v1

.gitignore

@@ -0,0 +1,29 @@
+# Paper LaTeX source (not needed for HF Space)
+/paper_latex
+
+# Python
+__pycache__/
+*.py[cod]
+*.pyo
+*.egg-info/
+dist/
+build/
+*.egg
+
+# Virtual environments
+.venv/
+venv/
+env/
+
+# IDE
+.idea/
+.vscode/
+*.swp
+*.swo
+
+# OS
+.DS_Store
+Thumbs.db
+
+# Gradio
+flagged/

README.md

@@ -1,12 +1,12 @@
 ---
-title: M
-emoji: 🦀
-colorFrom: gray
-colorTo: pink
+title: M3-Bench Leaderboard
+emoji: 🎮
+colorFrom: indigo
+colorTo: purple
 sdk: gradio
-sdk_version: 6.9.0
+sdk_version: 6.11.0
 app_file: app.py
 pinned: false
+license: apache-2.0
+short_description: "Multi-Model Multi-View Game-Theoretic Benchmark for LLMs"
 ---
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

app.py

@@ -0,0 +1,16 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+M3-Bench Leaderboard – HF Space entry point
+"""
+
+from __future__ import annotations
+
+from create_leaderboard import demo
+
+if __name__ == "__main__":
+    demo.launch(
+        server_name="0.0.0.0",
+        server_port=7860,
+        share=False,
+    )

create_leaderboard.py

@@ -0,0 +1,216 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+M3-Bench Leaderboard – Gradio UI
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+from datetime import datetime
+
+import pandas as pd
+import gradio as gr
+
+from tabs.leaderboard_tab import create_leaderboard_tab
+
+
+def get_leaderboard_info():
+    leaderboard_path = Path(__file__).parent / "data" / "leaderboard.csv"
+    if leaderboard_path.exists():
+        try:
+            df = pd.read_csv(leaderboard_path)
+            model_count = len(df["model"].unique())
+            mtime = leaderboard_path.stat().st_mtime
+            last_update = datetime.fromtimestamp(mtime).strftime("%d %B %Y")
+            return model_count, last_update
+        except Exception:
+            pass
+    return 0, "Unknown"
+
+
+model_count, last_update = get_leaderboard_info()
+
+# ---------------------------------------------------------------------------
+# UI
+# ---------------------------------------------------------------------------
+
+with gr.Blocks(title="M3-Bench Leaderboard") as demo:
+
+    gr.HTML("""
+    <style>
+        .title-block {
+            background: linear-gradient(135deg, #667eea 0%, #764ba2 100%);
+            -webkit-background-clip: text;
+            -webkit-text-fill-color: transparent;
+            background-clip: text;
+            color: transparent;
+            text-align: center;
+            font-size: 2.2rem;
+            font-weight: 800;
+            margin: 0.5rem 0 0.5rem 0;
+            display: inline-block;
+            width: 100%;
+        }
+        .subtitle-block {
+            text-align: center;
+            font-size: 1.1rem;
+            color: #555;
+            margin-bottom: 0.5rem;
+        }
+        .intro-block {
+            text-align: center;
+            margin-bottom: 1.25rem;
+            line-height: 2;
+        }
+        .intro-block a {
+            color: #667eea;
+            text-decoration: none;
+            font-weight: 600;
+            margin: 0 0.3rem;
+        }
+        .intro-block a:hover { text-decoration: underline; }
+        .info-badge {
+            display: inline-block;
+            background: #f0f0ff;
+            border: 1px solid #ddd;
+            border-radius: 12px;
+            padding: 2px 10px;
+            margin: 0 4px;
+            font-size: 0.9em;
+            color: #444;
+        }
+    </style>
+    """)
+
+    gr.HTML(f"""
+    <div class="title-block">
+        M3-Bench Leaderboard
+    </div>
+    <div class="subtitle-block">
+        Multi-Model Multi-View Game-Theoretic Benchmark for LLMs
+    </div>
+    <div class="intro-block">
+        Evaluating LLMs through game-theoretic interactions across 4 difficulty levels and 3 evaluation views.<br>
+        <a href="#" target="_blank">📄 Paper</a> |
+        <a href="#" target="_blank">💻 Code</a> |
+        <a href="#" target="_blank">🌐 Website</a>
+        <br>
+        <span class="info-badge">🤖 Models: {model_count}</span>
+        <span class="info-badge">📅 Last Update: {last_update}</span>
+    </div>
+    """)
+
+    with gr.Tabs():
+        create_leaderboard_tab()
+
+        with gr.Tab("📊 About M3-Bench"):
+            gr.Markdown("""
+## About M3-Bench
+
+**M3-Bench** is a comprehensive benchmark for evaluating Large Language Models (LLMs) through
+game-theoretic interactions. It assesses model capabilities across **4 difficulty levels** using
+**3 complementary evaluation views**.
+
+### Evaluation Views
+
+| View | Description |
+|------|-------------|
+| **BTA** (Birds-eye, Technically-grounded, Artistically-aware) | Holistic evaluation combining multiple assessment perspectives |
+| **RPA** (Role-Play Assessment) | Evaluation through role-playing scenarios within games |
+| **CCA** (Compositional Capability Assessment) | Fine-grained assessment of compositional reasoning abilities |
+
+### Difficulty Levels
+
+| Level | Games | Description |
+|-------|-------|-------------|
+| **Level 1** | PD, SH, UG | Basic strategic reasoning (Prisoner's Dilemma, Stag Hunt, Ultimatum Game) |
+| **Level 2** | RPD, GE, AOB | Iterated/extended games (Repeated PD, Gift Exchange, All-or-Nothing Bargaining) |
+| **Level 3** | PGG, VD, CPR | Multi-agent coordination (Public Goods Game, Volunteer's Dilemma, Common Pool Resource) |
+| **Level 4** | AC, WW, KP | Complex strategic reasoning (Auction Competition, Wage War, Keynesian Beauty Contest) |
+
+### Model Categories
+
+- 🔒 **Closed-Source**: Proprietary frontier models (GPT-5.1, Claude Opus 4.5, Gemini-3 Pro, Grok-4.1)
+- 🔓 **Open-Weight**: Open-weight models (GPT-oss 120b, DeepSeek V3.2, Kimi-K2 Thinking, Mistral Large 3)
+- 🧠 **Reasoning**: Reasoning-oriented models (GPT-5 pro, Gemini-3 Deep Think, DeepSeek-V3.2 Speciale)
+- 👤 **Human**: Human participant baseline
+""")
+
+    # Citation block
+    gr.HTML("""
+    <style>
+        .citation-block {
+            margin-top: 2rem;
+            padding: 1.5rem;
+            border: 1px solid #e0e0e0;
+            border-radius: 8px;
+            background-color: #f9f9f9;
+        }
+        .citation-title {
+            font-size: 1.15rem;
+            font-weight: 600;
+            margin-bottom: 1rem;
+            color: #333;
+        }
+        .citation-content {
+            background-color: #fff;
+            border: 1px solid #ddd;
+            border-radius: 4px;
+            padding: 1rem;
+            font-family: monospace;
+            font-size: 0.85rem;
+            white-space: pre-wrap;
+            line-height: 1.5;
+            position: relative;
+        }
+        .copy-btn {
+            position: absolute;
+            top: 8px;
+            right: 8px;
+            padding: 6px 12px;
+            background-color: #667eea;
+            color: white;
+            border: none;
+            border-radius: 4px;
+            cursor: pointer;
+            font-size: 0.85rem;
+        }
+        .copy-btn:hover { background-color: #5a6fd6; }
+        .copy-btn.copied { background-color: #198754; }
+    </style>
+    <div class="citation-block">
+        <div class="citation-title">📚 Citation</div>
+        <div class="citation-content" id="citation-text">
+            <button class="copy-btn" onclick="copyCitation()">Copy</button>
+@article{m3bench2025,
+    title     = {M3-Bench: Multi-Model Multi-View Game-Theoretic Benchmark for LLMs},
+    author    = {Authors},
+    journal   = {arXiv preprint},
+    year      = {2025},
+}</div>
+    </div>
+    <script>
+    function copyCitation() {
+        const citationText = `@article{m3bench2025,
+    title     = {M3-Bench: Multi-Model Multi-View Game-Theoretic Benchmark for LLMs},
+    author    = {Authors},
+    journal   = {arXiv preprint},
+    year      = {2025},
+}`;
+        navigator.clipboard.writeText(citationText).then(function() {
+            const btn = document.querySelector('.copy-btn');
+            btn.textContent = 'Copied!';
+            btn.classList.add('copied');
+            setTimeout(function() {
+                btn.textContent = 'Copy';
+                btn.classList.remove('copied');
+            }, 2000);
+        });
+    }
+    </script>
+    """)
+
+
+if __name__ == "__main__":
+    demo.launch()

data/leaderboard.csv

@@ -0,0 +1,37 @@
+model,category,view,PD,SH,UG,RPD,GE,AOB,PGG,VD,CPR,AC,WW,KP,L1,L2,L3,L4
+GPT-5.1,Closed-Source,BTA,97,94,95,96,82,93,93,88,94,91,80,90,93.0,89.0,90.3,86.2
+GPT-5.1,Closed-Source,RPA,95,92,93,93,84,90,90,86,91,88,84,87,91.7,88.0,88.5,85.8
+GPT-5.1,Closed-Source,CCA,90,88,96,88,91,93,88,84,85,80,90,76,89.3,86.8,87.2,83.7
+Claude Opus 4.5,Closed-Source,BTA,95,92,91,93,85,95,91,86,90,87,83,85,91.0,88.8,89.3,85.5
+Claude Opus 4.5,Closed-Source,RPA,93,90,94,91,82,93,88,84,88,85,87,83,89.7,86.5,87.5,84.7
+Claude Opus 4.5,Closed-Source,CCA,92,94,97,90,95,94,90,87,83,82,93,78,90.3,89.0,88.0,86.8
+Gemini-3 Pro,Closed-Source,BTA,93,95,89,90,86,88,88,90,91,86,78,84,89.7,86.8,87.3,83.2
+Gemini-3 Pro,Closed-Source,RPA,91,90,91,88,83,85,86,87,89,84,82,85,89.3,85.7,85.8,83.5
+Gemini-3 Pro,Closed-Source,CCA,87,85,93,84,88,90,84,85,82,78,86,74,86.8,83.5,84.3,81.0
+Grok-4.1,Closed-Source,BTA,90,88,83,87,81,93,85,83,89,82,72,86,88.7,86.3,84.5,79.0
+Grok-4.1,Closed-Source,RPA,87,85,86,84,79,90,83,81,86,79,76,83,86.7,84.2,82.8,78.7
+Grok-4.1,Closed-Source,CCA,84,82,90,80,85,92,81,80,80,74,82,71,86.2,82.7,81.8,77.0
+GPT-oss 120b,Open-Weight,BTA,89,91,86,88,90,84,84,87,85,79,75,78,87.0,85.3,84.0,77.5
+GPT-oss 120b,Open-Weight,RPA,86,88,89,85,86,82,82,84,83,77,79,76,86.3,83.2,82.3,77.5
+GPT-oss 120b,Open-Weight,CCA,83,86,91,82,89,86,80,83,79,75,84,70,85.0,81.8,81.8,77.7
+DeepSeek V3.2,Open-Weight,BTA,88,90,85,85,88,83,87,81,92,84,68,80,86.7,84.2,83.8,76.2
+DeepSeek V3.2,Open-Weight,RPA,85,87,88,83,85,81,84,79,89,81,73,77,85.5,82.5,82.2,76.2
+DeepSeek V3.2,Open-Weight,CCA,82,84,90,79,87,85,82,78,84,72,79,69,84.0,80.5,81.0,74.8
+Kimi-K2 Thinking,Open-Weight,BTA,92,86,90,84,79,86,86,80,83,73,82,68,87.0,81.2,82.0,74.3
+Kimi-K2 Thinking,Open-Weight,RPA,89,84,92,81,77,83,83,78,81,80,86,76,86.2,79.3,80.3,81.0
+Kimi-K2 Thinking,Open-Weight,CCA,85,81,94,77,83,87,79,76,77,60,72,53,84.5,78.5,79.2,63.2
+Mistral Large 3,Open-Weight,BTA,86,84,82,82,76,80,79,76,82,74,70,71,81.8,77.8,77.7,70.5
+Mistral Large 3,Open-Weight,RPA,83,81,85,79,74,78,77,74,79,71,72,69,81.0,76.2,76.2,69.8
+Mistral Large 3,Open-Weight,CCA,79,77,88,75,80,82,74,72,75,62,74,58,79.3,75.3,74.7,66.0
+GPT-5 pro,Reasoning,BTA,96,93,94,94,85,91,91,86,92,86,76,88,92.7,89.0,88.2,82.2
+GPT-5 pro,Reasoning,RPA,97,96,93,96,93,95,93,92,94,93,91,92,95.2,94.2,92.2,92.0
+GPT-5 pro,Reasoning,CCA,78,76,88,79,72,84,77,74,80,65,70,60,78.5,75.5,77.2,66.8
+Gemini-3 Deep Think,Reasoning,BTA,91,89,88,89,83,87,87,85,90,83,84,80,89.7,85.7,86.8,81.5
+Gemini-3 Deep Think,Reasoning,RPA,94,93,95,93,91,93,88,90,91,91,88,90,93.0,92.2,90.2,90.0
+Gemini-3 Deep Think,Reasoning,CCA,76,80,85,76,80,82,78,76,77,63,78,58,78.8,75.5,77.8,67.8
+DeepSeek-V3.2 Speciale,Reasoning,BTA,94,91,92,92,80,89,89,82,91,84,70,85,91.3,86.7,85.0,78.2
+DeepSeek-V3.2 Speciale,Reasoning,RPA,97,96,94,97,94,96,90,92,93,95,92,94,95.2,95.2,92.5,94.0
+DeepSeek-V3.2 Speciale,Reasoning,CCA,74,72,83,72,68,78,71,68,73,55,62,50,74.8,69.7,71.5,58.0
+Human,Human,BTA,91,88,93,86,78,84,83,80,79,78,87,72,86.3,82.2,82.8,80.7
+Human,Human,RPA,86,84,90,82,76,81,80,78,77,75,83,70,84.2,79.5,80.2,77.5
+Human,Human,CCA,93,90,94,89,92,91,90,85,76,84,92,68,89.0,85.0,85.3,84.7

requirements.txt

@@ -0,0 +1,2 @@
+gradio>=6.0.0
+pandas>=1.5

tabs/leaderboard_tab.py

@@ -0,0 +1,304 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+M3-Bench Leaderboard – Leaderboard Tab
+"""
+
+from __future__ import annotations
+
+import gradio as gr
+import pandas as pd
+from pathlib import Path
+
+BASE_DIR = Path(__file__).resolve().parent.parent
+DATA_PATH = BASE_DIR / "data" / "leaderboard.csv"
+
+# Category colors for row shading
+CATEGORY_COLORS = {
+    "Closed-Source": "#e8f0fe",    # cool blue
+    "Open-Weight": "#e8f5e9",      # sage green
+    "Reasoning": "#fce4ec",        # warm peach
+    "Human": "#e0f7fa",            # light cyan
+}
+
+CATEGORY_EMOJI = {
+    "Closed-Source": "🔒",
+    "Open-Weight": "🔓",
+    "Reasoning": "🧠",
+    "Human": "👤",
+}
+
+# View full names
+VIEW_NAMES = {
+    "BTA": "Birds-eye, Technically-grounded, Artistically-aware",
+    "RPA": "Role-Play Assessment",
+    "CCA": "Compositional Capability Assessment",
+}
+
+# Game abbreviation full names
+GAME_FULL_NAMES = {
+    "PD": "Prisoner's Dilemma",
+    "SH": "Stag Hunt",
+    "UG": "Ultimatum Game",
+    "RPD": "Repeated PD",
+    "GE": "Gift Exchange",
+    "AOB": "All-or-Nothing Bargaining",
+    "PGG": "Public Goods Game",
+    "VD": "Volunteer's Dilemma",
+    "CPR": "Common Pool Resource",
+    "AC": "Auction Competition",
+    "WW": "Wage War",
+    "KP": "Keynesian Beauty Contest",
+}
+
+
+def load_leaderboard() -> pd.DataFrame:
+    if not DATA_PATH.exists():
+        raise FileNotFoundError(f"Leaderboard file not found: {DATA_PATH}")
+    df = pd.read_csv(DATA_PATH)
+    df.columns = [c.strip() for c in df.columns]
+    return df
+
+
+def compute_overall(df: pd.DataFrame) -> pd.DataFrame:
+    """Compute overall average from the 4 level averages."""
+    df["Overall"] = df[["L1", "L2", "L3", "L4"]].mean(axis=1).round(2)
+    return df
+
+
+def aggregate_by_model(df: pd.DataFrame) -> pd.DataFrame:
+    """Average across all 3 views per model to get a single row per model."""
+    numeric_cols = ["PD", "SH", "UG", "RPD", "GE", "AOB", "PGG", "VD", "CPR",
+                    "AC", "WW", "KP", "L1", "L2", "L3", "L4"]
+    agg = df.groupby(["model", "category"])[numeric_cols].mean().reset_index()
+    agg["Overall"] = agg[["L1", "L2", "L3", "L4"]].mean(axis=1).round(2)
+    return agg
+
+
+def make_ranked_html(df: pd.DataFrame, sort_col: str = "Overall") -> str:
+    """Build an HTML table from the DataFrame."""
+    df = df.sort_values(by=sort_col, ascending=False).reset_index(drop=True)
+
+    # Assign ranks with ties
+    df["Rank"] = df[sort_col].rank(method="min", ascending=False).astype(int)
+
+    medal = {1: "🥇", 2: "🥈", 3: "🥉"}
+
+    # Start building HTML
+    html = """
+    <style>
+        .lb-table {
+            width: 100%;
+            border-collapse: collapse;
+            font-family: -apple-system, BlinkMacSystemFont, 'Segoe UI', Roboto, sans-serif;
+            font-size: 14px;
+        }
+        .lb-table th {
+            background: #1a1a2e;
+            color: #fff;
+            padding: 10px 8px;
+            text-align: center;
+            font-weight: 600;
+            position: sticky;
+            top: 0;
+            z-index: 10;
+        }
+        .lb-table th.group-header {
+            border-bottom: 2px solid #16213e;
+        }
+        .lb-table td {
+            padding: 8px 8px;
+            text-align: center;
+            border-bottom: 1px solid #eee;
+        }
+        .lb-table td.model-name {
+            text-align: left;
+            font-weight: 600;
+            padding-left: 12px;
+        }
+        .lb-table td.category-cell {
+            font-size: 12px;
+        }
+        .lb-table tr:hover {
+            background-color: #f0f0ff !important;
+        }
+        .lb-table .rank-cell {
+            font-weight: 700;
+            font-size: 15px;
+        }
+        .score-best {
+            font-weight: 700;
+            color: #d32f2f;
+        }
+        .score-second {
+            font-weight: 600;
+            text-decoration: underline;
+        }
+    </style>
+    """
+
+    # Determine best and second-best per column
+    score_cols = [c for c in df.columns if c not in ["Rank", "model", "category"]]
+    best_vals = {}
+    second_vals = {}
+    for col in score_cols:
+        vals = df[col].dropna().unique()
+        sorted_vals = sorted(vals, reverse=True)
+        best_vals[col] = sorted_vals[0] if len(sorted_vals) > 0 else None
+        second_vals[col] = sorted_vals[1] if len(sorted_vals) > 1 else None
+
+    html += '<table class="lb-table">'
+    # Header row
+    html += "<thead><tr>"
+    html += '<th style="width:50px">Rank</th>'
+    html += '<th style="text-align:left;width:180px">Model</th>'
+    html += '<th style="width:80px">Category</th>'
+    html += '<th style="width:70px;background:#16213e">Overall</th>'
+
+    level_groups = [
+        ("Level 1", ["PD", "SH", "UG"]),
+        ("Level 2", ["RPD", "GE", "AOB"]),
+        ("Level 3", ["PGG", "VD", "CPR"]),
+        ("Level 4", ["AC", "WW", "KP"]),
+        ("Level Avg.", ["L1", "L2", "L3", "L4"]),
+    ]
+
+    for _, cols in level_groups:
+        for c in cols:
+            html += f'<th style="width:55px">{c}</th>'
+    html += "</tr>"
+
+    # Second header row with group labels
+    html += "<tr>"
+    html += '<th colspan="4"></th>'
+    colors = ["#2a2a4a", "#2a3a2a", "#3a2a2a", "#2a2a3a", "#1a1a2e"]
+    for i, (name, cols) in enumerate(level_groups):
+        html += f'<th colspan="{len(cols)}" class="group-header" style="background:{colors[i]};font-size:12px">{name}</th>'
+    html += "</tr></thead>"
+
+    html += "<tbody>"
+    for _, row in df.iterrows():
+        cat = row["category"]
+        bg = CATEGORY_COLORS.get(cat, "#fff")
+        emoji = CATEGORY_EMOJI.get(cat, "")
+        rank = row["Rank"]
+        rank_display = f'{medal.get(rank, "")} {rank}' if rank in medal else str(rank)
+
+        html += f'<tr style="background-color:{bg}">'
+        html += f'<td class="rank-cell">{rank_display}</td>'
+        html += f'<td class="model-name">{row["model"]}</td>'
+        html += f'<td class="category-cell">{emoji} {cat}</td>'
+
+        # Overall with special styling
+        ov = row["Overall"]
+        ov_class = ""
+        if ov == best_vals.get("Overall"):
+            ov_class = "score-best"
+        elif ov == second_vals.get("Overall"):
+            ov_class = "score-second"
+        html += f'<td class="{ov_class}" style="font-weight:700;background:rgba(0,0,0,0.03)">{ov}</td>'
+
+        for _, cols in level_groups:
+            for c in cols:
+                val = row[c]
+                cls = ""
+                if val == best_vals.get(c):
+                    cls = "score-best"
+                elif val == second_vals.get(c):
+                    cls = "score-second"
+                display = f"{val:.1f}" if isinstance(val, float) and val != int(val) else str(int(val)) if pd.notna(val) else "-"
+                html += f'<td class="{cls}">{display}</td>'
+
+        html += "</tr>"
+
+    html += "</tbody></table>"
+    return html
+
+
+def make_view_html(df: pd.DataFrame, view: str, sort_col: str = "Overall") -> str:
+    """Filter to one view and produce an HTML table."""
+    filtered = df[df["view"] == view].copy()
+    filtered = compute_overall(filtered)
+    filtered = filtered.drop(columns=["view"])
+    return make_ranked_html(filtered, sort_col)
+
+
+def make_aggregated_html(df: pd.DataFrame, sort_col: str = "Overall") -> str:
+    """Aggregate across all views and produce an HTML table."""
+    agg = aggregate_by_model(df)
+    return make_ranked_html(agg, sort_col)
+
+
+def create_leaderboard_tab():
+    with gr.Tab("🏆 Leaderboard"):
+
+        gr.Markdown("""
+        Select a view to see per-view scores, or choose **All Views (Averaged)** for the overall ranking.
+        Scores are multiplied by 100. **Bold red** = best in column, <u>underlined</u> = second best.
+        """)
+
+        with gr.Row():
+            view_selector = gr.Radio(
+                choices=["All Views (Averaged)", "BTA", "RPA", "CCA"],
+                value="All Views (Averaged)",
+                label="View",
+                interactive=True,
+            )
+            sort_selector = gr.Dropdown(
+                choices=["Overall", "L1", "L2", "L3", "L4",
+                         "PD", "SH", "UG", "RPD", "GE", "AOB",
+                         "PGG", "VD", "CPR", "AC", "WW", "KP"],
+                value="Overall",
+                label="Sort by",
+                interactive=True,
+            )
+
+        with gr.Row():
+            category_filter = gr.CheckboxGroup(
+                choices=["Closed-Source", "Open-Weight", "Reasoning", "Human"],
+                value=["Closed-Source", "Open-Weight", "Reasoning", "Human"],
+                label="Model Categories",
+                interactive=True,
+            )
+
+        table_html = gr.HTML()
+
+        def update_table(view_choice, sort_col, categories):
+            df = load_leaderboard()
+            if categories:
+                df = df[df["category"].isin(categories)]
+            if df.empty:
+                return "<p>No data to display.</p>"
+            if view_choice == "All Views (Averaged)":
+                return make_aggregated_html(df, sort_col)
+            else:
+                return make_view_html(df, view_choice, sort_col)
+
+        # Set initial value
+        initial_html = update_table("All Views (Averaged)", "Overall",
+                                     ["Closed-Source", "Open-Weight", "Reasoning", "Human"])
+        table_html.value = initial_html
+
+        # Bind events
+        for component in [view_selector, sort_selector, category_filter]:
+            component.change(
+                fn=update_table,
+                inputs=[view_selector, sort_selector, category_filter],
+                outputs=table_html,
+            )
+
+        # Column descriptions
+        with gr.Accordion("📖 Column Descriptions & Game Abbreviations", open=False):
+            game_desc = "\n".join([f"- **{k}**: {v}" for k, v in GAME_FULL_NAMES.items()])
+            view_desc = "\n".join([f"- **{k}**: {v}" for k, v in VIEW_NAMES.items()])
+            gr.Markdown(f"""
+**Views:**
+{view_desc}
+
+**Game Abbreviations (Level 1–4):**
+{game_desc}
+
+**Level Averages (L1–L4):** Computed over all tasks in each level (not just the 3 shown).
+
+**Overall:** Average of L1, L2, L3, L4.
+""")