File size: 1,668 Bytes
ec4ae03
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
@startuml reward_system
!theme plain
top to bottom direction
skinparam backgroundColor #FEFEFE
skinparam defaultFontName Arial
skinparam defaultFontSize 14
skinparam ArrowColor #334155
skinparam RectangleBorderColor #64748B
skinparam RectangleFontColor #0F172A
skinparam roundcorner 10
skinparam linetype ortho
skinparam packageStyle rectangle
skinparam nodesep 54
skinparam ranksep 60

title AxiomForgeAI - Reward System

rectangle "Sampled Solution Attempt" as ATTEMPT #DBEAFE

rectangle "Grounded Reward\nknown-answer problem" as GROUNDED #ECFDF5 {
    rectangle "Final answer\nmatches gold" as GOLD #CCFBF1
    rectangle "PRM process score\nreasoning quality" as GPRM #CCFBF1
    rectangle "Chain consistency\ncorrect prefix + final check" as GCHAIN #CCFBF1
    rectangle "Format score\nparseable final answer" as GFORMAT #CCFBF1
}

rectangle "Self-Play Reward\ngenerated challenge" as SELFPLAY #EEF2FF {
    rectangle "Question quality\nclarity, novelty, solvability" as QUALITY #E0E7FF
    rectangle "Solution quality\nPRM + chain checks" as SOLUTION #E0E7FF
    rectangle "Format score\nparseable final answer" as SFORMAT #E0E7FF
}

rectangle "Combined Reward\none score per attempt" as SCORE #F1F5F9
rectangle "GRPO Group Comparison\nrank attempts within the same problem" as COMPARE #EDE9FE
rectangle "Step-Answer Alignment\nreward paths where reasoning supports the result" as ALIGN #DDD6FE

ATTEMPT -left-> GROUNDED : grounded
ATTEMPT -right-> SELFPLAY : self-play

GOLD --> GPRM
GPRM --> GCHAIN
GCHAIN --> GFORMAT

QUALITY --> SOLUTION
SOLUTION --> SFORMAT

GFORMAT -down-> SCORE
SFORMAT -down-> SCORE
SCORE -right-> COMPARE
COMPARE -right-> ALIGN
@enduml