MFMDQwen: Financial Misinformation Detection LLM

Updated 5 July 2026

MFMDQwen is an open-source LLM adapted for multilingual financial misinformation detection, handling heterogeneous task settings effectively.
It leverages the MFMD4Instruction dataset for fine-tuning and the MFMDBench benchmark for robust evaluation, achieving superior Macro-F1 scores.
Its methodology emphasizes instruction tuning and domain adaptation to accurately identify subtle numerical distortions and contextual manipulations.

Searching arXiv for the MFMDQwen paper and closely related references. MFMDQwen is a Qwen-based, instruction-tuned LLM for Multilingual Financial Misinformation Detection (MFMD), introduced as the first open-source LLM designed for MFMD tasks. It is presented together with MFMD4Instruction, the first instruction dataset supporting MFMD with LLMs, and MFMDBench, a benchmark for evaluating MFMD capabilities across multiple languages and task formulations. The work is positioned against prior financial misinformation systems that are primarily English-centric and single-task, and frames MFMD as a setting in which subtle numerical changes, amplified sentiment, reversed causality, and multilingual variation make detection materially harder than generic misinformation classification (Liu et al., 20 Apr 2026).

1. Problem setting and claimed novelty

MFMDQwen addresses multilingual financial misinformation detection as a domain-specific LLM problem in which a model must detect or verify false, manipulated, or misleading financial information across multiple languages and heterogeneous task schemas. The paper motivates the problem by emphasizing that financial misinformation can affect investment behavior, asset prices, economic sentiment, and market stability, while also noting that financial content often contains minor numerical distortions or context-sensitive causal claims that are difficult to detect through surface semantics alone (Liu et al., 20 Apr 2026).

The authors identify two deficiencies in prior work. First, existing LLM-based approaches for financial misinformation detection primarily focus on English. Second, they usually support only a single task rather than a broader multilingual, multi-task setup. In this framing, MFMDQwen’s novelty does not lie in a new transformer architecture. Rather, it lies in the combination of three artifacts: MFMD4Instruction as a multilingual supervised instruction corpus, MFMDQwen as the adapted Qwen-family model, and MFMDBench as the held-out multilingual benchmark.

The paper also distinguishes MFMDQwen from earlier domain-specific models such as FMDLlama. The claim is not merely that a finance-specialized model exists, but that multilinguality and task heterogeneity are treated as first-class design constraints. This suggests that the contribution is best understood as a domain adaptation and evaluation framework for multilingual finance misinformation rather than as an architectural advance.

2. Model design, backbone, and objective

MFMDQwen is built on Qwen-3-8B. The methods section states, “We then built MFMD-R based on Qwen-3-8B using the MFMD4Instruction dataset,” while the title, abstract, tables, and conclusions consistently use the name MFMDQwen. The model does not modify the Qwen transformer internals; instead, it uses post-training adaptation, primarily supervised fine-tuning, followed according to the methods section by RL, although the RL stage is not adequately documented (Liu et al., 20 Apr 2026).

The task is formalized as a generative conditional modeling problem over multiple tasks. For each task $t$ , the dataset is written as

$D_t = \{(q_i^t, r_i^t)\}_{i=1}^{N_t},$

where $q_i^t$ is the context token sequence containing the task description, input text, and query, and $r_i^t$ is the target token sequence representing the final answer. Training maximizes

$\max_{\phi} \sum_{t} \sum_{i=1}^{N_t} \log \mathbf{P}_{\phi}(r_i^t \mid q_i^t).$

This is a standard instruction-tuning objective expressed at the task level rather than as a new loss design.

A central design choice is the conversion of heterogeneous datasets into a unified instruction format. Each instance contains a task description, the input fields, and a short label string as output. The appendix includes task-specific prompt templates for multiple schemas: True or False for Global4Languages, entailed or refuted for FinDVer, real or false for MDFEND, Supported, Refuted, or Not enough information for CHEF, MANI or NO_MANI for BanMANI, and [true/false] for RFC. MFMDQwen therefore operates over multiple task-specific label spaces through instruction conditioning rather than through a single universal classifier head.

The multilingual capability is attributed to the combination of a multilingual-capable base model and training data in English, Chinese, Greek, and Bengali. The paper does not report translation-based augmentation, tokenizer modification, multilingual adapters, or language-specific heads. A plausible implication is that cross-lingual transfer is expected to emerge from instruction tuning on mixed-language task data rather than from explicit multilingual architectural mechanisms.

3. MFMD4Instruction and MFMDBench

MFMD4Instruction is the supervised instruction-tuning corpus used to train MFMDQwen, while MFMDBench is the evaluation benchmark. The paper states that MFMD4Instruction is assembled from nine existing sources covering four languages: English, Chinese, Bengali, and Greek. These sources span claim verification, fact checking, fake news or misinformation detection, manipulated social media detection, and reference-free counterfactual financial misinformation detection (Liu et al., 20 Apr 2026).

The source datasets are:

FinDVer: 700 claim verification instances from financial reports, balanced as 350 entailed and 350 refuted.
RFC-Bench / RFC: 1,826 original-perturbed paragraph pairs from Yahoo Finance articles on 223 U.S. stocks, with perturbations under Directional Flipping, Numerical Perturbation, Sentiment Amplification, and Causal Distortion.
CHEF: 1,188 Chinese fact-checking samples with labels supported, refuted, and not enough information.
MDFEND: 1,321 Weibo posts labeled real (959) and fake (362).
BanMANI / Bengali Manipulation: 101 samples labeled MANI (52) and NO_MANI (49).
Global4Languages: aligned claim verification data in English, Chinese, Bengali, and Greek, with 144 instances per language and labels true (23) and false (121).

The allocation between training and evaluation is explicit:

Component	Instances	Role
MFMD4Instruction	4,837	SFT / training
MFMDBench	2,432	Evaluation
Total	7,269	Combined corpus

At the dataset level, MFMDBench includes GlobalGr, GlobalBe, GlobalCh, GlobalEn, CHEF, MDFEND, Bengali, FinDVer, and RFC. The paper reports limited preprocessing detail beyond conversion into a unified instruction format. It does not report deduplication, contamination checks, class rebalancing beyond source datasets, translation pipelines, or multilingual balancing strategies. For source-level quality control, only RFC is specifically described as using multi-stage expert review and dual-annotator evaluation.

This design makes MFMDBench a multilingual, multi-task held-out suite rather than a single-schema benchmark. Its significance is methodological: it evaluates whether a single open-source LLM can support multiple financial misinformation tasks with task-conditioned label semantics.

4. Training pipeline and experimental protocol

The reported pipeline has two stages: SFT and RL. However, only SFT is specified in reproducible detail. The reported SFT settings are: base model Qwen-3-8B, learning rate $1 \times 10^{-5}$ , warmup ratio 0.1, 2 epochs, batch size 128, max input length 24k tokens, max output length 8k tokens, DeepSpeed ZeRO-3, CPU parameter offloading, full-parameter training, and hardware consisting of 4 × NVIDIA L40s GPUs with 48 GB each (Liu et al., 20 Apr 2026).

These sequence limits are notable because some tasks, especially FinDVer, involve long financial reports and evidence paragraphs. The paper does not report the optimizer type, weight decay, gradient accumulation, checkpointing strategy, validation selection criterion, or test-time decoding configuration. The RL stage is mentioned, but the reinforcement learning algorithm, reward design, RL hyperparameters, training duration, and ablations isolating RL’s contribution are not reported.

The baseline set is broad within open-source LLMs. It includes Qwen3 reasoning variants (Qwen3-8B-R, Qwen3-14B-R, Qwen3-32B-R), Qwen3 non-reasoning variants (Qwen3-8B, Qwen3-14B, Qwen3-32B), Qwen2.5-72B-Instruct, Llama-3.1-8B-Instruct, Llama-3.3-70B-Instruct, and the domain-specific baseline FMDLlama. Evaluation uses Accuracy and Macro-F1, with the analysis mainly focusing on F1.

The partial reporting of the training stack is itself consequential. It means that MFMDQwen is described as a two-stage post-trained specialist model, but the paper fully documents only the SFT portion of that claim.

5. Empirical results and error profile

The main empirical result is that MFMDQwen achieves the best overall performance on MFMDBench among the tested open-source models, with an average Macro-F1 of 0.818, and the top score on 6 out of 9 benchmarks (Liu et al., 20 Apr 2026).

Model	Average Macro-F1	Note
MFMDQwen	0.818	Best overall
Qwen3-32B-R	0.674	Baseline
Qwen3-32B	0.668	Baseline
Llama3.3-70B	0.663	Baseline
Qwen2.5-72B	0.662	Baseline
FMDLlama	0.293	Domain-specific baseline

The paper reports the following MFMDQwen per-dataset scores:

Dataset	ACC	F1
GlobalGr	0.862	0.758
GlobalBe	0.897	0.803
GlobalCh	0.862	0.710
GlobalEn	0.897	0.756
CHEF	0.849	0.843
MDFEND	0.932	0.913
Bengali	0.952	0.952
FinDVer	0.700	0.673
RFC	0.954	0.954

The strongest gains are reported on Chinese datasets, especially CHEF and MDFEND. On English datasets the margin is smaller, and FinDVer is a relative weakness: MFMDQwen obtains 0.673 F1, while larger general-purpose models score higher there, including Qwen3-32B-R at 0.857, Qwen3-32B at 0.793, Qwen2.5-72B at 0.841, and Llama3.3-70B at 0.818. The authors interpret this as evidence that structured evidence-based judgment over financial reports remains a relative strength of larger general-purpose LLMs.

The appendix identifies five recurring error types: label polarity reversal or prior-driven guessing, numerical reasoning and arithmetic errors, table parsing and financial statement grounding errors, partial evidence matching, and manipulation detection under subtle lexical mismatch. The paper also notes that some baseline LLMs produce uncertain or irrelevant responses, and that reasoning models do not always outperform non-reasoning models of the same size because extra reasoning can induce overthinking.

These failure modes are consistent with the observed FinDVer weakness. They indicate that MFMDQwen’s task specialization improves multilingual robustness and broad benchmark performance, but does not eliminate bottlenecks in arithmetic verification, semi-structured table grounding, or precise evidence alignment.

6. Distinction from financial QA and relation to multimodal Qwen work

MFMDQwen concerns multilingual financial misinformation detection, not multilingual multimodal financial question answering. That distinction matters because adjacent finance-Qwen work can appear superficially similar. For example, the FAMMA benchmark defines Financial Multilingual Multimodal Question Answering and evaluates Qwen2-VL 70B, but it does not introduce a new Qwen-based model, does not fine-tune Qwen, and is best characterized as a benchmark and evaluation paper rather than a model-training paper (Xue et al., 2024).

This separation clarifies a common source of confusion. FAMMA targets multilingual, multimodal finance QA with charts, tables, and screenshots, whereas MFMDQwen targets multilingual misinformation detection over textual claims, reports, posts, and evidence. In other words, MFMDQwen is not a finance-QA system and is not presented as a multimodal VLM.

At the same time, later Qwen-centric multimodal post-training work suggests a nearby methodological trajectory. MMFineReason shows that large-scale Qwen-native rationale distillation, quality filtering, and difficulty-aware selection can substantially strengthen Qwen3-VL reasoning without changing the underlying architecture (Lin et al., 29 Jan 2026). A plausible implication is that a future multimodal extension of MFMDQwen could adopt a comparable data-centric recipe if financial misinformation detection were expanded to tables, charts, screenshots, or evidence-rich document images.

7. Limitations, reproducibility, and significance

The paper explicitly states that, due to compute and cost constraints, evaluation was limited to 8B-, 14B-, and 32B-scale LLMs. Additional limitations are evident from the reported methodology. Performance is not uniformly best across all tasks, especially on FinDVer; the model still exhibits weaknesses in arithmetic verification, table grounding, and subtle factual distinctions; the RL stage is underreported; and the multilingual scope covers only four languages: English, Chinese, Greek, and Bengali (Liu et al., 20 Apr 2026).

The work also does not include formal ablations for SFT versus SFT+RL, multilingual versus monolingual training, source dataset contributions, prompt template effects, or sequence length effects. It does not report fairness analyses, subgroup bias evaluations, or an extended ethics section on misuse and deployment risk. As a result, the paper’s central empirical claim is strong, but the causal decomposition of that gain remains limited.

Implementation availability is summarized through the project page at https://github.com/lzw108/FMD. The paper describes MFMDQwen as open source and presents MFMD4Instruction and MFMDBench as released resources, although it does not enumerate repository contents in detail.

In the literature on finance-specific LLM adaptation, MFMDQwen is most significant as an end-to-end multilingual specialization pipeline. Its contribution is not a new architecture, but a concrete synthesis of task formalization, instruction tuning, multilingual benchmarking, and domain-specific evaluation for financial misinformation detection. Within that scope, it establishes an open-source baseline showing that targeted multilingual post-training can outperform both larger general-purpose open LLMs and an earlier finance-specific misinformation model on a dedicated multilingual benchmark (Liu et al., 20 Apr 2026).