+    {
+      "title": "6. Analysis & Discussion",
+      "heading_level": null,
+      "page_id": 6,
+      "polygon": [
+        [
+          52.59375,
+          403.734375
+        ],
+        [
+          180.32403564453125,
+          403.734375
+        ],
+        [
+          180.32403564453125,
+          415.7253723144531
+        ],
+        [
+          52.59375,
+          415.7253723144531
+        ]
+      ]
+    },
+    {
+      "title": "6.1. Factor Analysis of LLM",
+      "heading_level": null,
+      "page_id": 6,
+      "polygon": [
+        [
+          52.59375,
+          424.6725158691406
+        ],
+        [
+          175.29006958007812,
+          424.6725158691406
+        ],
+        [
+          175.29006958007812,
+          434.6351013183594
+        ],
+        [
+          52.59375,
+          434.6351013183594
+        ]
+      ]
+    },
+    {
+      "title": "6.2. Correlation with Other Benchmarks",
+      "heading_level": null,
+      "page_id": 6,
+      "polygon": [
+        [
+          306.0,
+          294.24151611328125
+        ],
+        [
+          479.9324645996094,
+          294.24151611328125
+        ],
+        [
+          479.9324645996094,
+          304.2041015625
+        ],
+        [
+          306.0,
+          304.2041015625
+        ]
+      ]
+    },
+    {
+      "title": "7. Conclusion",
+      "heading_level": null,
+      "page_id": 7,
+      "polygon": [
+        [
+          54.0,
+          348.0
+        ],
+        [
+          125.25,
+          348.0
+        ],
+        [
+          125.25,
+          357.0
+        ],
+        [
+          54.0,
+          357.0
+        ]
+      ]
+    },
+    {
+      "title": "8. Future Work",
+      "heading_level": null,
+      "page_id": 7,
+      "polygon": [
+        [
+          53.7890625,
+          645.43359375
+        ],
+        [
+          135.0,
+          645.43359375
+        ],
+        [
+          135.0,
+          656.25
+        ],
+        [
+          53.7890625,
+          656.25
+        ]
+      ]
+    },
+    {
+      "title": "Limitations",
+      "heading_level": null,
+      "page_id": 7,
+      "polygon": [
+        [
+          306.0,
+          455.25
+        ],
+        [
+          367.5,
+          455.25
+        ],
+        [
+          367.5,
+          465.75
+        ],
+        [
+          306.0,
+          465.75
+        ]
+      ]
+    },
+    {
+      "title": "Impact Statement",
+      "heading_level": null,
+      "page_id": 8,
+      "polygon": [
+        [
+          53.19140625,
+          155.84765625
+        ],
+        [
+          146.74185180664062,
+          155.84765625
+        ],
+        [
+          146.74185180664062,
+          168.14337158203125
+        ],
+        [
+          53.19140625,
+          168.14337158203125
+        ]
+      ]
+    },
+    {
+      "title": "A. RAPM Text Data Generation",
+      "heading_level": null,
+      "page_id": 9,
+      "polygon": [
+        [
+          53.19140625,
+          67.67578125
+        ],
+        [
+          219.3577423095703,
+          67.67578125
+        ],
+        [
+          219.3577423095703,
+          79.80230712890625
+        ],
+        [
+          53.19140625,
+          79.80230712890625
+        ]
+      ]
+    },
+    {
+      "title": "A.1. Rule Space and Attributes",
+      "heading_level": null,
+      "page_id": 9,
+      "polygon": [
+        [
+          52.294921875,
+          88.74945068359375
+        ],
+        [
+          186.96624755859375,
+          88.74945068359375
+        ],
+        [
+          186.96624755859375,
+          99.38671875
+        ],
+        [
+          52.294921875,
+          99.38671875
+        ]
+      ]
+    },
+    {
+      "title": "A.2. Generation Algorithm",
+      "heading_level": null,
+      "page_id": 9,
+      "polygon": [
+        [
+          53.7890625,
+          461.301513671875
+        ],
+        [
+          169.9999237060547,
+          461.301513671875
+        ],
+        [
+          169.9999237060547,
+          471.41015625
+        ],
+        [
+          53.7890625,
+          471.41015625
+        ]
+      ]
+    },
+    {
+      "title": "Algorithm 1 RAPM Text Matrix Generation",
+      "heading_level": null,
+      "page_id": 10,
+      "polygon": [
+        [
+          51.99609375,
+          68.8359375
+        ],
+        [
+          233.8209686279297,
+          68.8359375
+        ],
+        [
+          233.8209686279297,
+          79.12103271484375
+        ],
+        [
+          51.99609375,
+          79.12103271484375
+        ]
+      ]
+    },
+    {
+      "title": "A.3. Multiple-Choice and Distractor Generation",
+      "heading_level": null,
+      "page_id": 10,
+      "polygon": [
+        [
+          53.7890625,
+          407.21484375
+        ],
+        [
+          259.65338134765625,
+          407.21484375
+        ],
+        [
+          259.65338134765625,
+          417.37408447265625
+        ],
+        [
+          53.7890625,
+          417.37408447265625
+        ]
+      ]
+    },
+    {
+      "title": "B. RAPM Text Gen Automatic Rule Validation",
+      "heading_level": null,
+      "page_id": 11,
+      "polygon": [
+        [
+          53.7890625,
+          67.67578125
+        ],
+        [
+          293.39630126953125,
+          67.67578125
+        ],
+        [
+          293.39630126953125,
+          79.80230712890625
+        ],
+        [
+          53.7890625,
+          79.80230712890625
+        ]
+      ]
+    },
+    {
+      "title": "Algorithm 2 RAPM Text Gen Validation (CellSatisfies)",
+      "heading_level": null,
+      "page_id": 11,
+      "polygon": [
+        [
+          52.59375,
+          129.031494140625
+        ],
+        [
+          278.9416198730469,
+          129.031494140625
+        ],
+        [
+          278.9416198730469,
+          139.08502197265625
+        ],
+        [
+          52.59375,
+          139.08502197265625
+        ]
+      ]
+    },
+    {
+      "title": "C. Models & Experimental Setup",
+      "heading_level": null,
+      "page_id": 12,
+      "polygon": [
+        [
+          51.99609375,
+          67.84710693359375
+        ],
+        [
+          225.1321258544922,
+          67.84710693359375
+        ],
+        [
+          225.1321258544922,
+          79.80230712890625
+        ],
+        [
+          51.99609375,
+          79.80230712890625
+        ]
+      ]
+    },
+    {
+      "title": "C.1. Model Access Names and Notes",
+      "heading_level": null,
+      "page_id": 12,
+      "polygon": [
+        [
+          53.7890625,
+          87.01171875
+        ],
+        [
+          209.1796875,
+          87.01171875
+        ],
+        [
+          209.1796875,
+          98.7120361328125
+        ],
+        [
+          53.7890625,
+          98.7120361328125
+        ]
+      ]
+    },
+    {
+      "title": "Google AI Studio",
+      "heading_level": null,
+      "page_id": 12,
+      "polygon": [
+        [
+          53.490234375,
+          160.48828125
+        ],
+        [
+          129.06361389160156,
+          160.48828125
+        ],
+        [
+          129.06361389160156,
+          172.26416015625
+        ],
+        [
+          53.490234375,
+          172.26416015625
+        ]
+      ]
+    },
+    {
+      "title": "OpenAI API",
+      "heading_level": null,
+      "page_id": 12,
+      "polygon": [
+        [
+          53.19140625,
+          225.17156982421875
+        ],
+        [
+          109.40740203857422,
+          225.17156982421875
+        ],
+        [
+          109.40740203857422,
+          235.1341552734375
+        ],
+        [
+          53.19140625,
+          235.1341552734375
+        ]
+      ]
+    },
+    {
+      "title": "OpenRouter",
+      "heading_level": null,
+      "page_id": 12,
+      "polygon": [
+        [
+          52.892578125,
+          287.71875
+        ],
+        [
+          108.57054901123047,
+          287.71875
+        ],
+        [
+          108.57054901123047,
+          298.004150390625
+        ],
+        [
+          52.892578125,
+          298.004150390625
+        ]
+      ]
+    },
+    {
+      "title": "Local / vLLM",
+      "heading_level": null,
+      "page_id": 12,
+      "polygon": [
+        [
+          54.38671875,
+          388.65234375
+        ],
+        [
+          115.34765625,
+          388.65234375
+        ],
+        [
+          115.34765625,
+          399.7291564941406
+        ],
+        [
+          54.38671875,
+          399.7291564941406
+        ]
+      ]
+    },
+    {
+      "title": "C.2. vLLM Inference Settings",
+      "heading_level": null,
+      "page_id": 12,
+      "polygon": [
+        [
+          53.7890625,
+          433.82855224609375
+        ],
+        [
+          181.71597290039062,
+          433.82855224609375
+        ],
+        [
+          181.71597290039062,
+          443.7911376953125
+        ],
+        [
+          53.7890625,
+          443.7911376953125
+        ]
+      ]
+    },
+    {
+      "title": "C.3. Evaluation Protocol Details",
+      "heading_level": null,
+      "page_id": 12,
+      "polygon": [
+        [
+          54.087890625,
+          500.02734375
+        ],
+        [
+          191.58888244628906,
+          500.02734375
+        ],
+        [
+          191.58888244628906,
+          511.48712158203125
+        ],
+        [
+          54.087890625,
+          511.48712158203125
+        ]
+      ]
+    },
+    {
+      "title": "C.4. Reasoning Configuration",
+      "heading_level": null,
+      "page_id": 13,
+      "polygon": [
+        [
+          54.38671875,
+          68.44921875
+        ],
+        [
+          181.64620971679688,
+          68.44921875
+        ],
+        [
+          181.64620971679688,
+          79.33905029296875
+        ],
+        [
+          54.38671875,
+          79.33905029296875
+        ]
+      ]
+    },
+    {
+      "title": "Standardized CoT Prompt Addition",
+      "heading_level": null,
+      "page_id": 13,
+      "polygon": [
+        [
+          58.869140625,
+          326.00390625
+        ],
+        [
+          202.655029296875,
+          326.00390625
+        ],
+        [
+          202.655029296875,
+          336.445068359375
+        ],
+        [
+          58.869140625,
+          336.445068359375
+        ]
+      ]
+    },
+    {
+      "title": "SWM & WCST:",
+      "heading_level": null,
+      "page_id": 13,
+      "polygon": [
+        [
+          57.673828125,
+          342.6328125
+        ],
+        [
+          123.78455352783203,
+          342.6328125
+        ],
+        [
+          123.78455352783203,
+          352.5404968261719
+        ],
+        [
+          57.673828125,
+          352.5404968261719
+        ]
+      ]
+    },
+    {
+      "title": "RAPM:",
+      "heading_level": null,
+      "page_id": 13,
+      "polygon": [
+        [
+          60.36328125,
+          372.41015625
+        ],
+        [
+          90.56404113769531,
+          372.41015625
+        ],
+        [
+          90.56404113769531,
+          382.4284973144531
+        ],
+        [
+          60.36328125,
+          382.4284973144531
+        ]
+      ]
+    },
+    {
+      "title": "C.5. Disabling Reasoning",
+      "heading_level": null,
+      "page_id": 13,
+      "polygon": [
+        [
+          55.439998626708984,
+          444.33984375
+        ],
+        [
+          162.27891540527344,
+          444.33984375
+        ],
+        [
+          162.27891540527344,
+          456.1661071777344
+        ],
+        [
+          55.439998626708984,
+          456.1661071777344
+        ]
+      ]
+    },
+    {
+      "title": "Standardized Non-CoT Instruction\nAnswer only with your final answer.",
+      "heading_level": null,
+      "page_id": 13,
+      "polygon": [
+        [
+          59.765625,
+          577.37109375
+        ],
+        [
+          199.6171875,
+          577.37109375
+        ],
+        [
+          199.6171875,
+          603.8653869628906
+        ],
+        [
+          59.765625,
+          603.8653869628906
+        ]
+      ]
+    },
+    {
+      "title": "Gemini 2.5 Pro Specific Prompt. For Gemini 2.5 Pro, the model does not natively support disabling reasoning via API\nparameters. However, we found a way to enforce the no-reasoning condition by appending a specific critical instruction to\noverride the model's default behavior.",
+      "heading_level": null,
+      "page_id": 13,
+      "polygon": [
+        [
+          54.984375,
+          632.1055603027344
+        ],
+        [
+          541.4478149414062,
+          632.1055603027344
+        ],
+        [
+          541.4478149414062,
+          666.703125
+        ],
+        [
+          54.984375,
+          666.703125
+        ]
+      ]
+    },
+    {
+      "title": "Gemini 2.5 Pro Non-CoT Instruction",
+      "heading_level": null,
+      "page_id": 13,
+      "polygon": [
+        [
+          60.064453125,
+          681.3594665527344
+        ],
+        [
+          208.18429565429688,
+          681.3594665527344
+        ],
+        [
+          208.18429565429688,
+          691.3220672607422
+        ],
+        [
+          60.064453125,
+          691.3220672607422
+        ]
+      ]
+    },
+    {
+      "title": "D. External Benchmark Setup & Results",
+      "heading_level": null,
+      "page_id": 14,
+      "polygon": [
+        [
+          52.59375,
+          67.84710693359375
+        ],
+        [
+          262.2410583496094,
+          67.84710693359375
+        ],
+        [
+          262.2410583496094,
+          79.80230712890625
+        ],
+        [
+          52.59375,
+          79.80230712890625
+        ]
+      ]
+    },
+    {
+      "title": "E. Full Results",
+      "heading_level": null,
+      "page_id": 15,
+      "polygon": [
+        [
+          53.490234375,
+          67.84710693359375
+        ],
+        [
+          130.1719512939453,
+          67.84710693359375
+        ],
+        [
+          130.1719512939453,
+          79.80230712890625
+        ],
+        [
+          53.490234375,
+          79.80230712890625
+        ]
+      ]
+    },
+    {
+      "title": "E.1. Raven's Advanced Progressive Matrices (RAPM)",
+      "heading_level": null,
+      "page_id": 15,
+      "polygon": [
+        [
+          53.19140625,
+          88.74945068359375
+        ],
+        [
+          283.4341125488281,
+          88.74945068359375
+        ],
+        [
+          283.4341125488281,
+          98.7120361328125
+        ],
+        [
+          53.19140625,
+          98.7120361328125
+        ]
+      ]
+    },
+    {
+      "title": "E.2. Spatial Working Memory (SWM)",
+      "heading_level": null,
+      "page_id": 16,
+      "polygon": [
+        [
+          53.19140625,
+          69.37646484375
+        ],
+        [
+          217.41195678710938,
+          69.37646484375
+        ],
+        [
+          217.41195678710938,
+          79.33905029296875
+        ],
+        [
+          53.19140625,
+          79.33905029296875
+        ]
+      ]
+    },
+    {
+      "title": "A Neuropsychologically Grounded Evaluation of LLM Cognitive Abilities",
+      "heading_level": null,
+      "page_id": 17,
+      "polygon": [
+        [
+          155.25,
+          45.0
+        ],
+        [
+          441.0,
+          45.0
+        ],
+        [
+          441.0,
+          56.25
+        ],
+        [
+          155.25,
+          56.25
+        ]
+      ]
+    },
+    {
+      "title": "E.3. Wisconsin Card Sorting Test (WCST)",
+      "heading_level": null,
+      "page_id": 18,
+      "polygon": [
+        [
+          52.59375,
+          69.37646484375
+        ],
+        [
+          235.06565856933594,
+          69.37646484375
+        ],
+        [
+          235.06565856933594,
+          79.33905029296875
+        ],
+        [
+          52.59375,
+          79.33905029296875
+        ]
+      ]
+    },
+    {
+      "title": "F. Prompts",
+      "heading_level": null,
+      "page_id": 19,
+      "polygon": [
+        [
+          53.19140625,
+          67.84710693359375
+        ],
+        [
+          111.01972961425781,
+          67.84710693359375
+        ],
+        [
+          111.01972961425781,
+          79.80230712890625
+        ],
+        [
+          53.19140625,
+          79.80230712890625
+        ]
+      ]
+    },
+    {
+      "title": "System Prompt for Image RAPM",
+      "heading_level": null,
+      "page_id": 19,
+      "polygon": [
+        [
+          57.673828125,
+          128.00390625
+        ],
+        [
+          194.07723999023438,
+          128.00390625
+        ],
+        [
+          194.07723999023438,
+          139.31805419921875
+        ],
+        [
+          57.673828125,
+          139.31805419921875
+        ]
+      ]
+    },
+    {
+      "title": "/* PATTERN HINT */",
+      "heading_level": null,
+      "page_id": 19,
+      "polygon": [
+        [
+          56.77734375,
+          245.56640625
+        ],
+        [
+          141.0539093017578,
+          245.56640625
+        ],
+        [
+          141.0539093017578,
+          255.05841064453125
+        ],
+        [
+          56.77734375,
+          255.05841064453125
+        ]
+      ]
+    },
+    {
+      "title": "System Prompt for Text RAPM",
+      "heading_level": null,
+      "page_id": 19,
+      "polygon": [
+        [
+          57.97265625,
+          434.886474609375
+        ],
+        [
+          186.5953369140625,
+          434.886474609375
+        ],
+        [
+          186.5953369140625,
+          444.8490905761719
+        ],
+        [
+          57.97265625,
+          444.8490905761719
+        ]
+      ]
+    },
+    {
+      "title": "/* PATTERN HINT */",
+      "heading_level": null,
+      "page_id": 19,
+      "polygon": [
+        [
+          56.1796875,
+          490.74609375
+        ],
+        [
+          142.2421875,
+          490.74609375
+        ],
+        [
+          142.2421875,
+          500.8133544921875
+        ],
+        [
+          56.1796875,
+          500.8133544921875
+        ]
+      ]
+    },
+    {
+      "title": "1089\n1090\n/* MC VERSION */",
+      "heading_level": null,
+      "page_id": 19,
+      "polygon": [
+        [
+          25.552000045776367,
+          595.1964416503906
+        ],
+        [
+          133.91664123535156,
+          595.1964416503906
+        ],
+        [
+          133.91664123535156,
+          617.1140441894531
+        ],
+        [
+          25.552000045776367,
+          617.1140441894531
+        ]
+      ]
+    },
+    {
+      "title": "1093\n1094\n/* GEN VERSION */",
+      "heading_level": null,
+      "page_id": 19,
+      "polygon": [
+        [
+          25.552000045776367,
+          643.0174407958984
+        ],
+        [
+          138.39088439941406,
+          643.0174407958984
+        ],
+        [
+          138.39088439941406,
+          664.9350433349609
+        ],
+        [
+          25.552000045776367,
+          664.9350433349609
+        ]
+      ]
+    },
+    {
+      "title": "1100\n1101\n1102\n1103\n1104\n1105\n1106\n1107\n1108\n1109\n1110\n1111\n1112\n1113\n1114\n1115\n1116\n1117\n1118\n1119\n1120\n1121\n1122\n1123\n1124\n1125\n1126\n1127\n1128\n1129\n1130\n1131\n1132\n1133\n1134\n1135\n1136\n1137\n1138\n1139\n1140\nA Neuropsychologically Grounded Evaluation of LLM Cognitive Abilities\nSystem Prompt for SWM\nYou will be performing the Spatial Working Memory task.\nYou will be given an image containing {N BOXES} yellow boxes in a grid.\nThere are {N TOKENS} types of tokens, hidden in any one of {N BOXES} boxes.\nEach token type is represented by a distinct color.\nYour goal is to find the {N TOKENS} types of tokens {N BOXES} times each, by repeatedly selecting a box to open.\nA box can contain multiple types of tokens, but only one token of each type.\nIf the box contains multiple tokens, a token with mixed colors corresponding to the tokens will be shown.\nOnce the token is found, another will be generated in another box.\nThe token will be generated in a box that has never contained a token of that type before in the trial.\nThe token may be generated in a box that has been opened and found empty before, as long as it never contained that type of token\npreviously.\n/* TEXT VERSION */\nYour final answer should be a number from 1-{N BOXES}, the index of the box you selected\n/* IMAGE VERSION */\nYour final answer should be a grid coordinate (x, y), the grid coordinate of the box you choose\nSystem Prompt for WCST\nYou are performing the Wisconsin Card Sorting Test (WCST).\nYou will be shown a given card with a symbol on it, and you will have to match it to one of four option cards according to an attribute\nthat you have to figure out.\nThe cards will be described by the following attributes:\n1. Number of symbols\n2. Color of symbols\n3. Shape of symbols\n/* HARD VERSION */\n4. Background color of the card\nYou will be told \"Correct!\" if you are correct and \"Incorrect. Please try again.\" if you are incorrect.\nIf you are incorrect, you either made a mistake or the rule has changed.\nIf you believe you have made a mistake, correct it and try again.\nIf you believe the rule has changed, you have to figure out the correct rule to match the cards.\nIf you are correct, you have to stick with the same attribute until you are incorrect.\nThere is always a true answer in the task, and you have to keep performing the task until the end of the test.\nYour final answer should be a number between 1-4 corresponding to the index of the card you think is the correct match.",
+      "heading_level": null,
+      "page_id": 20,
+      "polygon": [
+        [
+          23.90625,
+          40.21875
+        ],
+        [
+          536.22802734375,
+          40.21875
+        ],
+        [
+          536.22802734375,
+          557.3380432128906
+        ],
+        [
+          23.90625,
+          557.3380432128906
+        ]
+      ]
+    },
+    {
+      "title": "G. Turn Examples",
+      "heading_level": null,
+      "page_id": 21,
+      "polygon": [
+        [
+          52.59375,
+          67.84710693359375
+        ],
+        [
+          148.8220672607422,
+          67.84710693359375
+        ],
+        [
+          148.8220672607422,
+          79.80230712890625
+        ],
+        [
+          52.59375,
+          79.80230712890625
+        ]
+      ]
+    },
+    {
+      "title": "",
+      "heading_level": null,
+      "page_id": 21,
+      "polygon": [
+        [
+          71.71875,
+          435.05859375
+        ],
+        [
+          175.7109375,
+          435.05859375
+        ],
+        [
+          175.7109375,
+          443.56640625
+        ],
+        [
+          71.71875,
+          443.56640625
+        ]
+      ]
+    },
+    {
+      "title": "",
+      "heading_level": null,
+      "page_id": 22,
+      "polygon": [
+        [
+          70.224609375,
+          462.90234375
+        ],
+        [
+          140.150390625,
+          462.90234375
+        ],
+        [
+          140.150390625,
+          471.41015625
+        ],
+        [
+          70.224609375,
+          471.41015625
+        ]
+      ]
+    }
+  ],
+  "page_stats": [
+    {
+      "page_id": 0,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          294
+        ],
+        [
+          "Line",
+          141
+        ],
+        [
+          "Text",
+          11
+        ],
+        [
+          "PageHeader",
+          9
+        ],
+        [
+          "SectionHeader",
+          4
+        ],
+        [
+          "ListItem",
+          3
+        ],
+        [
+          "Footnote",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "ListGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 1,
+      "text_extraction_method": "surya",
+      "block_counts": [
+        [
+          "Line",
+          159
+        ],
+        [
+          "Span",
+          56
+        ],
+        [
+          "Text",
+          37
+        ],
+        [
+          "Reference",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Figure",
+          1
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "SectionHeader",
+          1
+        ],
+        [
+          "Footnote",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "FigureGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 2,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          433
+        ],
+        [
+          "Line",
+          155
+        ],
+        [
+          "Text",
+          15
+        ],
+        [
+          "SectionHeader",
+          4
+        ],
+        [
+          "ListItem",
+          3
+        ],
+        [
+          "Equation",
+          2
+        ],
+        [
+          "Reference",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "ListGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 3,
+      "text_extraction_method": "surya",
+      "block_counts": [
+        [
+          "Line",
+          218
+        ],
+        [
+          "Span",
+          136
+        ],
+        [
+          "Text",
+          67
+        ],
+        [
+          "Equation",
+          5
+        ],
+        [
+          "SectionHeader",
+          3
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 4,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          497
+        ],
+        [
+          "TableCell",
+          241
+        ],
+        [
+          "Line",
+          123
+        ],
+        [
+          "Text",
+          9
+        ],
+        [
+          "PageHeader",
+          2
+        ],
+        [
+          "Caption",
+          2
+        ],
+        [
+          "Table",
+          2
+        ],
+        [
+          "TableGroup",
+          2
+        ],
+        [
+          "Reference",
+          2
+        ],
+        [
+          "SectionHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 5,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          764
+        ],
+        [
+          "Line",
+          133
+        ],
+        [
+          "TableCell",
+          109
+        ],
+        [
+          "Text",
+          14
+        ],
+        [
+          "SectionHeader",
+          4
+        ],
+        [
+          "Reference",
+          4
+        ],
+        [
+          "Caption",
+          3
+        ],
+        [
+          "Table",
+          3
+        ],
+        [
+          "PageHeader",
+          2
+        ],
+        [
+          "TableGroup",
+          2
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 6,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          676
+        ],
+        [
+          "Line",
+          140
+        ],
+        [
+          "TableCell",
+          74
+        ],
+        [
+          "Text",
+          12
+        ],
+        [
+          "Reference",
+          4
+        ],
+        [
+          "Caption",
+          3
+        ],
+        [
+          "Table",
+          3
+        ],
+        [
+          "SectionHeader",
+          3
+        ],
+        [
+          "ListItem",
+          3
+        ],
+        [
+          "PageHeader",
+          2
+        ],
+        [
+          "TableGroup",
+          2
+        ],
+        [
+          "Footnote",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "ListGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 7,
+      "text_extraction_method": "surya",
+      "block_counts": [
+        [
+          "Line",
+          223
+        ],
+        [
+          "Span",
+          111
+        ],
+        [
+          "Text",
+          51
+        ],
+        [
+          "SectionHeader",
+          3
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Figure",
+          1
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "FigureGroup",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 8,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          161
+        ],
+        [
+          "Line",
+          80
+        ],
+        [
+          "Text",
+          3
+        ],
+        [
+          "Footnote",
+          3
+        ],
+        [
+          "Reference",
+          3
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "SectionHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 9,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          332
+        ],
+        [
+          "Line",
+          86
+        ],
+        [
+          "ListItem",
+          25
+        ],
+        [
+          "Text",
+          15
+        ],
+        [
+          "SectionHeader",
+          3
+        ],
+        [
+          "ListGroup",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 10,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          354
+        ],
+        [
+          "Line",
+          78
+        ],
+        [
+          "Text",
+          11
+        ],
+        [
+          "ListItem",
+          6
+        ],
+        [
+          "SectionHeader",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Code",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "ListGroup",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 11,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          533
+        ],
+        [
+          "Line",
+          106
+        ],
+        [
+          "ListItem",
+          6
+        ],
+        [
+          "Text",
+          4
+        ],
+        [
+          "SectionHeader",
+          2
+        ],
+        [
+          "Reference",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Code",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "ListGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 12,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          301
+        ],
+        [
+          "Line",
+          94
+        ],
+        [
+          "ListItem",
+          16
+        ],
+        [
+          "SectionHeader",
+          8
+        ],
+        [
+          "Text",
+          8
+        ],
+        [
+          "ListGroup",
+          4
+        ],
+        [
+          "Reference",
+          4
+        ],
+        [
+          "Footnote",
+          3
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 13,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          331
+        ],
+        [
+          "Line",
+          81
+        ],
+        [
+          "Text",
+          30
+        ],
+        [
+          "SectionHeader",
+          8
+        ],
+        [
+          "ListItem",
+          4
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "ListGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 14,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          175
+        ],
+        [
+          "Line",
+          77
+        ],
+        [
+          "TableCell",
+          44
+        ],
+        [
+          "Text",
+          10
+        ],
+        [
+          "Reference",
+          3
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "SectionHeader",
+          1
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "Footnote",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "TableGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 15,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          337
+        ],
+        [
+          "Line",
+          150
+        ],
+        [
+          "TableCell",
+          123
+        ],
+        [
+          "Text",
+          3
+        ],
+        [
+          "Reference",
+          3
+        ],
+        [
+          "SectionHeader",
+          2
+        ],
+        [
+          "Caption",
+          2
+        ],
+        [
+          "Table",
+          2
+        ],
+        [
+          "TableGroup",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 16,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          2895
+        ],
+        [
+          "TableCell",
+          449
+        ],
+        [
+          "Line",
+          119
+        ],
+        [
+          "Text",
+          3
+        ],
+        [
+          "Caption",
+          2
+        ],
+        [
+          "Table",
+          2
+        ],
+        [
+          "TableGroup",
+          2
+        ],
+        [
+          "Reference",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "SectionHeader",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 17,
+      "text_extraction_method": "surya",
+      "block_counts": [
+        [
+          "Line",
+          62
+        ],
+        [
+          "Span",
+          58
+        ],
+        [
+          "TableCell",
+          36
+        ],
+        [
+          "Text",
+          2
+        ],
+        [
+          "SectionHeader",
+          1
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 18,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          1975
+        ],
+        [
+          "TableCell",
+          300
+        ],
+        [
+          "Line",
+          97
+        ],
+        [
+          "Text",
+          13
+        ],
+        [
+          "Caption",
+          2
+        ],
+        [
+          "Table",
+          2
+        ],
+        [
+          "TableGroup",
+          2
+        ],
+        [
+          "Reference",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "SectionHeader",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 19,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          202
+        ],
+        [
+          "Line",
+          92
+        ],
+        [
+          "Text",
+          25
+        ],
+        [
+          "ListItem",
+          22
+        ],
+        [
+          "SectionHeader",
+          7
+        ],
+        [
+          "ListGroup",
+          3
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 20,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          225
+        ],
+        [
+          "Line",
+          90
+        ],
+        [
+          "SectionHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "Text",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 21,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          244
+        ],
+        [
+          "Line",
+          83
+        ],
+        [
+          "Text",
+          42
+        ],
+        [
+          "SectionHeader",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Figure",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 22,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          197
+        ],
+        [
+          "Line",
+          88
+        ],
+        [
+          "Text",
+          35
+        ],
+        [
+          "ListItem",
+          10
+        ],
+        [
+          "ListGroup",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Picture",
+          1
+        ],
+        [
+          "SectionHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    }
+  ],
+  "debug_data_path": "debug_data/a4461009-05b7-42b6-b207-5e6e0c2e0731"
+}

icml26/a4461009-05b7-42b6-b207-5e6e0c2e0731/model_text_v3.txt

@@ -0,0 +1,527 @@
+[p. 1 | section: Abstract | type: Text]
+Large language models (LLMs) display a unified "general factor" of capability across 10 benchmarks—a finding confirmed by our factor analysis of 156 models—yet they still struggle with simple, trivial tasks for humans. This is because current benchmarks focus on task completion, failing to probe the foundational cognitive abilities that highlight these behaviors. We address this by introducing the NeuroCognition benchmark, grounded in three adapted neuropsychological tests: Raven's Progressive Matrices (abstract relational reasoning), Spatial Working Memory (maintenance and systematic search), and the Wisconsin Card Sorting Test (cognitive flexibility). Our evaluation reveals that while models perform strongly on text, their performance degrades for images and with increased complexity. Furthermore, we observe that complex reasoning is not universally beneficial, whereas simple, humanlike strategies yield partial gains. We also find that NeuroCognition correlates positively with standard general-capability benchmarks, while still measuring distinct cognitive abilities beyond them. Overall, NeuroCognition emphasizes where current LLMs align with human-like intelligence and where they lack core adaptive cognition, showing the potential to serve as a verifiable, scalable source for improving LLMs.
+
+[p. 1 | section: 1. Introduction | type: Text]
+Large language models (LLMs) show increasingly powerful capabilities (?), further enhanced by the emergence of reasoning abilities (?). This is demonstrated by their beyond human-level capabilities in tasks such as olympiad-level problems (?). However, even with these impressive capabilities, they still stumble on trivial or straightforward tasks
+
+[p. 1 | section: 1. Introduction | type: Text]
+(??). This raises a curiosity about the underlying cognitive mechanism of LLMs and how it compares to human cognitive abilities.
+
+[p. 1 | section: 1. Introduction | type: Text]
+Most current LLM benchmarks are task-specific, limiting their analysis and insights to the task performance itself. These benchmarks are not able to uncover the underlying cognitive abilities required to solve the task (??) because task performance is not a reliable measure of intelligence, but rather heavily influenced by prior knowledge and experience (?). One interesting test that is often overlooked is neuropsychological tests, which are a well-established method for measuring specific cognitive abilities in humans, originally designed to isolate specific cognitive functions using controlled stimuli and interpretable outcomes. Previous works have attempted to apply neuropsychological tests to LLMs (??), but they are limited to a single task or modality, or they lack a deeper analysis of the model's underlying cognitive abilities.
+
+[p. 1 | section: 1. Introduction | type: Text]
+We present the NeuroCognition benchmark, built from three well-established neuropsychological tests: Raven's Progressive Matrices (????) (RPM) targeting abstract relational reasoning (fluid intelligence), Spatial Working Memory (?) (SWM) targeting maintenance and systematic search across working memory, and the Wisconsin Card Sorting Test (??) (WCST) targeting cognitive flexibility and task switching. Each test is adapted into multiple text and image variants (including harder versions) to vary interaction length, modality, and complexity. We also include experiments using assistance inspired by simple strategies that humans use (e.g., taking notes), a process known as cognitive offloading (??), to see if they reduce observed errors and improve overall performance.
+
+[p. 1 | section: 1. Introduction | type: Text]
+Our main contributions are:
+
+[p. 1 | section: 1. Introduction | type: ListGroup]
+NeuroCognition, a benchmark repurposed from wellestablished neuropsychological tests, into a simple, multimodal, scalable, robust assessment, and easily reproducible suite with various setups. Evaluation results show that state-of-the-art models perform strongly in baseline text settings, but struggle in image or when task complexity increases, and that explicit reasoning is not always beneficial. Experiments with human-like strategies (pattern
+
+[p. 2 | section: 1. Introduction | type: FigureGroup]
+Figure 1. The three neuropsychological tests used in our work.
+
+[p. 2 | section: 1. Introduction | type: Text]
+hints and notes) show mixed effects—limited on RAPM/SWM but more consistent on WCST—suggesting some deficits are fundamental.
+
+[p. 2 | section: 1. Introduction | type: Text]
+Factor analysis and correlation results show that NeuroCognition measures specific cognitive abilities that differ from, yet correlate with, the general capabilities captured by standard benchmarks.
+
+[p. 2 | section: 1. Introduction | type: Text]
+Overall, NeuroCognition is a benchmark that can identify gaps in LLMs' cognitive abilities relative to humans, offering insights into how LLMs operate and informing the design and training of future LLMs. Our code<sup>1</sup>, data, and results will be made publicly available.
+
+[p. 2 | section: 2. Related Works | type: Text]
+The neuropsychological tests included in our work measure different aspects of human cognitive abilities. The RPM (????) measures abstract, relational reasoning capabilities (?). RPM has been used to identify intellectually gifted individuals, and has been shown to be widely reliable across different geographical boundaries (?). The SWM test (?), on the other hand, measures the working memory, the ability to hold and mentally work with information, which may no longer be perceptually present (?). This test has been used to detect at-risk individuals for psychosis (?), and highlight the working memory deficits in individuals with autism (?). Finally, WCST (??) measures cognitive flexibility, i.e., the
+
+[p. 2 | section: 2. Related Works | type: Text]
+ability to adapt to a change in perspective and reallocate resources to process relevant information (??). Patients with schizophrenia are shown to perform worse on this test (?), and are shown to be robust even in clinical settings (?).
+
+[p. 2 | section: 2. Related Works | type: Text]
+Recent advances in LLMs have led many works seeking to establish stronger correlations between LLMs and human intelligence. Many works focus on developing a cognitive benchmark aiming to measure fluid intelligence (????), emphasizing abstractness and dynamic evaluation to avoid the data contamination issue (?). Other works applied human cognitive tests to LLMs (?) and showcased cognitive behaviors in LLMs, such as similar working memory capacity between humans and GPT-4 in the n-back test (?), and above human-level cognitive flexibility in the WCST (?). Some works also revealed common factors of models' performance across different benchmarks (??), indicating the existence of a common underlying intelligence.
+
+[p. 2 | section: 2. Related Works | type: Text]
+Aside from cognitive abilities, several studies have also quantified specific areas where LLMs struggle. In multi-turn evaluation, LLMs tend to perform worse than single-turn evaluations (?), and still perform sub-optimally in multi-turn tasks (??). LLMs also struggle in longer multi-step reasoning (?) and in complex multi-stage tasks (?). LLMs also exhibit cross-modal hallucination (?), and struggle in multi-image, multi-turn benchmarks (??). LLMs also fail in simple pattern tasks using modified alphabets (?), indicating the lack of general analogical reasoning in LLMs.
+
+[p. 2 | section: 2. Related Works | type: Footnote]
+& lt;sup>1</sup>Code available at
+
+[p. 3 | section: 3.1. Raven's Progressive Matrices | type: Text]
+Task. RPM presents a 3x3 matrix of construct patterns with the bottom-right cell missing. The participant needs to infer abstract relations across rows and columns, and fill the missing cell by either selecting one of the 8 available options or constructing it. In our benchmark, we use the Raven's Advanced Progressive Matrices (RAPM) (?) form.
+
+[p. 3 | section: 3.1. Raven's Progressive Matrices | type: Text]
+Setup. For the image version, we use the RAVEN dataset (?), implementation of RAPM with 7 different figure configurations licensed with GPL-3.0. We take 20 items per configuration, resulting in a total of 140 test data. We combine the 3x3 question matrix with the 8 available options into a single image input, as shown in Figure 1, and present it to the model in multiple-choice (MC) format. For the text setting, we programmatically generate symbolic matrices using character and string attributes (see Appendix A for generation details). We provide two answer formats for text setting: multiple-choice (MC, same 8 options) and generative (Gen, free-form string). A stored rule representation enables automatic validation of Gen outputs (see Appendix B for validation details). We generate 200 test data for the text setting with random configurations.
+
+[p. 3 | section: 3.1. Raven's Progressive Matrices | type: Text]
+Metrics. We report accuracy for both MC and Gen. Average human performance for images is 0.814, reported by ?, and not yet established for the text variant.
+
+[p. 3 | section: 3.2. Spatial Working Memory | type: Text]
+Task. In the SWM test, participants are presented with an image with several boxes. A token is hidden in one of the boxes, and participants are asked to search for it by elimination, opening one box at a time. After the token is found, it is regenerated in another box that has not previously contained it in the current trial. This is repeated until all boxes have contained all tokens, with the goal of locating all tokens with minimal redundant checking and without revisiting boxes unnecessarily.
+
+[p. 3 | section: 3.2. Spatial Working Memory | type: Text]
+Setup. We implement two difficulty levels. Easy: 8 boxes, 1 token active at a time, early stop after all 8 tokens found or a maximum of 64-guess cap. Hard: 12 boxes with 2 distinct tokens active concurrently (24 tokens found required, maximum 144 guesses), with each token directly reappearing after being found, and one box can contain both tokens at once. We experiment with three types of feedback setups: text-only, image-only (a grid with no additional textual information), and image+text variants. The image input uses a grid system for easier answer verification, as shown in Figure 1. The text-only setup identifies the boxes by number and ask the model to choose a box number.
+
+[p. 3 | section: 3.2. Spatial Working Memory | type: Text]
+Metrics. Since SWM is highly affected by luck, we define
+
+[p. 3 | section: 3.2. Spatial Working Memory | type: Text]
+some metrics that focus on penalizing errors and unfinished tests. We classify three avoidable error types:
+
+[p. 3 | section: 3.2. Spatial Working Memory | type: ListGroup]
+illegal: opens a box that can no longer contain a token because it has already yielded all available tokens for the trial. no-box: refers to a box index/identifier that does not exist. repeated: re-opens a box that was previously opened and found empty, before any new token was regenerated (i.e., no state change that could justify revisiting).
+
+[p. 3 | section: 3.2. Spatial Working Memory | type: Text]
+These three error types sum to nerr. Let nvalid be the total valid guesses. A consistency term Sc:
+
+[p. 3 | section: 3.2. Spatial Working Memory | type: Equation]
+S_c = 1 - \frac{n_{err}}{n_{valid}} \tag{1}
+
+[p. 3 | section: 3.2. Spatial Working Memory | type: Text]
+measures avoidance of structurally unjustified moves. Note that we omit invalid guesses (no parseable answer) to ensure fair evaluation of real guessing attempts. Let T f be number of tokens found out of T required; the final score Sswm is
+
+[p. 3 | section: 3.2. Spatial Working Memory | type: Equation]
+S_{swm} = \frac{T_f}{T} \times S_c \tag{2}
+
+[p. 3 | section: 3.2. Spatial Working Memory | type: Text]
+Human worst-case. A cognitively normal human, given a complete history of their own moves (like conversation history in LLMs), will be able to achieve full completion by systematic elimination and yield nerr = 0 regardless of the setup and difficulty, therefore obtaining Sswm = 1.
+
+[p. 3 | section: 3.3. Wisconsin Card Sorting Test | type: Text]
+Task. In WCST, four option cards containing shapes are shown to participants. They are then asked to match the given card to one of the four option cards using one of three rules: shape, color, or number. Importantly, participants are not informed of the correct rule for matching the cards and must eliminate potential rules to find the correct one. After choosing the correct answer, a different card is shown while maintaining the same rule. The matching rule is considered complete after a fixed number of consecutive correct answers and is changed upon completion without expressly informing the participant. The test finishes after a fixed number of rules are completed or a maximum number of guesses is reached.
+
+[p. 3 | section: 3.3. Wisconsin Card Sorting Test | type: Text]
+Setup. We require 5 consecutive correct matches (c = 5) to complete a rule instance. We define two difficulty levels. Easy: 3 candidate rules (color, shape, number), each repeated twice (6 total rule instances) or until the 64-guess cap. Hard: adds background color (4 candidate rules, 8 total rule instances) with a maximum of 96 guesses. We also construct controlled ambiguity variants for the hard setting: (i) off (no ambiguous cards), (ii) first (only the first card
+
+[p. 4 | section: 3.3. Wisconsin Card Sorting Test | type: Text]
+under a new rule is ambiguous), and (iii) rest (all subsequent cards after the first are ambiguous). Ambiguity occurs when multiple attributes match simultaneously, forcing the model to track disambiguating feedback over turns. The image-input setting provides images of the cards (Figure 1), while the text-only setting lists descriptions of the option cards.
+
+[p. 4 | section: 3.3. Wisconsin Card Sorting Test | type: Text]
+Metrics. Each completed rule i uses g_i guesses from the start of that rule until c consecutive correct response. Therefore, we can define a score S_{r_i} for each completed rule i:
+
+[p. 4 | section: 3.3. Wisconsin Card Sorting Test | type: Equation]
+S_{r_i} = \frac{c}{q_i}. (3)
+
+[p. 4 | section: 3.3. Wisconsin Card Sorting Test | type: Text]
+where unfinished rules contribute zero. Lower g_i means faster hypothesis convergence, which repeated reliance on an incorrect hypothesis inflates g_i and lowers S_{r_i} . With n_c completed out of N target rule instances, we can calculate the final score S_{west} as:
+
+[p. 4 | section: 3.3. Wisconsin Card Sorting Test | type: Equation]
+S_{wcst} = \frac{1}{N} \sum_{i=1}^{n_c} S_{r_i}. (4)
+
+[p. 4 | section: 3.3. Wisconsin Card Sorting Test | type: Text]
+We additionally report two specific error rates aggregated over a single trial with total turns L. Let y_t \in \{0,1\} denote correctness at turn t ( y_t = 1 if the model's match is correct), and let r_t \in \mathcal{R} be the rule the model is currently following at turn t (e.g., color/shape/number/background).
+
+[p. 4 | section: 3.3. Wisconsin Card Sorting Test | type: Text]
+Perseverative Response (PR). PR measures how often the model re-selects a rule that has already been ruled out by negative feedback within the current rule block. We use PR rather than the standard WCST perseverative error (PE) because it captures a broader range of errors while still including PE.
+
+[p. 4 | section: 3.3. Wisconsin Card Sorting Test | type: Text]
+Let k(t) index the current (ground-truth) rule block containing turn t, and let \mathcal{R}^t_{\neg} \subseteq \mathcal{R} be the set of rules eliminated by feedback from earlier turns in the same block. Then:
+
+[p. 4 | section: 3.3. Wisconsin Card Sorting Test | type: Equation]
+PR = \frac{\sum_{t=1}^{L} \mathbf{1} \{ \mathcal{R}_{\neg}^{t} \neq \emptyset \} \mathbf{1} \{ r_{t} \in \mathcal{R}_{\neg}^{t} \}}{\sum_{t=1}^{L} \mathbf{1} \{ \mathcal{R}_{\neg}^{t} \neq \emptyset \}}. (5)
+
+[p. 4 | section: 3.3. Wisconsin Card Sorting Test | type: Text]
+Failure-to-maintain-set (FMS). We mark a rule as conceptually acquired after m=3 consecutive correct responses within a block. For each block k, let \tau_k be the first turn index in that block at which this acquisition criterion is met (if it is never met, the block contributes no turns). Let \mathcal{T}_{post} be the set of all turns after acquisition and before the next rule switch:
+
+[p. 4 | section: 3.3. Wisconsin Card Sorting Test | type: Equation]
+\mathcal{T}_{\text{post}} = \bigcup_{k} \{ t : k(t) = k, \ t > \tau_k \}. \tag{6}
+
+[p. 4 | section: 3.3. Wisconsin Card Sorting Test | type: Text]
+FMS is the error rate on these post-acquisition turns:
+
+[p. 4 | section: 3.3. Wisconsin Card Sorting Test | type: Equation]
+FMS = \frac{\sum_{t \in \mathcal{T}_{post}} \mathbf{1}\{y_t = 0\}}{|\mathcal{T}_{post}|}. (7)
+
+[p. 4 | section: 3.3. Wisconsin Card Sorting Test | type: Text]
+Human worst-case. For the easy setup, a cognitively normal human needs at most two exploratory wrong guesses to find the correct rule, then continues with 5 correct answers \Rightarrow 5/7 \approx 0.714 . For hard, the worst case happened with the first ambiguity setup, where at most there are two exploratory wrong guesses, one ambiguous early correct, one disambiguating wrong, and 5 consecutive correct answers \Rightarrow 5/9 \approx 0.556 .
+
+[p. 4 | section: 3.4. Experiment Setup | type: Text]
+In our experiments, we benchmark several recent multimodal reasoning models: proprietary models (Gemini 3 Pro (?), Gemini 2.5 Pro (?), GPT-5 (?), o4-Mini (?), Claude Sonnet 4 (?), Grok 4.1 Fast (?), Grok 4 Fast (?) and openweight models (GLM 4.6V (?), GLM 4.5V (?), Qwen3-VL-235B (????)).
+
+[p. 4 | section: 3.4. Experiment Setup | type: Text]
+We use models' default reasoning capabilities via API parameters and chain-of-thought prompts. However, we disable reasoning for Claude Sonnet 4 and Grok 4 Fast on the RAPM test, due to overthinking issues that exhausted output limits and led to worse performance. For RAPM, we set thinking budgets close to the model's maximum output tokens. For SWM and WCST, we limit the output tokens to 8,192 tokens and the thinking budget to 4,000 tokens per turn. We repeat all SWM and WCST experiments three times, while RAPM is evaluated once. All evaluations are zero-shot with a temperature of 0 (if applicable). Further details are provided in Appendix C.
+
+[p. 4 | section: 4. Results | type: Text]
+Table 1 reports the baseline results for NeuroCognition, with the complete results available in Appendix E.
+
+[p. 4 | section: 4.1. Key Findings | type: Text]
+Models are stronger in text, struggle in image. Across all tests, models show a consistent advantage in text-only setups compared to image-based inputs. Even combined modalities (image+text) yield worse results than text alone, indicating difficulties in parsing and reasoning over visual stimuli and combining visual with textual information.
+
+[p. 4 | section: 4.1. Key Findings | type: Text]
+Performance drops with difficulty. Increasing task difficulty, particularly in SWM, causes a significant drop in performance. In the SWM text easy setup almost all models complete the task perfectly, while in the hard setting only three models manage to do so. Even the two best models are unable to do it consistently through all setups. This is interesting because the core ability required is still the same. The harder setting only increases demands on maintaining and updating internal state, proven to still be a challenge.
+
+[p. 4 | section: 4.1. Key Findings | type: Text]
+Models possess good cognitive flexibility. Compared to
+
+[p. 5 | section: 4.1. Key Findings | type: TableGroup]
+Table 1. Baseline NeuroCognition benchmark performance without hints, notes, or pattern hints. RAPM uses the reported human average performance from ?. SWM & WCST compares against a healthy human worst-case score. Full setup details for each Easy/Hard setup are provided in each test section. MC and Gen denote multiple-choice and generation setup, respectively. All results are zero shot and use each model's default reasoning level, except where marked with ∗ indicating reasoning was disabled due to some issues resulting in worse performance. Bold numbers and green cells denote scores that match or surpass human performance. ↑ means higher is better. RAPM (Accuracy ↑) SWM (Sswm ↑) WCST (Swcst ↑) Model Text Image Text Image+Text Image Text Image MC Gen MC Easy Hard Easy Hard Easy Hard Easy Hard Easy Hard GPT-5 0.540 0.605 0.279 1.000 1.000 0.985 0.989 0.927 0.631 0.848 0.815 0.824 0.794 Gemini 3 Pro 0.585 0.460 0.400 1.000 1.000 1.000 0.963 1.000 0.793 0.832 0.776 0.861 0.805 Gemini 2.5 Pro 0.525 0.560 0.186 1.000 0.980 0.733 0.480 0.408 0.622 0.825 0.806 0.823 0.776 Claude Sonnet 4 0.560∗ 0.385∗ 0.207∗ 1.000 0.866 0.256 0.265 0.630 0.005 0.764 0.773 0.800 0.710 Grok 4.1 Fast 0.505 0.495 0.100 1.000 0.983 0.145 0.017 0.014 0.009 0.841 0.825 0.107 0.060 Grok 4 Fast 0.690∗ 0.555∗ 0.086∗ 1.000 0.994 0.021 0.002 0.029 0.005 0.836 0.754 0.336 0.209 o4-Mini 0.450 0.410 0.207 0.951 0.811 0.055 0.032 0.042 0.005 0.789 0.792 0.347 0.277 GLM 4.6V 0.320 0.170 0.357 1.000 0.574 0.044 0.010 0.000 0.028 0.300 0.095 0.191 0.064 GLM 4.5V 0.035 0.000 0.128 1.000 0.536 0.044 0.000 0.000 0.000 0.773 0.309 0.108 0.083 Qwen3-VL-235B 0.485 0.570 0.493 1.000 0.760 0.026 0.016 0.031 0.135 0.230 0.233 0.420 0.071 Human - - 0.814 1.000 1.000 1.000 1.000 1.000 1.000 0.714 0.556 0.714 0.556
+
+[p. 5 | section: 4.1. Key Findings | type: Text]
+RAPM and SWM, models generally perform better on WCST, with more models passing the human worst-case baseline. This suggests current LLMs handle rule inference and task-switching behavior more robustly than abstract relational reasoning or complex working-memory maintenance, though image setup is still an issue for some models.
+
+[p. 5 | section: 4.1. Key Findings | type: Text]
+Answer format matters. In RAPM text, models exhibit different preferences for multiple-choice versus generative answer formats on the same items. This likely reflects differences in decision constraints and how models interpret the problem, which sometimes leads to overcomplication.
+
+[p. 5 | section: 4.2. Failure Cases & Patterns | type: Text]
+RAPM. Our analysis of RAPM reasoning traces (Appendix E Table 10, 11) reveals three persistent deficits across modalities. First, hallucination and mis-grounding: models fabricate visual features or impose non-existent numerical patterns on text that contradict the data. Second, semantic over-interpretation: models overlook simple local rules (XOR, concatenation) in favor of complex, global hypotheses like "symmetry grids." Third, specifically in Text Gen, models exhibit content-logic disconnect, where they frequently derive correct attribute constraints but generate hallucinated strings that satisfy these metadata rules while violating the actual constructive logic. These failures suggest a disconnect between low-level feature verification and abstract reasoning.
+
+[p. 5 | section: 4.2. Failure Cases & Patterns | type: Text]
+SWM. We investigate failures using four error types explained earlier in Section 3.2. Since models struggle most in the image modality, we focus on image and image-text
+
+[p. 5 | section: 4.2. Failure Cases & Patterns | type: TableGroup]
+Table 2. SWM image and image+text failure turns averaged across all runs (easy and hard). ↓ means lower is better. Model Illegal ↓ Repeat ↓ No-Box ↓ Invalid ↓ GPT-5 2.75 4.25 3.42 0.50 Gemini 3 Pro 5.67 1.17 0.00 0.58 Gemini 2.5 Pro 21.75 3.08 13.83 0.33 Claude Sonnet 4 1.42 4.17 47.08 18.42 Grok 4.1 Fast 0.08 4.42 78.00 9.00 Grok 4 Fast 0.00 6.00 68.92 21.58 o4-Mini 0.00 19.25 69.75 1.25 GLM 4.6V 0.50 4.92 9.67 82.50 GLM 4.5V 0.00 1.92 28.00 71.58 Qwen3-VL-235B 4.58 22.75 40.67 21.33
+
+[p. 5 | section: 4.2. Failure Cases & Patterns | type: Text]
+failure cases (Table 2) . From the data, we identify two main weaknesses, state tracking under sparse feedback (models forget key task state) and task execution (invalid and nobox actions instead of exploiting eliminations). High illegal and repeat counts indicate confusion and working-memory failures, high no-box rates suggest a visual grounding gap (difficulty identifying a valid box), and invalid turns reflect failures to follow the task instruction.
+
+[p. 5 | section: 4.2. Failure Cases & Patterns | type: Text]
+WCST. Table 3 summarizes WCST image error rates using FMS and PR as explained in Section 3.3. We focus on weaker models that struggle in WCST image. Overall, most errors happen after the model has found the right rule. Models either (i) fail to keep using the acquired rule (higher FMS) or (ii) fail to change their choice after negative feedback (higher PR), i.e., they choose a rule that was already ruled out within the current rule block. The background-
+
+[p. 6 | section: 4.2. Failure Cases & Patterns | type: TableGroup]
+Table 3. WCST error rates for models that struggle in WCST image tests. ↓ means lower is better. Model Easy Hard FMS ↓ PR ↓ FMS ↓ PR ↓ Grok 4.1 Fast 0.339 0.324 0.376 0.420 Grok 4 Fast 0.432 0.201 0.347 0.309 o4-Mini 0.320 0.092 0.437 0.352 GLM 4.6V 0.089 0.442 0.169 0.420 GLM 4.5V 0.404 0.310 0.303 0.432 Qwen3-VL-235B 0.229 0.619 0.500 0.794 Table 4. RAPM results with no reasoning. ↑ means higher is better. Paired t-test results show effects in Text MC is statistically significant (p < .05) while others are not (p > .05).
+
+[p. 6 | section: 4.2. Failure Cases & Patterns | type: Table]
+Model RAPM (Accuracy ↑) Text MC Text Gen Image MC Gemini 2.5 Pro 0.690∆+0.17 0.450∆−0.11 0.093∆−0.09 Grok 4.1 Fast 0.690∆+0.19 0.565∆+0.07 0.114∆+0.01 GLM 4.6V 0.635∆+0.32 0.175∆+0.01 0.336∆−0.02 GLM 4.5V 0.460∆+0.43 0.000∆0.00 0.014∆−0.11
+
+[p. 6 | section: 4.2. Failure Cases & Patterns | type: Text]
+color addition tends to make both problems worse. We provide more details and other error metrics in Appendix E.
+
+[p. 6 | section: 5. Ablation Study | type: Text]
+Following the baseline results, we explore aspects that could affect model performance. We present the change in performance compared to Table 1 baseline using a subscripted delta (∆) notation to indicate any improvements or declines.
+
+[p. 6 | section: 5.1. RAPM: Reasoning vs No Reasoning | type: Text]
+We compare RAPM performance when models run with their default reasoning behavior versus a constrained noreasoning setting, as shown in Table 4. This experiment was run only on models where reasoning can be disabled (excluding Claude Sonnet 4 and Grok 4 Fast because of reasoning issue). Disabling reasoning produces large, consistent gains in Text MC (p = .0207) across models, while Text Gen and Image MC shows mixed effects. The improvement in Text MC likely occurs because the decision space is explicit and limited, and avoiding extended chain-of-thought reduces overcomplication and distractor attraction. For Text Gen, using reasoning or not shows no clear effect (p = .83). For Image MC, disabling reasoning appears slightly worse (p = .173), likely because reasoning helps interpret visual inputs and detect patterns, though the effect may be small.
+
+[p. 6 | section: 5.1. RAPM: Reasoning vs No Reasoning | type: TableGroup]
+Table 5. RAPM results with pattern hints. ↑ means higher is better. Paired t-test results show no significant effect on all setups (p > .05). Model RAPM (Accuracy ↑) Text MC Text Gen Image MC GPT-5 0.620∆+0.08 0.720∆+0.12 0.257∆−0.02 Gemini 3 Pro 0.595∆+0.01 0.545∆+0.09 0.386∆−0.01 Gemini 2.5 Pro 0.485∆−0.04 0.585∆+0.03 0.221∆+0.04 Claude Sonnet 4 0.525∆−0.03 0.360∆−0.02 0.221∆+0.01 Grok 4.1 Fast 0.570∆+0.07 0.580∆+0.09 0.114∆+0.01 Grok 4 Fast 0.655∆−0.04 0.540∆−0.02 0.100∆+0.01 o4-Mini 0.435∆−0.02 0.525∆+0.12 0.236∆+0.03 GLM 4.6V 0.395∆+0.08 0.235∆+0.07 0.321∆−0.03 GLM 4.5V 0.085∆+0.05 0.000∆0.00 0.090∆−0.04 Qwen3-VL-235B 0.575∆+0.09 0.590∆+0.02 0.457∆−0.04
+
+[p. 6 | section: 5.2. RAPM: Pattern Hints | type: Text]
+In RAPM, we test whether providing hints regarding possible underlying patterns helps models narrow down their reasoning and identify the correct answer. The results shown in Table 5 vary across models and setups, and we find no significant effect (paired t-tests: Text MC p = .177, Text Gen p = .45, Image MC p = .666). Most results are similar to those obtained without using any patterns, and in some cases, they even decrease. Overall, pattern hints have a limited and inconsistent impact. Most changes are small (often positive), and the aggregate effect is not statistically significant, suggesting that many models still lack core abstract reasoning capabilities, even with these hints.
+
+[p. 6 | section: 5.3. SWM & WCST: Note-Taking Assistance | type: Text]
+For SWM and WCST, we explore assistance in the form of notes, which emulates cognitive offloading (??) where humans write down their observations to keep track of the state. The system provides the notes itself in every turn, explicitly stating the current state and actions taken by the model so far. From Table 6, we observe large gains for some models, but the average effect is not significant (Text Hard p = .161; Img+Text Easy p = .842), which may reflect limited sample size or generally small changes. This suggests that many models still struggle with working-memory maintenance, where notes are not even helpful and may even cause slight drops.
+
+[p. 6 | section: 5.3. SWM & WCST: Note-Taking Assistance | type: Text]
+For WCST, improvements from note-taking are generally consistent across models and setups, although they are close to reach statistical significance in our (Image Easy: p = .14; Image Hard: p = .0524). This indicates that note-taking provides practical benefits by helping models track tested attributes and maintain a consistent hypothesis.
+
+[p. 7 | section: 5.3. SWM & WCST: Note-Taking Assistance | type: TableGroup]
+Table 6. SWM results with note-taking assistance. Paired t-test results show no significant effects on all setups (p > .05). Model SWM (Sswm ↑) Text Hard Img+Txt Easy Gemini 2.5 Pro 0.965∆−0.02 0.704∆−0.03 Claude Sonnet 4 0.826∆−0.04 0.207∆−0.05 Grok 4.1 Fast 0.973∆−0.01 0.043∆−0.10 Grok 4 Fast 0.964∆−0.03 0.029∆+0.01 o4-Mini 0.935∆+0.12 0.012∆−0.04 GLM 4.6V 0.803∆+0.23 0.082∆+0.04 GLM 4.5V 0.851∆+0.32 0.000∆−0.04 Qwen3-VL-235B 0.782∆+0.02 0.318∆+0.29 Table 7. WCST image results with note-taking assistance. Paired t-test results show no significant effects on all setups (p > .05).
+
+[p. 7 | section: 5.3. SWM & WCST: Note-Taking Assistance | type: Table]
+Model WCST (Swcst ↑) Image Easy Image Hard Grok 4.1 Fast 0.135∆+0.028 0.197∆+0.137 Grok 4 Fast 0.215∆−0.120 0.257∆+0.048 o4-Mini 0.392∆+0.045 0.295∆+0.018 GLM 4.6V 0.430∆+0.239 0.411∆+0.347 GLM 4.5V 0.306∆+0.199 0.123∆+0.040 Qwen3-VL-235B 0.640∆+0.331 0.250∆+0.179
+
+[p. 7 | section: 6.1. Factor Analysis of LLM | type: Text]
+To understand the latent structure of capabilities across widely used LLM benchmarks inspired by ?, we conducted an exploratory factor analysis (EFA) on performance data obtained from Artificial Analysis 2 for 156 distinct LLMs. We selected 10 benchmarks widely recognized as standard measures of LLM "intelligence" and real-world capabilities, ranging from coding and math to agentic tool use. Importantly, this selection was driven by the statistical requirement for a large sample size (N > 100) to ensure valid factor analysis, which very few other benchmarks satisfy. We used Principal Axis Factoring for the factoring method.
+
+[p. 7 | section: 6.1. Factor Analysis of LLM | type: Text]
+Parallel analysis indicated a single-factor solution: a unidimensional latent variable explains most performance variation across these tasks. This factor (PA1) accounted for 75% of the total variance, providing strong evidence for a "general capability" factor. As shown in Table 8, all benchmarks had high loadings (λ > 0.76), with GPQA Diamond (0.943) and LiveCodeBench (0.926) as the strongest proxies. This coherence also exposes a limitation of current evaluation, where many benchmarks measure the same underlying construct. This makes NeuroCognition even more important, as
+
+[p. 7 | section: 6.1. Factor Analysis of LLM | type: TableGroup]
+Table 8. Factor analysis loadings for state-of-the-art LLM benchmarks. The single extracted factor (PA1) explains 75% of the total variance, indicating a strong general capability factor (g) across diverse tasks. Benchmark Loadings (λ) GPQA Diamond (?) 0.943 LiveCodeBench (?) 0.926 SciCode (?) 0.922 AA-LCR (?) 0.875 AIME 2025 (?) 0.870 IFBench (?) 0.861 Terminal Bench (?) 0.834 MMLU Pro (?) 0.825 τ -Bench (?) 0.818 Humanity's Last Exam (?) 0.768
+
+[p. 7 | section: 6.1. Factor Analysis of LLM | type: Text]
+it targets distinct, basic cognitive tasks.
+
+[p. 7 | section: 6.2. Correlation with Other Benchmarks | type: Text]
+To assess how NeuroCognition relates to real-world tasks, we correlate it with external benchmarks targeting specific cognitive demands. For working memory (SWM), we use a modified Long HotPotQA (?) that splits questions and passages at the word level and treats each word as a turn. For fluid intelligence (RAPM), we compare against ARC- AGI 2 (?). For cognitive flexibility (WCST), we use LLM Task Interference (?), which measures performance impacts under task switches. Detailed setups and results are in Appendix D. We also correlate the aggregate NeuroCognition score with the average of 11 state-of-the-art benchmarks (the 10 in Section 6.1 plus MMMU Pro (?)), as a proxy for "general capability". These scores are also taken and already normalized from Artificial Analysis. Pearson correlations are shown in Figure 2.
+
+[p. 7 | section: 6.2. Correlation with Other Benchmarks | type: Text]
+We observe a high correlation between the average NeuroCognition score and the 11-benchmark average (r = .86, p = .001, N = 10), validating our framework as a robust predictor of general model capability (g). More importantly, the heatmap provides a granular view of the functional relationships between cognitive primitives and downstream tasks:
+
+[p. 7 | section: 6.2. Correlation with Other Benchmarks | type: ListGroup]
+(1) Working Memory & Context: SWM Text Hard shows a moderate correlation with Long HotPotQA (r = .43, p = .21, N = 10). This suggests that working-memory maintenance is still somewhat relevant long-context retrieval. (2) Fluid Intelligence: RAPM Image MC is the strongest predictor of ARC-AGI 2 (r = .87, p = .024, N = 6), consistent with isolating an "abstract visual reasoning" factor that standard benchmarks often miss while also uncorrelated with the general benchmark average (r = .09). (3) Cognitive Flexibility: WCST Text Hard shows only a
+
+[p. 8 | section: 6.2. Correlation with Other Benchmarks | type: FigureGroup]
+Figure 2. Correlations between our NeuroCognition benchmark tests and corresponding real-world LLM benchmarks. Cells with bold text and highlighted borders indicate task pairs that are similar and rely on the same cognitive abilities (including averages). Statistical significance is denoted by asterisks (*p < .05, **p < .01).
+
+[p. 8 | section: 6.2. Correlation with Other Benchmarks | type: Text]
+weak, non-significant correlation with LLM Task Interference ( r=.24,\ p=.50,\ N=10 ). This suggests that the abstract rule-updating mechanisms measured by WCST may be distinct from practical task switching scenarios.
+
+[p. 8 | section: 7. Conclusion | type: Text]
+In this work, we introduce the NeuroCognition benchmark, a suite of neuropsychological tests (RAPM, SWM, WCST) adapted for LLMs to probe specific cognitive processes using multimodal, process-aware metrics. Across tasks, state-of-the-art models show meaningful but incomplete alignment with human cognition. They perform better in text, but performance drops for images and harder settings that demand intensive grounding and state tracking. Overall, explicit reasoning is not uniformly beneficial, and humanlike assistance shows mixed effects. The impact of pattern hints is limited, while note-taking helps more consistently in WCST but not in SWM. Complementing these results, our factor analysis suggests a dominant general-capability factor across standard benchmarks, highlighting redundancy in current evaluation. Correlations with external benchmarks further suggest that gains in NeuroCognition process scores translate into improvements in semantically related realworld tasks, while individual tests still measure distinct cognitive abilities. In summary, our results suggest that current LLMs do not possess the same kind of intelligence as humans, but still align with humans in some aspects, which can inform how we use and improve them.
+
+[p. 8 | section: 8. Future Work | type: Text]
+Since we apply neuropsychological tests designed for humans to LLMs, it is important for future work to validate whether the same cognitive abilities measured in humans are also captured in LLMs. This helps to further under-
+
+[p. 8 | section: 8. Future Work | type: Text]
+stand the relationship between human and LLM intelligence and improve the interpretability of the benchmark results. Another interesting direction is improving the LLM itself through neuropsychologically grounded training data. The data that we used in this benchmark captures basic cognitive processes that are foundational to human intelligence, and therefore can be used to improve the cognitive abilities of LLMs. Additionally, recent works investigate ways to compress context (??), or utilize some sort of memory in their architecture (???), and our benchmark could serve to evaluate such methods. Finally, this data can be easily generated synthetically with code, making it scalable to create a large amount of data for training purposes.
+
+[p. 8 | section: Limitations | type: Text]
+Our evaluation has three main limitations. First, sample sizes are constrained by cost: running state-of-the-art models with effectively unlimited reasoning is expensive. For RAPM, we capped runs at n = 200 (text) and n = 140(image), which may compromise the robustness of our results, especially in the original image modality. Second, our benchmark assumes that the same neuropsychological test used to assess a specific cognitive ability in humans can also be applied to LLMs to evaluate that same ability. In the neuropsychological field, this is not always the case, since a neuropsychological test is standardized on a specific population and must be validated based on psychometric properties (validity, reliability, and measurement invariance) when applied to a different population. Therefore, even when a model performs consistently as well as humans across all tasks in our benchmark, we can only claim that the model exhibits a similar level of intelligence and can simulate human-like cognitive abilities, but not that it thinks like humans or possesses the same underlying cognitive mechanism. Third, since the neuropsychological tests we
+
+[p. 9 | section: Limitations | type: Text]
+used are widely used and well-established, we assume they are valid and reliable based on their psychometric properties (validity, reliability, and measurement invariance). However, this assumption may not hold in all cases, and further validation across prompts, contexts, and a wide range of LLMs is needed.
+
+[p. 9 | section: Impact Statement | type: Footnote]
+github.com/features/copilot
+
+[p. 9 | section: Impact Statement | type: Footnote]
+gemini.google.com
+
+[p. 9 | section: Impact Statement | type: Footnote]
+chatgpt.com
+
+[p. 10 | section: A.1. Rule Space and Attributes | type: Text]
+The generation engine employs a compositional rule space defined by attributes that constrain string properties. For each matrix, two primary attributes are sampled—one for rows and one for columns. Additional consistency rules (leaks) are inferred if they naturally emerge from the intersection of constraints.
+
+[p. 10 | section: A.1. Rule Space and Attributes | type: Text]
+The implemented attributes and their parameters are:
+
+[p. 10 | section: A.1. Rule Space and Attributes | type: ListGroup]
+1. Character Set Type: Restricts the alphabet of the string. Values : letters, digits, symbols. Logic : All characters in the cell must belong to the specified set. 2. Type Count: Constraints on the count of specific character types. Types : vowels, consonants, uppercase/lowercase variants, etc. Rules : even, odd, multiple of N (for N ∈ {2, 3, 4}). Logic : The count of characters matching the type must satisfy the parity or divisibility rule. Zero counts are explicitly disallowed for these rules to ensure pattern salience. 3. Quantitative Constant: Fixes a metric to a specific integer value across the sequence. Metrics : String length, unique character count, or count of specific types. Range : Integers in [2, 5]. 4. Quantitative Progression: Enforces an arithmetic progression on a metric. Metrics : Same as Quantitative Constant. Parameters : Start value ∈ [1, 3], Step size ∈ [1, 3]. Logic : The metric value increases by the step size across the row or column (e.g., lengths 3, 5, 7). 5. Sorted Order: Enforces lexicographical ordering. Values : ascending, descending, mixed. Logic : Characters in the string must follow the specified order. mixed explicitly requires the string to be unsorted. 6. Positional Constraint: Fixes the character type at specific indices. Indices : first, last, even, odd. Logic : Characters at the specified positions must belong to a specific type (e.g., "first character must be a digit").
+
+[p. 10 | section: A.2. Generation Algorithm | type: Text]
+The matrix generation process follows a constraint-satisfaction approach with backtracking, as detailed in Algorithm 1. The core logic proceeds as follows:
+
+[p. 10 | section: A.2. Generation Algorithm | type: ListGroup]
+1. Attribute Sampling: Randomly sample one attribute specification for rows and one for columns. 2. Compatibility Check: Validate that the row and column attributes are compatible. For example, a "Sorted" row attribute conflicts with a "Positional" column attribute because sorting rearranges characters, potentially violating fixed-position constraints. Infeasible combinations trigger a re-sampling. 3. Constraint Propagation: The row and column attributes are combined to define specific constraints for each of the 3 × 3 cells. "Leak" detection is performed to identify additional properties that effectively become constant across rows or columns (e.g., if a row rule restricts characters to digits and a column rule restricts them to even numbers, a universal "digits" constraint might apply). 4. Grid Construction: Cells are generated sequentially. For each cell, a string is constructed to satisfy the merged constraints. Random backtracking is used if a cell cannot be generated given the current constraints and history. 5. Variation Enforcement: To prevent trivial patterns, we enforce a minimum Hamming distance between strings in the same row or column. This ensures that the progression relies on abstract rules rather than simple repetition.
+
+[p. 10 | section: A.2. Generation Algorithm | type: Text]
+Upon successful generation, the defining constraints for the missing cell (including "leaked" rules) are stored for automatic rule validation.
+
+[p. 11 | section: Algorithm 1 RAPM Text Matrix Generation | type: Code]
+Input: Random Seed S Output: 3 × 3 String Matrix M, Constraints C Initialize RNG with S repeat Sample row attribute Arow and column attribute Acol valid ← CheckCompatibility(Arow, Acol) until valid Initialize cell constraint grid C3×3 from Arow and Acol Apply inferred consistency rules (leaks) to C for i = 0 to 2 do for j = 0 to 2 do attempts ← 0 repeat Generate string s satisfying Ci,j vrow ← ∀k < j : Hamming(s,Mi,k) ≥ 2 vcol ← ∀k < i : Hamming(s,Mk,j ) ≥ 2 attempts ← attempts + 1 until (vrow ∧ vcol) or attempts > MAX RET RIES if not (vrow ∧ vcol) then fail and backtrack end if Mi,j ← s end for end for return M, C
+
+[p. 11 | section: A.3. Multiple-Choice and Distractor Generation | type: Text]
+For the text MC setup, we generate 8 options: 1 correct answer (ground truth for the bottom-right cell) and 7 distractors. Each distractor is guaranteed to violate at least one latent rule to ensure a unique solution.
+
+[p. 11 | section: A.3. Multiple-Choice and Distractor Generation | type: ListGroup]
+Correct Answer: The generated string for the bottom-right (3, 3) cell. Distractors: Generated by applying mutation strategies to the correct answer to violate latent rules. Strategies include: Break Ordering : Swapping adjacent characters to violate sort constraints. Break Positional : Changing a character at a critical index (e.g., first/last) to a forbidden type. Adjust Count : Adding or removing characters to violate parity or length constraints. Character Mutation : Randomly altering characters to valid but incorrect values.
+
+[p. 12 | section: B. RAPM Text Gen Automatic Rule Validation | type: Text]
+For the text Gen setup, we can programmatically check if a generated string S satisfies the stored underlying constraints C for the missing cell (3, 3). Algorithm 2 details this validation procedure.
+
+[p. 12 | section: Algorithm 2 RAPM Text Gen Validation (CellSatisfies) | type: Code]
+Input: Candidate String S, Constraints C Output: Boolean (True if valid) if C.f ixed length ̸= None and |S| ̸= C.f ixed length then return False end if for all (type, count) ∈ C.target counts do if Count(S, type) ̸= count then return False end if end for for all (type, rule) ∈ C.parity rules do c ← Count(S, type) if c = 0 then return False {Zero count forbidden for parity} end if if (rule = 'even' ∧ c is odd) or (rule = 'odd' ∧ c is even) then return False end if end for for all (type, k) ∈ C.multiple rules do c ← Count(S, type) if c = 0 or c (mod k) ̸= 0 then return False end if end for if C.unique count ̸= None and |Unique(S)| ̸= C.unique count then return False end if if C.ordering ̸= None and CheckOrder(S, C.ordering) = False then return False end if if C.positional ̸= None and CheckPositional(S, C.positional) = False then return False end if return True
+
+[p. 12 | section: Algorithm 2 RAPM Text Gen Validation (CellSatisfies) | type: Text]
+The validation function CellSatisfies(S, C3,3) performs the following checks:
+
+[p. 12 | section: Algorithm 2 RAPM Text Gen Validation (CellSatisfies) | type: ListGroup]
+1. Length Check: If a fixed length is enforced (from quantitative constraints), |S| must match. 2. Type Counts: Count(S, type) = k for all exact count constraints. 3. Parity/Divisibility: Count(S, type) must satisfy even/odd/multiple conditions. Crucially, as enforced during generation, zero counts are invalid for these rules to prevent trivial satisfactions (e.g., "0 vowels" satisfying "even vowels"). 4. Unique Characters: If unique count is constrained, |Unique(S)| is checked against the target. 5. Ordering: If sorted order is required, S is checked against its own sorted version. 6. Positional: Characters at constrained indices (first, last, even, odd) are verified against the allowed character set.
+
+[p. 12 | section: Algorithm 2 RAPM Text Gen Validation (CellSatisfies) | type: Text]
+Any string that satisfies all these constraints is considered a correct solution, providing a robust "verifiable" accuracy metric for generative models that goes beyond simple multiple-choice selection or exact-match.
+
+[p. 13 | section: Google AI Studio | type: ListGroup]
+Gemini 3 Pro (?): gemini-3-pro-preview Gemini 2.5 Pro (?): gemini-2.5-pro
+
+[p. 13 | section: OpenAI API | type: ListGroup]
+GPT-5 (?): gpt-5-2025-08-07 o4-Mini (?): o4-mini-2025-04-16
+
+[p. 13 | section: OpenRouter | type: ListGroup]
+Claude Sonnet 4 (?): anthropic/claude-sonnet-4 Grok 4.1 Fast (?): x-ai/grok-4.1-fast Grok 4 Fast (?): x-ai/grok-4-fast GLM 4.6V (?): z-ai/glm-4.6v 6 GLM 4.5V (?): z-ai/glm-4.5v 7
+
+[p. 13 | section: Local / vLLM | type: Text]
+• Qwen3-VL-235B (?): Qwen/Qwen3-VL-235B-A22B-Thinking 8
+
+[p. 13 | section: C.2. vLLM Inference Settings | type: Text]
+For local vLLM runs, we used a server with 4x NVIDIA H100 80GB GPUs and CUDA 12.8, and ran the model with FP8 quantization. Generations used the model's default Qwen3 configuration; temperature and other sampling settings were chosen per the evaluation protocol below.
+
+[p. 13 | section: C.3. Evaluation Protocol Details | type: ListGroup]
+Zero-shot: All evaluations are zero-shot. Temperature: We used a temperature of 0 when possible to reduce sampling variability; when a model's API does not allow overriding temperature, we used the model's default. Reasoning mode: We used each model's default reasoning configuration except for Claude Sonnet 4 and Grok 4 Fast, for which we adjusted reasoning as described in the main text (see Section 3.4) . Maximum output tokens: For RAPM, we allowed the model's maximum permitted output tokens. For SWM and WCST, we limited per-turn outputs to 8192 tokens. Thinking budget: The system prompt described a thinking budget (not enforced as a hard token limit). For RAPM, we used an effective thinking budget of 30,000 tokens for models with a context limit of 32K and 60,000 tokens for models with a context limit of 65K+. For SWM and WCST, the per-turn thinking budget was 4,000 tokens. Repetitions: SWM and WCST setups were run three times each to assess variability (data are generated on the fly); RAPM was run once per model. WCST (notes): For WCST runs that used note-taking assistance, we set the note window to 6.
+
+[p. 13 | section: C.3. Evaluation Protocol Details | type: Footnote]
+6
+
+[p. 13 | section: C.3. Evaluation Protocol Details | type: Footnote]
+7
+
+[p. 13 | section: C.3. Evaluation Protocol Details | type: Footnote]
+8
+
+[p. 14 | section: C.4. Reasoning Configuration | type: Text]
+Model reasoning capabilities were enabled through two complementary methods:
+
+[p. 14 | section: C.4. Reasoning Configuration | type: Text]
+API Configuration. In our ModelWrapper, we programmatically enabled reasoning features where supported by the provider APIs.
+
+[p. 14 | section: C.4. Reasoning Configuration | type: ListGroup]
+Google AI Studio: We accessed Gemini models via the OpenAI-compatible endpoint and enabled reasoning by specifically configuring the thinking config within the extra body parameter. We set include thoughts to True, thinking level for Gemini 3 Pro (default to high), and thinking budget for Gemini 2.5 Pro (default to 8192). OpenAI API: For OpenAI models, reasoning is handled through reasoning effort parameter in the chat completion calls. We use the default value medium for both GPT-5 and o4-Mini. OpenRouter: We set the reasoning parameter in the request body to {"exclude": False} to ensure reasoning tokens were generated and returned, and manage the thinking budget by setting max tokens within the reasoning configuration. For Grok, reasoning was explicitly enabled via {"enabled": True}. vLLM (Local): We enabled the enable thinking parameter in the chat template to activate reasoning generation.
+
+[p. 14 | section: C.4. Reasoning Configuration | type: Text]
+Prompt-Based Instructions. For all tasks, we included explicit instructions in the prompt to guide the reasoning process, specifically when Chain-of-Thought (CoT) was enabled.
+
+[p. 14 | section: SWM & WCST | type: Text]
+Explain your thought process regarding the problem and the feedbacks you received in maximum {think budget} tokens wrapped with <think> and </think>. Then, provide a really short summary of your reasoning after the closing </think> tag.
+
+[p. 14 | section: RAPM | type: Text]
+Explain your thought process (max {think budget} tokens) inside <think>...</think> then give final answer.
+
+[p. 14 | section: RAPM | type: Text]
+This dual approach ensured that models with native reasoning capabilities (activated via API) and standard models (guided via prompts) both engaged in structured reasoning before outputting a final answer.
+
+[p. 14 | section: C.5. Disabling Reasoning | type: Text]
+We disable the reasoning capabilities for some models that are able to do so to conduct the ablation study (see Section 5.1) and to address specific model issues (specifically, Claude Sonnet 4 and Grok 4 Fast with "overthinking" that exhausted token limits on RAPM).
+
+[p. 14 | section: C.5. Disabling Reasoning | type: Text]
+Configuration & Prompts. For most models, reasoning was disabled by reversing the API configurations described previously (e.g., setting include thoughts to False, reasoning enabled to False, or enable thinking to False) and replacing the CoT prompt instructions with a standard direction to provide only the final answer, as shown below.
+
+[p. 14 | section: Gemini 2.5 Pro Non-CoT Instruction | type: Text]
+CRITICAL INSTRUCTION: Do not use <thinking> tags or reasoning steps. Provide the response immediately.
+
+[p. 15 | section: D. External Benchmark Setup & Results | type: Text]
+Long HotPotQA. We work with the distractor development split of HotPotQA (?), sampling 500 question–passage pairs. To stress the models' turn tracking, we split each question and its supporting passages at the word level, treat every word as a separate turn, and fix the model's response on every intermediate turn to "Okay." Each sequence is ordered with the full question first, followed by the concatenated passages. The results for Long HotPotQA are shown in Table 9.
+
+[p. 15 | section: D. External Benchmark Setup & Results | type: Text]
+LLM Task Interference. Following the official implementation from ?, with minor updates, we evaluate two task/context pairings and sample 200 evaluation prompts each. The first pairing uses the MMLU AA task with Rotten Tomatoes exemplars as in-context data, while the second pairing is the Rotten Tomatoes task with TweetQA data. The results for LLM Task Interference are shown in Table 9.
+
+[p. 15 | section: D. External Benchmark Setup & Results | type: Text]
+ARC-AGI 2. Since the results for ARC-AGI 2 (?) are already publicly available in their leaderboard website 9 , we directly took the best results for each model, except for some models that are not available. Although the scores are very low, the results have been verified from the source. These are also summarized in Table 9.
+
+[p. 15 | section: D. External Benchmark Setup & Results | type: TableGroup]
+Table 9. Results for external benchmarks: Long HotPotQA, LLM Task Interference, and ARC-AGI 2. (-) indicates the model result is not available. Model Long HotPotQA Task Interference ARC-AGI 2 GPT-5 0.936 0.958 0.099 Gemini 3 Pro 0.932 0.962 0.311 Gemini 2.5 Pro 0.924 0.875 0.049 Claude Sonnet 4 0.930 0.948 0.059 Grok 4.1 Fast 0.910 0.913 - Grok 4 Fast 0.461 0.788 0.053 o4-Mini 0.802 0.933 0.061 GLM-4.6V 0.862 0.938 - GLM-4.5V 0.271 0.635 - Qwen3 VL 235B 0.916 0.940 -
+
+[p. 16 | section: E.1. Raven's Advanced Progressive Matrices (RAPM) | type: Text]
+The complete results for RAPM image tests, broken down by pattern type, are provided in Table 10. We also provide qualitative examples of reasoning failures across text and image modalities in Table 11, highlighting the prevalence of hallucination and over-interpretation.
+
+[p. 16 | section: E.1. Raven's Advanced Progressive Matrices (RAPM) | type: TableGroup]
+Table 10. RAPM image accuracy breakdown by pattern type. Model Dist. Four Dist. Nine Center In-Dist. Out-Center In-Out Center Left-Right Up-Down GPT-5 0.40 0.35 0.25 0.15 0.35 0.05 0.40 Gemini 3 Pro 0.55 0.65 0.55 0.35 0.20 0.25 0.25 Gemini 2.5 Pro 0.05 0.10 0.35 0.20 0.40 0.20 0.25 Claude Sonnet 4 0.30 0.45 0.15 0.20 0.20 0.10 0.15 Grok 4.1 Fast 0.15 0.20 0.15 0.05 0.00 0.00 0.15 Grok 4 Fast 0.05 0.10 0.05 0.25 0.10 0.10 0.05 o4-Mini 0.45 0.15 0.25 0.10 0.30 0.05 0.15 GLM 4.6V 0.45 0.30 0.40 0.35 0.30 0.25 0.45 GLM 4.5V 0.20 0.05 0.15 0.10 0.15 0.20 0.05 Qwen3-VL-235B 0.45 0.65 0.55 0.40 0.40 0.50 0.50
+
+[p. 16 | section: E.1. Raven's Advanced Progressive Matrices (RAPM) | type: TableGroup]
+Table 11. Selected reasoning failure examples in RAPM (image and text). Models frequently hallucinate attributes to fit a complex hypothesis ("visual/numerological Hallucination") or prefer abstract semantic rules over simple local operations ("semantic overinterpretation"). Task/Model ID Pattern/Context Reasoning Snippet (Failure Point) Outcome Image GPT-5 RAVEN 238 center Up-Down (Split layout) "Row 3: top becomes large solid black triangle. Bottom shapes by row show a sides-count sequence decreasing 5→4→3" (Visual Hallucination: No such triangle exists.) Pred: 2 Correct: 0 Image GPT-5 RAVEN 238 dist nine Dist. Nine (3x3 Grid logic) "Row 1: Left has all 9 positions filled middle emphasizes corners The combination is still all 9 positions" (Relational Over-complexity: Hallucinates a global superposition rule for a simple local pattern.) Pred: 5 Correct: 2 Text MC Gemini 3 Pro text rapm 0 Simple Arithmetic (Char length) "Row 1: 6, 10, 13 characters Row 2: 6, 9, 12 notice a potential arithmetic-like progression" (Numerological Hallucination: Forces arithmetic se ries on string lengths where rule is character types.) Pred: 3 Correct: 5 Text MC GPT-5 text rapm 4 Column Count (Uppercase) "Column 2 uppercase counts go 0 1 2. So for Column 3 they should be 0 1 2." (False Generalization: Hallucinates column-wise progression that doesn't exist in data.) Pred: 3 Correct: 6 Text Gen GPT-5 text rapm 2 Constraint Satisfac tion (Vowel Count) "Rows keep a constant number of uppercase let ters Column 3 needs 6 vowels A valid comple tion is 'aeiOuEsTH'." (Metadata vs Content: Hallucinates content string that fits metadata counts but fails derivation logic.) Pred: aeiOuEsTH Rules: 5 lowercase vowels, 4 uppercase letters Text Gen Gemini 3 Pro text rapm 3 String Length (Decreasing) "Column 2 seems to be decreasing length Row 3: 6, 9, ? Missing string should be length 8." (Numerological: Obsesses over length counts, ignor ing character content patterns.) Pred: T4Z]LBEBUM6 Rules: Col progres of uppercase sion vowels (2→3→4)
+
+[p. 17 | section: E.2. Spatial Working Memory (SWM) | type: Text]
+The complete results for SWM tests are provided in Table 12 for text modality, in Table 13 for image modality, and in 14 for image-text modality. All metrics are averaged over three runs.
+
+[p. 17 | section: E.2. Spatial Working Memory (SWM) | type: TableGroup]
+Table 12. Spatial Working Memory full results (text modality). All metrics are reported as mean with standard deviation (std) in a subscript. Difficulty Model Sswm Tokens Score Guesses Illegal Invalid Repeated No-box GPT-5 1.000±.00 1.000±.00 18.000±6.16 0.000±0.00 0.000±0.00 0.000±0.00 0.000±0.00 Gemini 3 Pro 1.000±.00 1.000±.00 18.000±2.94 0.000±0.00 0.000±0.00 0.000±0.00 0.000±0.00 Gemini 2.5 Pro 1.000±.00 1.000±.00 19.667±1.70 0.000±0.00 0.000±0.00 0.000±0.00 0.000±0.00 Claude Sonnet 4 1.000±.00 1.000±.00 23.000±3.56 0.000±0.00 0.000±0.00 0.000±0.00 0.000±0.00 Easy Grok 4.1 Fast 1.000±.00 1.000±.00 26.667±2.87 0.000±0.00 0.000±0.00 0.000±0.00 0.000±0.00 Grok 4 Fast 1.000±.00 1.000±.00 18.333±3.40 0.000±0.00 0.333±0.47 0.000±0.00 0.000±0.00 o4-Mini 0.951±.05 1.000±.00 23.667±4.03 0.000±0.00 0.667±0.94 1.000±0.82 0.000±0.00 GLM 4.6V 1.000±.00 1.000±.00 24.000±2.94 0.000±0.00 3.333±0.94 0.000±0.00 0.000±0.00 GLM 4.5V 1.000±.00 1.000±.00 20.333±1.70 0.000±0.00 0.000±0.00 0.000±0.00 0.000±0.00 Qwen3-VL-235B 1.000±.00 1.000±.00 20.000±2.45 0.000±0.00 0.000±0.00 0.000±0.00 0.000±0.00 GPT-5 1.000±.00 1.000±.00 56.333±1.25 0.000±0.00 0.000±0.00 0.000±0.00 0.000±0.00 Gemini 3 Pro 1.000±.00 1.000±.00 63.333±1.70 0.000±0.00 1.000±0.82 0.000±0.00 0.000±0.00 Gemini 2.5 Pro 0.980±.03 1.000±.00 82.333±13.57 0.000±0.00 0.000±0.00 2.000±2.83 0.000±0.00 Claude Sonnet 4 0.866±.18 1.000±.00 78.333±11.44 0.000±0.00 0.333±0.47 12.333±16.74 0.000±0.00 Hard Grok 4.1 Fast 0.983±.01 1.000±.00 52.667±9.29 0.000±0.00 0.000±0.00 1.000±0.82 0.000±0.00 Grok 4 Fast 0.994±.01 1.000±.00 63.000±4.55 0.333±0.47 0.333±0.47 0.000±0.00 0.000±0.00 o4-Mini 0.811±.07 1.000±.00 79.333±1.70 5.333±3.77 0.000±0.00 9.667±2.36 0.000±0.00 GLM 4.6V 0.574±.35 0.694±.27 129.333±20.74 3.000±3.56 33.333±34.50 32.333±44.32 0.000±0.00 GLM 4.5V 0.536±.24 0.736±.37 105.333±27.64 16.667±11.81 42.667±58.93 0.333±0.47 0.000±0.00 Qwen3-VL-235B 0.760±.06 1.000±.00 80.333±10.66 19.333±7.72 1.000±0.00 0.333±0.47 0.000±0.00 Gemini 2.5 Pro 0.965±.04 1.000±.00 66.667±9.18 0.000±0.00 0.000±0.00 2.333±2.62 0.000±0.00 Claude Sonnet 4 0.826±.06 1.000±.00 81.000±9.27 1.000±0.82 0.000±0.00 13.667±5.56 0.000±0.00 Grok 4.1 Fast 0.973±.01 1.000±.00 62.333±3.30 0.333±0.47 0.000±0.00 1.333±0.47 0.000±0.00 Hard - Notes Grok 4 Fast 0.964±.01 1.000±.00 65.333±1.25 0.333±0.47 0.000±0.00 2.000±0.00 0.000±0.00 o4-Mini 0.935±.03 1.000±.00 75.333±15.92 1.000±1.41 0.333±0.47 4.333±2.49 0.000±0.00 GLM 4.6V 0.803±.14 1.000±.00 85.000±20.22 0.333±0.47 1.000±0.00 19.000±16.87 0.000±0.00 GLM 4.5V 0.851±.09 1.000±.00 79.333±11.12 10.333±7.36 14.333±16.11 0.333±0.47 0.000±0.00 Qwen3-VL-235B 0.782±.23 0.972±.04 92.333±36.81 0.667±0.94 2.333±3.30 25.667±32.83 0.000±0.00
+
+[p. 17 | section: E.2. Spatial Working Memory (SWM) | type: TableGroup]
+Table 13. Spatial Working Memory full results (image modality). All metrics are reported as mean with standard deviation (std) in a subscript. Difficulty Model Sswm Tokens Score Guesses Illegal Invalid Repeated No-box GPT-5 0.927±0.08 1.000±0.00 27.667±4.64 0.000±0.00 0.000±0.00 2.000±2.83 0.333±0.47 Gemini 3 Pro 1.000±0.00 1.000±0.00 22.333±0.94 0.000±0.00 0.000±0.00 0.000±0.00 0.000±0.00 Gemini 2.5 Pro 0.408±0.38 0.625±0.44 51.333±17.91 0.000±0.00 0.000±0.00 2.667±3.77 30.667±20.42 Claude Sonnet 4 0.630±0.41 0.750±0.35 43.667±14.43 0.000±0.00 0.000±0.00 4.000±5.66 14.000±15.75 Grok 4.1 Fast 0.014±0.01 0.125±0.10 64.000±0.00 0.333±0.47 0.000±0.00 5.000±5.72 51.333±5.25 Easy Grok 4 Fast 0.029±0.04 0.125±0.18 64.000±0.00 0.000±0.00 0.333±0.47 14.333±13.27 42.000±10.61 o4-Mini 0.042±0.04 0.208±0.21 64.000±0.00 0.000±0.00 1.333±0.47 30.667±3.09 21.333±0.94 GLM 4.6V 0.000±0.00 0.000±0.00 64.000±0.00 0.000±0.00 63.333±0.47 0.000±0.00 0.000±0.00 GLM 4.5V 0.000±0.00 0.000±0.00 64.000±0.00 0.000±0.00 32.000±2.83 6.333±6.34 23.333±5.25 Qwen3-VL-235B 0.031±0.02 0.208±0.12 64.000±0.00 12.667±11.09 2.667±3.77 26.333±20.34 14.000±13.37 GPT-5 0.631±0.16 1.000±0.00 98.667±19.07 10.333±2.49 2.000±2.83 14.667±12.26 13.333±18.15 Gemini 3 Pro 0.793±0.24 0.958±0.06 90.000±38.88 20.000±27.58 1.000±0.82 4.667±4.11 0.000±0.00 Gemini 2.5 Pro 0.622±0.30 0.944±0.08 97.333±34.92 35.000±30.82 1.000±1.41 9.000±11.34 0.333±0.47 Claude Sonnet 4 0.005±0.00 0.056±0.02 144.000±0.00 0.000±0.00 44.333±0.47 8.000±3.74 82.667±2.87 Hard Grok 4.1 Fast 0.009±0.01 0.069±0.04 144.000±0.00 0.000±0.00 25.333±26.39 7.333±8.99 99.667±27.98 Grok 4 Fast 0.005±0.01 0.056±0.05 144.000±0.00 0.000±0.00 39.000±53.03 7.667±7.04 88.333±42.76 o4-Mini 0.005±0.00 0.056±0.02 144.000±0.00 0.000±0.00 1.333±1.25 32.000±13.49 98.333±13.47 GLM 4.6V 0.028±0.04 0.069±0.10 144.000±0.00 0.000±0.00 119.333±34.88 8.333±11.79 6.333±8.96 GLM 4.5V 0.000±0.00 0.000±0.00 144.000±0.00 0.000±0.00 111.667±0.94 0.667±0.94 30.667±2.05 Qwen3-VL-235B 0.135±0.12 0.347±0.27 144.000±0.00 4.333±6.13 63.000±1.41 30.667±14.06 21.000±19.87
+
+[p. 19 | section: E.3. Wisconsin Card Sorting Test (WCST) | type: Text]
+The complete results for WCST tests are provided in Table 15 for text modality, and in Table 16 for image modality. All metrics are averaged over three runs.
+
+[p. 19 | section: E.3. Wisconsin Card Sorting Test (WCST) | type: TableGroup]
+Table 15. Wisconsin Card Sorting Test full results (text modality). All metrics are reported as mean with standard deviation (std) in a subscript. Difficulty Model Swcst Accuracy FMS PR Completed Rules 1st Rule Trials Easy GPT-5 0.848±.02 0.842±.02 0.000±.00 0.000±.00 6.000±0.00 6.000±0.00 Gemini 3 Pro 0.804±.05 0.815±.04 0.000±.00 0.003±.01 7.000±1.00 6.833±1.07 Gemini 2.5 Pro 0.825±.02 0.819±.02 0.000±.00 0.000±.00 6.000±0.00 6.667±0.94 Claude Sonnet 4 0.764±.06 0.786±.03 0.000±.00 0.027±.02 6.000±0.00 7.000±0.82 Grok 4.1 Fast 0.841±.03 0.835±.03 0.000±.00 0.000±.00 6.000±0.00 6.000±0.00 Grok 4 Fast 0.836±.02 0.834±.02 0.000±.00 0.000±.00 6.000±0.00 6.667±0.47 o4-Mini 0.789±.01 0.805±.03 0.020±.03 0.011±.02 6.000±0.00 6.333±0.47 GLM 4.6V 0.300±.35 0.343±.33 0.305±.43 0.115±.09 2.333±2.62 29.000±25.66 GLM 4.5V 0.773±.03 0.789±.02 0.000±.00 0.019±.03 6.000±0.00 7.000±0.00 Qwen3-VL-235B 0.230±.14 0.243±.15 0.766±.17 0.731±.16 2.667±1.89 43.000±39.05 Hard GPT-5 0.815±.00 0.831±.02 0.000±.00 0.000±.00 8.000±0.00 6.667±0.47 Gemini 3 Pro 0.776±.06 0.803±.05 0.000±.00 0.007±.01 8.000±0.00 7.333±1.25 Gemini 2.5 Pro 0.806±.04 0.827±.01 0.000±.00 0.000±.00 8.000±0.00 7.000±0.00 Claude Sonnet 4 0.773±.06 0.805±.02 0.000±.00 0.007±.01 8.000±0.00 7.333±1.25 Grok 4.1 Fast 0.825±.03 0.835±.01 0.000±.00 0.009±.02 8.000±0.00 6.222±0.42 Grok 4 Fast 0.754±.04 0.768±.03 0.000±.00 0.007±.01 8.000±0.00 6.667±0.94 o4-Mini 0.792±.02 0.798±.02 0.000±.00 0.000±.00 8.000±0.00 6.667±0.47 GLM 4.6V 0.095±.12 0.175±.08 0.426±.41 0.214±.14 0.889±1.29 63.111±39.82 GLM 4.5V 0.309±.08 0.184±.09 0.000±.00 0.680±.04 3.000±1.63 6.667±0.47 Qwen3-VL-235B 0.233±.12 0.438±.12 0.475±.16 0.505±.11 3.333±1.25 9.667±4.50
+
+[p. 19 | section: E.3. Wisconsin Card Sorting Test (WCST) | type: TableGroup]
+Table 16. Wisconsin Card Sorting Test full results (image modality). All metrics are reported as mean with standard deviation (std) in a subscript. Difficulty Model Swcst Accuracy FMS PR Completed Rules 1st Rule Trials Easy GPT-5 0.824±.05 0.833±.03 0.000±.00 0.010±.01 6.000±0.00 6.000±0.00 Gemini 3 Pro 0.861±.00 0.857±.00 0.000±.00 0.000±.00 6.000±0.00 6.000±0.00 Gemini 2.5 Pro 0.823±.04 0.820±.04 0.000±.00 0.000±.00 6.000±0.00 7.000±0.82 Claude Sonnet 4 0.800±.02 0.797±.03 0.000±.00 0.010±.01 6.000±0.00 7.000±0.00 Grok 4.1 Fast 0.107±.07 0.432±.08 0.339±.24 0.324±.04 1.000±0.82 37.667±22.45 Grok 4 Fast 0.336±.08 0.602±.09 0.432±.10 0.201±.10 3.500±1.50 9.500±0.50 o4-Mini 0.347±.11 0.557±.03 0.320±.04 0.092±.05 3.333±0.94 7.000±0.82 GLM 4.6V 0.191±.11 0.297±.08 0.089±.13 0.442±.15 1.333±1.25 26.000±27.58 GLM 4.5V 0.108±.11 0.315±.22 0.404±.32 0.310±.21 0.333±0.47 25.667±13.22 Qwen3-VL-235B 0.420±.19 0.361±.09 0.229±.22 0.619±.09 4.000±0.82 7.333±1.89 Hard GPT-5 0.794±.05 0.819±.00 0.000±.00 0.000±.00 8.000±0.00 7.000±0.82 Gemini 3 Pro 0.806±.02 0.814±.04 0.000±.00 0.000±.00 8.000±0.00 6.667±0.47 Gemini 2.5 Pro 0.776±.01 0.795±.03 0.000±.00 0.000±.00 8.000±0.00 7.333±0.47 Claude Sonnet 4 0.710±.08 0.739±.06 0.134±.12 0.069±.06 8.000±0.00 6.667±0.47 Grok 4.1 Fast 0.077±.05 0.369±.09 0.376±.26 0.420±.07 0.714±0.70 44.143±20.58 Grok 4 Fast 0.209±.07 0.493±.04 0.347±.25 0.309±.03 3.000±1.41 11.333±2.36 o4-Mini 0.277±.13 0.528±.05 0.437±.09 0.352±.07 3.889±1.37 9.222±3.58 GLM 4.6V 0.064±.07 0.197±.09 0.169±.32 0.420±.12 0.444±0.96 64.556±26.88 GLM 4.5V 0.083±.07 0.398±.22 0.303±.25 0.432±.11 0.333±0.47 32.000±4.97 Qwen3-VL-235B 0.071±.07 0.208±.13 0.500±.38 0.794±.14 1.333±1.25 59.333±38.77
+
+[p. 20 | section: F. Prompts | type: Text]
+Below we list the system prompts used for each task. We also provide alternate or additional prompts used for the different setups within each task.
+
+[p. 20 | section: System Prompt for Image RAPM | type: Text]
+You are taking the Raven's Progressive Matrices (RAPM) test, a non-verbal intelligence test that measures abstract reasoning ability.
+
+[p. 20 | section: System Prompt for Image RAPM | type: Text]
+You will see a 3x3 matrix of images with the bottom-right image missing (shown as a question mark), followed by 8 answer choices numbered 1-8.
+
+[p. 20 | section: System Prompt for Image RAPM | type: Text]
+Your task is to:
+
+[p. 20 | section: System Prompt for Image RAPM | type: ListGroup]
+1. Analyze rows and columns 2. Infer the governing logical rule(s) 3. Select the answer choice (1-8) that correctly completes the matrix.
+
+[p. 20 | section: /* PATTERN HINT */ | type: Text]
+The patterns can involve:
+
+[p. 20 | section: /* PATTERN HINT */ | type: ListGroup]
+Shape transformations (rotation, reflection, scaling) 1062 - Position changes (movement, arrangement) 1063 - Attribute changes (color, size, number of elements) 1064 - Logical operations (addition, subtraction, intersection) Sequence progressions (systematic changes across rows/columns) 1065 Additional common rule types: 1066 1067 - Constant-in-row: Same value across a row; varies down columns. Quantitative step: Fixed +/- increment between adjacent cells (size / count / position offset). Figure add/subtract: Combine (overlay or juxtapose) or remove elements from two cells to form the third. Distribution-of-three: Three distinct categorical values appear once each per row (order may permute). Distribution-of-two: Two values each appear once; third slot is empty / null. Look horizontally and vertically; the missing piece must satisfy ALL relevant row and column rules.
+
+[p. 20 | section: /* PATTERN HINT */ | type: Text]
+Your final answer should be a number between 1-8 corresponding to the correct choice.
+
+[p. 20 | section: System Prompt for Text RAPM | type: Text]
+You are solving a TEXT-BASED 3x3 pattern matrix (Raven-style). Each cell contains a string; the bottom-right cell is missing ('?').
+
+[p. 20 | section: System Prompt for Text RAPM | type: Text]
+Goal: Infer the rule(s) acting across rows and columns.
+
+[p. 20 | section: /* PATTERN HINT */ | type: Text]
+Possible dimensions (one or more):
+
+[p. 20 | section: /* PATTERN HINT */ | type: ListGroup]
+Character set restriction (digits / letters / symbols) Quantitative constant (exact length / count / unique) Quantitative progression (arithmetic step across row/column) Parity / multiple rules (all even / all odd / multiples of N) Positional constraints (first/last/even/odd positions restricted) Ordering (ascending / descending / mixed) Layered combinations (e.g. constant + parity, progression + positional)
+
+[p. 20 | section: 1089 1090 /* MC VERSION */ | type: Text]
+You will be given 8 answer options (1-8). Select the single option that correctly fills the missing cell while satisfying ALL inferred row and column constraints.
+
+[p. 20 | section: 1089 1090 /* MC VERSION */ | type: Text]
+1092 Respond with <answer>NUMBER</answer> using just the chosen option number.
+
+[p. 20 | section: 1093 1094 /* GEN VERSION */ | type: Text]
+You must GENERATE the exact missing cell string that satisfies ALL inferred row and column constraints.
+
+[p. 20 | section: 1093 1094 /* GEN VERSION */ | type: Text]
+Respond with <answer>STRING</answer> containing only the candidate string (no quotes or extra text).
+
+[p. 22 | section: G. Turn Examples | type: Text]
+[. . . omitted turns . . . ]
+
+[p. 22 | section: G. Turn Examples | type: Text]
+User: No tokens found in box (3, 3). Answer concisely. Think step-by-step, utilizing information from previous feedbacks, and state your reasoning in maximum 4000 tokens, wrapped with <think> and </think>. Then, provide a really short summary of your reasoning after the closing </think> tag.
+
+[p. 22 | section: G. Turn Examples | type: Text]
+Which of the 8 boxes would you like to open?
+
+[p. 22 | section: G. Turn Examples | type: Text]
+Your final answer should be a grid coordinate (x, y), wrapped with <answer> and </answer>
+
+[p. 22 | section: G. Turn Examples | type: Text]
+Assistant: <think>I already found one, located at position (1, 0). [. . . ] I know that boxes (5, 0), (5, 1), (1, 2), and (3, 3) are empty. [. . . ] That leaves me with (10, 4), (0, 6), and (9, 6). To keep things organized and ensure I'm systematic, I'm going to follow my pre-defined search pattern – working from the top-left and moving to the right and down. That means the next box I need to check is (10, 4). [. . . ] <answer>(10, 4)</answer>
+
+[p. 22 | section: G. Turn Examples | type: Text]
+[. . . omitted turns . . . ]
+
+[p. 22 | section: G. Turn Examples | type: Text]
+User: No tokens found in box 1. Boxes that has contained token A: 11, Boxes that has contained token B: 8, 10,
+
+[p. 22 | section: G. Turn Examples | type: Text]
+Opened boxes: 12,
+
+[p. 22 | section: G. Turn Examples | type: Text]
+Answer concisely. Think step-by-step, utilizing information from previous feedbacks, and state your reasoning in maximum 4000 tokens, wrapped with <think> and </think>. Then, provide a really short summary of your reasoning after the closing </think> tag.
+
+[p. 22 | section: G. Turn Examples | type: Text]
+Which of the 12 boxes would you like to open?
+
+[p. 22 | section: G. Turn Examples | type: Text]
+Your final answer should be a box number, wrapped with <answer> and </answer>
+
+[p. 22 | section: G. Turn Examples | type: Text]
+Assistant: [. . . ]The message provides data on boxes: 11 contained token A, 8 and 10 contained token B, and 12 boxes were opened. [. . . ] The phrase "Opened boxes" suggests the total number of boxes accessed, with partial information on token contents.[. . . ] The phrase "Opened boxes: 12" likely means box 12 was last opened and found empty, given prior context.[. . . ]
+
+[p. 22 | section: G. Turn Examples | type: Text]
+**Summary:** 1 empty again; pure empties left 2-7,9,12; systematic next=2.
+
+[p. 22 | section: G. Turn Examples | type: Text]
+<answer>2</answer>
+
+[p. 23 | section: G. Turn Examples | type: Text]
+[. . . omitted turns . . . ]
+
+[p. 23 | section: G. Turn Examples | type: Text]
+User: Incorrect. Please try again.
+
+[p. 23 | section: G. Turn Examples | type: Text]
+Recent notes:
+
+[p. 23 | section: G. Turn Examples | type: ListGroup]
+Turn -1: matching number of symbols and symbol color Incorrect. Turn -2: matching number of symbols and symbol color Incorrect. Turn -3: matching symbol color Correct. Turn -4: matching symbol color Correct. Turn -5: matching symbol color Correct. Turn -6: matching symbol color Correct.
+
+[p. 23 | section: G. Turn Examples | type: Text]
+Look at the image showing 5 cards. Match the 'Given' card to one of cards 1-4 based on the rule you need to figure out.
+
+[p. 23 | section: G. Turn Examples | type: Text]
+Assistant: The previous attempts to match by symbol color (Turns 2-6) were correct until Turn 7, where the response was incorrect. [. . . ] However, the last incorrect attempt used "number of symbols and symbol color," which was wrong. The new rule might now be **background color**.[. . . ] Re-evaluating: The prior correct streak (Turns 2-6) used symbol color. After an incorrect response, the rule often shifts to **number of symbols**. [. . . ] Wait, the feedback says "matching number of symbols and symbol color – Incorrect" for Turn 7. The correct rule might now be **background color**. [. . . ]The rule shifted to background color; Given card has a purple background, matching Card 3.
+
+[p. 23 | section: G. Turn Examples | type: Text]
+<answer>3</answer>
+
+[p. 23 | section: G. Turn Examples | type: Text]
+[. . . omitted turns . . . ]
+
+[p. 23 | section: G. Turn Examples | type: Text]
+User: Incorrect. Please try again.
+
+[p. 23 | section: G. Turn Examples | type: Text]
+Given: two red triangle
+
+[p. 23 | section: G. Turn Examples | type: Text]
+Options:
+
+[p. 23 | section: G. Turn Examples | type: ListGroup]
+1. two green triangle 2. four yellow square 3. three blue star 4. one red circle
+
+[p. 23 | section: G. Turn Examples | type: Text]
+Assistant: [. . . ] My initial hypothesis (number and shape) was incorrect. The feedback suggests a different rule. The given card is red. Option 4 is also red, while others have different colors. I'll now hypothesize the rule is matching by color regardless of number or shape.<answer>4</answer>My initial hypothesis (number and shape) was incorrect. The feedback suggests a different rule. The given card is red. Option 4 is also red, while others have different colors. I'll now hypothesize the rule is matching by color regardless of number or shape.

icml26/a4461009-05b7-42b6-b207-5e6e0c2e0731/paper.md

@@ -0,0 +1,1140 @@
+
+
+{0}------------------------------------------------
+
+027
+
+035
+
+# A Neuropsychologically Grounded Evaluation of LLM Cognitive Abilities
+
+## Anonymous Authors<sup>1</sup>
+
+## Abstract
+
+Large language models (LLMs) display a unified "general factor" of capability across 10 benchmarks—a finding confirmed by our factor analysis of 156 models—yet they still struggle with simple, trivial tasks for humans. This is because current benchmarks focus on task completion, failing to probe the foundational cognitive abilities that highlight these behaviors. We address this by introducing the NeuroCognition benchmark, grounded in three adapted neuropsychological tests: Raven's Progressive Matrices (abstract relational reasoning), Spatial Working Memory (maintenance and systematic search), and the Wisconsin Card Sorting Test (cognitive flexibility). Our evaluation reveals that while models perform strongly on text, their performance degrades for images and with increased complexity. Furthermore, we observe that complex reasoning is not universally beneficial, whereas simple, humanlike strategies yield partial gains. We also find that NeuroCognition correlates positively with standard general-capability benchmarks, while still measuring distinct cognitive abilities beyond them. Overall, NeuroCognition emphasizes where current LLMs align with human-like intelligence and where they lack core adaptive cognition, showing the potential to serve as a verifiable, scalable source for improving LLMs.
+
+## 1. Introduction
+
+Large language models (LLMs) show increasingly powerful capabilities (?), further enhanced by the emergence of reasoning abilities (?). This is demonstrated by their beyond human-level capabilities in tasks such as olympiad-level problems (?). However, even with these impressive capabilities, they still stumble on trivial or straightforward tasks
+
+Preliminary work. Under review by the International Conference on Machine Learning (ICML). Do not distribute.
+
+(??). This raises a curiosity about the underlying cognitive mechanism of LLMs and how it compares to human cognitive abilities.
+
+Most current LLM benchmarks are task-specific, limiting their analysis and insights to the task performance itself. These benchmarks are not able to uncover the underlying cognitive abilities required to solve the task (??) because task performance is not a reliable measure of intelligence, but rather heavily influenced by prior knowledge and experience (?). One interesting test that is often overlooked is neuropsychological tests, which are a well-established method for measuring specific cognitive abilities in humans, originally designed to isolate specific cognitive functions using controlled stimuli and interpretable outcomes. Previous works have attempted to apply neuropsychological tests to LLMs (??), but they are limited to a single task or modality, or they lack a deeper analysis of the model's underlying cognitive abilities.
+
+We present the NeuroCognition benchmark, built from three well-established neuropsychological tests: Raven's Progressive Matrices (????) (RPM) targeting abstract relational reasoning (fluid intelligence), Spatial Working Memory (?) (SWM) targeting maintenance and systematic search across working memory, and the Wisconsin Card Sorting Test (??) (WCST) targeting cognitive flexibility and task switching. Each test is adapted into multiple text and image variants (including harder versions) to vary interaction length, modality, and complexity. We also include experiments using assistance inspired by simple strategies that humans use (e.g., taking notes), a process known as cognitive offloading (??), to see if they reduce observed errors and improve overall performance.
+
+Our main contributions are:
+
+- NeuroCognition, a benchmark repurposed from wellestablished neuropsychological tests, into a simple, multimodal, scalable, robust assessment, and easily reproducible suite with various setups.
+- Evaluation results show that state-of-the-art models perform strongly in baseline text settings, but struggle in image or when task complexity increases, and that explicit reasoning is not always beneficial.
+- Experiments with human-like strategies (pattern
+
+<sup>1</sup>Anonymous Institution, Anonymous City, Anonymous Region, Anonymous Country. Correspondence to: Anonymous Author <anon.email@domain.com>.
+
+{1}------------------------------------------------
+
+<span id="page-1-1"></span>![](_page_1_Figure_1.jpeg)
+
+Figure 1. The three neuropsychological tests used in our work.
+
+hints and notes) show mixed effects—limited on RAPM/SWM but more consistent on WCST—suggesting some deficits are fundamental.
+
+ Factor analysis and correlation results show that NeuroCognition measures specific cognitive abilities that differ from, yet correlate with, the general capabilities captured by standard benchmarks.
+
+Overall, NeuroCognition is a benchmark that can identify gaps in LLMs' cognitive abilities relative to humans, offering insights into how LLMs operate and informing the design and training of future LLMs. Our code<sup>1</sup>, data, and results will be made publicly available.
+
+#### 2. Related Works
+
+The neuropsychological tests included in our work measure different aspects of human cognitive abilities. The RPM (????) measures abstract, relational reasoning capabilities (?). RPM has been used to identify intellectually gifted individuals, and has been shown to be widely reliable across different geographical boundaries (?). The SWM test (?), on the other hand, measures the working memory, the ability to hold and mentally work with information, which may no longer be perceptually present (?). This test has been used to detect at-risk individuals for psychosis (?), and highlight the working memory deficits in individuals with autism (?). Finally, WCST (??) measures cognitive flexibility, i.e., the
+
+ability to adapt to a change in perspective and reallocate resources to process relevant information (??). Patients with schizophrenia are shown to perform worse on this test (?), and are shown to be robust even in clinical settings (?).
+
+Recent advances in LLMs have led many works seeking to establish stronger correlations between LLMs and human intelligence. Many works focus on developing a cognitive benchmark aiming to measure fluid intelligence (????), emphasizing abstractness and dynamic evaluation to avoid the data contamination issue (?). Other works applied human cognitive tests to LLMs (?) and showcased cognitive behaviors in LLMs, such as similar working memory capacity between humans and GPT-4 in the n-back test (?), and above human-level cognitive flexibility in the WCST (?). Some works also revealed common factors of models' performance across different benchmarks (??), indicating the existence of a common underlying intelligence.
+
+Aside from cognitive abilities, several studies have also quantified specific areas where LLMs struggle. In multi-turn evaluation, LLMs tend to perform worse than single-turn evaluations (?), and still perform sub-optimally in multi-turn tasks (??). LLMs also struggle in longer multi-step reasoning (?) and in complex multi-stage tasks (?). LLMs also exhibit cross-modal hallucination (?), and struggle in multi-image, multi-turn benchmarks (??). LLMs also fail in simple pattern tasks using modified alphabets (?), indicating the lack of general analogical reasoning in LLMs.
+
+<span id="page-1-0"></span><sup>&</sup>lt;sup>1</sup>Code available at https://github.com/reggans/CognitiveEval
+
+{2}------------------------------------------------
+
+## 3. NeuroCognition Benchmark Tests
+
+#### 3.1. Raven's Progressive Matrices
+
+153 154
+
+Task. RPM presents a 3x3 matrix of construct patterns with the bottom-right cell missing. The participant needs to infer abstract relations across rows and columns, and fill the missing cell by either selecting one of the 8 available options or constructing it. In our benchmark, we use the Raven's Advanced Progressive Matrices (RAPM) (?) form.
+
+Setup. For the image version, we use the RAVEN dataset (?), implementation of RAPM with 7 different figure configurations licensed with GPL-3.0. We take 20 items per configuration, resulting in a total of 140 test data. We combine the 3x3 question matrix with the 8 available options into a single image input, as shown in Figure [1,](#page-1-1) and present it to the model in multiple-choice (MC) format. For the text setting, we programmatically generate symbolic matrices using character and string attributes (see Appendix [A](#page-9-0) for generation details). We provide two answer formats for text setting: multiple-choice (MC, same 8 options) and generative (Gen, free-form string). A stored rule representation enables automatic validation of Gen outputs (see Appendix [B](#page-11-0) for validation details). We generate 200 test data for the text setting with random configurations.
+
+Metrics. We report accuracy for both MC and Gen. Average human performance for images is 0.814, reported by ?, and not yet established for the text variant.
+
+#### <span id="page-2-0"></span>3.2. Spatial Working Memory
+
+Task. In the SWM test, participants are presented with an image with several boxes. A token is hidden in one of the boxes, and participants are asked to search for it by elimination, opening one box at a time. After the token is found, it is regenerated in another box that has not previously contained it in the current trial. This is repeated until all boxes have contained all tokens, with the goal of locating all tokens with minimal redundant checking and without revisiting boxes unnecessarily.
+
+Setup. We implement two difficulty levels. Easy: 8 boxes, 1 token active at a time, early stop after all 8 tokens found or a maximum of 64-guess cap. Hard: 12 boxes with 2 distinct tokens active concurrently (24 tokens found required, maximum 144 guesses), with each token directly reappearing after being found, and one box can contain both tokens at once. We experiment with three types of feedback setups: text-only, image-only (a grid with no additional textual information), and image+text variants. The image input uses a grid system for easier answer verification, as shown in Figure [1.](#page-1-1) The text-only setup identifies the boxes by number and ask the model to choose a box number.
+
+Metrics. Since SWM is highly affected by luck, we define
+
+some metrics that focus on penalizing errors and unfinished tests. We classify three avoidable error types:
+
+- illegal: opens a box that can no longer contain a token because it has already yielded all available tokens for the trial.
+- no-box: refers to a box index/identifier that does not exist.
+- repeated: re-opens a box that was previously opened and found empty, before any new token was regenerated (i.e., no state change that could justify revisiting).
+
+These three error types sum to nerr. Let nvalid be the total valid guesses. A consistency term Sc:
+
+$$S_c = 1 - \frac{n_{err}}{n_{valid}} \tag{1}$$
+
+measures avoidance of structurally unjustified moves. Note that we omit *invalid* guesses (no parseable answer) to ensure fair evaluation of real guessing attempts. Let T<sup>f</sup> be number of tokens found out of T required; the final score Sswm is
+
+$$S_{swm} = \frac{T_f}{T} \times S_c \tag{2}$$
+
+Human worst-case. A cognitively normal human, given a complete history of their own moves (like conversation history in LLMs), will be able to achieve full completion by systematic elimination and yield nerr = 0 regardless of the setup and difficulty, therefore obtaining Sswm = 1.
+
+#### <span id="page-2-1"></span>3.3. Wisconsin Card Sorting Test
+
+Task. In WCST, four option cards containing shapes are shown to participants. They are then asked to match the given card to one of the four option cards using one of three rules: shape, color, or number. Importantly, participants are not informed of the correct rule for matching the cards and must eliminate potential rules to find the correct one. After choosing the correct answer, a different card is shown while maintaining the same rule. The matching rule is considered complete after a fixed number of consecutive correct answers and is changed upon completion without expressly informing the participant. The test finishes after a fixed number of rules are completed or a maximum number of guesses is reached.
+
+Setup. We require 5 consecutive correct matches (c = 5) to complete a rule instance. We define two difficulty levels. Easy: 3 candidate rules (color, shape, number), each repeated twice (6 total rule instances) or until the 64-guess cap. Hard: adds background color (4 candidate rules, 8 total rule instances) with a maximum of 96 guesses. We also construct controlled ambiguity variants for the hard setting: (i) *off* (no ambiguous cards), (ii) *first* (only the first card
+
+{3}------------------------------------------------
+
+under a new rule is ambiguous), and (iii) *rest* (all subsequent cards after the first are ambiguous). Ambiguity occurs when multiple attributes match simultaneously, forcing the model to track disambiguating feedback over turns. The image-input setting provides images of the cards (Figure 1), while the text-only setting lists descriptions of the option cards.
+
+**Metrics.** Each completed rule i uses  $g_i$  guesses from the start of that rule until c consecutive correct response. Therefore, we can define a score  $S_{r_i}$  for each completed rule i:
+
+$$S_{r_i} = \frac{c}{q_i}. (3)$$
+
+where unfinished rules contribute zero. Lower  $g_i$  means faster hypothesis convergence, which repeated reliance on an incorrect hypothesis inflates  $g_i$  and lowers  $S_{r_i}$ . With  $n_c$  completed out of N target rule instances, we can calculate the final score  $S_{west}$  as:
+
+$$S_{wcst} = \frac{1}{N} \sum_{i=1}^{n_c} S_{r_i}.$$
+ (4)
+
+We additionally report two specific error rates aggregated over a single trial with total turns L. Let  $y_t \in \{0,1\}$  denote correctness at turn t ( $y_t = 1$  if the model's match is correct), and let  $r_t \in \mathcal{R}$  be the rule the model is currently following at turn t (e.g., color/shape/number/background).
+
+**Perseverative Response (PR).** PR measures how often the model re-selects a rule that has already been ruled out by negative feedback within the current rule block. We use PR rather than the standard WCST perseverative error (PE) because it captures a broader range of errors while still including PE.
+
+Let k(t) index the current (ground-truth) rule block containing turn t, and let  $\mathcal{R}^t_{\neg} \subseteq \mathcal{R}$  be the set of rules eliminated by feedback from earlier turns in the same block. Then:
+
+$$PR = \frac{\sum_{t=1}^{L} \mathbf{1} \{ \mathcal{R}_{\neg}^{t} \neq \emptyset \} \mathbf{1} \{ r_{t} \in \mathcal{R}_{\neg}^{t} \}}{\sum_{t=1}^{L} \mathbf{1} \{ \mathcal{R}_{\neg}^{t} \neq \emptyset \}}.$$
+ (5)
+
+**Failure-to-maintain-set (FMS).** We mark a rule as *conceptually acquired* after m=3 consecutive correct responses within a block. For each block k, let  $\tau_k$  be the first turn index in that block at which this acquisition criterion is met (if it is never met, the block contributes no turns). Let  $\mathcal{T}_{post}$  be the set of all turns after acquisition and before the next rule switch:
+
+$$\mathcal{T}_{\text{post}} = \bigcup_{k} \{ t : k(t) = k, \ t > \tau_k \}. \tag{6}$$
+
+FMS is the error rate on these post-acquisition turns:
+
+$$FMS = \frac{\sum_{t \in \mathcal{T}_{post}} \mathbf{1}\{y_t = 0\}}{|\mathcal{T}_{post}|}.$$
+ (7)
+
+**Human worst-case.** For the easy setup, a cognitively normal human needs at most two exploratory wrong guesses to find the correct rule, then continues with 5 correct answers  $\Rightarrow 5/7 \approx 0.714$ . For hard, the worst case happened with the *first* ambiguity setup, where at most there are two exploratory wrong guesses, one ambiguous early correct, one disambiguating wrong, and 5 consecutive correct answers  $\Rightarrow 5/9 \approx 0.556$ .
+
+#### <span id="page-3-0"></span>3.4. Experiment Setup
+
+In our experiments, we benchmark several recent multimodal reasoning models: proprietary models (Gemini 3 Pro (?), Gemini 2.5 Pro (?), GPT-5 (?), o4-Mini (?), Claude Sonnet 4 (?), Grok 4.1 Fast (?), Grok 4 Fast (?) and openweight models (GLM 4.6V (?), GLM 4.5V (?), Qwen3-VL-235B (????)).
+
+We use models' default reasoning capabilities via API parameters and chain-of-thought prompts. However, we disable reasoning for Claude Sonnet 4 and Grok 4 Fast on the RAPM test, due to overthinking issues that exhausted output limits and led to worse performance. For RAPM, we set thinking budgets close to the model's maximum output tokens. For SWM and WCST, we limit the output tokens to 8,192 tokens and the thinking budget to 4,000 tokens per turn. We repeat all SWM and WCST experiments three times, while RAPM is evaluated once. All evaluations are zero-shot with a temperature of 0 (if applicable). Further details are provided in Appendix C.
+
+#### 4. Results
+
+Table 1 reports the baseline results for NeuroCognition, with the complete results available in Appendix E.
+
+#### 4.1. Key Findings
+
+Models are stronger in text, struggle in image. Across all tests, models show a consistent advantage in text-only setups compared to image-based inputs. Even combined modalities (image+text) yield worse results than text alone, indicating difficulties in parsing and reasoning over visual stimuli and combining visual with textual information.
+
+**Performance drops with difficulty.** Increasing task difficulty, particularly in SWM, causes a significant drop in performance. In the SWM text easy setup almost all models complete the task perfectly, while in the hard setting only three models manage to do so. Even the two best models are unable to do it consistently through all setups. This is interesting because the core ability required is still the same. The harder setting only increases demands on maintaining and updating internal state, proven to still be a challenge.
+
+Models possess good cognitive flexibility. Compared to
+
+{4}------------------------------------------------
+
+274
+
+<span id="page-4-0"></span>*Table 1.* Baseline NeuroCognition benchmark performance without hints, notes, or pattern hints. RAPM uses the reported human average performance from ?. SWM & WCST compares against a healthy human worst-case score. Full setup details for each Easy/Hard setup are provided in each test section. MC and Gen denote multiple-choice and generation setup, respectively. All results are zero shot and use each model's default reasoning level, except where marked with <sup>∗</sup> indicating reasoning was disabled due to some issues resulting in worse performance. Bold numbers and green cells denote scores that match or surpass human performance. ↑ means higher is better.
+
+|                 | RAPM (Accuracy ↑) |        |        |       |       | SWM (Sswm  | ↑)    |       |       | WCST (Swcst<br>↑) |       |       |       |
+|-----------------|-------------------|--------|--------|-------|-------|------------|-------|-------|-------|-------------------|-------|-------|-------|
+| Model           | Text<br>Image     |        |        | Text  |       | Image+Text |       |       | Image |                   | Text  | Image |       |
+|                 | MC                | Gen    | MC     | Easy  | Hard  | Easy       | Hard  | Easy  | Hard  | Easy              | Hard  | Easy  | Hard  |
+| GPT-5           | 0.540             | 0.605  | 0.279  | 1.000 | 1.000 | 0.985      | 0.989 | 0.927 | 0.631 | 0.848             | 0.815 | 0.824 | 0.794 |
+| Gemini 3 Pro    | 0.585             | 0.460  | 0.400  | 1.000 | 1.000 | 1.000      | 0.963 | 1.000 | 0.793 | 0.832             | 0.776 | 0.861 | 0.805 |
+| Gemini 2.5 Pro  | 0.525             | 0.560  | 0.186  | 1.000 | 0.980 | 0.733      | 0.480 | 0.408 | 0.622 | 0.825             | 0.806 | 0.823 | 0.776 |
+| Claude Sonnet 4 | 0.560∗            | 0.385∗ | 0.207∗ | 1.000 | 0.866 | 0.256      | 0.265 | 0.630 | 0.005 | 0.764             | 0.773 | 0.800 | 0.710 |
+| Grok 4.1 Fast   | 0.505             | 0.495  | 0.100  | 1.000 | 0.983 | 0.145      | 0.017 | 0.014 | 0.009 | 0.841             | 0.825 | 0.107 | 0.060 |
+| Grok 4 Fast     | 0.690∗            | 0.555∗ | 0.086∗ | 1.000 | 0.994 | 0.021      | 0.002 | 0.029 | 0.005 | 0.836             | 0.754 | 0.336 | 0.209 |
+| o4-Mini         | 0.450             | 0.410  | 0.207  | 0.951 | 0.811 | 0.055      | 0.032 | 0.042 | 0.005 | 0.789             | 0.792 | 0.347 | 0.277 |
+| GLM 4.6V        | 0.320             | 0.170  | 0.357  | 1.000 | 0.574 | 0.044      | 0.010 | 0.000 | 0.028 | 0.300             | 0.095 | 0.191 | 0.064 |
+| GLM 4.5V        | 0.035             | 0.000  | 0.128  | 1.000 | 0.536 | 0.044      | 0.000 | 0.000 | 0.000 | 0.773             | 0.309 | 0.108 | 0.083 |
+| Qwen3-VL-235B   | 0.485             | 0.570  | 0.493  | 1.000 | 0.760 | 0.026      | 0.016 | 0.031 | 0.135 | 0.230             | 0.233 | 0.420 | 0.071 |
+| Human           | -                 | -      | 0.814  | 1.000 | 1.000 | 1.000      | 1.000 | 1.000 | 1.000 | 0.714             | 0.556 | 0.714 | 0.556 |
+
+RAPM and SWM, models generally perform better on WCST, with more models passing the human worst-case baseline. This suggests current LLMs handle rule inference and task-switching behavior more robustly than abstract relational reasoning or complex working-memory maintenance, though image setup is still an issue for some models.
+
+Answer format matters. In RAPM text, models exhibit different preferences for multiple-choice versus generative answer formats on the same items. This likely reflects differences in decision constraints and how models interpret the problem, which sometimes leads to overcomplication.
+
+#### 4.2. Failure Cases & Patterns
+
+RAPM. Our analysis of RAPM reasoning traces (Appendix [E](#page-15-0) Table [10,](#page-15-1) [11\)](#page-15-2) reveals three persistent deficits across modalities. First, hallucination and mis-grounding: models fabricate visual features or impose non-existent numerical patterns on text that contradict the data. Second, semantic over-interpretation: models overlook simple local rules (XOR, concatenation) in favor of complex, global hypotheses like "symmetry grids." Third, specifically in Text Gen, models exhibit content-logic disconnect, where they frequently derive correct attribute constraints but generate hallucinated strings that satisfy these metadata rules while violating the actual constructive logic. These failures suggest a disconnect between low-level feature verification and abstract reasoning.
+
+SWM. We investigate failures using four error types explained earlier in Section [3.2.](#page-2-0) Since models struggle most in the image modality, we focus on image and image-text
+
+<span id="page-4-1"></span>*Table 2.* SWM image and image+text failure turns averaged across all runs (easy and hard). ↓ means lower is better.
+
+| Model           | Illegal ↓ | Repeat ↓ | No-Box ↓ | Invalid ↓ |
+|-----------------|-----------|----------|----------|-----------|
+| GPT-5           | 2.75      | 4.25     | 3.42     | 0.50      |
+| Gemini 3 Pro    | 5.67      | 1.17     | 0.00     | 0.58      |
+| Gemini 2.5 Pro  | 21.75     | 3.08     | 13.83    | 0.33      |
+| Claude Sonnet 4 | 1.42      | 4.17     | 47.08    | 18.42     |
+| Grok 4.1 Fast   | 0.08      | 4.42     | 78.00    | 9.00      |
+| Grok 4 Fast     | 0.00      | 6.00     | 68.92    | 21.58     |
+| o4-Mini         | 0.00      | 19.25    | 69.75    | 1.25      |
+| GLM 4.6V        | 0.50      | 4.92     | 9.67     | 82.50     |
+| GLM 4.5V        | 0.00      | 1.92     | 28.00    | 71.58     |
+| Qwen3-VL-235B   | 4.58      | 22.75    | 40.67    | 21.33     |
+
+failure cases (Table [2\)](#page-4-1). From the data, we identify two main weaknesses, state tracking under sparse feedback (models forget key task state) and task execution (invalid and nobox actions instead of exploiting eliminations). High illegal and repeat counts indicate confusion and working-memory failures, high no-box rates suggest a visual grounding gap (difficulty identifying a valid box), and invalid turns reflect failures to follow the task instruction.
+
+WCST. Table [3](#page-5-0) summarizes WCST image error rates using FMS and PR as explained in Section [3.3.](#page-2-1) We focus on weaker models that struggle in WCST image. Overall, most errors happen after the model has found the right rule. Models either (i) fail to keep using the acquired rule (higher FMS) or (ii) fail to change their choice after negative feedback (higher PR), i.e., they choose a rule that was already ruled out within the current rule block. The background-
+
+{5}------------------------------------------------
+
+313 314
+
+316
+
+324
+
+326
+
+328 329
+
+<span id="page-5-0"></span>*Table 3.* WCST error rates for models that struggle in WCST image tests. ↓ means lower is better.
+
+| Model         | Easy  |       | Hard  |       |  |
+|---------------|-------|-------|-------|-------|--|
+|               | FMS ↓ | PR ↓  | FMS ↓ | PR ↓  |  |
+| Grok 4.1 Fast | 0.339 | 0.324 | 0.376 | 0.420 |  |
+| Grok 4 Fast   | 0.432 | 0.201 | 0.347 | 0.309 |  |
+| o4-Mini       | 0.320 | 0.092 | 0.437 | 0.352 |  |
+| GLM 4.6V      | 0.089 | 0.442 | 0.169 | 0.420 |  |
+| GLM 4.5V      | 0.404 | 0.310 | 0.303 | 0.432 |  |
+| Qwen3-VL-235B | 0.229 | 0.619 | 0.500 | 0.794 |  |
+
+<span id="page-5-1"></span>*Table 4.* RAPM results with no reasoning. ↑ means higher is better. Paired t-test results show effects in Text MC is statistically significant (p < .05) while others are not (p > .05).
+
+| Model          | RAPM (Accuracy ↑) |             |             |  |  |  |  |
+|----------------|-------------------|-------------|-------------|--|--|--|--|
+|                | Text MC           | Text Gen    | Image MC    |  |  |  |  |
+| Gemini 2.5 Pro | 0.690∆+0.17       | 0.450∆−0.11 | 0.093∆−0.09 |  |  |  |  |
+| Grok 4.1 Fast  | 0.690∆+0.19       | 0.565∆+0.07 | 0.114∆+0.01 |  |  |  |  |
+| GLM 4.6V       | 0.635∆+0.32       | 0.175∆+0.01 | 0.336∆−0.02 |  |  |  |  |
+| GLM 4.5V       | 0.460∆+0.43       | 0.000∆0.00  | 0.014∆−0.11 |  |  |  |  |
+
+color addition tends to make both problems worse. We provide more details and other error metrics in Appendix [E.](#page-15-0)
+
+## 5. Ablation Study
+
+Following the baseline results, we explore aspects that could affect model performance. We present the change in performance compared to Table [1](#page-4-0) baseline using a subscripted delta (∆) notation to indicate any improvements or declines.
+
+#### <span id="page-5-3"></span>5.1. RAPM: Reasoning vs No Reasoning
+
+We compare RAPM performance when models run with their default reasoning behavior versus a constrained noreasoning setting, as shown in Table [4.](#page-5-1) This experiment was run only on models where reasoning can be disabled (excluding Claude Sonnet 4 and Grok 4 Fast because of reasoning issue). Disabling reasoning produces large, consistent gains in Text MC (p = .0207) across models, while Text Gen and Image MC shows mixed effects. The improvement in Text MC likely occurs because the decision space is explicit and limited, and avoiding extended chain-of-thought reduces overcomplication and distractor attraction. For Text Gen, using reasoning or not shows no clear effect (p = .83). For Image MC, disabling reasoning appears slightly worse (p = .173), likely because reasoning helps interpret visual inputs and detect patterns, though the effect may be small.
+
+<span id="page-5-2"></span>*Table 5.* RAPM results with pattern hints. ↑ means higher is better. Paired t-test results show no significant effect on all setups (p > .05).
+
+| Model           |             | RAPM (Accuracy ↑) |             |
+|-----------------|-------------|-------------------|-------------|
+|                 | Text MC     | Text Gen          | Image MC    |
+| GPT-5           | 0.620∆+0.08 | 0.720∆+0.12       | 0.257∆−0.02 |
+| Gemini 3 Pro    | 0.595∆+0.01 | 0.545∆+0.09       | 0.386∆−0.01 |
+| Gemini 2.5 Pro  | 0.485∆−0.04 | 0.585∆+0.03       | 0.221∆+0.04 |
+| Claude Sonnet 4 | 0.525∆−0.03 | 0.360∆−0.02       | 0.221∆+0.01 |
+| Grok 4.1 Fast   | 0.570∆+0.07 | 0.580∆+0.09       | 0.114∆+0.01 |
+| Grok 4 Fast     | 0.655∆−0.04 | 0.540∆−0.02       | 0.100∆+0.01 |
+| o4-Mini         | 0.435∆−0.02 | 0.525∆+0.12       | 0.236∆+0.03 |
+| GLM 4.6V        | 0.395∆+0.08 | 0.235∆+0.07       | 0.321∆−0.03 |
+| GLM 4.5V        | 0.085∆+0.05 | 0.000∆0.00        | 0.090∆−0.04 |
+| Qwen3-VL-235B   | 0.575∆+0.09 | 0.590∆+0.02       | 0.457∆−0.04 |
+
+#### 5.2. RAPM: Pattern Hints
+
+In RAPM, we test whether providing hints regarding possible underlying patterns helps models narrow down their reasoning and identify the correct answer. The results shown in Table [5](#page-5-2) vary across models and setups, and we find no significant effect (paired t-tests: Text MC p = .177, Text Gen p = .45, Image MC p = .666). Most results are similar to those obtained without using any patterns, and in some cases, they even decrease. Overall, pattern hints have a limited and inconsistent impact. Most changes are small (often positive), and the aggregate effect is not statistically significant, suggesting that many models still lack core abstract reasoning capabilities, even with these hints.
+
+#### 5.3. SWM & WCST: Note-Taking Assistance
+
+For SWM and WCST, we explore assistance in the form of notes, which emulates cognitive offloading (??) where humans write down their observations to keep track of the state. The system provides the notes itself in every turn, explicitly stating the current state and actions taken by the model so far. From Table [6,](#page-6-0) we observe large gains for some models, but the average effect is not significant (Text Hard p = .161; Img+Text Easy p = .842), which may reflect limited sample size or generally small changes. This suggests that many models still struggle with working-memory maintenance, where notes are not even helpful and may even cause slight drops.
+
+For WCST, improvements from note-taking are generally consistent across models and setups, although they are close to reach statistical significance in our (Image Easy: p = .14; Image Hard: p = .0524). This indicates that note-taking provides practical benefits by helping models track tested attributes and maintain a consistent hypothesis.
+
+{6}------------------------------------------------
+
+354
+
+356
+
+374
+
+376
+
+<span id="page-6-0"></span>*Table 6.* SWM results with note-taking assistance. Paired t-test results show no significant effects on all setups (p > .05).
+
+| Model           | SWM (Sswm<br>↑) |              |  |  |  |
+|-----------------|-----------------|--------------|--|--|--|
+|                 | Text Hard       | Img+Txt Easy |  |  |  |
+| Gemini 2.5 Pro  | 0.965∆−0.02     | 0.704∆−0.03  |  |  |  |
+| Claude Sonnet 4 | 0.826∆−0.04     | 0.207∆−0.05  |  |  |  |
+| Grok 4.1 Fast   | 0.973∆−0.01     | 0.043∆−0.10  |  |  |  |
+| Grok 4 Fast     | 0.964∆−0.03     | 0.029∆+0.01  |  |  |  |
+| o4-Mini         | 0.935∆+0.12     | 0.012∆−0.04  |  |  |  |
+| GLM 4.6V        | 0.803∆+0.23     | 0.082∆+0.04  |  |  |  |
+| GLM 4.5V        | 0.851∆+0.32     | 0.000∆−0.04  |  |  |  |
+| Qwen3-VL-235B   | 0.782∆+0.02     | 0.318∆+0.29  |  |  |  |
+
+*Table 7.* WCST image results with note-taking assistance. Paired t-test results show no significant effects on all setups (p > .05).
+
+| Model         | WCST (Swcst<br>↑) |              |  |  |  |  |
+|---------------|-------------------|--------------|--|--|--|--|
+|               | Image Easy        | Image Hard   |  |  |  |  |
+| Grok 4.1 Fast | 0.135∆+0.028      | 0.197∆+0.137 |  |  |  |  |
+| Grok 4 Fast   | 0.215∆−0.120      | 0.257∆+0.048 |  |  |  |  |
+| o4-Mini       | 0.392∆+0.045      | 0.295∆+0.018 |  |  |  |  |
+| GLM 4.6V      | 0.430∆+0.239      | 0.411∆+0.347 |  |  |  |  |
+| GLM 4.5V      | 0.306∆+0.199      | 0.123∆+0.040 |  |  |  |  |
+| Qwen3-VL-235B | 0.640∆+0.331      | 0.250∆+0.179 |  |  |  |  |
+
+## 6. Analysis & Discussion
+
+#### <span id="page-6-3"></span>6.1. Factor Analysis of LLM
+
+To understand the latent structure of capabilities across widely used LLM benchmarks inspired by ?, we conducted an exploratory factor analysis (EFA) on performance data obtained from Artificial Analysis[2](#page-6-1) for 156 distinct LLMs. We selected 10 benchmarks widely recognized as standard measures of LLM "intelligence" and real-world capabilities, ranging from coding and math to agentic tool use. Importantly, this selection was driven by the statistical requirement for a large sample size (N > 100) to ensure valid factor analysis, which very few other benchmarks satisfy. We used Principal Axis Factoring for the factoring method.
+
+Parallel analysis indicated a single-factor solution: a unidimensional latent variable explains most performance variation across these tasks. This factor (PA1) accounted for 75% of the total variance, providing strong evidence for a "general capability" factor. As shown in Table [8,](#page-6-2) all benchmarks had high loadings (λ > 0.76), with *GPQA Diamond* (0.943) and *LiveCodeBench* (0.926) as the strongest proxies. This coherence also exposes a limitation of current evaluation, where many benchmarks measure the same underlying construct. This makes NeuroCognition even more important, as
+
+<span id="page-6-2"></span>*Table 8.* Factor analysis loadings for state-of-the-art LLM benchmarks. The single extracted factor (PA1) explains 75% of the total variance, indicating a strong general capability factor (g) across diverse tasks.
+
+| Benchmark                | Loadings (λ) |
+|--------------------------|--------------|
+| GPQA Diamond (?)         | 0.943        |
+| LiveCodeBench (?)        | 0.926        |
+| SciCode (?)              | 0.922        |
+| AA-LCR (?)               | 0.875        |
+| AIME 2025 (?)            | 0.870        |
+| IFBench (?)              | 0.861        |
+| Terminal Bench (?)       | 0.834        |
+| MMLU Pro (?)             | 0.825        |
+| τ -Bench (?)             | 0.818        |
+| Humanity's Last Exam (?) | 0.768        |
+
+it targets distinct, basic cognitive tasks.
+
+#### 6.2. Correlation with Other Benchmarks
+
+To assess how NeuroCognition relates to real-world tasks, we correlate it with external benchmarks targeting specific cognitive demands. For working memory (SWM), we use a modified *Long HotPotQA* (?) that splits questions and passages at the word level and treats each word as a turn. For fluid intelligence (RAPM), we compare against *ARC-AGI 2* (?). For cognitive flexibility (WCST), we use *LLM Task Interference* (?), which measures performance impacts under task switches. Detailed setups and results are in Appendix [D.](#page-14-0) We also correlate the aggregate NeuroCognition score with the average of 11 state-of-the-art benchmarks (the 10 in Section [6.1](#page-6-3) plus MMMU Pro (?)), as a proxy for "general capability". These scores are also taken and already normalized from Artificial Analysis. Pearson correlations are shown in Figure [2.](#page-7-0)
+
+We observe a high correlation between the average NeuroCognition score and the 11-benchmark average (r = .86, p = .001, N = 10), validating our framework as a robust predictor of general model capability (g). More importantly, the heatmap provides a granular view of the functional relationships between cognitive primitives and downstream tasks:
+
+- (1) Working Memory & Context: *SWM Text Hard* shows a moderate correlation with *Long HotPotQA* (r = .43, p = .21, N = 10). This suggests that working-memory maintenance is still somewhat relevant long-context retrieval.
+- (2) Fluid Intelligence: *RAPM Image MC* is the strongest predictor of *ARC-AGI 2* (r = .87, p = .024, N = 6), consistent with isolating an "abstract visual reasoning" factor that standard benchmarks often miss while also uncorrelated with the general benchmark average (r = .09).
+- (3) Cognitive Flexibility: *WCST Text Hard* shows only a
+
+<span id="page-6-1"></span>https://artificialanalysis.ai/leaderboards/models
+
+{7}------------------------------------------------
+
+<span id="page-7-0"></span>![](_page_7_Figure_1.jpeg)
+
+Figure 2. Correlations between our NeuroCognition benchmark tests and corresponding real-world LLM benchmarks. Cells with bold text and highlighted borders indicate task pairs that are similar and rely on the same cognitive abilities (including averages). Statistical significance is denoted by asterisks (\*p < .05, \*\*p < .01).
+
+weak, non-significant correlation with *LLM Task Interference* ( $r=.24,\ p=.50,\ N=10$ ). This suggests that the abstract rule-updating mechanisms measured by WCST may be distinct from practical task switching scenarios.
+
+#### 7. Conclusion
+
+385
+
+387
+
+388
+
+389
+
+390
+
+398
+
+399
+
+400
+
+401 402 403
+
+404
+
+405
+
+406
+
+407
+
+408
+
+409
+
+410
+
+411
+
+412
+
+413
+
+414
+
+415
+
+416
+
+417
+
+418
+
+419
+
+420
+
+421
+
+422
+
+423
+
+424
+
+425
+
+426
+
+427
+
+428
+
+429
+
+430
+
+431
+
+432
+
+433
+
+434
+
+435
+
+436
+
+437
+
+438
+
+439
+
+In this work, we introduce the NeuroCognition benchmark, a suite of neuropsychological tests (RAPM, SWM, WCST) adapted for LLMs to probe specific cognitive processes using multimodal, process-aware metrics. Across tasks, state-of-the-art models show meaningful but incomplete alignment with human cognition. They perform better in text, but performance drops for images and harder settings that demand intensive grounding and state tracking. Overall, explicit reasoning is not uniformly beneficial, and humanlike assistance shows mixed effects. The impact of pattern hints is limited, while note-taking helps more consistently in WCST but not in SWM. Complementing these results, our factor analysis suggests a dominant general-capability factor across standard benchmarks, highlighting redundancy in current evaluation. Correlations with external benchmarks further suggest that gains in NeuroCognition process scores translate into improvements in semantically related realworld tasks, while individual tests still measure distinct cognitive abilities. In summary, our results suggest that current LLMs do not possess the same kind of intelligence as humans, but still align with humans in some aspects, which can inform how we use and improve them.
+
+## 8. Future Work
+
+Since we apply neuropsychological tests designed for humans to LLMs, it is important for future work to validate whether the same cognitive abilities measured in humans are also captured in LLMs. This helps to further under-
+
+stand the relationship between human and LLM intelligence and improve the interpretability of the benchmark results. Another interesting direction is improving the LLM itself through neuropsychologically grounded training data. The data that we used in this benchmark captures basic cognitive processes that are foundational to human intelligence, and therefore can be used to improve the cognitive abilities of LLMs. Additionally, recent works investigate ways to compress context (??), or utilize some sort of memory in their architecture (???), and our benchmark could serve to evaluate such methods. Finally, this data can be easily generated synthetically with code, making it scalable to create a large amount of data for training purposes.
+
+#### Limitations
+
+Our evaluation has three main limitations. First, sample sizes are constrained by cost: running state-of-the-art models with effectively unlimited reasoning is expensive. For RAPM, we capped runs at n = 200 (text) and n = 140(image), which may compromise the robustness of our results, especially in the original image modality. Second, our benchmark assumes that the same neuropsychological test used to assess a specific cognitive ability in humans can also be applied to LLMs to evaluate that same ability. In the neuropsychological field, this is not always the case, since a neuropsychological test is standardized on a specific population and must be validated based on psychometric properties (validity, reliability, and measurement invariance) when applied to a different population. Therefore, even when a model performs consistently as well as humans across all tasks in our benchmark, we can only claim that the model exhibits a similar level of intelligence and can simulate human-like cognitive abilities, but not that it thinks like humans or possesses the same underlying cognitive mechanism. Third, since the neuropsychological tests we
+
+{8}------------------------------------------------
+
+used are widely used and well-established, we assume they are valid and reliable based on their psychometric properties (validity, reliability, and measurement invariance). However, this assumption may not hold in all cases, and further validation across prompts, contexts, and a wide range of LLMs is needed.
+
+## Impact Statement
+
+In our work, we evaluate LLMs under the assumption that their behavior can be interpreted using human cognitive processes. However, we do not claim that LLMs possess consciousness, self-awareness, or the same fundamental cognitive architecture as humans. We acknowledge the use of GitHub Copilot[3](#page-8-0) , Gemini[4](#page-8-1) , and ChatGPT[5](#page-8-2) for coding and writing assistance.
+
+<span id="page-8-0"></span> github.com/features/copilot
+
+<span id="page-8-1"></span> gemini.google.com
+
+<span id="page-8-2"></span> chatgpt.com
+
+{9}------------------------------------------------
+
+## <span id="page-9-0"></span>A. RAPM Text Data Generation
+
+### A.1. Rule Space and Attributes
+
+504
+
+513 514
+
+516
+
+518 519 520
+
+524 525 526
+
+528 529 530
+
+534
+
+536
+
+The generation engine employs a compositional rule space defined by attributes that constrain string properties. For each matrix, two primary attributes are sampled—one for rows and one for columns. Additional consistency rules (leaks) are inferred if they naturally emerge from the intersection of constraints.
+
+The implemented attributes and their parameters are:
+
+- 1. Character Set Type: Restricts the alphabet of the string.
+  - *Values*: letters, digits, symbols.
+  - *Logic*: All characters in the cell must belong to the specified set.
+- 2. Type Count: Constraints on the count of specific character types.
+  - *Types*: vowels, consonants, uppercase/lowercase variants, etc.
+  - *Rules*: even, odd, multiple of N (for N ∈ {2, 3, 4}).
+  - *Logic*: The count of characters matching the type must satisfy the parity or divisibility rule. Zero counts are explicitly disallowed for these rules to ensure pattern salience.
+- 3. Quantitative Constant: Fixes a metric to a specific integer value across the sequence.
+  - *Metrics*: String length, unique character count, or count of specific types.
+  - *Range*: Integers in [2, 5].
+- 4. Quantitative Progression: Enforces an arithmetic progression on a metric.
+  - *Metrics*: Same as Quantitative Constant.
+  - *Parameters*: Start value ∈ [1, 3], Step size ∈ [1, 3].
+  - *Logic*: The metric value increases by the step size across the row or column (e.g., lengths 3, 5, 7).
+- 5. Sorted Order: Enforces lexicographical ordering.
+  - *Values*: ascending, descending, mixed.
+  - *Logic*: Characters in the string must follow the specified order. mixed explicitly requires the string to be unsorted.
+- 6. Positional Constraint: Fixes the character type at specific indices.
+  - *Indices*: first, last, even, odd.
+  - *Logic*: Characters at the specified positions must belong to a specific type (e.g., "first character must be a digit").
+
+#### A.2. Generation Algorithm
+
+The matrix generation process follows a constraint-satisfaction approach with backtracking, as detailed in Algorithm [1.](#page-10-0) The core logic proceeds as follows:
+
+- 1. Attribute Sampling: Randomly sample one attribute specification for rows and one for columns.
+- 2. Compatibility Check: Validate that the row and column attributes are compatible. For example, a "Sorted" row attribute conflicts with a "Positional" column attribute because sorting rearranges characters, potentially violating fixed-position constraints. Infeasible combinations trigger a re-sampling.
+- 3. Constraint Propagation: The row and column attributes are combined to define specific constraints for each of the 3 × 3 cells. "Leak" detection is performed to identify additional properties that effectively become constant across rows or columns (e.g., if a row rule restricts characters to digits and a column rule restricts them to even numbers, a universal "digits" constraint might apply).
+- 4. Grid Construction: Cells are generated sequentially. For each cell, a string is constructed to satisfy the merged constraints. Random backtracking is used if a cell cannot be generated given the current constraints and history.
+- 5. Variation Enforcement: To prevent trivial patterns, we enforce a minimum Hamming distance between strings in the same row or column. This ensures that the progression relies on abstract rules rather than simple repetition.
+
+Upon successful generation, the defining constraints for the missing cell (including "leaked" rules) are stored for automatic rule validation.
+
+{10}------------------------------------------------
+
+#### <span id="page-10-0"></span>Algorithm 1 RAPM Text Matrix Generation
+
+550
+
+554
+
+556
+
+574
+
+576
+
+594
+
+596
+
+```
+Input: Random Seed S
+Output: 3 × 3 String Matrix M, Constraints C
+Initialize RNG with S
+repeat
+  Sample row attribute Arow and column attribute Acol
+  valid ← CheckCompatibility(Arow, Acol)
+until valid
+Initialize cell constraint grid C3×3 from Arow and Acol
+Apply inferred consistency rules (leaks) to C
+for i = 0 to 2 do
+  for j = 0 to 2 do
+    attempts ← 0
+    repeat
+       Generate string s satisfying Ci,j
+       vrow ← ∀k < j : Hamming(s,Mi,k) ≥ 2
+       vcol ← ∀k < i : Hamming(s,Mk,j ) ≥ 2
+       attempts ← attempts + 1
+    until (vrow ∧ vcol) or attempts > MAX RET RIES
+    if not (vrow ∧ vcol) then
+       fail and backtrack
+    end if
+    Mi,j ← s
+  end for
+end for
+return M, C
+```
+
+#### A.3. Multiple-Choice and Distractor Generation
+
+For the text MC setup, we generate 8 options: 1 correct answer (ground truth for the bottom-right cell) and 7 distractors. Each distractor is guaranteed to violate at least one latent rule to ensure a unique solution.
+
+- Correct Answer: The generated string for the bottom-right (3, 3) cell.
+- Distractors: Generated by applying mutation strategies to the correct answer to violate latent rules. Strategies include:
+  - *Break Ordering*: Swapping adjacent characters to violate sort constraints.
+  - *Break Positional*: Changing a character at a critical index (e.g., first/last) to a forbidden type.
+  - *Adjust Count*: Adding or removing characters to violate parity or length constraints.
+  - *Character Mutation*: Randomly altering characters to valid but incorrect values.
+
+{11}------------------------------------------------
+
+## <span id="page-11-0"></span>B. RAPM Text Gen Automatic Rule Validation
+
+For the text Gen setup, we can programmatically check if a generated string S satisfies the stored underlying constraints C for the missing cell (3, 3). Algorithm [2](#page-11-1) details this validation procedure.
+
+#### <span id="page-11-1"></span>Algorithm 2 RAPM Text Gen Validation (CellSatisfies)
+
+```
+Input: Candidate String S, Constraints C
+Output: Boolean (True if valid)
+if C.f ixed length ̸= None and |S| ̸= C.f ixed length then
+  return False
+end if
+for all (type, count) ∈ C.target counts do
+  if Count(S, type) ̸= count then
+    return False
+  end if
+end for
+for all (type, rule) ∈ C.parity rules do
+  c ← Count(S, type)
+  if c = 0 then
+    return False {Zero count forbidden for parity}
+  end if
+  if (rule = 'even' ∧ c is odd) or (rule = 'odd' ∧ c is even) then
+    return False
+  end if
+end for
+for all (type, k) ∈ C.multiple rules do
+  c ← Count(S, type)
+  if c = 0 or c (mod k) ̸= 0 then
+    return False
+  end if
+end for
+if C.unique count ̸= None and |Unique(S)| ̸= C.unique count then
+  return False
+end if
+if C.ordering ̸= None and CheckOrder(S, C.ordering) = False then
+  return False
+end if
+if C.positional ̸= None and CheckPositional(S, C.positional) = False then
+  return False
+end if
+return True
+```
+
+The validation function CellSatisfies(S, C3,3) performs the following checks:
+
+- 1. Length Check: If a fixed length is enforced (from quantitative constraints), |S| must match.
+- 2. Type Counts: Count(S, type) = k for all exact count constraints.
+- 3. Parity/Divisibility: Count(S, type) must satisfy even/odd/multiple conditions. Crucially, as enforced during generation, zero counts are invalid for these rules to prevent trivial satisfactions (e.g., "0 vowels" satisfying "even vowels").
+- 4. Unique Characters: If unique count is constrained, |Unique(S)| is checked against the target.
+- 5. Ordering: If sorted order is required, S is checked against its own sorted version.
+- 6. Positional: Characters at constrained indices (first, last, even, odd) are verified against the allowed character set.
+
+Any string that satisfies all these constraints is considered a correct solution, providing a robust "verifiable" accuracy metric for generative models that goes beyond simple multiple-choice selection or exact-match.
+
+{12}------------------------------------------------
+
+## <span id="page-12-0"></span>C. Models & Experimental Setup
+
+## C.1. Model Access Names and Notes
+
+Below we list the canonical access names or placeholders used when calling each model, grouped by access method (Google, OpenAI, OpenRouter, Local). All models listed here are verified and accessible during the submission and release of this paper.
+
+## Google AI Studio
+
+- Gemini 3 Pro (?): gemini-3-pro-preview
+- Gemini 2.5 Pro (?): gemini-2.5-pro
+
+#### OpenAI API
+
+- GPT-5 (?): gpt-5-2025-08-07
+- o4-Mini (?): o4-mini-2025-04-16
+
+#### OpenRouter
+
+- Claude Sonnet 4 (?): anthropic/claude-sonnet-4
+- Grok 4.1 Fast (?): x-ai/grok-4.1-fast
+- Grok 4 Fast (?): x-ai/grok-4-fast
+- GLM 4.6V (?): z-ai/glm-4.6v[6](#page-12-1)
+- GLM 4.5V (?): z-ai/glm-4.5v[7](#page-12-2)
+
+## Local / vLLM
+
+704
+
+706
+
+713 714 • Qwen3-VL-235B (?): Qwen/Qwen3-VL-235B-A22B-Thinking[8](#page-12-3)
+
+#### C.2. vLLM Inference Settings
+
+For local vLLM runs, we used a server with 4x NVIDIA H100 80GB GPUs and CUDA 12.8, and ran the model with FP8 quantization. Generations used the model's default Qwen3 configuration; temperature and other sampling settings were chosen per the evaluation protocol below.
+
+## C.3. Evaluation Protocol Details
+
+- Zero-shot: All evaluations are zero-shot.
+- Temperature: We used a temperature of 0 when possible to reduce sampling variability; when a model's API does not allow overriding temperature, we used the model's default.
+- Reasoning mode: We used each model's default reasoning configuration except for Claude Sonnet 4 and Grok 4 Fast, for which we adjusted reasoning as described in the main text (see Section [3.4\)](#page-3-0).
+- Maximum output tokens: For RAPM, we allowed the model's maximum permitted output tokens. For SWM and WCST, we limited per-turn outputs to 8192 tokens.
+- Thinking budget: The system prompt described a thinking budget (not enforced as a hard token limit). For RAPM, we used an effective thinking budget of 30,000 tokens for models with a context limit of 32K and 60,000 tokens for models with a context limit of 65K+. For SWM and WCST, the per-turn thinking budget was 4,000 tokens.
+- Repetitions: SWM and WCST setups were run three times each to assess variability (data are generated on the fly); RAPM was run once per model.
+- WCST (notes): For WCST runs that used note-taking assistance, we set the note window to 6.
+
+<span id="page-12-1"></span><sup>6</sup> https://huggingface.co/zai-org/GLM-4.6V
+
+<span id="page-12-2"></span><sup>7</sup> https://huggingface.co/zai-org/GLM-4.5V
+
+<span id="page-12-3"></span><sup>8</sup> https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Thinking
+
+{13}------------------------------------------------
+
+## C.4. Reasoning Configuration
+
+Model reasoning capabilities were enabled through two complementary methods:
+
+718 719
+
+720
+
+716
+
+API Configuration. In our ModelWrapper, we programmatically enabled reasoning features where supported by the provider APIs.
+
+724
+
+726
+
+734
+
+736
+
+- Google AI Studio: We accessed Gemini models via the OpenAI-compatible endpoint and enabled reasoning by specifically configuring the thinking config within the extra body parameter. We set include thoughts to True, thinking level for Gemini 3 Pro (default to high), and thinking budget for Gemini 2.5 Pro (default to 8192).
+- OpenAI API: For OpenAI models, reasoning is handled through reasoning effort parameter in the chat completion calls. We use the default value medium for both GPT-5 and o4-Mini.
+- OpenRouter: We set the reasoning parameter in the request body to {"exclude": False} to ensure reasoning tokens were generated and returned, and manage the thinking budget by setting max tokens within the reasoning configuration. For Grok, reasoning was explicitly enabled via {"enabled": True}.
+- vLLM (Local): We enabled the enable thinking parameter in the chat template to activate reasoning generation.
+
+Prompt-Based Instructions. For all tasks, we included explicit instructions in the prompt to guide the reasoning process, specifically when Chain-of-Thought (CoT) was enabled.
+
+#### Standardized CoT Prompt Addition
+
+#### SWM & WCST:
+
+Explain your thought process regarding the problem and the feedbacks you received in maximum {think budget} tokens wrapped with <think> and </think>. Then, provide a really short summary of your reasoning after the closing </think> tag.
+
+#### RAPM:
+
+Explain your thought process (max {think budget} tokens) inside <think>...</think> then give final answer.
+
+743 744 745
+
+This dual approach ensured that models with native reasoning capabilities (activated via API) and standard models (guided via prompts) both engaged in structured reasoning before outputting a final answer.
+
+746 747
+
+748 749
+
+## C.5. Disabling Reasoning
+
+750 751 We disable the reasoning capabilities for some models that are able to do so to conduct the ablation study (see Section [5.1\)](#page-5-3) and to address specific model issues (specifically, Claude Sonnet 4 and Grok 4 Fast with "overthinking" that exhausted token limits on RAPM).
+
+754 756 Configuration & Prompts. For most models, reasoning was disabled by reversing the API configurations described previously (e.g., setting include thoughts to False, reasoning enabled to False, or enable thinking to False) and replacing the CoT prompt instructions with a standard direction to provide only the final answer, as shown below.
+
+758 759
+
+## Standardized Non-CoT Instruction Answer only with your final answer.
+
+760 761
+
+762
+
+## Gemini 2.5 Pro Specific Prompt. For Gemini 2.5 Pro, the model does not natively support disabling reasoning via API parameters. However, we found a way to enforce the no-reasoning condition by appending a specific critical instruction to override the model's default behavior.
+
+#### Gemini 2.5 Pro Non-CoT Instruction
+
+768 769 CRITICAL INSTRUCTION: Do not use <thinking> tags or reasoning steps. Provide the response immediately.
+
+{14}------------------------------------------------
+
+## <span id="page-14-0"></span>D. External Benchmark Setup & Results
+
+Long HotPotQA. We work with the distractor development split of HotPotQA (?), sampling 500 question–passage pairs. To stress the models' turn tracking, we split each question and its supporting passages at the word level, treat every word as a separate turn, and fix the model's response on every intermediate turn to "Okay." Each sequence is ordered with the full question first, followed by the concatenated passages. The results for Long HotPotQA are shown in Table [9.](#page-14-1)
+
+LLM Task Interference. Following the official implementation from ?, with minor updates, we evaluate two task/context pairings and sample 200 evaluation prompts each. The first pairing uses the MMLU AA task with Rotten Tomatoes exemplars as in-context data, while the second pairing is the Rotten Tomatoes task with TweetQA data. The results for LLM Task Interference are shown in Table [9.](#page-14-1)
+
+ARC-AGI 2. Since the results for ARC-AGI 2 (?) are already publicly available in their leaderboard website[9](#page-14-2) , we directly took the best results for each model, except for some models that are not available. Although the scores are very low, the results have been verified from the source. These are also summarized in Table [9.](#page-14-1)
+
+<span id="page-14-1"></span>*Table 9.* Results for external benchmarks: Long HotPotQA, LLM Task Interference, and ARC-AGI 2. (-) indicates the model result is not available.
+
+| Model           | Long HotPotQA | Task Interference | ARC-AGI 2 |
+|-----------------|---------------|-------------------|-----------|
+| GPT-5           | 0.936         | 0.958             | 0.099     |
+| Gemini 3 Pro    | 0.932         | 0.962             | 0.311     |
+| Gemini 2.5 Pro  | 0.924         | 0.875             | 0.049     |
+| Claude Sonnet 4 | 0.930         | 0.948             | 0.059     |
+| Grok 4.1 Fast   | 0.910         | 0.913             | -         |
+| Grok 4 Fast     | 0.461         | 0.788             | 0.053     |
+| o4-Mini         | 0.802         | 0.933             | 0.061     |
+| GLM-4.6V        | 0.862         | 0.938             | -         |
+| GLM-4.5V        | 0.271         | 0.635             | -         |
+| Qwen3 VL 235B   | 0.916         | 0.940             | -         |
+
+<span id="page-14-2"></span> https://arcprize.org/leaderboard
+
+{15}------------------------------------------------
+
+## <span id="page-15-0"></span>E. Full Results
+
+#### E.1. Raven's Advanced Progressive Matrices (RAPM)
+
+The complete results for RAPM image tests, broken down by pattern type, are provided in Table [10.](#page-15-1) We also provide qualitative examples of reasoning failures across text and image modalities in Table [11,](#page-15-2) highlighting the prevalence of hallucination and over-interpretation.
+
+*Table 10.* RAPM image accuracy breakdown by pattern type.
+
+<span id="page-15-1"></span>
+
+| Model           | Dist. Four | Dist. Nine | Center | In-Dist. Out-Center | In-Out Center | Left-Right | Up-Down |
+|-----------------|------------|------------|--------|---------------------|---------------|------------|---------|
+| GPT-5           | 0.40       | 0.35       | 0.25   | 0.15                | 0.35          | 0.05       | 0.40    |
+| Gemini 3 Pro    | 0.55       | 0.65       | 0.55   | 0.35                | 0.20          | 0.25       | 0.25    |
+| Gemini 2.5 Pro  | 0.05       | 0.10       | 0.35   | 0.20                | 0.40          | 0.20       | 0.25    |
+| Claude Sonnet 4 | 0.30       | 0.45       | 0.15   | 0.20                | 0.20          | 0.10       | 0.15    |
+| Grok 4.1 Fast   | 0.15       | 0.20       | 0.15   | 0.05                | 0.00          | 0.00       | 0.15    |
+| Grok 4 Fast     | 0.05       | 0.10       | 0.05   | 0.25                | 0.10          | 0.10       | 0.05    |
+| o4-Mini         | 0.45       | 0.15       | 0.25   | 0.10                | 0.30          | 0.05       | 0.15    |
+| GLM 4.6V        | 0.45       | 0.30       | 0.40   | 0.35                | 0.30          | 0.25       | 0.45    |
+| GLM 4.5V        | 0.20       | 0.05       | 0.15   | 0.10                | 0.15          | 0.20       | 0.05    |
+| Qwen3-VL-235B   | 0.45       | 0.65       | 0.55   | 0.40                | 0.40          | 0.50       | 0.50    |
+
+<span id="page-15-2"></span>*Table 11.* Selected reasoning failure examples in RAPM (image and text). Models frequently hallucinate attributes to fit a complex hypothesis ("visual/numerological Hallucination") or prefer abstract semantic rules over simple local operations ("semantic overinterpretation").
+
+| Task/Model               | ID                     | Pattern/Context                              | Reasoning Snippet (Failure Point)                                                                                                                                                                                                    | Outcome                                                                                    |
+|--------------------------|------------------------|----------------------------------------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|--------------------------------------------------------------------------------------------|
+| Image<br>GPT-5           | RAVEN 238<br>center    | Up-Down<br>(Split layout)                    | "Row 3: top becomes large solid black triangle.<br>Bottom shapes by row show a sides-count sequence<br>decreasing 5→4→3"<br>(Visual Hallucination: No such triangle exists.)                                                         | Pred: 2<br>Correct: 0                                                                      |
+| Image<br>GPT-5           | RAVEN 238<br>dist nine | Dist. Nine<br>(3x3 Grid logic)               | "Row 1: Left has all 9 positions filled middle<br>emphasizes corners The combination is still all 9<br>positions"<br>(Relational Over-complexity: Hallucinates a global<br>superposition rule for a simple local pattern.)           | Pred: 5<br>Correct: 2                                                                      |
+| Text MC<br>Gemini 3 Pro  | text rapm 0            | Simple Arithmetic<br>(Char length)           | "Row 1: 6, 10, 13 characters Row 2: 6, 9, 12<br>notice a potential arithmetic-like progression"<br>(Numerological Hallucination: Forces arithmetic se<br>ries on string lengths where rule is character types.)                      | Pred: 3<br>Correct: 5                                                                      |
+| Text MC<br>GPT-5         | text rapm 4            | Column Count<br>(Uppercase)                  | "Column 2 uppercase counts go 0 1 2. So for<br>Column 3 they should be 0 1 2."<br>(False Generalization: Hallucinates column-wise<br>progression that doesn't exist in data.)                                                        | Pred: 3<br>Correct: 6                                                                      |
+| Text Gen<br>GPT-5        | text rapm 2            | Constraint Satisfac<br>tion<br>(Vowel Count) | "Rows keep a constant number of uppercase let<br>ters Column 3 needs 6 vowels A valid comple<br>tion is 'aeiOuEsTH'."<br>(Metadata vs Content: Hallucinates content string<br>that fits metadata counts but fails derivation logic.) | Pred: aeiOuEsTH<br>Rules:<br>5 lowercase<br>vowels, 4 uppercase<br>letters                 |
+| Text Gen<br>Gemini 3 Pro | text rapm 3            | String Length<br>(Decreasing)                | "Column 2 seems to be decreasing length Row 3:<br>6, 9, ? Missing string should be length 8."<br>(Numerological: Obsesses over length counts, ignor<br>ing character content patterns.)                                              | Pred:<br>T4Z]LBEBUM6<br>Rules:<br>Col progres<br>of<br>uppercase<br>sion<br>vowels (2→3→4) |
+
+{16}------------------------------------------------
+
+#### E.2. Spatial Working Memory (SWM)
+
+The complete results for SWM tests are provided in Table [12](#page-16-0) for text modality, in Table [13](#page-16-1) for image modality, and in [14](#page-17-0) for image-text modality. All metrics are averaged over three runs.
+
+<span id="page-16-0"></span>*Table 12.* Spatial Working Memory full results (text modality). All metrics are reported as mean with standard deviation (std) in a subscript.
+
+| Difficulty   | Model           | Sswm      | Tokens Score | Guesses       | Illegal      | Invalid      | Repeated     | No-box     |
+|--------------|-----------------|-----------|--------------|---------------|--------------|--------------|--------------|------------|
+|              | GPT-5           | 1.000±.00 | 1.000±.00    | 18.000±6.16   | 0.000±0.00   | 0.000±0.00   | 0.000±0.00   | 0.000±0.00 |
+|              | Gemini 3 Pro    | 1.000±.00 | 1.000±.00    | 18.000±2.94   | 0.000±0.00   | 0.000±0.00   | 0.000±0.00   | 0.000±0.00 |
+|              | Gemini 2.5 Pro  | 1.000±.00 | 1.000±.00    | 19.667±1.70   | 0.000±0.00   | 0.000±0.00   | 0.000±0.00   | 0.000±0.00 |
+|              | Claude Sonnet 4 | 1.000±.00 | 1.000±.00    | 23.000±3.56   | 0.000±0.00   | 0.000±0.00   | 0.000±0.00   | 0.000±0.00 |
+| Easy         | Grok 4.1 Fast   | 1.000±.00 | 1.000±.00    | 26.667±2.87   | 0.000±0.00   | 0.000±0.00   | 0.000±0.00   | 0.000±0.00 |
+|              | Grok 4 Fast     | 1.000±.00 | 1.000±.00    | 18.333±3.40   | 0.000±0.00   | 0.333±0.47   | 0.000±0.00   | 0.000±0.00 |
+|              | o4-Mini         | 0.951±.05 | 1.000±.00    | 23.667±4.03   | 0.000±0.00   | 0.667±0.94   | 1.000±0.82   | 0.000±0.00 |
+|              | GLM 4.6V        | 1.000±.00 | 1.000±.00    | 24.000±2.94   | 0.000±0.00   | 3.333±0.94   | 0.000±0.00   | 0.000±0.00 |
+|              | GLM 4.5V        | 1.000±.00 | 1.000±.00    | 20.333±1.70   | 0.000±0.00   | 0.000±0.00   | 0.000±0.00   | 0.000±0.00 |
+|              | Qwen3-VL-235B   | 1.000±.00 | 1.000±.00    | 20.000±2.45   | 0.000±0.00   | 0.000±0.00   | 0.000±0.00   | 0.000±0.00 |
+|              | GPT-5           | 1.000±.00 | 1.000±.00    | 56.333±1.25   | 0.000±0.00   | 0.000±0.00   | 0.000±0.00   | 0.000±0.00 |
+|              | Gemini 3 Pro    | 1.000±.00 | 1.000±.00    | 63.333±1.70   | 0.000±0.00   | 1.000±0.82   | 0.000±0.00   | 0.000±0.00 |
+|              | Gemini 2.5 Pro  | 0.980±.03 | 1.000±.00    | 82.333±13.57  | 0.000±0.00   | 0.000±0.00   | 2.000±2.83   | 0.000±0.00 |
+|              | Claude Sonnet 4 | 0.866±.18 | 1.000±.00    | 78.333±11.44  | 0.000±0.00   | 0.333±0.47   | 12.333±16.74 | 0.000±0.00 |
+| Hard         | Grok 4.1 Fast   | 0.983±.01 | 1.000±.00    | 52.667±9.29   | 0.000±0.00   | 0.000±0.00   | 1.000±0.82   | 0.000±0.00 |
+|              | Grok 4 Fast     | 0.994±.01 | 1.000±.00    | 63.000±4.55   | 0.333±0.47   | 0.333±0.47   | 0.000±0.00   | 0.000±0.00 |
+|              | o4-Mini         | 0.811±.07 | 1.000±.00    | 79.333±1.70   | 5.333±3.77   | 0.000±0.00   | 9.667±2.36   | 0.000±0.00 |
+|              | GLM 4.6V        | 0.574±.35 | 0.694±.27    | 129.333±20.74 | 3.000±3.56   | 33.333±34.50 | 32.333±44.32 | 0.000±0.00 |
+|              | GLM 4.5V        | 0.536±.24 | 0.736±.37    | 105.333±27.64 | 16.667±11.81 | 42.667±58.93 | 0.333±0.47   | 0.000±0.00 |
+|              | Qwen3-VL-235B   | 0.760±.06 | 1.000±.00    | 80.333±10.66  | 19.333±7.72  | 1.000±0.00   | 0.333±0.47   | 0.000±0.00 |
+|              | Gemini 2.5 Pro  | 0.965±.04 | 1.000±.00    | 66.667±9.18   | 0.000±0.00   | 0.000±0.00   | 2.333±2.62   | 0.000±0.00 |
+|              | Claude Sonnet 4 | 0.826±.06 | 1.000±.00    | 81.000±9.27   | 1.000±0.82   | 0.000±0.00   | 13.667±5.56  | 0.000±0.00 |
+|              | Grok 4.1 Fast   | 0.973±.01 | 1.000±.00    | 62.333±3.30   | 0.333±0.47   | 0.000±0.00   | 1.333±0.47   | 0.000±0.00 |
+| Hard - Notes | Grok 4 Fast     | 0.964±.01 | 1.000±.00    | 65.333±1.25   | 0.333±0.47   | 0.000±0.00   | 2.000±0.00   | 0.000±0.00 |
+|              | o4-Mini         | 0.935±.03 | 1.000±.00    | 75.333±15.92  | 1.000±1.41   | 0.333±0.47   | 4.333±2.49   | 0.000±0.00 |
+|              | GLM 4.6V        | 0.803±.14 | 1.000±.00    | 85.000±20.22  | 0.333±0.47   | 1.000±0.00   | 19.000±16.87 | 0.000±0.00 |
+|              | GLM 4.5V        | 0.851±.09 | 1.000±.00    | 79.333±11.12  | 10.333±7.36  | 14.333±16.11 | 0.333±0.47   | 0.000±0.00 |
+|              | Qwen3-VL-235B   | 0.782±.23 | 0.972±.04    | 92.333±36.81  | 0.667±0.94   | 2.333±3.30   | 25.667±32.83 | 0.000±0.00 |
+
+<span id="page-16-1"></span>*Table 13.* Spatial Working Memory full results (image modality). All metrics are reported as mean with standard deviation (std) in a subscript.
+
+| Difficulty | Model           | Sswm       | Tokens Score | Guesses      | Illegal      | Invalid       | Repeated     | No-box       |
+|------------|-----------------|------------|--------------|--------------|--------------|---------------|--------------|--------------|
+|            | GPT-5           | 0.927±0.08 | 1.000±0.00   | 27.667±4.64  | 0.000±0.00   | 0.000±0.00    | 2.000±2.83   | 0.333±0.47   |
+|            | Gemini 3 Pro    | 1.000±0.00 | 1.000±0.00   | 22.333±0.94  | 0.000±0.00   | 0.000±0.00    | 0.000±0.00   | 0.000±0.00   |
+|            | Gemini 2.5 Pro  | 0.408±0.38 | 0.625±0.44   | 51.333±17.91 | 0.000±0.00   | 0.000±0.00    | 2.667±3.77   | 30.667±20.42 |
+|            | Claude Sonnet 4 | 0.630±0.41 | 0.750±0.35   | 43.667±14.43 | 0.000±0.00   | 0.000±0.00    | 4.000±5.66   | 14.000±15.75 |
+|            | Grok 4.1 Fast   | 0.014±0.01 | 0.125±0.10   | 64.000±0.00  | 0.333±0.47   | 0.000±0.00    | 5.000±5.72   | 51.333±5.25  |
+| Easy       | Grok 4 Fast     | 0.029±0.04 | 0.125±0.18   | 64.000±0.00  | 0.000±0.00   | 0.333±0.47    | 14.333±13.27 | 42.000±10.61 |
+|            | o4-Mini         | 0.042±0.04 | 0.208±0.21   | 64.000±0.00  | 0.000±0.00   | 1.333±0.47    | 30.667±3.09  | 21.333±0.94  |
+|            | GLM 4.6V        | 0.000±0.00 | 0.000±0.00   | 64.000±0.00  | 0.000±0.00   | 63.333±0.47   | 0.000±0.00   | 0.000±0.00   |
+|            | GLM 4.5V        | 0.000±0.00 | 0.000±0.00   | 64.000±0.00  | 0.000±0.00   | 32.000±2.83   | 6.333±6.34   | 23.333±5.25  |
+|            | Qwen3-VL-235B   | 0.031±0.02 | 0.208±0.12   | 64.000±0.00  | 12.667±11.09 | 2.667±3.77    | 26.333±20.34 | 14.000±13.37 |
+|            | GPT-5           | 0.631±0.16 | 1.000±0.00   | 98.667±19.07 | 10.333±2.49  | 2.000±2.83    | 14.667±12.26 | 13.333±18.15 |
+|            | Gemini 3 Pro    | 0.793±0.24 | 0.958±0.06   | 90.000±38.88 | 20.000±27.58 | 1.000±0.82    | 4.667±4.11   | 0.000±0.00   |
+|            | Gemini 2.5 Pro  | 0.622±0.30 | 0.944±0.08   | 97.333±34.92 | 35.000±30.82 | 1.000±1.41    | 9.000±11.34  | 0.333±0.47   |
+|            | Claude Sonnet 4 | 0.005±0.00 | 0.056±0.02   | 144.000±0.00 | 0.000±0.00   | 44.333±0.47   | 8.000±3.74   | 82.667±2.87  |
+| Hard       | Grok 4.1 Fast   | 0.009±0.01 | 0.069±0.04   | 144.000±0.00 | 0.000±0.00   | 25.333±26.39  | 7.333±8.99   | 99.667±27.98 |
+|            | Grok 4 Fast     | 0.005±0.01 | 0.056±0.05   | 144.000±0.00 | 0.000±0.00   | 39.000±53.03  | 7.667±7.04   | 88.333±42.76 |
+|            | o4-Mini         | 0.005±0.00 | 0.056±0.02   | 144.000±0.00 | 0.000±0.00   | 1.333±1.25    | 32.000±13.49 | 98.333±13.47 |
+|            | GLM 4.6V        | 0.028±0.04 | 0.069±0.10   | 144.000±0.00 | 0.000±0.00   | 119.333±34.88 | 8.333±11.79  | 6.333±8.96   |
+|            | GLM 4.5V        | 0.000±0.00 | 0.000±0.00   | 144.000±0.00 | 0.000±0.00   | 111.667±0.94  | 0.667±0.94   | 30.667±2.05  |
+|            | Qwen3-VL-235B   | 0.135±0.12 | 0.347±0.27   | 144.000±0.00 | 4.333±6.13   | 63.000±1.41   | 30.667±14.06 | 21.000±19.87 |
+
+{17}------------------------------------------------
+
+## A Neuropsychologically Grounded Evaluation of LLM Cognitive Abilities
+
+<span id="page-17-0"></span>Table 14. Spatial Working Memory full results (image-text modality). All metrics are reported as mean with standard deviation (std) in a subscript.
+
+| Difficulty   | Model                                                                                                                                          | $\boldsymbol{S}_{swm}$                                                                                                                                                                                                    | Tokens Score                                                                                                                                                                                                            | Guesses                                                                                                                                                                                                                                                                          | Illegal                                                                                                                                                                                                                                                                     | Invalid                                                                                                                                                                                                                                                    | Repeated                                                                                                                                                                                                                               | No-box                                                                                                                                                                                                                                              |
+|--------------|------------------------------------------------------------------------------------------------------------------------------------------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+| Easy         | GPT-5<br>Gemini 3 Pro<br>Gemini 2.5 Pro<br>Claude Sonnet 4<br>Grok 4.1 Fast<br>Grok 4 Fast<br>o4-Mini<br>GLM 4.6V<br>GLM 4.5V<br>Qwen3-VL-235B | $\begin{array}{c} 0.986_{\pm .02} \\ 1.000_{\pm .00} \\ 0.733_{\pm .28} \\ 0.256_{\pm .18} \\ 0.145_{\pm .20} \\ 0.021_{\pm .03} \\ 0.055_{\pm .03} \\ 0.044_{\pm .05} \\ 0.044_{\pm .02} \\ 0.026_{\pm .05} \end{array}$ | $\begin{array}{c} 1.000_{\pm.00} \\ 1.000_{\pm.00} \\ 0.958_{\pm0.06} \\ 0.625_{\pm0.37} \\ 0.375_{\pm0.44} \\ 0.125_{\pm0.18} \\ 0.292_{\pm0.12} \\ 0.167_{\pm0.16} \\ 0.208_{\pm0.12} \\ 0.125_{\pm.10} \end{array}$  | $\begin{array}{c} 24.000_{\pm 2.16} \\ 20.667_{\pm 1.70} \\ 36.667_{\pm 19.60} \\ 59.333_{\pm 6.60} \\ 53.667_{\pm 14.61} \\ 64.000_{\pm 0.00} \\ 64.000_{\pm 0.00} \\ 64.000_{\pm 0.00} \\ 64.000_{\pm 0.00} \\ 64.000_{\pm 0.00} \end{array}$                                  | $\begin{array}{c} 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 1.000_{\pm 0.10} \\ 0.000_{\pm 0.10} \\ 1.000_{\pm 1.41} \end{array}$ | $\begin{array}{c} 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.667_{\pm 0.94} \\ 0.333_{\pm 0.47} \\ 0.333_{\pm 0.47} \\ 27.667_{\pm 26.04} \\ 32.333_{\pm 4.03} \\ 0.000_{\pm 0.00} \end{array}$ | $\begin{array}{c} 0.333_{\pm 0.47} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 1.667_{\pm 2.36} \\ 4.333_{\pm 5.44} \\ 0.000_{\pm 0.00} \\ 3.333_{\pm 3.40} \\ 6.333_{\pm 5.31} \\ 0.667_{\pm 0.47} \\ 5.333_{\pm 6.85} \end{array}$    | $\begin{array}{c} 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 14.000_{\pm 17.68} \\ 39.000_{\pm 12.36} \\ 40.333_{\pm 16.94} \\ 56.333_{\pm 2.49} \\ 49.000_{\pm 4.97} \\ 19.000_{\pm 13.74} \\ 24.333_{\pm 4.50} \\ 47.000_{\pm 7.48} \end{array}$     |
+| Easy - Notes | Gemini 2.5 Pro<br>Claude Sonnet 4<br>Grok 4.1 Fast<br>Grok 4 Fast<br>o4-Mini<br>GLM 4.6V<br>GLM 4.5V<br>Qwen3-VL-235B                          | $\begin{array}{c} 0.704_{\pm.42} \\ 0.207_{\pm.18} \\ 0.043_{\pm.06} \\ 0.029_{\pm.02} \\ 0.012_{\pm.01} \\ 0.082_{\pm.09} \\ 0.000_{\pm.00} \\ 0.318_{\pm.45} \end{array}$                                               | $\begin{array}{c} 0.792_{\pm .29} \\ 0.542_{\pm .41} \\ 0.167_{\pm .24} \\ 0.208_{\pm .06} \\ 0.125_{\pm .10} \\ 0.333_{\pm .29} \\ 0.000_{\pm .00} \\ 0.333_{\pm .47} \end{array}$                                     | $\begin{array}{c} 32.667_{\pm 22.17} \\ 60.667_{\pm 4.71} \\ 64.000_{\pm 0.00} \\ 64.000_{\pm 0.00} \\ 64.000_{\pm 0.00} \\ 64.000_{\pm 0.00} \\ 64.000_{\pm 0.00} \\ 50.000_{\pm 19.80} \end{array}$                                                                            | $\begin{array}{c} 0.000_{\pm 0.00} \\ 0.333_{\pm 0.47} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \end{array}$                                                             | $\begin{array}{c} 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 7.000_{\pm 5.35} \\ 0.333_{\pm 0.47} \\ 1.333_{\pm 0.94} \\ 3.333_{\pm 3.40} \\ 30.667_{\pm 1.25} \\ 1.667_{\pm 2.36} \end{array}$                                                               | $\begin{array}{c} 0.333_{\pm 0.47} \\ 0.667_{\pm 0.94} \\ 7.333_{\pm 10.37} \\ 5.333_{\pm 2.87} \\ 0.333_{\pm 0.47} \\ 16.000_{\pm 22.63} \\ 0.000_{\pm 0.00} \\ 0.667_{\pm 0.47} \end{array}$                                         | $\begin{array}{c} 14.667_{\pm 20.74} \\ 43.333_{\pm 11.03} \\ 42.667_{\pm 16.50} \\ 50.333_{\pm 5.56} \\ 56.667_{\pm 0.47} \\ 33.333_{\pm 19.34} \\ 32.667_{\pm 1.25} \\ 38.000_{\pm 26.98} \end{array}$                                            |
+| Hard         | GPT-5<br>Gemini 3 Pro<br>Gemini 2.5 Pro<br>Claude Sonnet 4<br>Grok 4.1 Fast<br>Grok 4 Fast<br>o4-Mini<br>GLM 4.6V<br>GLM 4.5V<br>Qwen3-VL-235B | $\begin{array}{c} 0.989_{\pm .02} \\ 0.963_{\pm .05} \\ 0.480_{\pm .12} \\ 0.265_{\pm .31} \\ 0.017_{\pm .02} \\ 0.002_{\pm .00} \\ 0.032_{\pm .02} \\ 0.010_{\pm .01} \\ 0.000_{\pm .00} \\ 0.016_{\pm .01} \end{array}$ | $\begin{array}{c} 1.000_{\pm.00} \\ 1.000_{\pm.00} \\ 0.972_{\pm0.04} \\ 0.472_{\pm0.38} \\ 0.097_{\pm0.14} \\ 0.028_{\pm0.02} \\ 0.181_{\pm0.10} \\ 0.042_{\pm0.06} \\ 0.000_{\pm0.00} \\ 0.111_{\pm0.10} \end{array}$ | $\begin{array}{c} 59.333_{\pm 5.79} \\ 66.000_{\pm 5.10} \\ 120.000_{\pm 18.49} \\ 122.333_{\pm 30.64} \\ 144.000_{\pm 0.00} \\ 144.000_{\pm 0.00} \\ 144.000_{\pm 0.00} \\ 144.000_{\pm 0.00} \\ 144.000_{\pm 0.00} \\ 144.000_{\pm 0.00} \\ 144.000_{\pm 0.00} \\ \end{array}$ | $\begin{array}{c} 0.667_{\pm 0.94} \\ 2.667_{\pm 3.77} \\ 52.000_{\pm 29.63} \\ 5.667_{\pm 7.32} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.333_{\pm 0.47} \end{array}$                                       | $\begin{array}{c} 0.000_{\pm 0.00} \\ 1.333_{\pm 0.47} \\ 0.333_{\pm 0.47} \\ 29.333_{\pm 20.74} \\ 10.000_{\pm 7.48} \\ 46.667_{\pm 58.39} \\ 2.000_{\pm 0.82} \\ 119.667_{\pm 34.41} \\ 110.333_{\pm 1.70} \\ 19.667_{\pm 0.47} \end{array}$             | $\begin{array}{c} 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.0667_{\pm 0.94} \\ 3.000_{\pm 0.82} \\ 1.000_{\pm 0.82} \\ 2.000_{\pm 2.83} \\ 11.000_{\pm 8.04} \\ 5.000_{\pm 7.07} \\ 0.000_{\pm 0.00} \\ 28.667_{\pm 6.60} \end{array}$ | $\begin{array}{c} 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 10.333_{\pm 13.91} \\ 52.667_{\pm 34.57} \\ 120.667_{\pm 14.64} \\ 89.000_{\pm 53.11} \\ 110.333_{\pm 13.07} \\ 13.333_{\pm 18.86} \\ 33.667_{\pm 1.70} \\ 80.667_{\pm 3.68} \end{array}$ |
+
+{18}------------------------------------------------
+
+#### E.3. Wisconsin Card Sorting Test (WCST)
+
+1014 1015 1016
+
+1024
+
+1026
+
+1028
+
+1030 1031
+
+1034
+
+1036
+
+1038
+
+The complete results for WCST tests are provided in Table [15](#page-18-0) for text modality, and in Table [16](#page-18-1) for image modality. All metrics are averaged over three runs.
+
+<span id="page-18-0"></span>*Table 15.* Wisconsin Card Sorting Test full results (text modality). All metrics are reported as mean with standard deviation (std) in a subscript.
+
+| Difficulty | Model           | Swcst     | Accuracy  | FMS       | PR        | Completed Rules | 1st Rule Trials |
+|------------|-----------------|-----------|-----------|-----------|-----------|-----------------|-----------------|
+| Easy       | GPT-5           | 0.848±.02 | 0.842±.02 | 0.000±.00 | 0.000±.00 | 6.000±0.00      | 6.000±0.00      |
+|            | Gemini 3 Pro    | 0.804±.05 | 0.815±.04 | 0.000±.00 | 0.003±.01 | 7.000±1.00      | 6.833±1.07      |
+|            | Gemini 2.5 Pro  | 0.825±.02 | 0.819±.02 | 0.000±.00 | 0.000±.00 | 6.000±0.00      | 6.667±0.94      |
+|            | Claude Sonnet 4 | 0.764±.06 | 0.786±.03 | 0.000±.00 | 0.027±.02 | 6.000±0.00      | 7.000±0.82      |
+|            | Grok 4.1 Fast   | 0.841±.03 | 0.835±.03 | 0.000±.00 | 0.000±.00 | 6.000±0.00      | 6.000±0.00      |
+|            | Grok 4 Fast     | 0.836±.02 | 0.834±.02 | 0.000±.00 | 0.000±.00 | 6.000±0.00      | 6.667±0.47      |
+|            | o4-Mini         | 0.789±.01 | 0.805±.03 | 0.020±.03 | 0.011±.02 | 6.000±0.00      | 6.333±0.47      |
+|            | GLM 4.6V        | 0.300±.35 | 0.343±.33 | 0.305±.43 | 0.115±.09 | 2.333±2.62      | 29.000±25.66    |
+|            | GLM 4.5V        | 0.773±.03 | 0.789±.02 | 0.000±.00 | 0.019±.03 | 6.000±0.00      | 7.000±0.00      |
+|            | Qwen3-VL-235B   | 0.230±.14 | 0.243±.15 | 0.766±.17 | 0.731±.16 | 2.667±1.89      | 43.000±39.05    |
+| Hard       | GPT-5           | 0.815±.00 | 0.831±.02 | 0.000±.00 | 0.000±.00 | 8.000±0.00      | 6.667±0.47      |
+|            | Gemini 3 Pro    | 0.776±.06 | 0.803±.05 | 0.000±.00 | 0.007±.01 | 8.000±0.00      | 7.333±1.25      |
+|            | Gemini 2.5 Pro  | 0.806±.04 | 0.827±.01 | 0.000±.00 | 0.000±.00 | 8.000±0.00      | 7.000±0.00      |
+|            | Claude Sonnet 4 | 0.773±.06 | 0.805±.02 | 0.000±.00 | 0.007±.01 | 8.000±0.00      | 7.333±1.25      |
+|            | Grok 4.1 Fast   | 0.825±.03 | 0.835±.01 | 0.000±.00 | 0.009±.02 | 8.000±0.00      | 6.222±0.42      |
+|            | Grok 4 Fast     | 0.754±.04 | 0.768±.03 | 0.000±.00 | 0.007±.01 | 8.000±0.00      | 6.667±0.94      |
+|            | o4-Mini         | 0.792±.02 | 0.798±.02 | 0.000±.00 | 0.000±.00 | 8.000±0.00      | 6.667±0.47      |
+|            | GLM 4.6V        | 0.095±.12 | 0.175±.08 | 0.426±.41 | 0.214±.14 | 0.889±1.29      | 63.111±39.82    |
+|            | GLM 4.5V        | 0.309±.08 | 0.184±.09 | 0.000±.00 | 0.680±.04 | 3.000±1.63      | 6.667±0.47      |
+|            | Qwen3-VL-235B   | 0.233±.12 | 0.438±.12 | 0.475±.16 | 0.505±.11 | 3.333±1.25      | 9.667±4.50      |
+
+<span id="page-18-1"></span>*Table 16.* Wisconsin Card Sorting Test full results (image modality). All metrics are reported as mean with standard deviation (std) in a subscript.
+
+| Difficulty | Model           | Swcst     | Accuracy  | FMS       | PR        | Completed Rules | 1st Rule Trials |
+|------------|-----------------|-----------|-----------|-----------|-----------|-----------------|-----------------|
+| Easy       | GPT-5           | 0.824±.05 | 0.833±.03 | 0.000±.00 | 0.010±.01 | 6.000±0.00      | 6.000±0.00      |
+|            | Gemini 3 Pro    | 0.861±.00 | 0.857±.00 | 0.000±.00 | 0.000±.00 | 6.000±0.00      | 6.000±0.00      |
+|            | Gemini 2.5 Pro  | 0.823±.04 | 0.820±.04 | 0.000±.00 | 0.000±.00 | 6.000±0.00      | 7.000±0.82      |
+|            | Claude Sonnet 4 | 0.800±.02 | 0.797±.03 | 0.000±.00 | 0.010±.01 | 6.000±0.00      | 7.000±0.00      |
+|            | Grok 4.1 Fast   | 0.107±.07 | 0.432±.08 | 0.339±.24 | 0.324±.04 | 1.000±0.82      | 37.667±22.45    |
+|            | Grok 4 Fast     | 0.336±.08 | 0.602±.09 | 0.432±.10 | 0.201±.10 | 3.500±1.50      | 9.500±0.50      |
+|            | o4-Mini         | 0.347±.11 | 0.557±.03 | 0.320±.04 | 0.092±.05 | 3.333±0.94      | 7.000±0.82      |
+|            | GLM 4.6V        | 0.191±.11 | 0.297±.08 | 0.089±.13 | 0.442±.15 | 1.333±1.25      | 26.000±27.58    |
+|            | GLM 4.5V        | 0.108±.11 | 0.315±.22 | 0.404±.32 | 0.310±.21 | 0.333±0.47      | 25.667±13.22    |
+|            | Qwen3-VL-235B   | 0.420±.19 | 0.361±.09 | 0.229±.22 | 0.619±.09 | 4.000±0.82      | 7.333±1.89      |
+| Hard       | GPT-5           | 0.794±.05 | 0.819±.00 | 0.000±.00 | 0.000±.00 | 8.000±0.00      | 7.000±0.82      |
+|            | Gemini 3 Pro    | 0.806±.02 | 0.814±.04 | 0.000±.00 | 0.000±.00 | 8.000±0.00      | 6.667±0.47      |
+|            | Gemini 2.5 Pro  | 0.776±.01 | 0.795±.03 | 0.000±.00 | 0.000±.00 | 8.000±0.00      | 7.333±0.47      |
+|            | Claude Sonnet 4 | 0.710±.08 | 0.739±.06 | 0.134±.12 | 0.069±.06 | 8.000±0.00      | 6.667±0.47      |
+|            | Grok 4.1 Fast   | 0.077±.05 | 0.369±.09 | 0.376±.26 | 0.420±.07 | 0.714±0.70      | 44.143±20.58    |
+|            | Grok 4 Fast     | 0.209±.07 | 0.493±.04 | 0.347±.25 | 0.309±.03 | 3.000±1.41      | 11.333±2.36     |
+|            | o4-Mini         | 0.277±.13 | 0.528±.05 | 0.437±.09 | 0.352±.07 | 3.889±1.37      | 9.222±3.58      |
+|            | GLM 4.6V        | 0.064±.07 | 0.197±.09 | 0.169±.32 | 0.420±.12 | 0.444±0.96      | 64.556±26.88    |
+|            | GLM 4.5V        | 0.083±.07 | 0.398±.22 | 0.303±.25 | 0.432±.11 | 0.333±0.47      | 32.000±4.97     |
+|            | Qwen3-VL-235B   | 0.071±.07 | 0.208±.13 | 0.500±.38 | 0.794±.14 | 1.333±1.25      | 59.333±38.77    |
+
+{19}------------------------------------------------
+
+## F. Prompts
+
+1054
+
+1056
+
+1058
+
+1060 1061
+
+1068 1069
+
+1071
+
+1074 1075 1076
+
+1078
+
+1091
+
+Below we list the system prompts used for each task. We also provide alternate or additional prompts used for the different setups within each task.
+
+## System Prompt for Image RAPM
+
+You are taking the Raven's Progressive Matrices (RAPM) test, a non-verbal intelligence test that measures abstract reasoning ability.
+
+You will see a 3x3 matrix of images with the bottom-right image missing (shown as a question mark), followed by 8 answer choices numbered 1-8.
+
+Your task is to:
+
+- 1. Analyze rows and columns
+- 2. Infer the governing logical rule(s)
+- 3. Select the answer choice (1-8) that correctly completes the matrix.
+
+#### /\* PATTERN HINT \*/
+
+The patterns can involve:
+
+- Shape transformations (rotation, reflection, scaling)
+- 1062 - Position changes (movement, arrangement)
+- 1063 - Attribute changes (color, size, number of elements)
+- 1064 - Logical operations (addition, subtraction, intersection)
+  - Sequence progressions (systematic changes across rows/columns)
+- 1065 Additional common rule types:
+- 1066 1067 - Constant-in-row: Same value across a row; varies down columns.
+  - Quantitative step: Fixed +/- increment between adjacent cells (size / count / position offset).
+  - Figure add/subtract: Combine (overlay or juxtapose) or remove elements from two cells to form the third.
+  - Distribution-of-three: Three distinct categorical values appear once each per row (order may permute).
+  - Distribution-of-two: Two values each appear once; third slot is empty / null.
+    - Look horizontally and vertically; the missing piece must satisfy ALL relevant row and column rules.
+
+Your final answer should be a number between 1-8 corresponding to the correct choice.
+
+#### System Prompt for Text RAPM
+
+You are solving a TEXT-BASED 3x3 pattern matrix (Raven-style). Each cell contains a string; the bottom-right cell is missing ('?').
+
+Goal: Infer the rule(s) acting across rows and columns.
+
+#### /\* PATTERN HINT \*/
+
+Possible dimensions (one or more):
+
+- Character set restriction (digits / letters / symbols)
+- Quantitative constant (exact length / count / unique)
+- Quantitative progression (arithmetic step across row/column)
+- Parity / multiple rules (all even / all odd / multiples of N)
+- Positional constraints (first/last/even/odd positions restricted)
+- Ordering (ascending / descending / mixed)
+- Layered combinations (e.g. constant + parity, progression + positional)
+
+#### 1089 1090 /\* MC VERSION \*/
+
+You will be given 8 answer options (1-8). Select the single option that correctly fills the missing cell while satisfying ALL inferred row and column constraints.
+
+1092 Respond with <answer>NUMBER</answer> using just the chosen option number.
+
+#### 1093 1094 /\* GEN VERSION \*/
+
+You must GENERATE the exact missing cell string that satisfies ALL inferred row and column constraints.
+
+Respond with <answer>STRING</answer> containing only the candidate string (no quotes or extra text).
+
+{20}------------------------------------------------
+
+#### A Neuropsychologically Grounded Evaluation of LLM Cognitive Abilities System Prompt for SWM You will be performing the Spatial Working Memory task. You will be given an image containing {N BOXES} yellow boxes in a grid. There are {N TOKENS} types of tokens, hidden in any one of {N BOXES} boxes. Each token type is represented by a distinct color. Your goal is to find the {N TOKENS} types of tokens {N BOXES} times each, by repeatedly selecting a box to open. A box can contain multiple types of tokens, but only one token of each type. If the box contains multiple tokens, a token with mixed colors corresponding to the tokens will be shown. Once the token is found, another will be generated in another box. The token will be generated in a box that has never contained a token of that type before in the trial. The token may be generated in a box that has been opened and found empty before, as long as it never contained that type of token previously. /\* TEXT VERSION \*/ Your final answer should be a number from 1-{N BOXES}, the index of the box you selected /\* IMAGE VERSION \*/ Your final answer should be a grid coordinate (x, y), the grid coordinate of the box you choose System Prompt for WCST You are performing the Wisconsin Card Sorting Test (WCST). You will be shown a given card with a symbol on it, and you will have to match it to one of four option cards according to an attribute that you have to figure out. The cards will be described by the following attributes: 1. Number of symbols 2. Color of symbols 3. Shape of symbols /\* HARD VERSION \*/ 4. Background color of the card You will be told "Correct!" if you are correct and "Incorrect. Please try again." if you are incorrect. If you are incorrect, you either made a mistake or the rule has changed. If you believe you have made a mistake, correct it and try again. If you believe the rule has changed, you have to figure out the correct rule to match the cards. If you are correct, you have to stick with the same attribute until you are incorrect. There is always a true answer in the task, and you have to keep performing the task until the end of the test. Your final answer should be a number between 1-4 corresponding to the index of the card you think is the correct match.
+
+{21}------------------------------------------------
+
+## G. Turn Examples
+
+*[. . . omitted turns . . . ]*
+
+User: No tokens found in box (3, 3). Answer concisely. Think step-by-step, utilizing information from previous feedbacks, and state your reasoning in maximum 4000 tokens, wrapped with <think> and </think>. Then, provide a really short summary of your reasoning after the closing </think> tag.
+
+Which of the 8 boxes would you like to open?
+
+Your final answer should be a grid coordinate (x, y), wrapped with <answer> and </answer>
+
+![](_page_21_Figure_34.jpeg)
+
+Assistant: <think>I already found one, located at position (1, 0). [. . . ] I know that boxes (5, 0), (5, 1), (1, 2), and (3, 3) are empty. [. . . ] That leaves me with (10, 4), (0, 6), and (9, 6). To keep things organized and ensure I'm systematic, I'm going to follow my pre-defined search pattern – working from the top-left and moving to the right and down. That means the next box I need to check is (10, 4). [. . . ] <answer>(10, 4)</answer>
+
+*[. . . omitted turns . . . ]*
+
+User: No tokens found in box 1. Boxes that has contained token A: 11, Boxes that has contained token B: 8, 10,
+
+Opened boxes: 12,
+
+Answer concisely. Think step-by-step, utilizing information from previous feedbacks, and state your reasoning in maximum 4000 tokens, wrapped with <think> and </think>. Then, provide a really short summary of your reasoning after the closing </think> tag.
+
+Which of the 12 boxes would you like to open?
+
+Your final answer should be a box number, wrapped with <answer> and </answer>
+
+Assistant: [. . . ]The message provides data on boxes: 11 contained token A, 8 and 10 contained token B, and 12 boxes were opened. [. . . ] The phrase "Opened boxes" suggests the total number of boxes accessed, with partial information on token contents.[. . . ] The phrase "Opened boxes: 12" likely means box 12 was last opened and found empty, given prior context.[. . . ]
+
+\*\*Summary:\*\* 1 empty again; pure empties left 2-7,9,12; systematic next=2.
+
+<answer>2</answer>
+
+{22}------------------------------------------------
+
+*[. . . omitted turns . . . ]*
+
+User: Incorrect. Please try again.
+
+Recent notes:
+
+- Turn -1: matching number of symbols and symbol color Incorrect.
+- Turn -2: matching number of symbols and symbol color Incorrect.
+- Turn -3: matching symbol color Correct.
+- Turn -4: matching symbol color Correct.
+- Turn -5: matching symbol color Correct.
+- Turn -6: matching symbol color Correct.
+
+Look at the image showing 5 cards. Match the 'Given' card to one of cards 1-4 based on the rule you need to figure out.
+
+![](_page_22_Picture_32.jpeg)
+
+Assistant: The previous attempts to match by symbol color (Turns 2-6) were correct until Turn 7, where the response was incorrect. [. . . ] However, the last incorrect attempt used "number of symbols and symbol color," which was wrong. The new rule might now be \*\*background color\*\*.[. . . ] Re-evaluating: The prior correct streak (Turns 2-6) used symbol color. After an incorrect response, the rule often shifts to \*\*number of symbols\*\*. [. . . ] Wait, the feedback says "matching number of symbols and symbol color – Incorrect" for Turn 7. The correct rule might now be \*\*background color\*\*. [. . . ]The rule shifted to background color; Given card has a purple background, matching Card 3.
+
+<answer>3</answer>
+
+*[. . . omitted turns . . . ]*
+
+User: Incorrect. Please try again.
+
+Given: two red triangle
+
+Options:
+
+- 1. two green triangle
+- 2. four yellow square
+- 3. three blue star
+- 4. one red circle
+
+Assistant: [. . . ] My initial hypothesis (number and shape) was incorrect. The feedback suggests a different rule. The given card is red. Option 4 is also red, while others have different colors. I'll now hypothesize the rule is matching by color regardless of number or shape.<answer>4</answer>My initial hypothesis (number and shape) was incorrect. The feedback suggests a different rule. The given card is red. Option 4 is also red, while others have different colors. I'll now hypothesize the rule is matching by color regardless of number or shape.

icml26/a4461009-05b7-42b6-b207-5e6e0c2e0731/parse_report.json

@@ -0,0 +1,78 @@
+{
+  "paper_id": "a4461009-05b7-42b6-b207-5e6e0c2e0731",
+  "pipeline": "marker_non_llm_v3",
+  "parser": "marker_single",
+  "formats": [
+    "markdown",
+    "chunks"
+  ],
+  "llm_enabled": false,
+  "pdf_path": "/network/scratch/j/jianan.zhao/ReviewAgent/data/processed_papers/icml26_20260429_1952_duequeue/raw/a4461009-05b7-42b6-b207-5e6e0c2e0731.pdf",
+  "pdf_sha256": "e66a6229bfecbc1d417f98752cb6af183d3b62150e86069cdc44ff6ab5e583fb",
+  "bytes": 487832,
+  "source": "https://koala.science/storage/pdfs/a4461009-05b7-42b6-b207-5e6e0c2e0731.pdf",
+  "page_count": 23,
+  "ok": true,
+  "elapsed_seconds": 163.32,
+  "paper2markdown_v3": {
+    "paper_id": "a4461009-05b7-42b6-b207-5e6e0c2e0731",
+    "pipeline": "Paper2Markdown-V3",
+    "ok": true,
+    "page_count": 23,
+    "chunk_count": 178,
+    "main_body_chunk_count": 176,
+    "appendix_chunk_count": 2,
+    "reference_chunk_count": 0,
+    "model_text_chars": 75805,
+    "raw_markdown_chars": 95965,
+    "sanitized_chars": 85481,
+    "page_provenance": {
+      "min_page": 1,
+      "max_page": 23,
+      "invalid_count": 0
+    },
+    "marker_block_type_counts": {
+      "Caption": 1,
+      "Code": 2,
+      "Equation": 7,
+      "Figure": 1,
+      "FigureGroup": 2,
+      "Footnote": 10,
+      "ListGroup": 17,
+      "PageFooter": 20,
+      "PageHeader": 32,
+      "Picture": 1,
+      "SectionHeader": 64,
+      "Table": 3,
+      "TableGroup": 13,
+      "Text": 421
+    },
+    "asset_count_raw": 4,
+    "asset_count_model_kept": 4,
+    "asset_count_rejected": 0,
+    "asset_reject_reasons": {
+      "kept": 4
+    },
+    "artifact_leak_audit": {
+      "ok": true,
+      "hits": {
+        "Anonymous Authors": [],
+        "ACKNOWLEDGMENT": [],
+        "OpenReview": [],
+        "\"accept_label\"": [],
+        "\"decision\"": [],
+        "\"decision_tier\"": [],
+        "\"source_status\"": [],
+        "Meta-review": [],
+        "Official Review": [],
+        "official_reviews": [],
+        "meta_reviews": [],
+        "suggested_verdict_score": []
+      },
+      "artifact_count": 2
+    },
+    "default_model_input": "model_text_v3.txt",
+    "appendix_input": "appendix_text_v3.txt",
+    "reference_input": "reference_text_v3.txt"
+  }
+}

icml26/a4461009-05b7-42b6-b207-5e6e0c2e0731/reference_text_v3.txt

@@ -0,0 +1 @@
+

icml26/a4461009-05b7-42b6-b207-5e6e0c2e0731/sanitization_report.json

@@ -0,0 +1,61 @@
+{
+  "paper_id": "a4461009-05b7-42b6-b207-5e6e0c2e0731",
+  "pipeline": "Paper2Markdown-V3",
+  "ok": true,
+  "page_count": 23,
+  "chunk_count": 178,
+  "main_body_chunk_count": 176,
+  "appendix_chunk_count": 2,
+  "reference_chunk_count": 0,
+  "model_text_chars": 75805,
+  "raw_markdown_chars": 95965,
+  "sanitized_chars": 85481,
+  "page_provenance": {
+    "min_page": 1,
+    "max_page": 23,
+    "invalid_count": 0
+  },
+  "marker_block_type_counts": {
+    "Caption": 1,
+    "Code": 2,
+    "Equation": 7,
+    "Figure": 1,
+    "FigureGroup": 2,
+    "Footnote": 10,
+    "ListGroup": 17,
+    "PageFooter": 20,
+    "PageHeader": 32,
+    "Picture": 1,
+    "SectionHeader": 64,
+    "Table": 3,
+    "TableGroup": 13,
+    "Text": 421
+  },
+  "asset_count_raw": 4,
+  "asset_count_model_kept": 4,
+  "asset_count_rejected": 0,
+  "asset_reject_reasons": {
+    "kept": 4
+  },
+  "artifact_leak_audit": {
+    "ok": true,
+    "hits": {
+      "Anonymous Authors": [],
+      "ACKNOWLEDGMENT": [],
+      "OpenReview": [],
+      "\"accept_label\"": [],
+      "\"decision\"": [],
+      "\"decision_tier\"": [],
+      "\"source_status\"": [],
+      "Meta-review": [],
+      "Official Review": [],
+      "official_reviews": [],
+      "meta_reviews": [],
+      "suggested_verdict_score": []
+    },
+    "artifact_count": 2
+  },
+  "default_model_input": "model_text_v3.txt",
+  "appendix_input": "appendix_text_v3.txt",
+  "reference_input": "reference_text_v3.txt"
+}

icml26/a4461009-05b7-42b6-b207-5e6e0c2e0731/sanitized_v3.txt

@@ -0,0 +1,634 @@
+{0}
+## Abstract
+Large language models (LLMs) display a unified "general factor" of capability across 10 benchmarks—a finding confirmed by our factor analysis of 156 models—yet they still struggle with simple, trivial tasks for humans. This is because current benchmarks focus on task completion, failing to probe the foundational cognitive abilities that highlight these behaviors. We address this by introducing the NeuroCognition benchmark, grounded in three adapted neuropsychological tests: Raven's Progressive Matrices (abstract relational reasoning), Spatial Working Memory (maintenance and systematic search), and the Wisconsin Card Sorting Test (cognitive flexibility). Our evaluation reveals that while models perform strongly on text, their performance degrades for images and with increased complexity. Furthermore, we observe that complex reasoning is not universally beneficial, whereas simple, humanlike strategies yield partial gains. We also find that NeuroCognition correlates positively with standard general-capability benchmarks, while still measuring distinct cognitive abilities beyond them. Overall, NeuroCognition emphasizes where current LLMs align with human-like intelligence and where they lack core adaptive cognition, showing the potential to serve as a verifiable, scalable source for improving LLMs.
+## 1. Introduction
+Large language models (LLMs) show increasingly powerful capabilities (?), further enhanced by the emergence of reasoning abilities (?). This is demonstrated by their beyond human-level capabilities in tasks such as olympiad-level problems (?). However, even with these impressive capabilities, they still stumble on trivial or straightforward tasks
+(??). This raises a curiosity about the underlying cognitive mechanism of LLMs and how it compares to human cognitive abilities.
+Most current LLM benchmarks are task-specific, limiting their analysis and insights to the task performance itself. These benchmarks are not able to uncover the underlying cognitive abilities required to solve the task (??) because task performance is not a reliable measure of intelligence, but rather heavily influenced by prior knowledge and experience (?). One interesting test that is often overlooked is neuropsychological tests, which are a well-established method for measuring specific cognitive abilities in humans, originally designed to isolate specific cognitive functions using controlled stimuli and interpretable outcomes. Previous works have attempted to apply neuropsychological tests to LLMs (??), but they are limited to a single task or modality, or they lack a deeper analysis of the model's underlying cognitive abilities.
+We present the NeuroCognition benchmark, built from three well-established neuropsychological tests: Raven's Progressive Matrices (????) (RPM) targeting abstract relational reasoning (fluid intelligence), Spatial Working Memory (?) (SWM) targeting maintenance and systematic search across working memory, and the Wisconsin Card Sorting Test (??) (WCST) targeting cognitive flexibility and task switching. Each test is adapted into multiple text and image variants (including harder versions) to vary interaction length, modality, and complexity. We also include experiments using assistance inspired by simple strategies that humans use (e.g., taking notes), a process known as cognitive offloading (??), to see if they reduce observed errors and improve overall performance.
+Our main contributions are:
+- NeuroCognition, a benchmark repurposed from wellestablished neuropsychological tests, into a simple, multimodal, scalable, robust assessment, and easily reproducible suite with various setups.
+- Evaluation results show that state-of-the-art models perform strongly in baseline text settings, but struggle in image or when task complexity increases, and that explicit reasoning is not always beneficial.
+- Experiments with human-like strategies (pattern
+{1}------------------------------------------------
+<span id="page-1-1"></span>![](_page_1_Figure_1.jpeg)
+Figure 1. The three neuropsychological tests used in our work.
+hints and notes) show mixed effects—limited on RAPM/SWM but more consistent on WCST—suggesting some deficits are fundamental.
+Factor analysis and correlation results show that NeuroCognition measures specific cognitive abilities that differ from, yet correlate with, the general capabilities captured by standard benchmarks.
+Overall, NeuroCognition is a benchmark that can identify gaps in LLMs' cognitive abilities relative to humans, offering insights into how LLMs operate and informing the design and training of future LLMs. Our code<sup>1</sup>, data, and results will be made publicly available.
+#### 2. Related Works
+The neuropsychological tests included in our work measure different aspects of human cognitive abilities. The RPM (????) measures abstract, relational reasoning capabilities (?). RPM has been used to identify intellectually gifted individuals, and has been shown to be widely reliable across different geographical boundaries (?). The SWM test (?), on the other hand, measures the working memory, the ability to hold and mentally work with information, which may no longer be perceptually present (?). This test has been used to detect at-risk individuals for psychosis (?), and highlight the working memory deficits in individuals with autism (?). Finally, WCST (??) measures cognitive flexibility, i.e., the
+ability to adapt to a change in perspective and reallocate resources to process relevant information (??). Patients with schizophrenia are shown to perform worse on this test (?), and are shown to be robust even in clinical settings (?).
+Recent advances in LLMs have led many works seeking to establish stronger correlations between LLMs and human intelligence. Many works focus on developing a cognitive benchmark aiming to measure fluid intelligence (????), emphasizing abstractness and dynamic evaluation to avoid the data contamination issue (?). Other works applied human cognitive tests to LLMs (?) and showcased cognitive behaviors in LLMs, such as similar working memory capacity between humans and GPT-4 in the n-back test (?), and above human-level cognitive flexibility in the WCST (?). Some works also revealed common factors of models' performance across different benchmarks (??), indicating the existence of a common underlying intelligence.
+Aside from cognitive abilities, several studies have also quantified specific areas where LLMs struggle. In multi-turn evaluation, LLMs tend to perform worse than single-turn evaluations (?), and still perform sub-optimally in multi-turn tasks (??). LLMs also struggle in longer multi-step reasoning (?) and in complex multi-stage tasks (?). LLMs also exhibit cross-modal hallucination (?), and struggle in multi-image, multi-turn benchmarks (??). LLMs also fail in simple pattern tasks using modified alphabets (?), indicating the lack of general analogical reasoning in LLMs.
+<span id="page-1-0"></span><sup>&</sup>lt;sup>1</sup>Code available at
+{2}------------------------------------------------
+## 3. NeuroCognition Benchmark Tests
+#### 3.1. Raven's Progressive Matrices
+Task. RPM presents a 3x3 matrix of construct patterns with the bottom-right cell missing. The participant needs to infer abstract relations across rows and columns, and fill the missing cell by either selecting one of the 8 available options or constructing it. In our benchmark, we use the Raven's Advanced Progressive Matrices (RAPM) (?) form.
+Setup. For the image version, we use the RAVEN dataset (?), implementation of RAPM with 7 different figure configurations licensed with GPL-3.0. We take 20 items per configuration, resulting in a total of 140 test data. We combine the 3x3 question matrix with the 8 available options into a single image input, as shown in Figure [1,](#page-1-1) and present it to the model in multiple-choice (MC) format. For the text setting, we programmatically generate symbolic matrices using character and string attributes (see Appendix [A](#page-9-0) for generation details). We provide two answer formats for text setting: multiple-choice (MC, same 8 options) and generative (Gen, free-form string). A stored rule representation enables automatic validation of Gen outputs (see Appendix [B](#page-11-0) for validation details). We generate 200 test data for the text setting with random configurations.
+Metrics. We report accuracy for both MC and Gen. Average human performance for images is 0.814, reported by ?, and not yet established for the text variant.
+#### <span id="page-2-0"></span>3.2. Spatial Working Memory
+Task. In the SWM test, participants are presented with an image with several boxes. A token is hidden in one of the boxes, and participants are asked to search for it by elimination, opening one box at a time. After the token is found, it is regenerated in another box that has not previously contained it in the current trial. This is repeated until all boxes have contained all tokens, with the goal of locating all tokens with minimal redundant checking and without revisiting boxes unnecessarily.
+Setup. We implement two difficulty levels. Easy: 8 boxes, 1 token active at a time, early stop after all 8 tokens found or a maximum of 64-guess cap. Hard: 12 boxes with 2 distinct tokens active concurrently (24 tokens found required, maximum 144 guesses), with each token directly reappearing after being found, and one box can contain both tokens at once. We experiment with three types of feedback setups: text-only, image-only (a grid with no additional textual information), and image+text variants. The image input uses a grid system for easier answer verification, as shown in Figure [1.](#page-1-1) The text-only setup identifies the boxes by number and ask the model to choose a box number.
+Metrics. Since SWM is highly affected by luck, we define
+some metrics that focus on penalizing errors and unfinished tests. We classify three avoidable error types:
+- illegal: opens a box that can no longer contain a token because it has already yielded all available tokens for the trial.
+- no-box: refers to a box index/identifier that does not exist.
+- repeated: re-opens a box that was previously opened and found empty, before any new token was regenerated (i.e., no state change that could justify revisiting).
+These three error types sum to nerr. Let nvalid be the total valid guesses. A consistency term Sc:
+$$S_c = 1 - \frac{n_{err}}{n_{valid}} \tag{1}$$
+measures avoidance of structurally unjustified moves. Note that we omit *invalid* guesses (no parseable answer) to ensure fair evaluation of real guessing attempts. Let T<sup>f</sup> be number of tokens found out of T required; the final score Sswm is
+$$S_{swm} = \frac{T_f}{T} \times S_c \tag{2}$$
+Human worst-case. A cognitively normal human, given a complete history of their own moves (like conversation history in LLMs), will be able to achieve full completion by systematic elimination and yield nerr = 0 regardless of the setup and difficulty, therefore obtaining Sswm = 1.
+#### <span id="page-2-1"></span>3.3. Wisconsin Card Sorting Test
+Task. In WCST, four option cards containing shapes are shown to participants. They are then asked to match the given card to one of the four option cards using one of three rules: shape, color, or number. Importantly, participants are not informed of the correct rule for matching the cards and must eliminate potential rules to find the correct one. After choosing the correct answer, a different card is shown while maintaining the same rule. The matching rule is considered complete after a fixed number of consecutive correct answers and is changed upon completion without expressly informing the participant. The test finishes after a fixed number of rules are completed or a maximum number of guesses is reached.
+Setup. We require 5 consecutive correct matches (c = 5) to complete a rule instance. We define two difficulty levels. Easy: 3 candidate rules (color, shape, number), each repeated twice (6 total rule instances) or until the 64-guess cap. Hard: adds background color (4 candidate rules, 8 total rule instances) with a maximum of 96 guesses. We also construct controlled ambiguity variants for the hard setting: (i) *off* (no ambiguous cards), (ii) *first* (only the first card
+{3}------------------------------------------------
+under a new rule is ambiguous), and (iii) *rest* (all subsequent cards after the first are ambiguous). Ambiguity occurs when multiple attributes match simultaneously, forcing the model to track disambiguating feedback over turns. The image-input setting provides images of the cards (Figure 1), while the text-only setting lists descriptions of the option cards.
+**Metrics.** Each completed rule i uses $g_i$ guesses from the start of that rule until c consecutive correct response. Therefore, we can define a score $S_{r_i}$ for each completed rule i:
+$$S_{r_i} = \frac{c}{q_i}. (3)$$
+where unfinished rules contribute zero. Lower $g_i$ means faster hypothesis convergence, which repeated reliance on an incorrect hypothesis inflates $g_i$ and lowers $S_{r_i}$ . With $n_c$ completed out of N target rule instances, we can calculate the final score $S_{west}$ as:
+$$S_{wcst} = \frac{1}{N} \sum_{i=1}^{n_c} S_{r_i}.$$
+(4)
+We additionally report two specific error rates aggregated over a single trial with total turns L. Let $y_t \in \{0,1\}$ denote correctness at turn t ( $y_t = 1$ if the model's match is correct), and let $r_t \in \mathcal{R}$ be the rule the model is currently following at turn t (e.g., color/shape/number/background).
+**Perseverative Response (PR).** PR measures how often the model re-selects a rule that has already been ruled out by negative feedback within the current rule block. We use PR rather than the standard WCST perseverative error (PE) because it captures a broader range of errors while still including PE.
+Let k(t) index the current (ground-truth) rule block containing turn t, and let $\mathcal{R}^t_{\neg} \subseteq \mathcal{R}$ be the set of rules eliminated by feedback from earlier turns in the same block. Then:
+$$PR = \frac{\sum_{t=1}^{L} \mathbf{1} \{ \mathcal{R}_{\neg}^{t} \neq \emptyset \} \mathbf{1} \{ r_{t} \in \mathcal{R}_{\neg}^{t} \}}{\sum_{t=1}^{L} \mathbf{1} \{ \mathcal{R}_{\neg}^{t} \neq \emptyset \}}.$$
+(5)
+**Failure-to-maintain-set (FMS).** We mark a rule as *conceptually acquired* after m=3 consecutive correct responses within a block. For each block k, let $\tau_k$ be the first turn index in that block at which this acquisition criterion is met (if it is never met, the block contributes no turns). Let $\mathcal{T}_{post}$ be the set of all turns after acquisition and before the next rule switch:
+$$\mathcal{T}_{\text{post}} = \bigcup_{k} \{ t : k(t) = k, \ t > \tau_k \}. \tag{6}$$
+FMS is the error rate on these post-acquisition turns:
+$$FMS = \frac{\sum_{t \in \mathcal{T}_{post}} \mathbf{1}\{y_t = 0\}}{|\mathcal{T}_{post}|}.$$
+(7)
+**Human worst-case.** For the easy setup, a cognitively normal human needs at most two exploratory wrong guesses to find the correct rule, then continues with 5 correct answers $\Rightarrow 5/7 \approx 0.714$ . For hard, the worst case happened with the *first* ambiguity setup, where at most there are two exploratory wrong guesses, one ambiguous early correct, one disambiguating wrong, and 5 consecutive correct answers $\Rightarrow 5/9 \approx 0.556$ .
+#### <span id="page-3-0"></span>3.4. Experiment Setup
+In our experiments, we benchmark several recent multimodal reasoning models: proprietary models (Gemini 3 Pro (?), Gemini 2.5 Pro (?), GPT-5 (?), o4-Mini (?), Claude Sonnet 4 (?), Grok 4.1 Fast (?), Grok 4 Fast (?) and openweight models (GLM 4.6V (?), GLM 4.5V (?), Qwen3-VL-235B (????)).
+We use models' default reasoning capabilities via API parameters and chain-of-thought prompts. However, we disable reasoning for Claude Sonnet 4 and Grok 4 Fast on the RAPM test, due to overthinking issues that exhausted output limits and led to worse performance. For RAPM, we set thinking budgets close to the model's maximum output tokens. For SWM and WCST, we limit the output tokens to 8,192 tokens and the thinking budget to 4,000 tokens per turn. We repeat all SWM and WCST experiments three times, while RAPM is evaluated once. All evaluations are zero-shot with a temperature of 0 (if applicable). Further details are provided in Appendix C.
+#### 4. Results
+Table 1 reports the baseline results for NeuroCognition, with the complete results available in Appendix E.
+#### 4.1. Key Findings
+Models are stronger in text, struggle in image. Across all tests, models show a consistent advantage in text-only setups compared to image-based inputs. Even combined modalities (image+text) yield worse results than text alone, indicating difficulties in parsing and reasoning over visual stimuli and combining visual with textual information.
+**Performance drops with difficulty.** Increasing task difficulty, particularly in SWM, causes a significant drop in performance. In the SWM text easy setup almost all models complete the task perfectly, while in the hard setting only three models manage to do so. Even the two best models are unable to do it consistently through all setups. This is interesting because the core ability required is still the same. The harder setting only increases demands on maintaining and updating internal state, proven to still be a challenge.
+Models possess good cognitive flexibility. Compared to
+{4}------------------------------------------------
+<span id="page-4-0"></span>*Table 1.* Baseline NeuroCognition benchmark performance without hints, notes, or pattern hints. RAPM uses the reported human average performance from ?. SWM & WCST compares against a healthy human worst-case score. Full setup details for each Easy/Hard setup are provided in each test section. MC and Gen denote multiple-choice and generation setup, respectively. All results are zero shot and use each model's default reasoning level, except where marked with <sup>∗</sup> indicating reasoning was disabled due to some issues resulting in worse performance. Bold numbers and green cells denote scores that match or surpass human performance. ↑ means higher is better.
+| | RAPM (Accuracy ↑) | | | | | SWM (Sswm | ↑) | | | WCST (Swcst<br>↑) | | | |
+|-----------------|-------------------|--------|--------|-------|-------|------------|-------|-------|-------|-------------------|-------|-------|-------|
+| Model | Text<br>Image | | | Text | | Image+Text | | | Image | | Text | Image | |
+| | MC | Gen | MC | Easy | Hard | Easy | Hard | Easy | Hard | Easy | Hard | Easy | Hard |
+| GPT-5 | 0.540 | 0.605 | 0.279 | 1.000 | 1.000 | 0.985 | 0.989 | 0.927 | 0.631 | 0.848 | 0.815 | 0.824 | 0.794 |
+| Gemini 3 Pro | 0.585 | 0.460 | 0.400 | 1.000 | 1.000 | 1.000 | 0.963 | 1.000 | 0.793 | 0.832 | 0.776 | 0.861 | 0.805 |
+| Gemini 2.5 Pro | 0.525 | 0.560 | 0.186 | 1.000 | 0.980 | 0.733 | 0.480 | 0.408 | 0.622 | 0.825 | 0.806 | 0.823 | 0.776 |
+| Claude Sonnet 4 | 0.560∗ | 0.385∗ | 0.207∗ | 1.000 | 0.866 | 0.256 | 0.265 | 0.630 | 0.005 | 0.764 | 0.773 | 0.800 | 0.710 |
+| Grok 4.1 Fast | 0.505 | 0.495 | 0.100 | 1.000 | 0.983 | 0.145 | 0.017 | 0.014 | 0.009 | 0.841 | 0.825 | 0.107 | 0.060 |
+| Grok 4 Fast | 0.690∗ | 0.555∗ | 0.086∗ | 1.000 | 0.994 | 0.021 | 0.002 | 0.029 | 0.005 | 0.836 | 0.754 | 0.336 | 0.209 |
+| o4-Mini | 0.450 | 0.410 | 0.207 | 0.951 | 0.811 | 0.055 | 0.032 | 0.042 | 0.005 | 0.789 | 0.792 | 0.347 | 0.277 |
+| GLM 4.6V | 0.320 | 0.170 | 0.357 | 1.000 | 0.574 | 0.044 | 0.010 | 0.000 | 0.028 | 0.300 | 0.095 | 0.191 | 0.064 |
+| GLM 4.5V | 0.035 | 0.000 | 0.128 | 1.000 | 0.536 | 0.044 | 0.000 | 0.000 | 0.000 | 0.773 | 0.309 | 0.108 | 0.083 |
+| Qwen3-VL-235B | 0.485 | 0.570 | 0.493 | 1.000 | 0.760 | 0.026 | 0.016 | 0.031 | 0.135 | 0.230 | 0.233 | 0.420 | 0.071 |
+| Human | - | - | 0.814 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 0.714 | 0.556 | 0.714 | 0.556 |
+RAPM and SWM, models generally perform better on WCST, with more models passing the human worst-case baseline. This suggests current LLMs handle rule inference and task-switching behavior more robustly than abstract relational reasoning or complex working-memory maintenance, though image setup is still an issue for some models.
+Answer format matters. In RAPM text, models exhibit different preferences for multiple-choice versus generative answer formats on the same items. This likely reflects differences in decision constraints and how models interpret the problem, which sometimes leads to overcomplication.
+#### 4.2. Failure Cases & Patterns
+RAPM. Our analysis of RAPM reasoning traces (Appendix [E](#page-15-0) Table [10,](#page-15-1) [11\)](#page-15-2) reveals three persistent deficits across modalities. First, hallucination and mis-grounding: models fabricate visual features or impose non-existent numerical patterns on text that contradict the data. Second, semantic over-interpretation: models overlook simple local rules (XOR, concatenation) in favor of complex, global hypotheses like "symmetry grids." Third, specifically in Text Gen, models exhibit content-logic disconnect, where they frequently derive correct attribute constraints but generate hallucinated strings that satisfy these metadata rules while violating the actual constructive logic. These failures suggest a disconnect between low-level feature verification and abstract reasoning.
+SWM. We investigate failures using four error types explained earlier in Section [3.2.](#page-2-0) Since models struggle most in the image modality, we focus on image and image-text
+<span id="page-4-1"></span>*Table 2.* SWM image and image+text failure turns averaged across all runs (easy and hard). ↓ means lower is better.
+| Model | Illegal ↓ | Repeat ↓ | No-Box ↓ | Invalid ↓ |
+|-----------------|-----------|----------|----------|-----------|
+| GPT-5 | 2.75 | 4.25 | 3.42 | 0.50 |
+| Gemini 3 Pro | 5.67 | 1.17 | 0.00 | 0.58 |
+| Gemini 2.5 Pro | 21.75 | 3.08 | 13.83 | 0.33 |
+| Claude Sonnet 4 | 1.42 | 4.17 | 47.08 | 18.42 |
+| Grok 4.1 Fast | 0.08 | 4.42 | 78.00 | 9.00 |
+| Grok 4 Fast | 0.00 | 6.00 | 68.92 | 21.58 |
+| o4-Mini | 0.00 | 19.25 | 69.75 | 1.25 |
+| GLM 4.6V | 0.50 | 4.92 | 9.67 | 82.50 |
+| GLM 4.5V | 0.00 | 1.92 | 28.00 | 71.58 |
+| Qwen3-VL-235B | 4.58 | 22.75 | 40.67 | 21.33 |
+failure cases (Table [2\)](#page-4-1). From the data, we identify two main weaknesses, state tracking under sparse feedback (models forget key task state) and task execution (invalid and nobox actions instead of exploiting eliminations). High illegal and repeat counts indicate confusion and working-memory failures, high no-box rates suggest a visual grounding gap (difficulty identifying a valid box), and invalid turns reflect failures to follow the task instruction.
+WCST. Table [3](#page-5-0) summarizes WCST image error rates using FMS and PR as explained in Section [3.3.](#page-2-1) We focus on weaker models that struggle in WCST image. Overall, most errors happen after the model has found the right rule. Models either (i) fail to keep using the acquired rule (higher FMS) or (ii) fail to change their choice after negative feedback (higher PR), i.e., they choose a rule that was already ruled out within the current rule block. The background-
+{5}------------------------------------------------
+<span id="page-5-0"></span>*Table 3.* WCST error rates for models that struggle in WCST image tests. ↓ means lower is better.
+| Model | Easy | | Hard | | |
+|---------------|-------|-------|-------|-------|--|
+| | FMS ↓ | PR ↓ | FMS ↓ | PR ↓ | |
+| Grok 4.1 Fast | 0.339 | 0.324 | 0.376 | 0.420 | |
+| Grok 4 Fast | 0.432 | 0.201 | 0.347 | 0.309 | |
+| o4-Mini | 0.320 | 0.092 | 0.437 | 0.352 | |
+| GLM 4.6V | 0.089 | 0.442 | 0.169 | 0.420 | |
+| GLM 4.5V | 0.404 | 0.310 | 0.303 | 0.432 | |
+| Qwen3-VL-235B | 0.229 | 0.619 | 0.500 | 0.794 | |
+<span id="page-5-1"></span>*Table 4.* RAPM results with no reasoning. ↑ means higher is better. Paired t-test results show effects in Text MC is statistically significant (p < .05) while others are not (p > .05).
+| Model | RAPM (Accuracy ↑) | | | | | | |
+|----------------|-------------------|-------------|-------------|--|--|--|--|
+| | Text MC | Text Gen | Image MC | | | | |
+| Gemini 2.5 Pro | 0.690∆+0.17 | 0.450∆−0.11 | 0.093∆−0.09 | | | | |
+| Grok 4.1 Fast | 0.690∆+0.19 | 0.565∆+0.07 | 0.114∆+0.01 | | | | |
+| GLM 4.6V | 0.635∆+0.32 | 0.175∆+0.01 | 0.336∆−0.02 | | | | |
+| GLM 4.5V | 0.460∆+0.43 | 0.000∆0.00 | 0.014∆−0.11 | | | | |
+color addition tends to make both problems worse. We provide more details and other error metrics in Appendix [E.](#page-15-0)
+## 5. Ablation Study
+Following the baseline results, we explore aspects that could affect model performance. We present the change in performance compared to Table [1](#page-4-0) baseline using a subscripted delta (∆) notation to indicate any improvements or declines.
+#### <span id="page-5-3"></span>5.1. RAPM: Reasoning vs No Reasoning
+We compare RAPM performance when models run with their default reasoning behavior versus a constrained noreasoning setting, as shown in Table [4.](#page-5-1) This experiment was run only on models where reasoning can be disabled (excluding Claude Sonnet 4 and Grok 4 Fast because of reasoning issue). Disabling reasoning produces large, consistent gains in Text MC (p = .0207) across models, while Text Gen and Image MC shows mixed effects. The improvement in Text MC likely occurs because the decision space is explicit and limited, and avoiding extended chain-of-thought reduces overcomplication and distractor attraction. For Text Gen, using reasoning or not shows no clear effect (p = .83). For Image MC, disabling reasoning appears slightly worse (p = .173), likely because reasoning helps interpret visual inputs and detect patterns, though the effect may be small.
+<span id="page-5-2"></span>*Table 5.* RAPM results with pattern hints. ↑ means higher is better. Paired t-test results show no significant effect on all setups (p > .05).
+| Model | | RAPM (Accuracy ↑) | |
+|-----------------|-------------|-------------------|-------------|
+| | Text MC | Text Gen | Image MC |
+| GPT-5 | 0.620∆+0.08 | 0.720∆+0.12 | 0.257∆−0.02 |
+| Gemini 3 Pro | 0.595∆+0.01 | 0.545∆+0.09 | 0.386∆−0.01 |
+| Gemini 2.5 Pro | 0.485∆−0.04 | 0.585∆+0.03 | 0.221∆+0.04 |
+| Claude Sonnet 4 | 0.525∆−0.03 | 0.360∆−0.02 | 0.221∆+0.01 |
+| Grok 4.1 Fast | 0.570∆+0.07 | 0.580∆+0.09 | 0.114∆+0.01 |
+| Grok 4 Fast | 0.655∆−0.04 | 0.540∆−0.02 | 0.100∆+0.01 |
+| o4-Mini | 0.435∆−0.02 | 0.525∆+0.12 | 0.236∆+0.03 |
+| GLM 4.6V | 0.395∆+0.08 | 0.235∆+0.07 | 0.321∆−0.03 |
+| GLM 4.5V | 0.085∆+0.05 | 0.000∆0.00 | 0.090∆−0.04 |
+| Qwen3-VL-235B | 0.575∆+0.09 | 0.590∆+0.02 | 0.457∆−0.04 |
+#### 5.2. RAPM: Pattern Hints
+In RAPM, we test whether providing hints regarding possible underlying patterns helps models narrow down their reasoning and identify the correct answer. The results shown in Table [5](#page-5-2) vary across models and setups, and we find no significant effect (paired t-tests: Text MC p = .177, Text Gen p = .45, Image MC p = .666). Most results are similar to those obtained without using any patterns, and in some cases, they even decrease. Overall, pattern hints have a limited and inconsistent impact. Most changes are small (often positive), and the aggregate effect is not statistically significant, suggesting that many models still lack core abstract reasoning capabilities, even with these hints.
+#### 5.3. SWM & WCST: Note-Taking Assistance
+For SWM and WCST, we explore assistance in the form of notes, which emulates cognitive offloading (??) where humans write down their observations to keep track of the state. The system provides the notes itself in every turn, explicitly stating the current state and actions taken by the model so far. From Table [6,](#page-6-0) we observe large gains for some models, but the average effect is not significant (Text Hard p = .161; Img+Text Easy p = .842), which may reflect limited sample size or generally small changes. This suggests that many models still struggle with working-memory maintenance, where notes are not even helpful and may even cause slight drops.
+For WCST, improvements from note-taking are generally consistent across models and setups, although they are close to reach statistical significance in our (Image Easy: p = .14; Image Hard: p = .0524). This indicates that note-taking provides practical benefits by helping models track tested attributes and maintain a consistent hypothesis.
+{6}------------------------------------------------
+<span id="page-6-0"></span>*Table 6.* SWM results with note-taking assistance. Paired t-test results show no significant effects on all setups (p > .05).
+| Model | SWM (Sswm<br>↑) | | | | |
+|-----------------|-----------------|--------------|--|--|--|
+| | Text Hard | Img+Txt Easy | | | |
+| Gemini 2.5 Pro | 0.965∆−0.02 | 0.704∆−0.03 | | | |
+| Claude Sonnet 4 | 0.826∆−0.04 | 0.207∆−0.05 | | | |
+| Grok 4.1 Fast | 0.973∆−0.01 | 0.043∆−0.10 | | | |
+| Grok 4 Fast | 0.964∆−0.03 | 0.029∆+0.01 | | | |
+| o4-Mini | 0.935∆+0.12 | 0.012∆−0.04 | | | |
+| GLM 4.6V | 0.803∆+0.23 | 0.082∆+0.04 | | | |
+| GLM 4.5V | 0.851∆+0.32 | 0.000∆−0.04 | | | |
+| Qwen3-VL-235B | 0.782∆+0.02 | 0.318∆+0.29 | | | |
+*Table 7.* WCST image results with note-taking assistance. Paired t-test results show no significant effects on all setups (p > .05).
+| Model | WCST (Swcst<br>↑) | | | | | |
+|---------------|-------------------|--------------|--|--|--|--|
+| | Image Easy | Image Hard | | | | |
+| Grok 4.1 Fast | 0.135∆+0.028 | 0.197∆+0.137 | | | | |
+| Grok 4 Fast | 0.215∆−0.120 | 0.257∆+0.048 | | | | |
+| o4-Mini | 0.392∆+0.045 | 0.295∆+0.018 | | | | |
+| GLM 4.6V | 0.430∆+0.239 | 0.411∆+0.347 | | | | |
+| GLM 4.5V | 0.306∆+0.199 | 0.123∆+0.040 | | | | |
+| Qwen3-VL-235B | 0.640∆+0.331 | 0.250∆+0.179 | | | | |
+## 6. Analysis & Discussion
+#### <span id="page-6-3"></span>6.1. Factor Analysis of LLM
+To understand the latent structure of capabilities across widely used LLM benchmarks inspired by ?, we conducted an exploratory factor analysis (EFA) on performance data obtained from Artificial Analysis[2](#page-6-1) for 156 distinct LLMs. We selected 10 benchmarks widely recognized as standard measures of LLM "intelligence" and real-world capabilities, ranging from coding and math to agentic tool use. Importantly, this selection was driven by the statistical requirement for a large sample size (N > 100) to ensure valid factor analysis, which very few other benchmarks satisfy. We used Principal Axis Factoring for the factoring method.
+Parallel analysis indicated a single-factor solution: a unidimensional latent variable explains most performance variation across these tasks. This factor (PA1) accounted for 75% of the total variance, providing strong evidence for a "general capability" factor. As shown in Table [8,](#page-6-2) all benchmarks had high loadings (λ > 0.76), with *GPQA Diamond* (0.943) and *LiveCodeBench* (0.926) as the strongest proxies. This coherence also exposes a limitation of current evaluation, where many benchmarks measure the same underlying construct. This makes NeuroCognition even more important, as
+<span id="page-6-2"></span>*Table 8.* Factor analysis loadings for state-of-the-art LLM benchmarks. The single extracted factor (PA1) explains 75% of the total variance, indicating a strong general capability factor (g) across diverse tasks.
+| Benchmark | Loadings (λ) |
+|--------------------------|--------------|
+| GPQA Diamond (?) | 0.943 |
+| LiveCodeBench (?) | 0.926 |
+| SciCode (?) | 0.922 |
+| AA-LCR (?) | 0.875 |
+| AIME 2025 (?) | 0.870 |
+| IFBench (?) | 0.861 |
+| Terminal Bench (?) | 0.834 |
+| MMLU Pro (?) | 0.825 |
+| τ -Bench (?) | 0.818 |
+| Humanity's Last Exam (?) | 0.768 |
+it targets distinct, basic cognitive tasks.
+#### 6.2. Correlation with Other Benchmarks
+To assess how NeuroCognition relates to real-world tasks, we correlate it with external benchmarks targeting specific cognitive demands. For working memory (SWM), we use a modified *Long HotPotQA* (?) that splits questions and passages at the word level and treats each word as a turn. For fluid intelligence (RAPM), we compare against *ARC-AGI 2* (?). For cognitive flexibility (WCST), we use *LLM Task Interference* (?), which measures performance impacts under task switches. Detailed setups and results are in Appendix [D.](#page-14-0) We also correlate the aggregate NeuroCognition score with the average of 11 state-of-the-art benchmarks (the 10 in Section [6.1](#page-6-3) plus MMMU Pro (?)), as a proxy for "general capability". These scores are also taken and already normalized from Artificial Analysis. Pearson correlations are shown in Figure [2.](#page-7-0)
+We observe a high correlation between the average NeuroCognition score and the 11-benchmark average (r = .86, p = .001, N = 10), validating our framework as a robust predictor of general model capability (g). More importantly, the heatmap provides a granular view of the functional relationships between cognitive primitives and downstream tasks:
+- (1) Working Memory & Context: *SWM Text Hard* shows a moderate correlation with *Long HotPotQA* (r = .43, p = .21, N = 10). This suggests that working-memory maintenance is still somewhat relevant long-context retrieval.
+- (2) Fluid Intelligence: *RAPM Image MC* is the strongest predictor of *ARC-AGI 2* (r = .87, p = .024, N = 6), consistent with isolating an "abstract visual reasoning" factor that standard benchmarks often miss while also uncorrelated with the general benchmark average (r = .09).
+- (3) Cognitive Flexibility: *WCST Text Hard* shows only a
+<span id="page-6-1"></span>
+{7}------------------------------------------------
+<span id="page-7-0"></span>![](_page_7_Figure_1.jpeg)
+Figure 2. Correlations between our NeuroCognition benchmark tests and corresponding real-world LLM benchmarks. Cells with bold text and highlighted borders indicate task pairs that are similar and rely on the same cognitive abilities (including averages). Statistical significance is denoted by asterisks (\*p < .05, \*\*p < .01).
+weak, non-significant correlation with *LLM Task Interference* ( $r=.24,\ p=.50,\ N=10$ ). This suggests that the abstract rule-updating mechanisms measured by WCST may be distinct from practical task switching scenarios.
+#### 7. Conclusion
+In this work, we introduce the NeuroCognition benchmark, a suite of neuropsychological tests (RAPM, SWM, WCST) adapted for LLMs to probe specific cognitive processes using multimodal, process-aware metrics. Across tasks, state-of-the-art models show meaningful but incomplete alignment with human cognition. They perform better in text, but performance drops for images and harder settings that demand intensive grounding and state tracking. Overall, explicit reasoning is not uniformly beneficial, and humanlike assistance shows mixed effects. The impact of pattern hints is limited, while note-taking helps more consistently in WCST but not in SWM. Complementing these results, our factor analysis suggests a dominant general-capability factor across standard benchmarks, highlighting redundancy in current evaluation. Correlations with external benchmarks further suggest that gains in NeuroCognition process scores translate into improvements in semantically related realworld tasks, while individual tests still measure distinct cognitive abilities. In summary, our results suggest that current LLMs do not possess the same kind of intelligence as humans, but still align with humans in some aspects, which can inform how we use and improve them.
+## 8. Future Work
+Since we apply neuropsychological tests designed for humans to LLMs, it is important for future work to validate whether the same cognitive abilities measured in humans are also captured in LLMs. This helps to further under-
+stand the relationship between human and LLM intelligence and improve the interpretability of the benchmark results. Another interesting direction is improving the LLM itself through neuropsychologically grounded training data. The data that we used in this benchmark captures basic cognitive processes that are foundational to human intelligence, and therefore can be used to improve the cognitive abilities of LLMs. Additionally, recent works investigate ways to compress context (??), or utilize some sort of memory in their architecture (???), and our benchmark could serve to evaluate such methods. Finally, this data can be easily generated synthetically with code, making it scalable to create a large amount of data for training purposes.
+#### Limitations
+Our evaluation has three main limitations. First, sample sizes are constrained by cost: running state-of-the-art models with effectively unlimited reasoning is expensive. For RAPM, we capped runs at n = 200 (text) and n = 140(image), which may compromise the robustness of our results, especially in the original image modality. Second, our benchmark assumes that the same neuropsychological test used to assess a specific cognitive ability in humans can also be applied to LLMs to evaluate that same ability. In the neuropsychological field, this is not always the case, since a neuropsychological test is standardized on a specific population and must be validated based on psychometric properties (validity, reliability, and measurement invariance) when applied to a different population. Therefore, even when a model performs consistently as well as humans across all tasks in our benchmark, we can only claim that the model exhibits a similar level of intelligence and can simulate human-like cognitive abilities, but not that it thinks like humans or possesses the same underlying cognitive mechanism. Third, since the neuropsychological tests we
+{8}------------------------------------------------
+used are widely used and well-established, we assume they are valid and reliable based on their psychometric properties (validity, reliability, and measurement invariance). However, this assumption may not hold in all cases, and further validation across prompts, contexts, and a wide range of LLMs is needed.
+## Impact Statement
+<span id="page-8-0"></span> github.com/features/copilot
+<span id="page-8-1"></span> gemini.google.com
+<span id="page-8-2"></span> chatgpt.com
+{9}------------------------------------------------
+## <span id="page-9-0"></span>A. RAPM Text Data Generation
+### A.1. Rule Space and Attributes
+The generation engine employs a compositional rule space defined by attributes that constrain string properties. For each matrix, two primary attributes are sampled—one for rows and one for columns. Additional consistency rules (leaks) are inferred if they naturally emerge from the intersection of constraints.
+The implemented attributes and their parameters are:
+- 1. Character Set Type: Restricts the alphabet of the string.
+- *Values*: letters, digits, symbols.
+- *Logic*: All characters in the cell must belong to the specified set.
+- 2. Type Count: Constraints on the count of specific character types.
+- *Types*: vowels, consonants, uppercase/lowercase variants, etc.
+- *Rules*: even, odd, multiple of N (for N ∈ {2, 3, 4}).
+- *Logic*: The count of characters matching the type must satisfy the parity or divisibility rule. Zero counts are explicitly disallowed for these rules to ensure pattern salience.
+- 3. Quantitative Constant: Fixes a metric to a specific integer value across the sequence.
+- *Metrics*: String length, unique character count, or count of specific types.
+- *Range*: Integers in [2, 5].
+- 4. Quantitative Progression: Enforces an arithmetic progression on a metric.
+- *Metrics*: Same as Quantitative Constant.
+- *Parameters*: Start value ∈ [1, 3], Step size ∈ [1, 3].
+- *Logic*: The metric value increases by the step size across the row or column (e.g., lengths 3, 5, 7).
+- 5. Sorted Order: Enforces lexicographical ordering.
+- *Values*: ascending, descending, mixed.
+- *Logic*: Characters in the string must follow the specified order. mixed explicitly requires the string to be unsorted.
+- 6. Positional Constraint: Fixes the character type at specific indices.
+- *Indices*: first, last, even, odd.
+- *Logic*: Characters at the specified positions must belong to a specific type (e.g., "first character must be a digit").
+#### A.2. Generation Algorithm
+The matrix generation process follows a constraint-satisfaction approach with backtracking, as detailed in Algorithm [1.](#page-10-0) The core logic proceeds as follows:
+- 1. Attribute Sampling: Randomly sample one attribute specification for rows and one for columns.
+- 2. Compatibility Check: Validate that the row and column attributes are compatible. For example, a "Sorted" row attribute conflicts with a "Positional" column attribute because sorting rearranges characters, potentially violating fixed-position constraints. Infeasible combinations trigger a re-sampling.
+- 3. Constraint Propagation: The row and column attributes are combined to define specific constraints for each of the 3 × 3 cells. "Leak" detection is performed to identify additional properties that effectively become constant across rows or columns (e.g., if a row rule restricts characters to digits and a column rule restricts them to even numbers, a universal "digits" constraint might apply).
+- 4. Grid Construction: Cells are generated sequentially. For each cell, a string is constructed to satisfy the merged constraints. Random backtracking is used if a cell cannot be generated given the current constraints and history.
+- 5. Variation Enforcement: To prevent trivial patterns, we enforce a minimum Hamming distance between strings in the same row or column. This ensures that the progression relies on abstract rules rather than simple repetition.
+Upon successful generation, the defining constraints for the missing cell (including "leaked" rules) are stored for automatic rule validation.
+{10}------------------------------------------------
+#### <span id="page-10-0"></span>Algorithm 1 RAPM Text Matrix Generation
+```
+Input: Random Seed S
+Output: 3 × 3 String Matrix M, Constraints C
+Initialize RNG with S
+repeat
+Sample row attribute Arow and column attribute Acol
+valid ← CheckCompatibility(Arow, Acol)
+until valid
+Initialize cell constraint grid C3×3 from Arow and Acol
+Apply inferred consistency rules (leaks) to C
+for i = 0 to 2 do
+for j = 0 to 2 do
+attempts ← 0
+repeat
+Generate string s satisfying Ci,j
+vrow ← ∀k < j : Hamming(s,Mi,k) ≥ 2
+vcol ← ∀k < i : Hamming(s,Mk,j ) ≥ 2
+attempts ← attempts + 1
+until (vrow ∧ vcol) or attempts > MAX RET RIES
+if not (vrow ∧ vcol) then
+fail and backtrack
+end if
+Mi,j ← s
+end for
+end for
+return M, C
+```
+#### A.3. Multiple-Choice and Distractor Generation
+For the text MC setup, we generate 8 options: 1 correct answer (ground truth for the bottom-right cell) and 7 distractors. Each distractor is guaranteed to violate at least one latent rule to ensure a unique solution.
+- Correct Answer: The generated string for the bottom-right (3, 3) cell.
+- Distractors: Generated by applying mutation strategies to the correct answer to violate latent rules. Strategies include:
+- *Break Ordering*: Swapping adjacent characters to violate sort constraints.
+- *Break Positional*: Changing a character at a critical index (e.g., first/last) to a forbidden type.
+- *Adjust Count*: Adding or removing characters to violate parity or length constraints.
+- *Character Mutation*: Randomly altering characters to valid but incorrect values.
+{11}------------------------------------------------
+## <span id="page-11-0"></span>B. RAPM Text Gen Automatic Rule Validation
+For the text Gen setup, we can programmatically check if a generated string S satisfies the stored underlying constraints C for the missing cell (3, 3). Algorithm [2](#page-11-1) details this validation procedure.
+#### <span id="page-11-1"></span>Algorithm 2 RAPM Text Gen Validation (CellSatisfies)
+```
+Input: Candidate String S, Constraints C
+Output: Boolean (True if valid)
+if C.f ixed length ̸= None and |S| ̸= C.f ixed length then
+return False
+end if
+for all (type, count) ∈ C.target counts do
+if Count(S, type) ̸= count then
+return False
+end if
+end for
+for all (type, rule) ∈ C.parity rules do
+c ← Count(S, type)
+if c = 0 then
+return False {Zero count forbidden for parity}
+end if
+if (rule = 'even' ∧ c is odd) or (rule = 'odd' ∧ c is even) then
+return False
+end if
+end for
+for all (type, k) ∈ C.multiple rules do
+c ← Count(S, type)
+if c = 0 or c (mod k) ̸= 0 then
+return False
+end if
+end for
+if C.unique count ̸= None and |Unique(S)| ̸= C.unique count then
+return False
+end if
+if C.ordering ̸= None and CheckOrder(S, C.ordering) = False then
+return False
+end if
+if C.positional ̸= None and CheckPositional(S, C.positional) = False then
+return False
+end if
+return True
+```
+The validation function CellSatisfies(S, C3,3) performs the following checks:
+- 1. Length Check: If a fixed length is enforced (from quantitative constraints), |S| must match.
+- 2. Type Counts: Count(S, type) = k for all exact count constraints.
+- 3. Parity/Divisibility: Count(S, type) must satisfy even/odd/multiple conditions. Crucially, as enforced during generation, zero counts are invalid for these rules to prevent trivial satisfactions (e.g., "0 vowels" satisfying "even vowels").
+- 4. Unique Characters: If unique count is constrained, |Unique(S)| is checked against the target.
+- 5. Ordering: If sorted order is required, S is checked against its own sorted version.
+- 6. Positional: Characters at constrained indices (first, last, even, odd) are verified against the allowed character set.
+Any string that satisfies all these constraints is considered a correct solution, providing a robust "verifiable" accuracy metric for generative models that goes beyond simple multiple-choice selection or exact-match.
+{12}------------------------------------------------
+## <span id="page-12-0"></span>C. Models & Experimental Setup
+## C.1. Model Access Names and Notes
+## Google AI Studio
+- Gemini 3 Pro (?): gemini-3-pro-preview
+- Gemini 2.5 Pro (?): gemini-2.5-pro
+#### OpenAI API
+- GPT-5 (?): gpt-5-2025-08-07
+- o4-Mini (?): o4-mini-2025-04-16
+#### OpenRouter
+- Claude Sonnet 4 (?): anthropic/claude-sonnet-4
+- Grok 4.1 Fast (?): x-ai/grok-4.1-fast
+- Grok 4 Fast (?): x-ai/grok-4-fast
+- GLM 4.6V (?): z-ai/glm-4.6v[6](#page-12-1)
+- GLM 4.5V (?): z-ai/glm-4.5v[7](#page-12-2)
+## Local / vLLM
+713 714 • Qwen3-VL-235B (?): Qwen/Qwen3-VL-235B-A22B-Thinking[8](#page-12-3)
+#### C.2. vLLM Inference Settings
+For local vLLM runs, we used a server with 4x NVIDIA H100 80GB GPUs and CUDA 12.8, and ran the model with FP8 quantization. Generations used the model's default Qwen3 configuration; temperature and other sampling settings were chosen per the evaluation protocol below.
+## C.3. Evaluation Protocol Details
+- Zero-shot: All evaluations are zero-shot.
+- Temperature: We used a temperature of 0 when possible to reduce sampling variability; when a model's API does not allow overriding temperature, we used the model's default.
+- Reasoning mode: We used each model's default reasoning configuration except for Claude Sonnet 4 and Grok 4 Fast, for which we adjusted reasoning as described in the main text (see Section [3.4\)](#page-3-0).
+- Maximum output tokens: For RAPM, we allowed the model's maximum permitted output tokens. For SWM and WCST, we limited per-turn outputs to 8192 tokens.
+- Thinking budget: The system prompt described a thinking budget (not enforced as a hard token limit). For RAPM, we used an effective thinking budget of 30,000 tokens for models with a context limit of 32K and 60,000 tokens for models with a context limit of 65K+. For SWM and WCST, the per-turn thinking budget was 4,000 tokens.
+- Repetitions: SWM and WCST setups were run three times each to assess variability (data are generated on the fly); RAPM was run once per model.
+- WCST (notes): For WCST runs that used note-taking assistance, we set the note window to 6.
+<span id="page-12-1"></span><sup>6</sup>
+<span id="page-12-2"></span><sup>7</sup>
+<span id="page-12-3"></span><sup>8</sup>
+{13}------------------------------------------------
+## C.4. Reasoning Configuration
+Model reasoning capabilities were enabled through two complementary methods:
+API Configuration. In our ModelWrapper, we programmatically enabled reasoning features where supported by the provider APIs.
+- Google AI Studio: We accessed Gemini models via the OpenAI-compatible endpoint and enabled reasoning by specifically configuring the thinking config within the extra body parameter. We set include thoughts to True, thinking level for Gemini 3 Pro (default to high), and thinking budget for Gemini 2.5 Pro (default to 8192).
+- OpenAI API: For OpenAI models, reasoning is handled through reasoning effort parameter in the chat completion calls. We use the default value medium for both GPT-5 and o4-Mini.
+- OpenRouter: We set the reasoning parameter in the request body to {"exclude": False} to ensure reasoning tokens were generated and returned, and manage the thinking budget by setting max tokens within the reasoning configuration. For Grok, reasoning was explicitly enabled via {"enabled": True}.
+- vLLM (Local): We enabled the enable thinking parameter in the chat template to activate reasoning generation.
+Prompt-Based Instructions. For all tasks, we included explicit instructions in the prompt to guide the reasoning process, specifically when Chain-of-Thought (CoT) was enabled.
+#### Standardized CoT Prompt Addition
+#### SWM & WCST:
+Explain your thought process regarding the problem and the feedbacks you received in maximum {think budget} tokens wrapped with <think> and </think>. Then, provide a really short summary of your reasoning after the closing </think> tag.
+#### RAPM:
+Explain your thought process (max {think budget} tokens) inside <think>...</think> then give final answer.
+This dual approach ensured that models with native reasoning capabilities (activated via API) and standard models (guided via prompts) both engaged in structured reasoning before outputting a final answer.
+## C.5. Disabling Reasoning
+750 751 We disable the reasoning capabilities for some models that are able to do so to conduct the ablation study (see Section [5.1\)](#page-5-3) and to address specific model issues (specifically, Claude Sonnet 4 and Grok 4 Fast with "overthinking" that exhausted token limits on RAPM).
+754 756 Configuration & Prompts. For most models, reasoning was disabled by reversing the API configurations described previously (e.g., setting include thoughts to False, reasoning enabled to False, or enable thinking to False) and replacing the CoT prompt instructions with a standard direction to provide only the final answer, as shown below.
+## Standardized Non-CoT Instruction Answer only with your final answer.
+## Gemini 2.5 Pro Specific Prompt. For Gemini 2.5 Pro, the model does not natively support disabling reasoning via API parameters. However, we found a way to enforce the no-reasoning condition by appending a specific critical instruction to override the model's default behavior.
+#### Gemini 2.5 Pro Non-CoT Instruction
+768 769 CRITICAL INSTRUCTION: Do not use <thinking> tags or reasoning steps. Provide the response immediately.
+{14}------------------------------------------------
+## <span id="page-14-0"></span>D. External Benchmark Setup & Results
+Long HotPotQA. We work with the distractor development split of HotPotQA (?), sampling 500 question–passage pairs. To stress the models' turn tracking, we split each question and its supporting passages at the word level, treat every word as a separate turn, and fix the model's response on every intermediate turn to "Okay." Each sequence is ordered with the full question first, followed by the concatenated passages. The results for Long HotPotQA are shown in Table [9.](#page-14-1)
+LLM Task Interference. Following the official implementation from ?, with minor updates, we evaluate two task/context pairings and sample 200 evaluation prompts each. The first pairing uses the MMLU AA task with Rotten Tomatoes exemplars as in-context data, while the second pairing is the Rotten Tomatoes task with TweetQA data. The results for LLM Task Interference are shown in Table [9.](#page-14-1)
+ARC-AGI 2. Since the results for ARC-AGI 2 (?) are already publicly available in their leaderboard website[9](#page-14-2) , we directly took the best results for each model, except for some models that are not available. Although the scores are very low, the results have been verified from the source. These are also summarized in Table [9.](#page-14-1)
+<span id="page-14-1"></span>*Table 9.* Results for external benchmarks: Long HotPotQA, LLM Task Interference, and ARC-AGI 2. (-) indicates the model result is not available.
+| Model | Long HotPotQA | Task Interference | ARC-AGI 2 |
+|-----------------|---------------|-------------------|-----------|
+| GPT-5 | 0.936 | 0.958 | 0.099 |
+| Gemini 3 Pro | 0.932 | 0.962 | 0.311 |
+| Gemini 2.5 Pro | 0.924 | 0.875 | 0.049 |
+| Claude Sonnet 4 | 0.930 | 0.948 | 0.059 |
+| Grok 4.1 Fast | 0.910 | 0.913 | - |
+| Grok 4 Fast | 0.461 | 0.788 | 0.053 |
+| o4-Mini | 0.802 | 0.933 | 0.061 |
+| GLM-4.6V | 0.862 | 0.938 | - |
+| GLM-4.5V | 0.271 | 0.635 | - |
+| Qwen3 VL 235B | 0.916 | 0.940 | - |
+<span id="page-14-2"></span>
+{15}------------------------------------------------
+## <span id="page-15-0"></span>E. Full Results
+#### E.1. Raven's Advanced Progressive Matrices (RAPM)
+The complete results for RAPM image tests, broken down by pattern type, are provided in Table [10.](#page-15-1) We also provide qualitative examples of reasoning failures across text and image modalities in Table [11,](#page-15-2) highlighting the prevalence of hallucination and over-interpretation.
+*Table 10.* RAPM image accuracy breakdown by pattern type.
+<span id="page-15-1"></span>
+| Model | Dist. Four | Dist. Nine | Center | In-Dist. Out-Center | In-Out Center | Left-Right | Up-Down |
+|-----------------|------------|------------|--------|---------------------|---------------|------------|---------|
+| GPT-5 | 0.40 | 0.35 | 0.25 | 0.15 | 0.35 | 0.05 | 0.40 |
+| Gemini 3 Pro | 0.55 | 0.65 | 0.55 | 0.35 | 0.20 | 0.25 | 0.25 |
+| Gemini 2.5 Pro | 0.05 | 0.10 | 0.35 | 0.20 | 0.40 | 0.20 | 0.25 |
+| Claude Sonnet 4 | 0.30 | 0.45 | 0.15 | 0.20 | 0.20 | 0.10 | 0.15 |
+| Grok 4.1 Fast | 0.15 | 0.20 | 0.15 | 0.05 | 0.00 | 0.00 | 0.15 |
+| Grok 4 Fast | 0.05 | 0.10 | 0.05 | 0.25 | 0.10 | 0.10 | 0.05 |
+| o4-Mini | 0.45 | 0.15 | 0.25 | 0.10 | 0.30 | 0.05 | 0.15 |
+| GLM 4.6V | 0.45 | 0.30 | 0.40 | 0.35 | 0.30 | 0.25 | 0.45 |
+| GLM 4.5V | 0.20 | 0.05 | 0.15 | 0.10 | 0.15 | 0.20 | 0.05 |
+| Qwen3-VL-235B | 0.45 | 0.65 | 0.55 | 0.40 | 0.40 | 0.50 | 0.50 |
+<span id="page-15-2"></span>*Table 11.* Selected reasoning failure examples in RAPM (image and text). Models frequently hallucinate attributes to fit a complex hypothesis ("visual/numerological Hallucination") or prefer abstract semantic rules over simple local operations ("semantic overinterpretation").
+| Task/Model | ID | Pattern/Context | Reasoning Snippet (Failure Point) | Outcome |
+|--------------------------|------------------------|----------------------------------------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|--------------------------------------------------------------------------------------------|
+| Image<br>GPT-5 | RAVEN 238<br>center | Up-Down<br>(Split layout) | "Row 3: top becomes large solid black triangle.<br>Bottom shapes by row show a sides-count sequence<br>decreasing 5→4→3"<br>(Visual Hallucination: No such triangle exists.) | Pred: 2<br>Correct: 0 |
+| Image<br>GPT-5 | RAVEN 238<br>dist nine | Dist. Nine<br>(3x3 Grid logic) | "Row 1: Left has all 9 positions filled middle<br>emphasizes corners The combination is still all 9<br>positions"<br>(Relational Over-complexity: Hallucinates a global<br>superposition rule for a simple local pattern.) | Pred: 5<br>Correct: 2 |
+| Text MC<br>Gemini 3 Pro | text rapm 0 | Simple Arithmetic<br>(Char length) | "Row 1: 6, 10, 13 characters Row 2: 6, 9, 12<br>notice a potential arithmetic-like progression"<br>(Numerological Hallucination: Forces arithmetic se<br>ries on string lengths where rule is character types.) | Pred: 3<br>Correct: 5 |
+| Text MC<br>GPT-5 | text rapm 4 | Column Count<br>(Uppercase) | "Column 2 uppercase counts go 0 1 2. So for<br>Column 3 they should be 0 1 2."<br>(False Generalization: Hallucinates column-wise<br>progression that doesn't exist in data.) | Pred: 3<br>Correct: 6 |
+| Text Gen<br>GPT-5 | text rapm 2 | Constraint Satisfac<br>tion<br>(Vowel Count) | "Rows keep a constant number of uppercase let<br>ters Column 3 needs 6 vowels A valid comple<br>tion is 'aeiOuEsTH'."<br>(Metadata vs Content: Hallucinates content string<br>that fits metadata counts but fails derivation logic.) | Pred: aeiOuEsTH<br>Rules:<br>5 lowercase<br>vowels, 4 uppercase<br>letters |
+| Text Gen<br>Gemini 3 Pro | text rapm 3 | String Length<br>(Decreasing) | "Column 2 seems to be decreasing length Row 3:<br>6, 9, ? Missing string should be length 8."<br>(Numerological: Obsesses over length counts, ignor<br>ing character content patterns.) | Pred:<br>T4Z]LBEBUM6<br>Rules:<br>Col progres<br>of<br>uppercase<br>sion<br>vowels (2→3→4) |
+{16}------------------------------------------------
+#### E.2. Spatial Working Memory (SWM)
+The complete results for SWM tests are provided in Table [12](#page-16-0) for text modality, in Table [13](#page-16-1) for image modality, and in [14](#page-17-0) for image-text modality. All metrics are averaged over three runs.
+<span id="page-16-0"></span>*Table 12.* Spatial Working Memory full results (text modality). All metrics are reported as mean with standard deviation (std) in a subscript.
+| Difficulty | Model | Sswm | Tokens Score | Guesses | Illegal | Invalid | Repeated | No-box |
+|--------------|-----------------|-----------|--------------|---------------|--------------|--------------|--------------|------------|
+| | GPT-5 | 1.000±.00 | 1.000±.00 | 18.000±6.16 | 0.000±0.00 | 0.000±0.00 | 0.000±0.00 | 0.000±0.00 |
+| | Gemini 3 Pro | 1.000±.00 | 1.000±.00 | 18.000±2.94 | 0.000±0.00 | 0.000±0.00 | 0.000±0.00 | 0.000±0.00 |
+| | Gemini 2.5 Pro | 1.000±.00 | 1.000±.00 | 19.667±1.70 | 0.000±0.00 | 0.000±0.00 | 0.000±0.00 | 0.000±0.00 |
+| | Claude Sonnet 4 | 1.000±.00 | 1.000±.00 | 23.000±3.56 | 0.000±0.00 | 0.000±0.00 | 0.000±0.00 | 0.000±0.00 |
+| Easy | Grok 4.1 Fast | 1.000±.00 | 1.000±.00 | 26.667±2.87 | 0.000±0.00 | 0.000±0.00 | 0.000±0.00 | 0.000±0.00 |
+| | Grok 4 Fast | 1.000±.00 | 1.000±.00 | 18.333±3.40 | 0.000±0.00 | 0.333±0.47 | 0.000±0.00 | 0.000±0.00 |
+| | o4-Mini | 0.951±.05 | 1.000±.00 | 23.667±4.03 | 0.000±0.00 | 0.667±0.94 | 1.000±0.82 | 0.000±0.00 |
+| | GLM 4.6V | 1.000±.00 | 1.000±.00 | 24.000±2.94 | 0.000±0.00 | 3.333±0.94 | 0.000±0.00 | 0.000±0.00 |
+| | GLM 4.5V | 1.000±.00 | 1.000±.00 | 20.333±1.70 | 0.000±0.00 | 0.000±0.00 | 0.000±0.00 | 0.000±0.00 |
+| | Qwen3-VL-235B | 1.000±.00 | 1.000±.00 | 20.000±2.45 | 0.000±0.00 | 0.000±0.00 | 0.000±0.00 | 0.000±0.00 |
+| | GPT-5 | 1.000±.00 | 1.000±.00 | 56.333±1.25 | 0.000±0.00 | 0.000±0.00 | 0.000±0.00 | 0.000±0.00 |
+| | Gemini 3 Pro | 1.000±.00 | 1.000±.00 | 63.333±1.70 | 0.000±0.00 | 1.000±0.82 | 0.000±0.00 | 0.000±0.00 |
+| | Gemini 2.5 Pro | 0.980±.03 | 1.000±.00 | 82.333±13.57 | 0.000±0.00 | 0.000±0.00 | 2.000±2.83 | 0.000±0.00 |
+| | Claude Sonnet 4 | 0.866±.18 | 1.000±.00 | 78.333±11.44 | 0.000±0.00 | 0.333±0.47 | 12.333±16.74 | 0.000±0.00 |
+| Hard | Grok 4.1 Fast | 0.983±.01 | 1.000±.00 | 52.667±9.29 | 0.000±0.00 | 0.000±0.00 | 1.000±0.82 | 0.000±0.00 |
+| | Grok 4 Fast | 0.994±.01 | 1.000±.00 | 63.000±4.55 | 0.333±0.47 | 0.333±0.47 | 0.000±0.00 | 0.000±0.00 |
+| | o4-Mini | 0.811±.07 | 1.000±.00 | 79.333±1.70 | 5.333±3.77 | 0.000±0.00 | 9.667±2.36 | 0.000±0.00 |
+| | GLM 4.6V | 0.574±.35 | 0.694±.27 | 129.333±20.74 | 3.000±3.56 | 33.333±34.50 | 32.333±44.32 | 0.000±0.00 |
+| | GLM 4.5V | 0.536±.24 | 0.736±.37 | 105.333±27.64 | 16.667±11.81 | 42.667±58.93 | 0.333±0.47 | 0.000±0.00 |
+| | Qwen3-VL-235B | 0.760±.06 | 1.000±.00 | 80.333±10.66 | 19.333±7.72 | 1.000±0.00 | 0.333±0.47 | 0.000±0.00 |
+| | Gemini 2.5 Pro | 0.965±.04 | 1.000±.00 | 66.667±9.18 | 0.000±0.00 | 0.000±0.00 | 2.333±2.62 | 0.000±0.00 |
+| | Claude Sonnet 4 | 0.826±.06 | 1.000±.00 | 81.000±9.27 | 1.000±0.82 | 0.000±0.00 | 13.667±5.56 | 0.000±0.00 |
+| | Grok 4.1 Fast | 0.973±.01 | 1.000±.00 | 62.333±3.30 | 0.333±0.47 | 0.000±0.00 | 1.333±0.47 | 0.000±0.00 |
+| Hard - Notes | Grok 4 Fast | 0.964±.01 | 1.000±.00 | 65.333±1.25 | 0.333±0.47 | 0.000±0.00 | 2.000±0.00 | 0.000±0.00 |
+| | o4-Mini | 0.935±.03 | 1.000±.00 | 75.333±15.92 | 1.000±1.41 | 0.333±0.47 | 4.333±2.49 | 0.000±0.00 |
+| | GLM 4.6V | 0.803±.14 | 1.000±.00 | 85.000±20.22 | 0.333±0.47 | 1.000±0.00 | 19.000±16.87 | 0.000±0.00 |
+| | GLM 4.5V | 0.851±.09 | 1.000±.00 | 79.333±11.12 | 10.333±7.36 | 14.333±16.11 | 0.333±0.47 | 0.000±0.00 |
+| | Qwen3-VL-235B | 0.782±.23 | 0.972±.04 | 92.333±36.81 | 0.667±0.94 | 2.333±3.30 | 25.667±32.83 | 0.000±0.00 |
+<span id="page-16-1"></span>*Table 13.* Spatial Working Memory full results (image modality). All metrics are reported as mean with standard deviation (std) in a subscript.
+| Difficulty | Model | Sswm | Tokens Score | Guesses | Illegal | Invalid | Repeated | No-box |
+|------------|-----------------|------------|--------------|--------------|--------------|---------------|--------------|--------------|
+| | GPT-5 | 0.927±0.08 | 1.000±0.00 | 27.667±4.64 | 0.000±0.00 | 0.000±0.00 | 2.000±2.83 | 0.333±0.47 |
+| | Gemini 3 Pro | 1.000±0.00 | 1.000±0.00 | 22.333±0.94 | 0.000±0.00 | 0.000±0.00 | 0.000±0.00 | 0.000±0.00 |
+| | Gemini 2.5 Pro | 0.408±0.38 | 0.625±0.44 | 51.333±17.91 | 0.000±0.00 | 0.000±0.00 | 2.667±3.77 | 30.667±20.42 |
+| | Claude Sonnet 4 | 0.630±0.41 | 0.750±0.35 | 43.667±14.43 | 0.000±0.00 | 0.000±0.00 | 4.000±5.66 | 14.000±15.75 |
+| | Grok 4.1 Fast | 0.014±0.01 | 0.125±0.10 | 64.000±0.00 | 0.333±0.47 | 0.000±0.00 | 5.000±5.72 | 51.333±5.25 |
+| Easy | Grok 4 Fast | 0.029±0.04 | 0.125±0.18 | 64.000±0.00 | 0.000±0.00 | 0.333±0.47 | 14.333±13.27 | 42.000±10.61 |
+| | o4-Mini | 0.042±0.04 | 0.208±0.21 | 64.000±0.00 | 0.000±0.00 | 1.333±0.47 | 30.667±3.09 | 21.333±0.94 |
+| | GLM 4.6V | 0.000±0.00 | 0.000±0.00 | 64.000±0.00 | 0.000±0.00 | 63.333±0.47 | 0.000±0.00 | 0.000±0.00 |
+| | GLM 4.5V | 0.000±0.00 | 0.000±0.00 | 64.000±0.00 | 0.000±0.00 | 32.000±2.83 | 6.333±6.34 | 23.333±5.25 |
+| | Qwen3-VL-235B | 0.031±0.02 | 0.208±0.12 | 64.000±0.00 | 12.667±11.09 | 2.667±3.77 | 26.333±20.34 | 14.000±13.37 |
+| | GPT-5 | 0.631±0.16 | 1.000±0.00 | 98.667±19.07 | 10.333±2.49 | 2.000±2.83 | 14.667±12.26 | 13.333±18.15 |
+| | Gemini 3 Pro | 0.793±0.24 | 0.958±0.06 | 90.000±38.88 | 20.000±27.58 | 1.000±0.82 | 4.667±4.11 | 0.000±0.00 |
+| | Gemini 2.5 Pro | 0.622±0.30 | 0.944±0.08 | 97.333±34.92 | 35.000±30.82 | 1.000±1.41 | 9.000±11.34 | 0.333±0.47 |
+| | Claude Sonnet 4 | 0.005±0.00 | 0.056±0.02 | 144.000±0.00 | 0.000±0.00 | 44.333±0.47 | 8.000±3.74 | 82.667±2.87 |
+| Hard | Grok 4.1 Fast | 0.009±0.01 | 0.069±0.04 | 144.000±0.00 | 0.000±0.00 | 25.333±26.39 | 7.333±8.99 | 99.667±27.98 |
+| | Grok 4 Fast | 0.005±0.01 | 0.056±0.05 | 144.000±0.00 | 0.000±0.00 | 39.000±53.03 | 7.667±7.04 | 88.333±42.76 |
+| | o4-Mini | 0.005±0.00 | 0.056±0.02 | 144.000±0.00 | 0.000±0.00 | 1.333±1.25 | 32.000±13.49 | 98.333±13.47 |
+| | GLM 4.6V | 0.028±0.04 | 0.069±0.10 | 144.000±0.00 | 0.000±0.00 | 119.333±34.88 | 8.333±11.79 | 6.333±8.96 |
+| | GLM 4.5V | 0.000±0.00 | 0.000±0.00 | 144.000±0.00 | 0.000±0.00 | 111.667±0.94 | 0.667±0.94 | 30.667±2.05 |
+| | Qwen3-VL-235B | 0.135±0.12 | 0.347±0.27 | 144.000±0.00 | 4.333±6.13 | 63.000±1.41 | 30.667±14.06 | 21.000±19.87 |
+{17}------------------------------------------------
+## A Neuropsychologically Grounded Evaluation of LLM Cognitive Abilities
+<span id="page-17-0"></span>Table 14. Spatial Working Memory full results (image-text modality). All metrics are reported as mean with standard deviation (std) in a subscript.
+| Difficulty | Model | $\boldsymbol{S}_{swm}$ | Tokens Score | Guesses | Illegal | Invalid | Repeated | No-box |
+|--------------|------------------------------------------------------------------------------------------------------------------------------------------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+| Easy | GPT-5<br>Gemini 3 Pro<br>Gemini 2.5 Pro<br>Claude Sonnet 4<br>Grok 4.1 Fast<br>Grok 4 Fast<br>o4-Mini<br>GLM 4.6V<br>GLM 4.5V<br>Qwen3-VL-235B | $\begin{array}{c} 0.986_{\pm .02} \\ 1.000_{\pm .00} \\ 0.733_{\pm .28} \\ 0.256_{\pm .18} \\ 0.145_{\pm .20} \\ 0.021_{\pm .03} \\ 0.055_{\pm .03} \\ 0.044_{\pm .05} \\ 0.044_{\pm .02} \\ 0.026_{\pm .05} \end{array}$ | $\begin{array}{c} 1.000_{\pm.00} \\ 1.000_{\pm.00} \\ 0.958_{\pm0.06} \\ 0.625_{\pm0.37} \\ 0.375_{\pm0.44} \\ 0.125_{\pm0.18} \\ 0.292_{\pm0.12} \\ 0.167_{\pm0.16} \\ 0.208_{\pm0.12} \\ 0.125_{\pm.10} \end{array}$ | $\begin{array}{c} 24.000_{\pm 2.16} \\ 20.667_{\pm 1.70} \\ 36.667_{\pm 19.60} \\ 59.333_{\pm 6.60} \\ 53.667_{\pm 14.61} \\ 64.000_{\pm 0.00} \\ 64.000_{\pm 0.00} \\ 64.000_{\pm 0.00} \\ 64.000_{\pm 0.00} \\ 64.000_{\pm 0.00} \end{array}$ | $\begin{array}{c} 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 1.000_{\pm 0.10} \\ 0.000_{\pm 0.10} \\ 1.000_{\pm 1.41} \end{array}$ | $\begin{array}{c} 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.667_{\pm 0.94} \\ 0.333_{\pm 0.47} \\ 0.333_{\pm 0.47} \\ 27.667_{\pm 26.04} \\ 32.333_{\pm 4.03} \\ 0.000_{\pm 0.00} \end{array}$ | $\begin{array}{c} 0.333_{\pm 0.47} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 1.667_{\pm 2.36} \\ 4.333_{\pm 5.44} \\ 0.000_{\pm 0.00} \\ 3.333_{\pm 3.40} \\ 6.333_{\pm 5.31} \\ 0.667_{\pm 0.47} \\ 5.333_{\pm 6.85} \end{array}$ | $\begin{array}{c} 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 14.000_{\pm 17.68} \\ 39.000_{\pm 12.36} \\ 40.333_{\pm 16.94} \\ 56.333_{\pm 2.49} \\ 49.000_{\pm 4.97} \\ 19.000_{\pm 13.74} \\ 24.333_{\pm 4.50} \\ 47.000_{\pm 7.48} \end{array}$ |
+| Easy - Notes | Gemini 2.5 Pro<br>Claude Sonnet 4<br>Grok 4.1 Fast<br>Grok 4 Fast<br>o4-Mini<br>GLM 4.6V<br>GLM 4.5V<br>Qwen3-VL-235B | $\begin{array}{c} 0.704_{\pm.42} \\ 0.207_{\pm.18} \\ 0.043_{\pm.06} \\ 0.029_{\pm.02} \\ 0.012_{\pm.01} \\ 0.082_{\pm.09} \\ 0.000_{\pm.00} \\ 0.318_{\pm.45} \end{array}$ | $\begin{array}{c} 0.792_{\pm .29} \\ 0.542_{\pm .41} \\ 0.167_{\pm .24} \\ 0.208_{\pm .06} \\ 0.125_{\pm .10} \\ 0.333_{\pm .29} \\ 0.000_{\pm .00} \\ 0.333_{\pm .47} \end{array}$ | $\begin{array}{c} 32.667_{\pm 22.17} \\ 60.667_{\pm 4.71} \\ 64.000_{\pm 0.00} \\ 64.000_{\pm 0.00} \\ 64.000_{\pm 0.00} \\ 64.000_{\pm 0.00} \\ 64.000_{\pm 0.00} \\ 50.000_{\pm 19.80} \end{array}$ | $\begin{array}{c} 0.000_{\pm 0.00} \\ 0.333_{\pm 0.47} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \end{array}$ | $\begin{array}{c} 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 7.000_{\pm 5.35} \\ 0.333_{\pm 0.47} \\ 1.333_{\pm 0.94} \\ 3.333_{\pm 3.40} \\ 30.667_{\pm 1.25} \\ 1.667_{\pm 2.36} \end{array}$ | $\begin{array}{c} 0.333_{\pm 0.47} \\ 0.667_{\pm 0.94} \\ 7.333_{\pm 10.37} \\ 5.333_{\pm 2.87} \\ 0.333_{\pm 0.47} \\ 16.000_{\pm 22.63} \\ 0.000_{\pm 0.00} \\ 0.667_{\pm 0.47} \end{array}$ | $\begin{array}{c} 14.667_{\pm 20.74} \\ 43.333_{\pm 11.03} \\ 42.667_{\pm 16.50} \\ 50.333_{\pm 5.56} \\ 56.667_{\pm 0.47} \\ 33.333_{\pm 19.34} \\ 32.667_{\pm 1.25} \\ 38.000_{\pm 26.98} \end{array}$ |
+| Hard | GPT-5<br>Gemini 3 Pro<br>Gemini 2.5 Pro<br>Claude Sonnet 4<br>Grok 4.1 Fast<br>Grok 4 Fast<br>o4-Mini<br>GLM 4.6V<br>GLM 4.5V<br>Qwen3-VL-235B | $\begin{array}{c} 0.989_{\pm .02} \\ 0.963_{\pm .05} \\ 0.480_{\pm .12} \\ 0.265_{\pm .31} \\ 0.017_{\pm .02} \\ 0.002_{\pm .00} \\ 0.032_{\pm .02} \\ 0.010_{\pm .01} \\ 0.000_{\pm .00} \\ 0.016_{\pm .01} \end{array}$ | $\begin{array}{c} 1.000_{\pm.00} \\ 1.000_{\pm.00} \\ 0.972_{\pm0.04} \\ 0.472_{\pm0.38} \\ 0.097_{\pm0.14} \\ 0.028_{\pm0.02} \\ 0.181_{\pm0.10} \\ 0.042_{\pm0.06} \\ 0.000_{\pm0.00} \\ 0.111_{\pm0.10} \end{array}$ | $\begin{array}{c} 59.333_{\pm 5.79} \\ 66.000_{\pm 5.10} \\ 120.000_{\pm 18.49} \\ 122.333_{\pm 30.64} \\ 144.000_{\pm 0.00} \\ 144.000_{\pm 0.00} \\ 144.000_{\pm 0.00} \\ 144.000_{\pm 0.00} \\ 144.000_{\pm 0.00} \\ 144.000_{\pm 0.00} \\ 144.000_{\pm 0.00} \\ \end{array}$ | $\begin{array}{c} 0.667_{\pm 0.94} \\ 2.667_{\pm 3.77} \\ 52.000_{\pm 29.63} \\ 5.667_{\pm 7.32} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.333_{\pm 0.47} \end{array}$ | $\begin{array}{c} 0.000_{\pm 0.00} \\ 1.333_{\pm 0.47} \\ 0.333_{\pm 0.47} \\ 29.333_{\pm 20.74} \\ 10.000_{\pm 7.48} \\ 46.667_{\pm 58.39} \\ 2.000_{\pm 0.82} \\ 119.667_{\pm 34.41} \\ 110.333_{\pm 1.70} \\ 19.667_{\pm 0.47} \end{array}$ | $\begin{array}{c} 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 0.0667_{\pm 0.94} \\ 3.000_{\pm 0.82} \\ 1.000_{\pm 0.82} \\ 2.000_{\pm 2.83} \\ 11.000_{\pm 8.04} \\ 5.000_{\pm 7.07} \\ 0.000_{\pm 0.00} \\ 28.667_{\pm 6.60} \end{array}$ | $\begin{array}{c} 0.000_{\pm 0.00} \\ 0.000_{\pm 0.00} \\ 10.333_{\pm 13.91} \\ 52.667_{\pm 34.57} \\ 120.667_{\pm 14.64} \\ 89.000_{\pm 53.11} \\ 110.333_{\pm 13.07} \\ 13.333_{\pm 18.86} \\ 33.667_{\pm 1.70} \\ 80.667_{\pm 3.68} \end{array}$ |
+{18}------------------------------------------------
+#### E.3. Wisconsin Card Sorting Test (WCST)
+The complete results for WCST tests are provided in Table [15](#page-18-0) for text modality, and in Table [16](#page-18-1) for image modality. All metrics are averaged over three runs.
+<span id="page-18-0"></span>*Table 15.* Wisconsin Card Sorting Test full results (text modality). All metrics are reported as mean with standard deviation (std) in a subscript.
+| Difficulty | Model | Swcst | Accuracy | FMS | PR | Completed Rules | 1st Rule Trials |
+|------------|-----------------|-----------|-----------|-----------|-----------|-----------------|-----------------|
+| Easy | GPT-5 | 0.848±.02 | 0.842±.02 | 0.000±.00 | 0.000±.00 | 6.000±0.00 | 6.000±0.00 |
+| | Gemini 3 Pro | 0.804±.05 | 0.815±.04 | 0.000±.00 | 0.003±.01 | 7.000±1.00 | 6.833±1.07 |
+| | Gemini 2.5 Pro | 0.825±.02 | 0.819±.02 | 0.000±.00 | 0.000±.00 | 6.000±0.00 | 6.667±0.94 |
+| | Claude Sonnet 4 | 0.764±.06 | 0.786±.03 | 0.000±.00 | 0.027±.02 | 6.000±0.00 | 7.000±0.82 |
+| | Grok 4.1 Fast | 0.841±.03 | 0.835±.03 | 0.000±.00 | 0.000±.00 | 6.000±0.00 | 6.000±0.00 |
+| | Grok 4 Fast | 0.836±.02 | 0.834±.02 | 0.000±.00 | 0.000±.00 | 6.000±0.00 | 6.667±0.47 |
+| | o4-Mini | 0.789±.01 | 0.805±.03 | 0.020±.03 | 0.011±.02 | 6.000±0.00 | 6.333±0.47 |
+| | GLM 4.6V | 0.300±.35 | 0.343±.33 | 0.305±.43 | 0.115±.09 | 2.333±2.62 | 29.000±25.66 |
+| | GLM 4.5V | 0.773±.03 | 0.789±.02 | 0.000±.00 | 0.019±.03 | 6.000±0.00 | 7.000±0.00 |
+| | Qwen3-VL-235B | 0.230±.14 | 0.243±.15 | 0.766±.17 | 0.731±.16 | 2.667±1.89 | 43.000±39.05 |
+| Hard | GPT-5 | 0.815±.00 | 0.831±.02 | 0.000±.00 | 0.000±.00 | 8.000±0.00 | 6.667±0.47 |
+| | Gemini 3 Pro | 0.776±.06 | 0.803±.05 | 0.000±.00 | 0.007±.01 | 8.000±0.00 | 7.333±1.25 |
+| | Gemini 2.5 Pro | 0.806±.04 | 0.827±.01 | 0.000±.00 | 0.000±.00 | 8.000±0.00 | 7.000±0.00 |
+| | Claude Sonnet 4 | 0.773±.06 | 0.805±.02 | 0.000±.00 | 0.007±.01 | 8.000±0.00 | 7.333±1.25 |
+| | Grok 4.1 Fast | 0.825±.03 | 0.835±.01 | 0.000±.00 | 0.009±.02 | 8.000±0.00 | 6.222±0.42 |
+| | Grok 4 Fast | 0.754±.04 | 0.768±.03 | 0.000±.00 | 0.007±.01 | 8.000±0.00 | 6.667±0.94 |
+| | o4-Mini | 0.792±.02 | 0.798±.02 | 0.000±.00 | 0.000±.00 | 8.000±0.00 | 6.667±0.47 |
+| | GLM 4.6V | 0.095±.12 | 0.175±.08 | 0.426±.41 | 0.214±.14 | 0.889±1.29 | 63.111±39.82 |
+| | GLM 4.5V | 0.309±.08 | 0.184±.09 | 0.000±.00 | 0.680±.04 | 3.000±1.63 | 6.667±0.47 |
+| | Qwen3-VL-235B | 0.233±.12 | 0.438±.12 | 0.475±.16 | 0.505±.11 | 3.333±1.25 | 9.667±4.50 |
+<span id="page-18-1"></span>*Table 16.* Wisconsin Card Sorting Test full results (image modality). All metrics are reported as mean with standard deviation (std) in a subscript.
+| Difficulty | Model | Swcst | Accuracy | FMS | PR | Completed Rules | 1st Rule Trials |
+|------------|-----------------|-----------|-----------|-----------|-----------|-----------------|-----------------|
+| Easy | GPT-5 | 0.824±.05 | 0.833±.03 | 0.000±.00 | 0.010±.01 | 6.000±0.00 | 6.000±0.00 |
+| | Gemini 3 Pro | 0.861±.00 | 0.857±.00 | 0.000±.00 | 0.000±.00 | 6.000±0.00 | 6.000±0.00 |
+| | Gemini 2.5 Pro | 0.823±.04 | 0.820±.04 | 0.000±.00 | 0.000±.00 | 6.000±0.00 | 7.000±0.82 |
+| | Claude Sonnet 4 | 0.800±.02 | 0.797±.03 | 0.000±.00 | 0.010±.01 | 6.000±0.00 | 7.000±0.00 |
+| | Grok 4.1 Fast | 0.107±.07 | 0.432±.08 | 0.339±.24 | 0.324±.04 | 1.000±0.82 | 37.667±22.45 |
+| | Grok 4 Fast | 0.336±.08 | 0.602±.09 | 0.432±.10 | 0.201±.10 | 3.500±1.50 | 9.500±0.50 |
+| | o4-Mini | 0.347±.11 | 0.557±.03 | 0.320±.04 | 0.092±.05 | 3.333±0.94 | 7.000±0.82 |
+| | GLM 4.6V | 0.191±.11 | 0.297±.08 | 0.089±.13 | 0.442±.15 | 1.333±1.25 | 26.000±27.58 |
+| | GLM 4.5V | 0.108±.11 | 0.315±.22 | 0.404±.32 | 0.310±.21 | 0.333±0.47 | 25.667±13.22 |
+| | Qwen3-VL-235B | 0.420±.19 | 0.361±.09 | 0.229±.22 | 0.619±.09 | 4.000±0.82 | 7.333±1.89 |
+| Hard | GPT-5 | 0.794±.05 | 0.819±.00 | 0.000±.00 | 0.000±.00 | 8.000±0.00 | 7.000±0.82 |
+| | Gemini 3 Pro | 0.806±.02 | 0.814±.04 | 0.000±.00 | 0.000±.00 | 8.000±0.00 | 6.667±0.47 |
+| | Gemini 2.5 Pro | 0.776±.01 | 0.795±.03 | 0.000±.00 | 0.000±.00 | 8.000±0.00 | 7.333±0.47 |
+| | Claude Sonnet 4 | 0.710±.08 | 0.739±.06 | 0.134±.12 | 0.069±.06 | 8.000±0.00 | 6.667±0.47 |
+| | Grok 4.1 Fast | 0.077±.05 | 0.369±.09 | 0.376±.26 | 0.420±.07 | 0.714±0.70 | 44.143±20.58 |
+| | Grok 4 Fast | 0.209±.07 | 0.493±.04 | 0.347±.25 | 0.309±.03 | 3.000±1.41 | 11.333±2.36 |
+| | o4-Mini | 0.277±.13 | 0.528±.05 | 0.437±.09 | 0.352±.07 | 3.889±1.37 | 9.222±3.58 |
+| | GLM 4.6V | 0.064±.07 | 0.197±.09 | 0.169±.32 | 0.420±.12 | 0.444±0.96 | 64.556±26.88 |
+| | GLM 4.5V | 0.083±.07 | 0.398±.22 | 0.303±.25 | 0.432±.11 | 0.333±0.47 | 32.000±4.97 |
+| | Qwen3-VL-235B | 0.071±.07 | 0.208±.13 | 0.500±.38 | 0.794±.14 | 1.333±1.25 | 59.333±38.77 |
+{19}------------------------------------------------
+## F. Prompts
+Below we list the system prompts used for each task. We also provide alternate or additional prompts used for the different setups within each task.
+## System Prompt for Image RAPM
+You are taking the Raven's Progressive Matrices (RAPM) test, a non-verbal intelligence test that measures abstract reasoning ability.
+You will see a 3x3 matrix of images with the bottom-right image missing (shown as a question mark), followed by 8 answer choices numbered 1-8.
+Your task is to:
+- 1. Analyze rows and columns
+- 2. Infer the governing logical rule(s)
+- 3. Select the answer choice (1-8) that correctly completes the matrix.
+#### /\* PATTERN HINT \*/
+The patterns can involve:
+- Shape transformations (rotation, reflection, scaling)
+- 1062 - Position changes (movement, arrangement)
+- 1063 - Attribute changes (color, size, number of elements)
+- 1064 - Logical operations (addition, subtraction, intersection)
+- Sequence progressions (systematic changes across rows/columns)
+- 1065 Additional common rule types:
+- 1066 1067 - Constant-in-row: Same value across a row; varies down columns.
+- Quantitative step: Fixed +/- increment between adjacent cells (size / count / position offset).
+- Figure add/subtract: Combine (overlay or juxtapose) or remove elements from two cells to form the third.
+- Distribution-of-three: Three distinct categorical values appear once each per row (order may permute).
+- Distribution-of-two: Two values each appear once; third slot is empty / null.
+- Look horizontally and vertically; the missing piece must satisfy ALL relevant row and column rules.
+Your final answer should be a number between 1-8 corresponding to the correct choice.
+#### System Prompt for Text RAPM
+You are solving a TEXT-BASED 3x3 pattern matrix (Raven-style). Each cell contains a string; the bottom-right cell is missing ('?').
+Goal: Infer the rule(s) acting across rows and columns.
+#### /\* PATTERN HINT \*/
+Possible dimensions (one or more):
+- Character set restriction (digits / letters / symbols)
+- Quantitative constant (exact length / count / unique)
+- Quantitative progression (arithmetic step across row/column)
+- Parity / multiple rules (all even / all odd / multiples of N)
+- Positional constraints (first/last/even/odd positions restricted)
+- Ordering (ascending / descending / mixed)
+- Layered combinations (e.g. constant + parity, progression + positional)
+#### 1089 1090 /\* MC VERSION \*/
+You will be given 8 answer options (1-8). Select the single option that correctly fills the missing cell while satisfying ALL inferred row and column constraints.
+1092 Respond with <answer>NUMBER</answer> using just the chosen option number.
+#### 1093 1094 /\* GEN VERSION \*/
+You must GENERATE the exact missing cell string that satisfies ALL inferred row and column constraints.
+Respond with <answer>STRING</answer> containing only the candidate string (no quotes or extra text).
+{20}------------------------------------------------
+#### A Neuropsychologically Grounded Evaluation of LLM Cognitive Abilities System Prompt for SWM You will be performing the Spatial Working Memory task. You will be given an image containing {N BOXES} yellow boxes in a grid. There are {N TOKENS} types of tokens, hidden in any one of {N BOXES} boxes. Each token type is represented by a distinct color. Your goal is to find the {N TOKENS} types of tokens {N BOXES} times each, by repeatedly selecting a box to open. A box can contain multiple types of tokens, but only one token of each type. If the box contains multiple tokens, a token with mixed colors corresponding to the tokens will be shown. Once the token is found, another will be generated in another box. The token will be generated in a box that has never contained a token of that type before in the trial. The token may be generated in a box that has been opened and found empty before, as long as it never contained that type of token previously. /\* TEXT VERSION \*/ Your final answer should be a number from 1-{N BOXES}, the index of the box you selected /\* IMAGE VERSION \*/ Your final answer should be a grid coordinate (x, y), the grid coordinate of the box you choose System Prompt for WCST You are performing the Wisconsin Card Sorting Test (WCST). You will be shown a given card with a symbol on it, and you will have to match it to one of four option cards according to an attribute that you have to figure out. The cards will be described by the following attributes: 1. Number of symbols 2. Color of symbols 3. Shape of symbols /\* HARD VERSION \*/ 4. Background color of the card You will be told "Correct!" if you are correct and "Incorrect. Please try again." if you are incorrect. If you are incorrect, you either made a mistake or the rule has changed. If you believe you have made a mistake, correct it and try again. If you believe the rule has changed, you have to figure out the correct rule to match the cards. If you are correct, you have to stick with the same attribute until you are incorrect. There is always a true answer in the task, and you have to keep performing the task until the end of the test. Your final answer should be a number between 1-4 corresponding to the index of the card you think is the correct match.
+{21}------------------------------------------------
+## G. Turn Examples
+*[. . . omitted turns . . . ]*
+User: No tokens found in box (3, 3). Answer concisely. Think step-by-step, utilizing information from previous feedbacks, and state your reasoning in maximum 4000 tokens, wrapped with <think> and </think>. Then, provide a really short summary of your reasoning after the closing </think> tag.
+Which of the 8 boxes would you like to open?
+Your final answer should be a grid coordinate (x, y), wrapped with <answer> and </answer>
+![](_page_21_Figure_34.jpeg)
+Assistant: <think>I already found one, located at position (1, 0). [. . . ] I know that boxes (5, 0), (5, 1), (1, 2), and (3, 3) are empty. [. . . ] That leaves me with (10, 4), (0, 6), and (9, 6). To keep things organized and ensure I'm systematic, I'm going to follow my pre-defined search pattern – working from the top-left and moving to the right and down. That means the next box I need to check is (10, 4). [. . . ] <answer>(10, 4)</answer>
+*[. . . omitted turns . . . ]*
+User: No tokens found in box 1. Boxes that has contained token A: 11, Boxes that has contained token B: 8, 10,
+Opened boxes: 12,
+Answer concisely. Think step-by-step, utilizing information from previous feedbacks, and state your reasoning in maximum 4000 tokens, wrapped with <think> and </think>. Then, provide a really short summary of your reasoning after the closing </think> tag.
+Which of the 12 boxes would you like to open?
+Your final answer should be a box number, wrapped with <answer> and </answer>
+Assistant: [. . . ]The message provides data on boxes: 11 contained token A, 8 and 10 contained token B, and 12 boxes were opened. [. . . ] The phrase "Opened boxes" suggests the total number of boxes accessed, with partial information on token contents.[. . . ] The phrase "Opened boxes: 12" likely means box 12 was last opened and found empty, given prior context.[. . . ]
+\*\*Summary:\*\* 1 empty again; pure empties left 2-7,9,12; systematic next=2.
+<answer>2</answer>
+{22}------------------------------------------------
+*[. . . omitted turns . . . ]*
+User: Incorrect. Please try again.
+Recent notes:
+- Turn -1: matching number of symbols and symbol color Incorrect.
+- Turn -2: matching number of symbols and symbol color Incorrect.
+- Turn -3: matching symbol color Correct.
+- Turn -4: matching symbol color Correct.
+- Turn -5: matching symbol color Correct.
+- Turn -6: matching symbol color Correct.
+Look at the image showing 5 cards. Match the 'Given' card to one of cards 1-4 based on the rule you need to figure out.
+![](_page_22_Picture_32.jpeg)
+Assistant: The previous attempts to match by symbol color (Turns 2-6) were correct until Turn 7, where the response was incorrect. [. . . ] However, the last incorrect attempt used "number of symbols and symbol color," which was wrong. The new rule might now be \*\*background color\*\*.[. . . ] Re-evaluating: The prior correct streak (Turns 2-6) used symbol color. After an incorrect response, the rule often shifts to \*\*number of symbols\*\*. [. . . ] Wait, the feedback says "matching number of symbols and symbol color – Incorrect" for Turn 7. The correct rule might now be \*\*background color\*\*. [. . . ]The rule shifted to background color; Given card has a purple background, matching Card 3.
+<answer>3</answer>
+*[. . . omitted turns . . . ]*
+User: Incorrect. Please try again.
+Given: two red triangle
+Options:
+- 1. two green triangle
+- 2. four yellow square
+- 3. three blue star
+- 4. one red circle
+Assistant: [. . . ] My initial hypothesis (number and shape) was incorrect. The feedback suggests a different rule. The given card is red. Option 4 is also red, while others have different colors. I'll now hypothesize the rule is matching by color regardless of number or shape.<answer>4</answer>My initial hypothesis (number and shape) was incorrect. The feedback suggests a different rule. The given card is red. Option 4 is also red, while others have different colors. I'll now hypothesize the rule is matching by color regardless of number or shape.