@fengrru/process-reward

Experimental — API may break in minor versions. See STABILITY.md.

Process Reward Model with MC rollout labeling for AI agent reasoning.

Install

npm install @fengrru/process-reward

Quick Start

import { ProcessRewardModel } from "@fengrru/process-reward"

const prm = new ProcessRewardModel()

// Score a reasoning step
const score = prm.scoreStep({
  type: "math",
  content: "x = 5 + 3",
  previousSteps: ["Given: x + 2 = 10"],
})

console.log(score.score) // 0-1
console.log(score.details) // scoring breakdown

// Batch scoring
const scores = prm.batchScoreSteps(steps)

// Monte Carlo labeling
const labels = prm.labelSteps(steps, { outcome: "correct" })

Scoring Domains

| Domain | Heuristics | |--------|------------| | math | Equation validity, divide-by-zero, coherence | | code | Function definition, syntax errors | | logic | Premise introduction, conclusion markers |

Features

• MC rollout: confidence from multiple simulations
• Weak supervision: heuristic labels when no MC
• Cross-step coherence: boosts consistent reasoning
• Custom scorers: register domain-specific evaluators

Documentation

• API Reference — TypeDoc-generated API docs
• Source Code
• Examples

License

MIT