Instruction Faithfulness in VLMs

Updated 25 December 2025

Instruction faithfulness is a measure that evaluates how reliably vision-language model outputs adhere to user instructions and accompanying evidence.
Multiple metrics such as N-gram overlaps, CLIPScore, and chain-of-thought analysis are used to assess instruction-following accuracy, groundedness, and causal alignment.
Advancements like retrieval-augmented generation and self-reflection training help mitigate hallucinations and enhance model faithfulness in practical applications.

Instruction faithfulness (“VLM Score”) quantifies the degree to which outputs from vision-LLMs (VLMs) or instruction-following models reliably align with, and are supported by, the user’s input instruction and any accompanying evidence or context (e.g., images, reference text). As VLMs and instruction-tuned LLMs are deployed in high-stakes contexts, rigorous, granular, and domain-sensitive metrics for faithfulness have become foundational for both research diagnostics and risk mitigation.

1. Core Concepts and Definitions

Instruction faithfulness is formally the extent to which a model output $y$ remains consistent with the instruction $u$ (and, where applicable, ground-truth evidence $e$ such as an image or passage), i.e., $y \models (u, e)$ . Hallucinations—statements in $y$ not entailed by $u$ or $e$ —directly decrease faithfulness. Depending on modality and application, the “VLM Score” measures this property via a range of interpretable, often reference-free, metrics (Malin et al., 2024, Uppaal et al., 13 Dec 2025, Liu et al., 27 Oct 2025, Jing et al., 2023).

Key properties:

Instruction-following accuracy: Does the model perform the described task?
Groundedness: Are claims substantiated by evidence (visual, textual)?
Causal alignment: Are cited rationales, or reasoning steps, necessary and sufficient for predictions?
Faithful perception: In multimodal settings, are image/vision-derived steps actually supported by the input?

2. Metric Taxonomy

Instruction faithfulness is operationalized through diverse quantitative metrics, which are selected based on modality and desired granularity.

Traditional Metrics

N-gram overlap (EM, ROUGE, BLEU): Compute lexical overlap with reference. Limited correlation with human faithfulness judgments; prone to being “hacked” by copy-paste responses (Malin et al., 2024, Adlakha et al., 2023).
BERTScore, CLIPScore: Embedding or vision-LLM similarity to reference. Moderate positive correlation.
Token-based faithfulness (K-Precision, K-Recall, K-F1): Measures overlap between the answer and supporting knowledge tokens or passages (Adlakha et al., 2023). K-Precision correlates robustly with human annotations (ρ ≈ 0.50).

Granular and Automated Metrics

Atomic Fact FaithScore: Extracts fine-grained atomic facts from outputs and checks, via a visual entailment model, whether each is entailed by the corresponding image. The overall FaithScore is the fraction verified by the visual model (Jing et al., 2023).
Counterfactual Consistency Score (CCS): The EDCT framework makes targeted, minimal counterfactual edits to visual concepts in the image; a fully faithful model should (a) update its answer and (b) update its explanation accordingly. CCS is the average product of binary prediction-change and explanation-causality scores across all cited concepts (Ding et al., 27 Sep 2025).

Chain-of-Thought and Step-Level Metrics

Unfaithful Perception Rate (UPR) / Chain Faithfulness: Decomposes reasoning chains into perception and reasoning steps. UPR is the fraction of perception steps in a chain that are unsupported by evidence, yielding a chain-level faithfulness metric $F_{\text{chain}} = 1 - \text{UPR}$ (Uppaal et al., 13 Dec 2025).

Visual Information Faithfulness

Reliability: Fraction of examples where perturbing/removing visual cues does not change the model's answer. Low reliability indicates that visual evidence is ignored (Liu et al., 27 Oct 2025).
Sufficiency: Fraction where visual cues alone enable correct prediction (with a third-party judge). High sufficiency implies that the cited visual region is informative.

3. Evaluation Frameworks and Protocols

Numerous recent works instantiate end-to-end faithfulness evaluations via modular pipelines:

FaithScore Pipeline (Jing et al., 2023):
1. Identify descriptive sub-sentences via LLM.
2. Decompose each into atomic image facts (entity, count, color, relation, attribute).
3. For each fact, invoke a visual entailment model to predict support.
4. Return the proportion of facts that pass.
EDCT (Explanation-Driven Counterfactual Testing) (Ding et al., 27 Sep 2025):
1. Extract key visual concepts from model explanations.
2. Generate minimally altered, counterfactual images (e.g., change color/object).
3. Requery the VLM; collect answers and post-counterfactual explanations.
4. Score as faithful if both the answer and the cited reasoning track the counterfactual edit.
UPR/Step-Level Faithfulness (Uppaal et al., 13 Dec 2025):
1. Segment chain-of-thought responses into steps.
2. Annotate steps as Perception or Reasoning.
3. Prompt an off-the-shelf VLM judge (optionally grounded with a caption of the input image) to rate each perception step as Faithful/Unfaithful.
4. Aggregate across perception steps and report UPR.
Visual Reliability/Sufficiency (Liu et al., 27 Oct 2025):
1. Apply “visual intervention” (replace visual cue with noise) and “textual intervention” (inject contradiction) separately.
2. Measure change in outputs (accuracy) post-intervention to compute causal faithfulness metrics.

4. Model Training and Optimization for Instruction Faithfulness

Instruction faithfulness is not only an evaluation target, but an explicit optimization objective in state-of-the-art models.

Edit-GRPO in VIVA (Cong et al., 18 Dec 2025):
- Modifies Group Relative Policy Optimization to optimize a composite reward, including “Instruction Following” measured by the difference in CLIP similarity to edited vs. source descriptions, “Source Preservation” (CLIP consistency with source), and human/aesthetic preference. Only edits aligning with the instruction yield high reward. Fine-tuning with lightweight LoRA yields strong gains in “Instruction Following” scores.
SCCM for Chain Faithfulness (Liu et al., 27 Oct 2025):
- Sufficient-Component Cause Model learning combines answer accuracy, format compliance, and an SCCM reward (product of sufficiency and minimality of visual cues). This incentivizes minimal faithful evidence, yielding substantial increases in chain faithfulness and causal dependence on cited visual crops.
Self-Reflection for Chain Faithfulness (Uppaal et al., 13 Dec 2025):
- A training-free procedure that detects unfaithful perception steps (via a black-box VLM judge) and locally regenerates them, typically correcting >90% within three retries.

5. Empirical Calibration and Benchmarking

Quantitative experiments have calibrated the faithfulness of leading VLMs and LLMs using the above metrics.

On VIE-Bench (instruction-based video editing), VIVA achieves 9.72/10 on instruction-following, outperforming baselines by 1.3–4 points. Removing Edit-GRPO or the VLM instructor sharply degrades faithfulness (Cong et al., 18 Dec 2025).
On OK-VQA using CCS, Gemini 2.5 Flash achieves CCS = 0.674, whereas open-source VLMs score 0.43–0.56 (Ding et al., 27 Sep 2025).
On LLaVA-1k (FaithScore): LLaVA-1.5 >0.85, Multimodal-GPT ≈0.53, revealing major gaps in visual grounding (Jing et al., 2023).
For chain-of-thought tasks, SCCM raises reliability from ≈26% to ≈61% and sufficiency from ≈41% to ≈76% on fine-grained spatial benchmarks (Liu et al., 27 Oct 2025). UPR reductions via self-reflection correlate with both improved stepwise faithfulness and incremental gains in final-task accuracy (Uppaal et al., 13 Dec 2025).
Human evaluations consistently rate LLM entailment and K-Precision as the most aligned with actual groundedness, with lexical and embedding-based metrics lagging behind (Adlakha et al., 2023, Malin et al., 2024).

6. Mitigation of Hallucination and Faithfulness Failures

Mitigating instruction-level unfaithfulness involves both model design and use of faithfulness signals during training and inference.

Retrieval-Augmented Generation (RAG): Integrate retrieved evidence at generation time, decreasing hallucination by up to 50% (Malin et al., 2024, Adlakha et al., 2023).
Prompt-based Self-Critique/Refine: Iterative prompting to identify and repair factual inconsistencies in model outputs yields 15–20% improvements in faithfulness (Malin et al., 2024).
Targeted Fine-Tuning: SCCM or Edit-GRPO mechanisms encourage models to produce evidence-minimal, instruction-grounded chains and outputs (Liu et al., 27 Oct 2025, Cong et al., 18 Dec 2025).
Human–AI Loops: Annotator-in-the-loop interfaces allow for real-time correction of hallucinated facts (as mapped via FaithScore decomposition) (Jing et al., 2023).

7. Best Practices, Limitations, and Future Directions

Key principles for robust instruction faithfulness assessment include:

Multi-metric reporting: Combine lexical, model-based, and LLM-based faithfulness scores.
Regular human validation: Corroborate automated metrics via annotation (Pearson/Spearman correlation).
Step-level granularity: Evaluate and mitigate faithfulness per chain step, not just at final output (Uppaal et al., 13 Dec 2025).
Domain-specific calibration: Fine-tune faithfulness judges to match application requirements (Malin et al., 2024, Jing et al., 2023).
Comprehensive artifact logging: Store evaluation and judge transcripts for regulatory transparency (e.g., EU AI Act compliance) (Ding et al., 27 Sep 2025).
Holistic evaluation: Report both correctness (does the output satisfy the request?) and faithfulness (is it well-grounded?), as these are partially independent (Adlakha et al., 2023).

A persistent limitation is the potential for adversarial exploitation of token-overlap metrics and the imperfect alignment of LLM-based judges with human domain experts in edge cases. Ongoing research targets reference-free, step-level, and multimodal faithfulness diagnostics as robust solutions.

This overview reflects major threads in current research on instruction faithfulness and the VLM Score in instruction-following VLMs and LLMs, referencing established and emerging metrics, optimization strategies, empirical benchmarks, and recommended protocols (Cong et al., 18 Dec 2025, Uppaal et al., 13 Dec 2025, Liu et al., 27 Oct 2025, Ding et al., 27 Sep 2025, Jing et al., 2023, Malin et al., 2024, Adlakha et al., 2023).