File size: 8,555 Bytes
0772b5a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
import os
import sys
import time
import json
import re
import traceback
from typing import Dict, Any, List
from openai import OpenAI
import sympy as sp

API_KEY = os.getenv("GOOGLE_API_KEY", "")
BASE_URL = os.getenv("LLM_BASE_URL", "https://generativelanguage.googleapis.com/v1beta/openai/")
MODEL = os.getenv("LLM_MODEL", "gemma-4-31b-it")

client = OpenAI(
    api_key=API_KEY,
    base_url=BASE_URL,
    timeout=60.0,
)

TEST_PROBLEMS = [
    {
        "id": "case_1_easy",
        "name": "Hình chóp tứ giác đều (Easy)",
        "question": "Cho hình chóp S.ABCD có đáy ABCD là hình vuông cạnh 10. Chiều cao SO vuông góc với đáy tại tâm O, SO=15. Tính thể tích khối chóp S.ABCD.",
        "expected": "500",
    },
    {
        "id": "case_2_medium",
        "name": "Hình chóp tam giác đều (Medium)",
        "question": "Cho hình chóp tam giác đều S.ABC có cạnh đáy bằng 6, chiều cao SO = 8 vuông góc với đáy tại trọng tâm O của tam giác ABC. Tính thể tích khối chóp S.ABC.",
        "expected": "24*sqrt(3) ≈ 41.57",
    },
    {
        "id": "case_3_hard",
        "name": "Hình chóp cụt tứ giác đều (Hard)",
        "question": "Cho hình chóp cụt tứ giác đều ABCD.A1B1C1D1 có cạnh đáy dưới bằng 8, cạnh đáy trên bằng 4, chiều cao giữa hai đáy h=6. Tính thể tích khối chóp cụt.",
        "expected": "224",
    }
]

# ==============================================================================
# 1. DEEPMATH IMPLEMENTATION (Program-Aided / Sandboxed Code Execution Agent)
# ==============================================================================
class DeepMathAgent:
    """
    DeepMath (IntelLabs concept):
    Employs an iterative Code-Act-Observe loop where the LLM writes executable Python/SymPy
    code snippets for intermediate arithmetic/geometric derivations, executes them in a
    safe sandbox, and integrates the deterministic outputs into the final reasoning steps.
    """
    def __init__(self, client: OpenAI, model: str):
        self.client = client
        self.model = model

    def solve(self, question: str) -> Dict[str, Any]:
        start_time = time.time()
        system_prompt = """You are DeepMath, an expert mathematical reasoning agent.
When solving geometry and math problems:
1. Explain the geometric method step-by-step in Vietnamese.
2. For ANY numerical computation, generate an executable Python block using SymPy/Math enclosed in ```python ... ```.
3. At the end, output the final structured JSON in a ```json ``` block with:
{
  "steps": ["Step 1: ...", "Step 2: ..."],
  "python_code": "... combined python code ...",
  "evaluated_variables": {"var_name": "value"},
  "answer": "final numerical or exact symbolic answer"
}
"""
        messages = [
            {"role": "system", "content": system_prompt},
            {"role": "user", "content": f"Hãy giải bài toán hình học sau:\n{question}"}
        ]

        response = self.client.chat.completions.create(
            model=self.model,
            messages=messages,
            temperature=0.1,
        )
        content = response.choices[0].message.content or ""

        # Extract and execute Python code snippets in a safe SymPy environment
        code_blocks = re.findall(r"```python(.*?)```", content, re.DOTALL)
        exec_globals = {"sp": sp, "math": __import__("math"), "sqrt": sp.sqrt}
        exec_locals = {}
        for block in code_blocks:
            try:
                exec(block, exec_globals, exec_locals)
            except Exception as e:
                exec_locals["_error"] = str(e)

        # Extract JSON
        json_match = re.search(r"```json(.*?)```", content, re.DOTALL)
        if json_match:
            try:
                parsed_json = json.loads(json_match.group(1).strip())
            except Exception:
                parsed_json = {"raw": content}
        else:
            parsed_json = {"raw": content}

        elapsed = time.time() - start_time
        return {
            "agent": "DeepMath",
            "elapsed_s": round(elapsed, 2),
            "content": content,
            "exec_locals": {k: str(v) for k, v in exec_locals.items() if not k.startswith("_")},
            "parsed_json": parsed_json,
        }

# ==============================================================================
# 2. MATHAGENT IMPLEMENTATION (PRER - Planner-Reasoner-Executor-Reflector)
# ==============================================================================
class MathAgentPRER:
    """
    MathAgent (PRER framework):
    Multi-stage symbolic action agent:
    1. Preprocess: splits into Conditions & Sub-questions.
    2. Select & Act: selects reasoning actions (Calculate, Transform, Deduce).
    3. Check & Reflector: validates step correctness.
    4. Summary: synthesizes final proof and answer.
    """
    def __init__(self, client: OpenAI, model: str):
        self.client = client
        self.model = model

    def solve(self, question: str) -> Dict[str, Any]:
        start_time = time.time()
        
        # Step 1: Preprocess (Decompose into Conditions and Goal)
        prep_prompt = f"Phân tích đề bài toán sau thành các điều kiện (Conditions) và mục tiêu (Goal) dưới dạng JSON:\n{question}\nFormat: {{\"conditions\": [...], \"goal\": \"...\"}}"
        prep_res = self.client.chat.completions.create(
            model=self.model,
            messages=[{"role": "user", "content": prep_prompt}],
            temperature=0.1,
        )
        prep_content = prep_res.choices[0].message.content or ""

        # Step 2: Reasoner & Executor (Calculate + Deduce)
        reason_prompt = f"""Dựa trên bài toán: {question}
Thực hiện các bước giải toán hình học chi tiết, tính toán các công thức diện tích và thể tích chính xác.
Trả về JSON gồm:
{{
  "steps": ["Bước 1: ...", "Bước 2: ..."],
  "formulas": ["S_day = ...", "V = ..."],
  "answer": "kết quả cuối cùng"
}}"""
        reason_res = self.client.chat.completions.create(
            model=self.model,
            messages=[{"role": "user", "content": reason_prompt}],
            temperature=0.1,
        )
        reason_content = reason_res.choices[0].message.content or ""

        elapsed = time.time() - start_time
        return {
            "agent": "MathAgent (PRER)",
            "elapsed_s": round(elapsed, 2),
            "prep_content": prep_content,
            "reason_content": reason_content,
        }

def main():
    print("======================================================================", flush=True)
    print("    BENCHMARK & CAPABILITY COMPARISON: DeepMath vs MathAgent", flush=True)
    print(f"    Model: {MODEL} | Provider: Google Generative Language", flush=True)
    print("======================================================================", flush=True)

    deepmath = DeepMathAgent(client, MODEL)
    mathagent = MathAgentPRER(client, MODEL)

    for prob in TEST_PROBLEMS:
        print(f"\n" + "="*70, flush=True)
        print(f"🔥 PROBLEM: {prob['name']}", flush=True)
        print(f"Question: {prob['question']}", flush=True)
        print(f"Expected: {prob['expected']}", flush=True)
        print("="*70, flush=True)

        # 1. Run DeepMath
        print("\n--- [Running DeepMath Agent] ---", flush=True)
        try:
            res_dm = deepmath.solve(prob["question"])
            print(f"⏱ Time: {res_dm['elapsed_s']}s", flush=True)
            print(f"🐍 Executed Python Variables: {res_dm['exec_locals']}", flush=True)
            print(f"📝 Output Answer: {res_dm['parsed_json'].get('answer', 'N/A')}", flush=True)
            print(f"📋 Steps ({len(res_dm['parsed_json'].get('steps', []))}):", flush=True)
            for s in res_dm['parsed_json'].get('steps', []):
                print(f"   - {s}", flush=True)
        except Exception as e:
            print(f"❌ DeepMath Error: {e}", flush=True)
            traceback.print_exc()

        # 2. Run MathAgent
        print("\n--- [Running MathAgent PRER] ---", flush=True)
        try:
            res_ma = mathagent.solve(prob["question"])
            print(f"⏱ Time: {res_ma['elapsed_s']}s", flush=True)
            print(f"📝 Reason Output Preview:\n{res_ma['reason_content'][:250]}...", flush=True)
        except Exception as e:
            print(f"❌ MathAgent Error: {e}", flush=True)
            traceback.print_exc()

if __name__ == "__main__":
    main()