sxiong/MLR_planner_Qwen-1.5B-LoRA

MLR Planner LoRA for Qwen2.5-1.5B

This repository contains the high-level planner for Multi-Level Reasoning (MLR) in the paper [Enhancing Language Model Reasoning with Structured Multi-Level Modeling](https://proceedings.iclr.cc/paper_files/paper/2026/file/3db7d123a316fc690f02818b21967af4-Paper-Conference.pdf) (ICLR 26).

MLR decomposes long-horizon reasoning into an alternating plan--execute loop: the planner proposes a structured, abstract subgoal and the executor produces the detailed reasoning conditioned on it.

Base model and compatibility

• —Base architecture: Qwen2.5-1.5B
• —Required base checkpoint: sxiong/MLR_executor_Qwen-1.5B
• —LoRA rank: 16; alpha: 32
• —Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, and down_proj

Loading the adapter

Install compatible versions of PyTorch, Transformers, and PEFT, then load the adapter onto the executor checkpoint explicitly:

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_id = "sxiong/MLR_executor_Qwen-1.5B"
adapter_id = "sxiong/MLR_planner_Qwen-1.5B-LoRA"

tokenizer = AutoTokenizer.from_pretrained(base_id)
base_model = AutoModelForCausalLM.from_pretrained(
    base_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
planner = PeftModel.from_pretrained(base_model, adapter_id)
planner.eval()

For end-to-end MLR inference, use the accompanying MLR inference code. It applies the planner and executor with the required prompts, parser, stopping rule, and alternating plan--execute control flow. Direct free-form generation from this adapter is not the intended interface.

Citation

@inproceedings{xiong2026enhancing,
  title={Enhancing language model reasoning with structured multi-level modeling},
  author={Xiong, Siheng and Payani, Ali and Fekri, Faramarz},
  booktitle={International Conference on Learning Representations},
  volume={2026},
  pages={36557--36610},
  year={2026}
}