finqa

(EMNLP 2025) Weaver: A modular agentic pipeline that dynamically combines SQL and LLMs for advanced table-based question answering

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

CRDFAgent is a modular Python framework for multi-step question answering, combining clarification, routing, decomposition, retrieval, and aggregation to deliver accurate, reproducible answers for complex tasks like financial document QA.

AI analyst for SEC 10-K filings. RAG + LangGraph agent + GRPO fine-tuning on FinQA. 11.5% → 20.5% accuracy progression.

NumGuard-Fin provenance-constrained decoding implementation for verifiable numerical reasoning in financial question answering.

PUCT Monte Carlo tree search over executable programs for numerical QA on 10-K pages (FinQA): LightGBM action prior and program ranker, interpreter-verified candidates, BM25 + cross-encoder fact retrieval, and a measured simulations-vs-accuracy curve against greedy and beam search.

7.6% of FinQA questions require a number that appears nowhere in the filing they were written against

Self-improving financial-QA agent on FinQA: a governed loop that refines its own prompt + reasoning level over 3 cycles (no retraining), with a validation-guarded gate that verifies every change on held-out data.

Research framework for long-document financial QA using MapReduce and benchmark-driven evaluation, with a multi-document web app.

Controlled study: does continued pre-training on SEC 10-K filings help downstream financial QA? A clean negative result on a fair evaluation instrument. Qwen2.5-3B; FinQA/TAT-QA; CPT (LoRA and full-parameter), SFT, DPO.

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

Fine-tuning Phi-3-mini on FinQA and FinBench using QLoRA for financial reasoning. Before/after benchmarks included. Model published to Hugging Face Hub.

Evidence-grounded agent platform for auditable financial document analysis

Financial reasoning SLM: verified FinQA data, 4-bit QLoRA fine-tuning of Llama-3.2-3B on one T4, GGUF quantization study and llama.cpp deployment on a 4 GB GPU