Skip to content

Evian: Towards Explainable Visual Instruction-tuning Data Auditing

🕒 Published (v1): 2026-04-22 13:28 UTC · Source: Arxiv · Venue: ACL 2026 · link

Why this paper was selected

Explainable visual instruction data auditing for LVLM training quality

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

EVIAN is an automated pipeline for auditing visual instruction-tuning data that decomposes model responses into visual descriptions, subjective inferences, and factual claims, then evaluates each along three orthogonal axes: Image-Text Consistency, Logical Coherence, and Factual Accuracy. Training on a 10K EVIAN-curated subset outperforms training on the full 300K unfiltered pool, directly challenging the scale-centric paradigm in LVLM development. Ablations reveal that Logical Coherence is the single most critical dimension — its removal causes the largest performance drop.

Problem

Existing visual instruction-tuning datasets suffer from subtle semantic defects (object hallucination, factual errors, flawed reasoning) that coarse, uni-dimensional filters such as CLIPScore cannot detect. Current methods collapse diverse error types into a single opaque score, lacking the granularity to identify logical fallacies, attribute misattribution, or knowledge inconsistencies. This creates a bottleneck where scaling data volume does not guarantee training quality.

Method

EVIAN operates as a two-phase pipeline over image-instruction-response triples \(x_i = (I_i, P_i, R_i)\).

Phase 1 — Response Decomposition uses Qwen3-235B-A22B-Instruct via a three-step chain-of-thought: 1. Semantic Tagging: wraps subjective inferences in <INFER> tags and knowledge-dependent claims in <KNOW> tags, leaving pure visual text untagged. 2. Visual Distillation: rewrites or deletes tagged segments to produce a draft \(R_i^{\text{draft}}\) containing only image-grounded content. 3. Fluent Synthesis: reorganizes the draft into a coherent visual summary \(R_i^{\text{visual}}\) without adding new content.

Phase 2 — Multi-faceted Evaluation uses Qwen2.5-VL-7B-Instruct-AWQ to assign a score triplet \(S_i = (S_L, S_K, S_V) \in [1,5]^3\): - \(S_L\) (Logical Coherence): soundness of <INFER> reasoning relative to visual evidence. - \(S_K\) (Factual Accuracy): correctness of <KNOW> claims against internal knowledge. - \(S_V\) (Image-Text Consistency): fidelity of \(R_i^{\text{visual}}\) to the image; omissions tolerated, contradictions heavily penalized.

The aggregated scalar \(S_{\text{overall}} = \frac{S_L + S_K + S_V}{3}\) ranks samples for selection.

To benchmark the pipeline, the authors construct a 300K-sample testbed: 50K clean samples plus 250K samples with LLM-injected defects covering 14 error subtypes across the three auditing dimensions (e.g., spatial misattribution, causal fallacy, entity corruption), generated via a probabilistic cascade that prioritizes rarer knowledge/reasoning errors.

Key Contributions

  • A 300K-sample benchmark built by systematically injecting 14 categories of subtle semantic defects into samples drawn from eight datasets (ShareGPT-4V, LLaVA-1.5-Mix, DocVQA, ChartQA, etc.).
  • The Decomposition-then-Evaluation paradigm: structured separation of visual, inferential, and factual response components before scoring.
  • The EVIAN framework itself: fully automated, produces interpretable per-dimension scores with textual rationales rather than a single opaque scalar.
  • Empirical finding that Logical Coherence (\(S_L\)) is the most decisive quality dimension — removing it alone drops average performance from 70.20 to 57.27.

Results

  • EVIAN (10K subset) vs. full 300K data: average 70.20 vs. 63.77 — "less is more" holds.
  • EVIAN vs. SCALE (prior SOTA, 10K): 70.20 vs. 67.41 average.
  • EVIAN vs. Qwen2.5-VL direct scoring (same auditor, 10K): 70.20 vs. 66.34, confirming gains come from structured decomposition, not the model itself.
  • Discriminative power on benchmark: Jensen–Shannon divergence = 0.35, AUC = 0.86 separating clean from defect-injected samples; 92.3% of pristine entries score ≥ 3.0.
  • Ablation — removing \(S_L\): average drops to 57.27 (worst configuration); removing \(S_K\): 64.21; using only \(S_V\): 65.36.
  • Hallucination (POPE): EVIAN 79.87 vs. SCALE 73.81; using only \(S_V\) collapses POPE to 68.56, below random baseline (75.50).
  • Cross-architecture generalization: advantages persist on InternVL2-2B (reported in Appendix D).
  • Original-distribution validation: 10K EVIAN-curated subset rivals full 300K baseline on unmodified data (Appendix E).

Limitations

  • \(S_K\) (Factual Accuracy) relies on the auditor model's internal knowledge; claims beyond its knowledge cutoff or in specialized domains may be mis-scored.
  • The decomposition step depends on a large LLM (Qwen3-235B), making the pipeline computationally expensive for very large-scale corpora.
  • Defect injection for benchmark construction uses the same model family (Qwen3) as the decomposer, which may introduce distributional alignment that artificially inflates EVIAN's benchmark numbers.
  • Equal weighting of \(S_L\), \(S_K\), \(S_V\) in \(S_{\text{overall}}\) is a default; sensitivity to weight choices is deferred to an appendix, not fully explored in the main paper.
  • Evaluation is conducted on a single fine-tuning target (Qwen2-VL-2B) in the main table; cross-architecture results are relegated to the appendix.

Relevance to Vision-Language Models

EVIAN directly addresses one of the most persistent pain points in LVLM development: the inability of existing similarity-based filters (CLIP, BLIP) to catch logical and factual errors in instruction-tuning data. By demonstrating that a 10K high-quality subset beats a 300K unfiltered corpus, it provides strong empirical grounding for the "quality over quantity" principle that is increasingly shaping LVLM data practices. The finding that Logical Coherence is the dominant quality factor — more important than visual grounding or factual correctness in isolation — is a non-obvious and practically actionable insight for dataset curators. The framework also connects to the broader LLM-as-Judge literature by showing that structured decomposition before scoring significantly outperforms holistic model-based evaluation.