Evian: Towards Explainable Visual Instruction-tuning Data Auditing¶

Conference: ACL 2026
arXiv: 2604.20544
Code: None
Area: Interpretability
Keywords: Data Auditing, Visual Instruction Tuning, Explainable Evaluation, Data Quality, Multi-modal Large Language Models

TL;DR¶

This paper proposes the "Decomposition-then-Evaluation" paradigm and the EVIAN framework, which decomposes answers in visual instruction-tuning data into three components: visual descriptions, subjective reasoning, and factual claims. These are evaluated across three orthogonal dimensions: image-text consistency, logical coherence, and factual accuracy. The study finds that models trained on a small amount of high-quality data filtered by EVIAN outperform those trained on large-scale datasets.

Background & Motivation¶

Background: Large Vision-Language Models (LVLMs) rely on Visual Instruction Tuning (VIT) to achieve alignment between visual perception and language understanding, but the quality of training data remains inconsistent.

Limitations of Prior Work: (1) Large-scale data synthesis (e.g., LLaVA-Instruct-150K) improves instruction following but introduces noise; (2) Existing filtering methods (e.g., CLIP score) use coarse-grained single-dimensional scoring, failing to detect subtle semantic defects such as logical fallacies and factual errors; (3) The LLM-as-a-Judge paradigm suffers from bias, instability, and reasoning shortcut issues.

Key Challenge: Existing data filtering compresses various error types into a single opaque score, making it impossible to distinguish between different types of quality issues such as visual misrepresentation, factual inaccuracies, and reasoning defects.

Goal: To construct an explainable fine-grained data auditing framework that decomposes answers into verifiable cognitive components for multi-dimensional evaluation.

Key Insight: Treat responses as composite structures consisting of visual descriptions, subjective reasoning, and factual claims, rather than indivisible text blocks.

Core Idea: By decomposing complex auditing tasks into verifiable sub-tasks tailored to different cognitive components, more precise data quality assessment can be achieved compared to coarse-grained scoring. Furthermore, logical coherence is identified as the most critical factor in data quality.

Method¶

Overall Architecture¶

EVIAN audits an answer of a visual instruction sample as a "composite cognitive structure" rather than an indivisible text block. The process follows two phases: Phase 1 uses a three-step Chain-of-Thought (CoT) to decompose the answer into a structured form with labels and a pure visual summary, separating visual descriptions, subjective reasoning, and factual claims. Phase 2 scores the sample (1-5) along three non-overlapping dimensions: logical coherence \(S_L\), factual accuracy \(S_K\), and image-text consistency \(S_V\). The final quality score is the average: \(S_{\text{overall}} = (S_L + S_K + S_V) / 3\). Additionally, a controlled flaw injection benchmark is established to quantify the fine-grained detection capability of the auditing pipeline.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400, 'subGraphTitleMargin': {'top': 8, 'bottom': 16}}}}%%
flowchart TD
    IN["Visual Instruction Sample<br/>Image + Instruction + Answer"] --> P1
    subgraph P1["Phase 1: Three-step CoT Decomposition"]
        direction TB
        A1["Step 1 Semantic Labeling<br/>Insert INFER for reasoning, KNOW for facts"] --> A2["Step 2 Visual Distillation<br/>Remove labeled content, keep verifiable descriptions"] --> A3["Step 3 Smooth Synthesis<br/>Reorganize fragments into coherent visual summary"]
    end
    P1 --> MID["Structured Labeled Form + Pure Visual Summary"]
    MID --> P2
    subgraph P2["Phase 2: Three-dimensional Orthogonal Evaluation"]
        direction TB
        B1["S_L Logical Coherence<br/>Check if INFER has visual evidence"]
        B2["S_K Factual Accuracy<br/>Verify KNOW knowledge claims"]
        B3["S_V Image-Text Consistency<br/>Align visual summary with image"]
    end
    P2 --> SCORE["S_overall = (S_L + S_K + S_V) / 3<br/>Filter high-quality subset → Downstream SFT"]
    BENCH["Controlled Flaw Injection Benchmark<br/>15 flaw types, 300k samples"] -. Verification of detection capability .-> IN

Key Designs¶

1. Three-step CoT Decomposition: Breaking sentences into independently verifiable cognitive components

Coarse-grained scoring fails to identify logical fallacies or factual errors because it evaluates heterogeneous content together. EVIAN utilizes a three-step chain for isolation: Step 1 (Semantic Labeling) inserts <INFER> labels for subjective reasoning and <KNOW> labels for factual claims, leaving unlabeled parts as pure visual descriptions; Step 2 (Visual Distillation) removes or rewrites labeled content to leave only objective descriptions verifiable by the image; Step 3 (Smooth Synthesis) organizes these fragments into a coherent paragraph. This allows each component to be sent to its most appropriate evaluation dimension, avoiding the ambiguity inherent in mixed evaluation.

2. Three-dimensional Orthogonal Evaluation Scheme: Measuring different flaws with different metrics

Visual misrepresentation, factual inaccuracy, and reasoning defects are fundamentally different issues requiring distinct criteria. EVIAN orthogonalizes the evaluation: \(S_L\) focuses solely on the logical soundness of reasoning within <INFER> tags—specifically whether visual evidence supports it; \(S_K\) performs fact-checking on knowledge claims within <KNOW> tags; \(S_V\) measures the consistency between the pure visual summary and the image, explicitly prioritizing consistency over completeness. These non-overlapping dimensions ensure each flaw is pinpointed without being diluted by other signals.

3. Controlled Flaw Injection Benchmark: A 300,000-sample diagnostic tool for auditing pipelines

Existing datasets lack systematically injected controllable errors, making it difficult to quantify how well an auditing pipeline detects fine-grained defects. EVIAN constructs a benchmark with 15 semantic flaw types (5 subtypes each for visual consistency, logical coherence, and factual accuracy). These flaws are injected via a three-stage pipeline: content analysis, context-aware error type selection, and guided rewriting. This ensures the injected errors are subtle and contextually relevant rather than obvious blunders, allowing the "auditor's detection capability" to be measured statistically.

Loss & Training¶

Qwen3-235B is used for answer decomposition, and Qwen2.5-VL-7B serves as the automatic auditor for scoring. Downstream validation is performed by fine-tuning Qwen2-VL-2B on the filtered 10K subset. All comparison methods share the same architecture and SFT process to ensure that performance gains result solely from the data filtering strategy.

Key Experimental Results¶

Main Results (Qwen2-VL-2B fine-tuned on 10K subset)¶

Method	MME	MMBench	ScienceQA	A-OKVQA	POPE	Avg
Random	1475.76	0.5353	0.6614	0.7092	75.50	63.18
Full Data (300K)	1553.05	0.5953	0.6267	0.6934	78.17	63.77
SCALE (Prev. SOTA)	1814.97	0.6318	0.6916	0.7066	73.81	67.41
EVIAN (Ours)	1876.89	0.6463	0.7115	0.7493	79.87	70.20

Ablation Study¶

Configuration	Avg	Description
EVIAN (Full)	70.20	Optimal full framework
w/o Decomposition	67.93	Loss of 2.27 without decomposition phase
w/o \(S_L\) (Logical Coh.)	57.27	Largest loss without logical coherence (↓12.93)
w/o \(S_K\) (Factual Acc.)	64.21	Loss of 5.99 without factual accuracy
Only \(S_V\) (Visual Cons.)	65.36	Decent average but POPE drops to 68.56

Key Findings¶

Logical Coherence is Critical: Removing \(S_L\) caused the Average to plummet from 70.20 to 57.27, as relying only on \(S_K\) and \(S_V\) includes samples that are factually correct but logically inconsistent, creating contradictory supervision signals.
"Less is More": The 10K subset filtered by EVIAN (3.3% of 300K) outperforms the full 300K dataset.
In the score distribution, 92.3% of original high-quality samples scored \(\ge 3.0\), while flawed samples concentrated around 3.0 (JSD=0.35, AUC=0.86).
Cross-architecture validation (InternVL2-2B) indicates that improvements stem from data quality rather than inductive bias alignment between the auditor and target model.

Highlights & Insights¶

Core insight of the "Decomposition-then-Evaluation" paradigm: Decomposing auditing into verifiable sub-tasks makes complex auditing reliable.
Challenges the "the more data, the better" paradigm by surpassing full-scale training with only 3.3% of the data.
Discovers that logical coherence (rather than visual alignment or factual accuracy) is the most critical factor in data quality, a counter-intuitive and significant finding.
The taxonomy design of the flaw injection benchmark is systematic, covering 5 subtypes each for consistency, reasoning, and knowledge.

Limitations & Future Work¶

Reliability depends on large multi-modal models for decomposition and evaluation, which may inherit their biases and blind spots.
Errors in the decomposition phase propagate to subsequent evaluations, necessitating improved robustness.
High computational cost due to multiple calls to large models limits application to ultra-large-scale datasets.
Other dimensions of data quality, such as style diversity and pedagogical value, have not yet been modeled.

vs SCALE: SCALE employs multi-stage filtering (modality quality, relevance, clarity, task rarity) but lacks component-level decomposition; EVIAN achieves more precise fine-grained auditing via cognitive component decomposition.
vs CLIPScore/BLIP: Coarse-grained filtering based on similarity fails to capture logical fallacies and factual errors.
vs LLM-as-a-Judge: Direct holistic scoring by models is prone to bias and instability; EVIAN mitigates this through structured decomposition.

Rating¶

Novelty: ⭐⭐⭐⭐ "Decomposition-then-Evaluation" paradigm is novel; 15-type flaw taxonomy is systematic.
Experimental Thoroughness: ⭐⭐⭐⭐⭐ Multiple baselines, complete ablations, cross-architecture validation, and a 300k-sample benchmark.
Writing Quality: ⭐⭐⭐⭐ Clear structure, rich diagrams, and in-depth analysis.
Value: ⭐⭐⭐⭐ Significant guidance for multi-modal data curation; the findings on logical coherence priority have broad impact.