dLLM PRM Gap · Bidirectional PRM

News

2026-09: Accepted at NeurIPS 2026.

Public adapter release for dLLM PRM Gap. bidirectional attention with mask-aware mean pooling over intermediate denoising states

This repository contains only trainable adapter parameters and the reward head. It does not include the base model. The release configuration is recorded in config.json; the arXiv paper defines the release scope and citation.

Files

adapter.safetensors — compact adapter weights
config.json — public base-model id, architecture, and provenance

Load

from huggingface_hub import snapshot_download
from prm.checkpointing import load_diffusion_prm

path = snapshot_download("YanZhanPKU/dLLM-PRM-Gap-bidir-dream7b")
model, report = load_diffusion_prm(
    checkpoint=path,
    model_path="Dream-org/Dream-v0-Instruct-7B",
    local_files_only=False,
)

Role: process-reward model used by PRM Guided, Hybrid, and the snapshot diagnostics.

Paper: https://arxiv.org/abs/2609.35472.