File size: 10,719 Bytes
f8319a8
 
 
 
973cd6f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f8319a8
 
 
 
 
973cd6f
 
 
 
 
 
f8319a8
 
 
 
 
 
973cd6f
 
 
f8319a8
973cd6f
 
f8319a8
973cd6f
 
 
 
 
 
 
 
 
f8319a8
 
 
 
 
973cd6f
 
 
 
 
 
f8319a8
 
973cd6f
 
 
 
 
 
f8319a8
973cd6f
 
 
f8319a8
973cd6f
f8319a8
 
 
 
 
973cd6f
f8319a8
973cd6f
 
f8319a8
 
973cd6f
 
 
 
 
 
 
 
 
 
 
 
f8319a8
973cd6f
 
 
f8319a8
973cd6f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f8319a8
 
 
 
 
 
 
 
 
973cd6f
 
 
 
 
 
 
f8319a8
 
 
 
 
 
 
 
 
 
973cd6f
 
f8319a8
973cd6f
 
 
 
 
 
f8319a8
 
973cd6f
 
 
 
 
 
 
 
 
 
 
f8319a8
 
 
 
973cd6f
f8319a8
973cd6f
 
 
 
 
 
 
 
 
 
 
 
 
f8319a8
973cd6f
 
 
 
 
 
 
 
 
 
 
f8319a8
973cd6f
 
f8319a8
 
 
 
973cd6f
f8319a8
 
 
 
 
973cd6f
 
 
 
 
 
 
 
 
 
f8319a8
 
 
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
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
import math
from typing import Dict, Any, List, Tuple

class RewardSystem:
    """
    Dense, multi-component reward system for mathematical RL training.
    
    Key improvements over v1:
    1. All 7 reward components now contribute to the final score
    2. Partial credit support (continuous C ∈ [0,1] from verifier)
    3. Fixed history key mismatch (was breaking diversity detection)
    4. Adaptive efficiency curve that doesn't over-penalize reasonable lengths
    5. Removed random noise (adds variance without useful signal)
    6. Added format compliance reward for structured output
    
    Reward equation:
        R = α·C + β·Q + γ·P + δ·R_ref + η·D_norm + ζ·E_norm + λ·X + μ·F_fmt
    
    Weights: α=0.30, β=0.12, γ=0.10, δ=0.05, η=0.13, ζ=0.08, λ=0.07, μ=0.15
    Sum = 1.0
    
    References:
        - arxiv:2408.10215 (Reward shaping for RL convergence)
        - arxiv:2601.19100 (Reward engineering for software/code tasks)
        - DeepSeek-R1 GRPO (graduated correctness)
        - GRPO-λ (credit assignment)
    """
    
    # Reward component weights (sum to 1.0)
    W_CORRECTNESS = 0.30       # α: Primary — correctness drives learning
    W_REASONING = 0.12         # β: Reasoning quality
    W_PROCESS = 0.10           # γ: Step-by-step process supervision
    W_REFLECTION = 0.05        # δ: Self-correction behavior
    W_DIVERSITY = 0.13         # η: Answer diversity (prevents repetition)
    W_EFFICIENCY = 0.08        # ζ: Token efficiency
    W_EXPLORATION = 0.07       # λ: Exploration bonus
    W_FORMAT = 0.15            # μ: Format compliance (model must learn structure)
    
    def __init__(self, max_len: int = 1000):
        self.max_len = max_len

    def compute_diversity(self, current_answer: str, history: List[Dict[str, Any]]) -> float:
        """
        D = diversity (difference from past attempts).
        
        Graduated penalty instead of binary:
        - Exact repeat: -1.0 (steep penalty)
        - Similar to a past answer: -0.3
        - Unique: +1.0
        """
        if not history:
            return 1.0
            
        cur_ans_clean = current_answer.strip().lower()
        
        if not cur_ans_clean:
            return 0.0  # Empty answer gets no diversity credit
        
        for attempt in history:
            # BUGFIX: check both 'final_answer' and 'prediction' keys for compatibility
            prev_ans = attempt.get('final_answer', attempt.get('prediction', '')).strip().lower()
            if prev_ans == cur_ans_clean:
                return -1.0  # Exact repeat — strong penalty
            
            # Check for near-duplicates (edit distance heuristic)
            if prev_ans and cur_ans_clean:
                # Simple character overlap ratio
                overlap = sum(1 for a, b in zip(prev_ans, cur_ans_clean) if a == b)
                max_len = max(len(prev_ans), len(cur_ans_clean))
                if max_len > 0 and overlap / max_len > 0.85:
                    return -0.3  # Near-duplicate — moderate penalty
                
        return 1.0

    def compute_efficiency(self, action_string: str) -> float:
        """
        E = efficiency. Adaptive Gaussian penalty curve.
        
        Improved: wider optimal zone (30-120 tokens) to avoid penalizing
        legitimate mathematical reasoning that naturally needs more space.
        
        E ∈ [-0.5, 0.0] (always a penalty or neutral, never a bonus)
        """
        approx_tokens = len(action_string) / 4.0
        optimal_center = 80.0   # Wider center for math
        optimal_width = 60.0    # Generous width
        
        # Gentle Gaussian — penalizes only extreme lengths
        ratio = (approx_tokens - optimal_center) / optimal_width
        e = math.exp(-(ratio ** 2)) - 1.0
        
        # Additional penalty for very long outputs (anti-rambling)
        if approx_tokens > 300:
            e -= 0.3 * (approx_tokens - 300) / 300
        
        return max(-1.0, e)
        
    def compute_exploration_bonus(self, action_string: str, times_seen: int) -> float:
        """
        [PAPER TRACEABILITY: Exploration via Entropy Bonus]
        G. EXPLORATION VIA ENTROPY BONUS
        
        X = (entropy_bonus) / sqrt(1 + times_seen_problem)
        
        Improved with better entropy estimation using word-level diversity.
        """
        length = len(action_string)
        if length == 0:
            return 0.0
        
        # Character-level entropy
        unique_ratio = len(set(action_string)) / length
        char_entropy = math.log1p(unique_ratio)
        
        # Word-level diversity bonus (rewards varied vocabulary)
        words = action_string.lower().split()
        if words:
            unique_word_ratio = len(set(words)) / len(words)
            word_entropy = math.log1p(unique_word_ratio)
        else:
            word_entropy = 0.0
        
        combined = 0.6 * char_entropy + 0.4 * word_entropy
            
        return combined / math.sqrt(1.0 + times_seen)

    def compute_format_compliance(self, action_str: str, reasoning: str, final_answer: str) -> float:
        """
        Format compliance reward — teaches the model to output structured responses.
        
        Rewards:
        - Having both reasoning and answer sections
        - Using mathematical notation
        - Proper structure (reasoning before answer)
        
        F ∈ [0, 1]
        """
        score = 0.0
        
        # Has non-empty reasoning
        if reasoning and len(reasoning.strip()) > 10:
            score += 0.3
        
        # Has non-empty final answer
        if final_answer and len(final_answer.strip()) > 0:
            score += 0.3
        
        # Answer contains mathematical content
        math_indicators = ['x', '=', '+', '-', '*', '/', '^', 'sin', 'cos', 'exp', 'log', '(']
        math_count = sum(1 for m in math_indicators if m in final_answer.lower())
        if math_count >= 2:
            score += 0.2
        elif math_count >= 1:
            score += 0.1
        
        # Reasoning contains structured steps
        if any(marker in reasoning.lower() for marker in ['step', 'first', 'then', 'therefore', '=']):
            score += 0.2
        
        return min(1.0, score)

    def detect_trivial_output(self, action_string: str) -> bool:
        """Anti-reward hacking: detect trivial constant outputs"""
        # If the output is just a single character repeated or very low entropy
        if len(action_string) < 2:
            return True
        unique_chars = len(set(action_string))
        if unique_chars < 3 and len(action_string) > 10:
            return True
        # Detect repetitive patterns
        if len(action_string) > 20:
            # Check if a short pattern is repeated
            for plen in range(1, 6):
                pattern = action_string[:plen]
                if action_string == pattern * (len(action_string) // plen) + pattern[:len(action_string) % plen]:
                    return True
        return False

    def compute_reward(self, 
                      correctness: float, 
                      reasoning_quality: float,
                      process_supervision: float,
                      reflection_score: float,
                      action_str: str, 
                      final_answer: str,
                      history: List[Dict[str, Any]],
                      times_seen_problem: int,
                      reasoning: str = "") -> Tuple[float, Dict[str, float]]:
        """
        Dense composite reward using ALL 7 components + format compliance.
        
        R = α·C + β·Q_norm + γ·P_norm + δ·R_norm + η·D_norm + ζ·E_norm + λ·X + μ·F_fmt
        
        All components are normalized to [0, 1] before weighting.
        Final reward ∈ [0, 1].
        """
        if self.detect_trivial_output(action_str):
            components = {
                "total_reward": -0.5,
                "C_correctness": 0.0, "Q_reasoning": 0.0,
                "P_process_supervision": 0.0, "R_reflection": 0.0,
                "D_diversity": 0.0, "E_efficiency": -1.0,
                "X_exploration": 0.0, "F_format": 0.0,
            }
            return -0.5, components
        
        # --- Raw component computation ---
        c = correctness  # Already ∈ [0, 1] with graduated scoring
        q = reasoning_quality
        d = self.compute_diversity(final_answer, history)
        e = self.compute_efficiency(action_str)
        x = self.compute_exploration_bonus(action_str, times_seen_problem)
        f_fmt = self.compute_format_compliance(action_str, reasoning, final_answer)
        
        # If repeated answer, reduce correctness credit (anti-hacking)
        if d < -0.5:
            c = c * 0.3  # Steep discount but not full zeroing
            
        # --- Normalize all components to [0, 1] ---
        q_norm = min(1.0, max(0.0, math.tanh(q)))
        p_norm = (process_supervision + 1.0) / 2.0   # [-1, 1] → [0, 1]
        r_norm = (reflection_score + 1.0) / 2.0       # [-1, 1] → [0, 1]
        d_norm = (d + 1.0) / 2.0                      # [-1, 1] → [0, 1]
        e_norm = (e + 1.0) / 1.0                      # [-1, 0] → [0, 1]
        e_norm = min(1.0, max(0.0, e_norm))
        x_norm = min(1.0, max(0.0, x))
        f_norm = min(1.0, max(0.0, f_fmt))
        
        # --- Weighted composite ---
        total_r = (
            self.W_CORRECTNESS * c +
            self.W_REASONING * q_norm +
            self.W_PROCESS * p_norm +
            self.W_REFLECTION * r_norm +
            self.W_DIVERSITY * d_norm +
            self.W_EFFICIENCY * e_norm +
            self.W_EXPLORATION * x_norm +
            self.W_FORMAT * f_norm
        )
        
        # Clamp to [0, 1]
        total_r = min(1.0, max(0.0, total_r))
        
        components = {
            "total_reward": total_r,
            "C_correctness": c,
            "Q_reasoning": q_norm,
            "P_process_supervision": process_supervision,
            "R_reflection": reflection_score,
            "D_diversity": d,
            "E_efficiency": e,
            "X_exploration": x,
            "F_format": f_fmt,
            # Weighted contributions (for debugging)
            "_w_C": self.W_CORRECTNESS * c,
            "_w_Q": self.W_REASONING * q_norm,
            "_w_P": self.W_PROCESS * p_norm,
            "_w_R": self.W_REFLECTION * r_norm,
            "_w_D": self.W_DIVERSITY * d_norm,
            "_w_E": self.W_EFFICIENCY * e_norm,
            "_w_X": self.W_EXPLORATION * x_norm,
            "_w_F": self.W_FORMAT * f_norm,
        }
        
        return total_r, components