Research map / Vision-language and multimodal / Multimodal LLMs
CLER: Improving Multimodal Financial Reasoning by Cross-MLLM Error Reflection
Shuangyan Deng, Zhongsheng Wang, Rui Mao, Ciprian Doru Giurcăneanu, Jiamou Liu
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CLER, a cross-MLLM error reflection method that improves multimodal financial reasoning by using errors from stronger models to guide a smaller MLLM, aiming to cut inference costs.
Problem
MLLMs struggle with financial terminology, logical consistency, and numerical computation over joint text-image financial inputs, while commercial models that do well are too expensive for scalable financial use.
Approach
Uses cross-model error reflection: errors made by MLLMs are reflected upon to improve reasoning, with the goal of transferring stronger-model capability to a cheaper model. Details are not available beyond the abstract.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add cross-model error retrieval and reflection (CLER) over plain CoT | GPT-4o · FinMR (all) | accuracy (%): 54.17 -> 61.18 | Improved |
| add reflection on retrieved error pairs (ablation) | CLER (GPT-4o) · FinMR (expertise) | accuracy (%): 60.33 -> 65.90 | Improved |
| use contrastive retriever instead of semantic-similarity retriever (ablation) | CLER (GPT-4o) · FinMR (all) | accuracy (%): 58.52 -> 61.18 | Improved |
| use contrastive retriever instead of semantic-similarity retriever (ablation) | CLER (GPT-4o) · MMMU-Finance (math) | accuracy (%): 63.00 -> 66.78 | Improved |
| add retrieval without reflection over CoT (ablation) | GPT-4o · MMMU-Finance (math) | accuracy (%): 62.77 -> 62.99 | No clear effect |
| apply LEAP principles instead of CoT | GPT-4o · FinMR (math) | accuracy (%): 49.40 -> 42.97 | Hurt |
| apply LEAP principles instead of CoT | GPT-4o · FAMMA (all) | accuracy (%): 54.50 -> 54.50 | No clear effect |
| apply LEAP principles instead of CoT | Claude-4-sonnet · FAMMA (math) | accuracy (%): 59.00 -> 57.00 | Hurt |
| generate insights individually at test time instead of general pre-generated principles (ablation) | CLER · FinMR | accuracy (%): +3.0% for GPT-4o | Improved |
| reduce error database size to 80% and 50% (ablation) | CLER · FinMR, FAMMA, MMMU-Finance | accuracy | Hurt |
| upgrade backbone from Claude-3.7-sonnet to Claude-4-sonnet under CLER (ablation) | CLER · FinMR (all) | accuracy (%): 62.85 -> 65.10 | Improved |
Closest papers by meaning
- FinMMDocR: Benchmarking Financial Multimodal Reasoning with Scenario Awareness, Document Understanding, and Multi-Step Computation (AAAI 2026 · similarity 0.91)
- FinBen: A Holistic Financial Benchmark for Large Language Models (NeurIPS 2024 · similarity 0.87)
- STEER-ME: Assessing the Microeconomic Reasoning of Large Language Models (NeurIPS 2025 · similarity 0.86)
- FinSearchComp: Towards a Realistic, Expert-Level Evaluation of Financial Search and Reasoning (ICLR 2026 · similarity 0.86)
- MathSE: Improving Multimodal Mathematical Reasoning via Self-Evolving Iterative Reflection and Reward-Guided Fine-Tuning (AAAI 2026 · similarity 0.85)
- Conformal Thinking: Risk Control for Reasoning on a Compute Budget (ICML 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
