numerical-reasoning

A list of Numerical Multimodal reasoning papers and their implementation

Data and software artifacts for the EMNLP 2024 (Main) paper "What Are the Odds? Language Models Are Capable of Probabilistic Reasoning"

Thinking-aware baselines & low-data LoRA/QLoRA post-training on FinQA — a controlled Qwen3-4B vs Qwen3-8B numerical-reasoning study.

Code and data for the StarSem 2023 paper "Arithmetic-Based Pretraining -- Improvin Numeracy of Pretrained Language Models"

PROFIT (Processing and Reasoning Over Financial Information in Text): A benchmark designed to evaluate language models across five tasks of increasing complexity in numerical reasoning using SEC filings

Low-resource financial reasoning with Qwen3-4B/8B, LoRA/QLoRA SFT, controlled FinQA evaluation and ablation studies.

Empirical investigation of the representation–behavior gap in LLM numerical processing.

Decomposition-scored benchmark for multi-step numerical reasoning over Indian annual reports (retrieval / units / intermediate / final)

Numerical reasoning practice that generates every question from scratch. No question bank, no dependencies, no build step.

Program-of-Thought LLM agent for conversational numerical reasoning over financial documents (ConvFinQA). 82% exec-acc with the cheapest model; reproducible eval, executor-checked gold, documented negative-result analysis.