process-reward-model

[NeurIPS 2025 Spotlight] LLM post-training suite — featuring ReasonFlux, ReasonFlux-PRM, and ReasonFlux-Coder.

Official codebase for "Can 1B LLM Surpass 405B LLM? Rethinking Compute-Optimal Test-Time Scaling".

Fixing GRPO training collapse in long-horizon multi-tool agents. A lightweight PRM-Lite + LATA joint approach achieves +37% over vanilla GRPO on τ-bench airline (50-task, multi-turn).

A comprehensive collection of process reward models.

Curated papers, taxonomy, benchmarks, and decision guides for credit assignment in reasoning and agentic LLM reinforcement learning.

[AAAI 2026] Official codebase for "GenPRM: Scaling Test-Time Compute of Process Reward Models via Generative Reasoning".

[ICML 2025] Reward-guided Speculative Decoding (RSD) for efficiency and effectiveness.

Library for training process reward models

[ACL 2026 SAC Highlight Award] Can We Predict Before Executing Machine Learning Agents?

This repository includes code and materials for the paper "Efficient PRM Training Data Synthesis via Formal Verification" (ACL 2026 Findings).

[KDD 2025] Rewarding Graph Reasoning Process makes LLMs more Generalized Reasoners

This repository contains the official implementation of Error Typing for Smarter Rewards: Improving Process Reward Models with Error-Aware Hierarchical Supervision.

Multi-Agent System Process Reward Model (MASPRM): a lightweight process reward model guiding multi-agent systems at search time.

Step-level preference data construction toolkit for code-agent process reward models

Process-supervised RL for a multi-step reasoning agent — DAPO + a learned Process Reward Model (PRM) training a Qwen3-8B Planner. A modern, from-scratch rebuild of the AgentFlow paper (ICLR 2026).

Live proof of arXiv:2603.17815 — O(N) confirmed R²=0.952, 1,984 API calls

Interface for DeepSeek-R1 chain-of-thought model — exposes the internal reasoning trace, supports API and local Ollama backends, with streaming output and temperature control.

Experiments with test-time compute scaling — process reward models, MCTS, and beam search.

DPO-trained decision layer for coding agents — raises pass@1 from 0.419 → 0.639 on frozen 14B generator (p<0.000001)