Introduction

Task Description

This model was fine-tuned to improve its ability to perform spatial reasoning tasks. The objective is to enable the model to interpret natural language queries related to spatial relationships, directions, and locations and output actionable responses. The task addresses limitations in current LLMs, which often fail to perform precise spatial reasoning, such as determining relationships between points on a map, planning routes, or identifying locations based on bounding boxes.

Task Importance

Spatial reasoning is important for a wide range of applications such as navigation and geospatial analysis. Many smaller LLMs, while strong in general reasoning, often lack the ability to interpret spatial relationships with precision or utilize real-world geographic data effectively. For example, they struggle to answer queries like “What’s between Point A and Point B?” or “Find me the fastest route avoiding traffic at 8 AM tomorrow.” Even when the LLM has access to geospatial information, smaller models struggle to correctly interpret user questions.

Related Work/Gap Analysis

While there is ongoing research in integrating LLMs with geospatial systems, most existing solutions rely on symbolic AI or rule-based systems rather than leveraging the generalization capabilities of LLMs. Additionally, the paper “Advancing Spatial Reasoning in Large Language Models: An In-Depth Evaluation and Enhancement Using the StepGame Benchmark,” concluded that larger models like GPT-4 perform well in mapping natural language descriptions to spatial relations but struggle with multi-hop reasoning. This paper used the StepGame as a benchmark for spatial reasoning. Fine-tuning a model fills the gap identified in the paper, as the only solutions identified in their research was prompt engineering with Chain of Thought.

Research by organizations like OpenAI and Google has focused on improving contextual reasoning through fine-tuning, but there is limited work targeting spatial reasoning.

Main Results

The fine-tuned model slightly improved on general knowledge tasks such as MMLU Geography and Babi Task 17 compared to the original Mistral-7B base model. However, its performance on spatial reasoning benchmarks like SpatialEval significantly declined, suggesting that fine-tuning may have led to incompatibility between the prompt style used for training with StepGame and the multiple-choice formatting in SpatialEval.

Training Data

For this fine-tuning task the step-game dataset was used. This dataset is large and provides multi-step reasoning challenges for geospatial reasoning. The train-test split is predefined with 50,000 rows in the train split and 10,000 in the test split. It focuses on multi-step problem-solving with spatial relationships, such as directional logic, relative positioning, and route-based reasoning. It presents text-based tasks that require stepwise deductions, ensuring the model develops strong reasoning abilities beyond simple fact recall. This dataset follows the template of story, question and answer to assess spatial reasoning as depicted below.

Training Method

For this task of spatial reasoning LoRA (Low-Rank Adaptation) was used as the training method. LoRA allows for efficient fine-tuning of large language models by freezing the majority of the model weights and only updating small, low-rank adapter matrices within attention layers. It significantly reduces the computational cost and memory requirements of full fine-tuning, making it ideal for working with limited GPU resources. LoRA is especially effective for task-specific adaptation when the dataset is moderately sized and instruction formatting is consistent as in the case of this dataset of stepGame. In previous experiments with spatial reasoning fine-tuning, LoRA performed better than prompt tuning. While prompt tuning resulted in close to 0% accuracy on both the StepGame and MMLU evaluations, LoRA preserved partial task performance (18% accuracy) and retained some general knowledge ability (46% accuracy on MMLU geography vs. 52% before training). I used a learning rate of 1e-4, batch size of 4, and trained for 3 epochs. This setup preserved general reasoning ability while improving spatial accuracy.

Evaluation

Model	MMLU-Geography (%)	Spatial Eval (%)	Babi Task 17 (%)
mistralai/Mistral-7B-Instruct-v0.3 (base model)	75.63%	36.07%	51%
sareena/spatial_lora_mistral (fine-tuned)	76.17%	0%	53%
meta-llama/Llama-2-7b-hf	42.42%	18.25%	48.00%
google/gemma-7b	80.30%	7.01%	58.00%

Benchmark Tasks

SpatialQA Benchmark

This benchmark provides more realistic question answer pairs than the stepgame benchmark. It contains place names instead of abstracted concepts of letters and answers, but still requires the same multi-step geospatial reasoning capability. It complements stepGame by testing broader spatial logic and more realistic scenarios

bAbI Dataset (Task 17)

This benchmark was introduced by Facebook AI Research. It includes 20 synthetic question-answering tasks designed to evaluate various reasoning abilities in models. Tasks 17 ("Positional Reasoning") specifically assesses spatial reasoning through textual descriptions. Pathfinding in combination with spatial reasoning will potentially help assess the model’s performance in tasks such as calculating routes, which would be a common application for a geospatial reasoning fine-tuned model.

MMLU benchmark geography and environmental subset

This benchmark is a comprehensive evaluation suite designed to assess a model's reasoning and knowledge across a wide range of subjects. When focusing on geography and environmental science subsets, the benchmark offers an opportunity to test both domain-specific knowledge and broader reasoning abilities relevant to geospatial tasks. Input/Output: This benchmark consists of multiple-choice questions covering topics from elementary to advanced levels. This benchmark is intended to assess the model’s general performance post-processing, and its ability to apply knowledge across subjects most relevant to its fine-tuning task.

Comparison Models

LLaMA-2 and Gemma represent strong alternatives from Meta and Google respectively, offering diverse architectural approaches with a similar number of parameters and training data sources. Including these models allowed for a more meaningful evaluation of how my fine-tuned model performs not just against its own baseline, but also against state-of-the-art peers on spatial reasoning and general knowledge tasks.

Usage and Intended Uses

This model is designed to assist with natural language spatial reasoning, particularly in tasks that involve multi-step relational inference between objects or locations described in text. This could be implemented in agentic spatial systems and/or text-based game bots.

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model = AutoModelForCausalLM.from_pretrained("sareena/spatial_lora_mistral")
tokenizer = AutoTokenizer.from_pretrained("sareena/spatial_lora_mistral")

inputs = tokenizer("Q: The couch is to the left of the table. The lamp is on the couch. Where is the lamp?", return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Prompt Format

The model is trained on instruction-style input with a spatial reasoning question:

Q: The couch is to the left of the table. The lamp is on the couch. Where is the lamp in relation to the table?

Expected Output Format

The output is a short, natural language spatial answer:

A: left

Limitations

The model is still limited in deep reasoning capabilities and sometimes fails multi-hop spatial tasks. LoRA helps balance this trade-off, but fine-tuning on more diverse spatial tasks could yield stronger generalization. )erformance on the SpatialEval benchmark dropped drastically, due to incompatibility between the prompt style used for training and the multiple-choice formatting in SpatialEval. Future work to remediate this would be to test more prompt formats in training or use instruction-tuned datasets more similar to the downstream evaluations.m

Citation

Hendrycks, Dan, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt.
"Measuring Massive Multitask Language Understanding." arXiv preprint arXiv:2009.03300 (2020).
Li, Fangjun, et al.
“Advancing Spatial Reasoning in Large Language Models: An in-Depth Evaluation and Enhancement Using the StepGame Benchmark.”
arXiv.Org, 8 Jan. 2024. https://arxiv.org/abs/2401.03991
Mirzaee, Roshanak, Hossein Rajaby Faghihi, Qiang Ning, and Parisa Kordjamshidi.
"SpartQA: A Textual Question Answering Benchmark for Spatial Reasoning." arXiv preprint arXiv:2104.05832 (2021).
Shi, Zhengxiang, et al.
“StepGame: A New Benchmark for Robust Multi-Hop Spatial Reasoning in Texts.”
arXiv.Org, 18 Apr. 2022. https://arxiv.org/abs/2204.08292
Wang, Mila, Xiang Lorraine Li, and William Yang Wang.
"SpatialEval: A Benchmark for Spatial Reasoning Evaluation." arXiv preprint arXiv:2104.08635 (2021).
Weston, Jason, Antoine Bordes, Sumit Chopra, and Tomas Mikolov.
"Towards AI-Complete Question Answering: A Set of Prerequisite Toy Tasks." arXiv preprint arXiv:1502.05698 (2015).

sareena
/

spatial_lora_mistral