Module prm 

Process Reward Model (PRM) for Step-by-Step Verification

Implements step-level verification based on Math-Shepherd research achieving +6.2% GSM8K improvement through granular reasoning validation.

Scientific Foundation

Based on:

• Math-Shepherd (Wang et al., 2024): Process reward models for math reasoning
• Let’s Verify Step by Step (Lightman et al., 2023): Step-level human verification

Key Concepts

• Outcome Reward Model (ORM): Scores only final answer correctness
• Process Reward Model (PRM): Scores each reasoning step independently

PRM advantages:

1. Better credit assignment - identifies WHERE reasoning went wrong
2. More training signal - learns from partial success
3. Improved calibration - confidence per step

Usage

use reasonkit::thinktool::prm::{ProcessRewardModel, StepScore};

let prm = ProcessRewardModel::new();
let steps = vec!["Step 1: Given x + 2 = 5", "Step 2: x = 5 - 2 = 3"];
let scores = prm.score_steps(&steps).await?;

Structs

PrmConfig

Process Reward Model configuration

PrmMetrics

PrmReranker

Best-of-N with PRM reranking

PrmResult

Result of PRM evaluation

StepIssue

Issue identified in a reasoning step

StepParser

Step parser to extract reasoning steps from LLM output

StepScore

Individual step score from PRM

VerificationPrompts

Verification prompt templates for different step types

Enums

IssueType

ScoreAggregation

Severity

VerificationStrategy