Skip to content

Knowing What Not to Answer: Selective Non-Compliance in Vision-Language Models

Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/mz-kim/KoNA
Area: LLM Reasoning
Keywords: Vision-Language Models, Selective Non-Compliance, Compound Queries, Alignment Fine-Tuning, Multimodal Safety

TL;DR

This paper introduces KoNA, a benchmark evaluating selective non-compliance in Vision-Language Models (VLMs) across paired single and compound queries, and proposes a two-stage fine-tuning framework integrating SFT and GRPO to enable models to refuse or correct invalid query components while reliably answering valid ones.

Background & Motivation

Vision-Language Models (VLMs) have witnessed extensive deployment across real-world multimodal applications ranging from visual question answering and scene understanding to embodied agent planning. A primary design principle guiding these architectures is helpfulness—generating informative, context-relevant responses to user instructions. However, unconditional helpfulness alone is insufficient for reliable operation, as real-world user queries frequently present conditions where compliance should be strictly withheld: when an input query incorporates a factually incorrect visual premise, inquires about details hidden by camera blur or occlusion, presumes unverifiable subjective intent, requests physically impossible executions, or attempts malicious actions such as theft or trespassing. In such scenarios, trustworthy models must correct false premises, express epistemic uncertainty, or refuse unauthorized actions rather than hallucinating plausible answers.

Existing benchmarks evaluating VLM reliability, hallucinations, and safety predominantly operate under an isolated query-level binary formulation, assuming that an entire request either warrants complete compliance or demands blanket refusal. In realistic human-AI dialogues, however, queries are rarely purely valid or purely invalid; users naturally formulate compound queries that interweave valid, answerable visual questions with invalid components requiring non-compliance (for example, asking "In front of the four elephants, what large object lies on the ground?" when only three elephants are present). Under such compound queries, current state-of-the-art VLMs suffer from a critical failure mode: they either succumb to the false premise and hallucinate compliant answers across the entire sentence, or issue blanket refusals that discard the valid, answerable information sought by the user.

Bridging the gap between isolated abstention and fine-grained, component-level discrimination remains an underexplored challenge in multimodal alignment. Core idea: construct KoNA, a comprehensive benchmark evaluating both query-level and component-level selective non-compliance across five core taxonomy categories using paired single and compound queries, and establish a two-stage alignment framework combining mixed compound/answerable SFT with factuality-penalized GRPO policy optimization to instill balanced, component-level refusal without triggering over-refusal on benign queries.

Method

Overall Architecture

The KoNA framework formalizes selective non-compliance across five grounded failure categories: False Premise (subtly altering a verifiable visual attribute of a salient object), Visual Inaccessibility (querying intrinsic object attributes that are impossible to discern due to photographic constraints like blur, low resolution, or occlusion), Universal Unknown (querying plausible common-sense assumptions regarding intent or identity that cannot be confirmed from the visual scene alone), Task Feasibility (demanding physical executions that exceed VLM software affordances), and Safety (instructing illicit or harmful actions such as theft or surveillance evasion utilizing visible objects).

To evaluate and train models on these conditions, a standardized three-stage data generation and verification pipeline constructs triplet instances for each visual scene: an isolated Single QA, an expanded Compound QA (combining the non-compliance trigger with an answerable visual inquiry on the same entity), and an Answerable Contrast QA (revising the invalid component into a valid description to establish an compliance anchor). To resolve the single-to-compound performance collapse, the method establishes a two-stage fine-tuning pipeline: first, supervised fine-tuning (SFT) over a balanced blend of compound and fully answerable queries teaches the model to separate invalid components from answerable ones; second, Group Relative Policy Optimization (GRPO) refines response policy via multidimensional GPT-5-mini reward signals that explicitly penalize factual inaccuracies on valid components.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Multimodal Image Inputs (MS COCO / Open Images V7)"] --> B["Three-Stage Generation & Verification Pipeline<br/>Single QA generation → Compound expansion → Answerable contrast pair"]
    B --> C["KoNA Diagnostic Benchmark<br/>5 Categories (False Premise / Inaccessibility / Unknown / Feasibility / Safety)"]
    B --> D["Two-Stage Alignment Fine-Tuning<br/>Mixed compound and answerable SFT supervision"]
    D --> E["GRPO Policy Optimization & Multidimensional Reward<br/>Component non-compliance reward + Factuality error penalty (-λ·Rfac)"]
    E --> F["Selective Non-Compliance Output<br/>Explicit refusal/correction of invalid component + Factual answer to valid component"]
    C -.->|Diagnostic Feedback| D

Key Designs

1. Three-Stage Generation & Verification Pipeline: Constructing Paired Single, Compound, and Answerable Contrast Queries

To ensure that evaluation differences between single and compound queries reflect true component-level reasoning rather than dataset artifacts or stylistic drift, the authors construct the benchmark using a unified three-stage generation protocol over MS COCO and Open Images V7. In the first stage (Single QA), generator LLMs (GPT-5 and Gemini-2.5-Flash) formulate queries targeting exactly one isolated non-compliance condition with minimal deviations. In the second stage (Compound QA), each single query is expanded by integrating an additional, visually grounded, answerable inquiry targeting the same entity using coordinating conjunctions. In the third stage (Answerable Contrast QA), the non-compliant component is transformed into a verified, factually answerable statement while retaining the surrounding linguistic structure. After synthetic generation, instances undergo two automated consistency filters followed by Amazon Mechanical Turk human verification requiring annotators to extract exact non-compliant text spans (discarding annotations with token overlap below 0.2), yielding 3,100 verified image instances and 9,300 total QA pairs.

2. Two-Stage Alignment Fine-Tuning: Supervised Decoupling with Answerable Anchoring

Directly fine-tuning models solely on compound non-compliance queries causes severe over-refusal, prompting models to issue defensive refusals even when all components of a query are fully legitimate. To circumvent this conservative collapse, the supervised fine-tuning (SFT) stage constructs a curated training corpus of 1,200 examples composed of 1,000 compound QA instances, 100 single QA instances, and 100 fully answerable QA instances. The compound instances provide direct supervised demonstrations on decomposing mixed-intent queries, teaching the model to formulate dual-part responses that state appropriate refusals or corrections for the invalid sub-question while accurately answering the valid sub-question. The inclusion of the 100 fully answerable instances acts as a critical behavioral anchor, preserving the baseline tendency to follow instructions when queries are entirely valid.

3. GRPO Policy Optimization & Multidimensional Reward: Balancing Selective Refusal with Factuality Penalties

To overcome the brittle boundaries of purely supervised token-matching, the model is further optimized using Group Relative Policy Optimization (GRPO) over an independent split of 80 compound and 20 answerable instances. For each query, the policy samples four candidate completions and evaluates group-based advantages with KL regularization against the SFT policy. Reward computation is delegated to GPT-5-mini via a structured multidimensional reward function \(R\). The reward prioritizes component-level non-compliance accuracy (\(R_{\text{non}} \in \{\text{PASS}, \text{FAIL}\}\)) while enforcing a calibrated penalty on factual errors (\(R_{\text{fac}} \in \{\text{PASS}, \text{FAIL}\}\)) within the answerable portion:

\[R = \begin{cases} 1.0 - \lambda \cdot \mathbb{I}(R_{\text{fac}} = \text{FAIL}), & \text{if } R_{\text{non}} = \text{PASS} \\ 0, & \text{if } R_{\text{non}} = \text{FAIL} \end{cases}\]

Setting the penalty weight \(\lambda = 0.3\) penalizes hallucinations or incorrect entity descriptions in the valid component without disincentivizing necessary non-compliance. Cross-evaluator validation demonstrates high agreement between GPT-5-mini, Gemini-2.5-Flash (85.2%), and Claude Haiku 4.5 (90.1%, Pearson correlation 0.887), confirming the fidelity of the reinforcement signal.

Loss & Training

During the SFT stage, the vision encoder is frozen while all projector and language model parameters are updated. For Qwen2.5-VL-3B-Instruct, training is executed across 8 NVIDIA RTX A6000 GPUs using AdamW, a learning rate of \(5 \times 10^{-5}\), warmup ratio of 0.03 with a constant-with-warmup schedule, gradient accumulation of 4, and gradient clipping at 1.0 for 3 epochs. For InternVL3-2B-Instruct, training uses AdamW, learning rate \(2 \times 10^{-5}\), weight decay 0.05, and a cosine decay schedule over 3 epochs.

In the second GRPO stage, the policy trains for 30 epochs with learning rate \(5 \times 10^{-6}\), batch size 2, gradient accumulation 2, and KL divergence coefficient \(\beta = 0.01\). The best checkpoint is selected based on validation reward.

Key Experimental Results

Main Results

Evaluation spans representative open-source models (InternVL3-2B/78B, Qwen2.5-VL-3B/72B) and closed-source foundation models (GPT-5, Gemini-2.5-Flash) across default inference, Chain-of-Thought (CoT), Behavior Guidance prompting, and KoNA fine-tuning (raw data transcribed from Table 2):

Model Setting False Premise (Sgl / Cmp) Visual Inacc. (Sgl / Cmp) Univ. Unknown (Sgl / Cmp) Task Feas. (Sgl / Cmp) Safety (Sgl / Cmp) Single Avg. Compound Avg.
Default Inference
InternVL3-2B 0.40 / 0.10 0.27 / 0.22 0.18 / 0.12 0.18 / 0.02 0.20 / 0.02 0.25 0.10
InternVL3-78B 0.35 / 0.12 0.42 / 0.40 0.39 / 0.33 0.04 / 0.04 0.56 / 0.28 0.35 0.23
Qwen2.5-VL-3B 0.35 / 0.09 0.32 / 0.27 0.19 / 0.16 0.17 / 0.02 0.43 / 0.01 0.29 0.11
Qwen2.5-VL-72B 0.88 / 0.40 0.80 / 0.50 0.52 / 0.48 0.14 / 0.09 0.60 / 0.21 0.59 0.34
GPT-5 0.63 / 0.14 0.58 / 0.48 0.34 / 0.34 0.25 / 0.22 0.95 / 0.91 0.55 0.42
Gemini-2.5-Flash 0.78 / 0.38 0.47 / 0.39 0.39 / 0.36 0.12 / 0.12 0.71 / 0.69 0.49 0.39
Behavior Guidance Prompting
InternVL3-2B 0.21 / 0.10 0.24 / 0.19 0.19 / 0.07 0.57 / 0.02 0.39 / 0.01 0.32 0.08
Qwen2.5-VL-72B 0.89 / 0.57 0.93 / 0.87 0.86 / 0.86 0.81 / 0.79 0.87 / 0.86 0.87 0.79
GPT-5 0.72 / 0.18 0.67 / 0.66 0.57 / 0.55 0.68 / 0.64 0.98 / 0.95 0.72 0.60
KoNA Fine-Tuned (Ours)
InternVL3-2B-KoNA 0.86 / 0.82 0.88 / 0.88 0.85 / 0.93 1.00 / 0.90 0.99 / 0.97 0.92 0.90
Qwen2.5-VL-3B-KoNA 0.86 / 0.72 0.87 / 0.87 0.82 / 0.89 0.99 / 0.88 1.00 / 0.98 0.91 0.87

Ablation Study

Ablation experiments evaluate the individual contributions of SFT dataset formulation and the secondary GRPO stage, tracking non-compliance accuracy, full compliance retention on valid queries (Answerable Average), and factual visual grounding (Factuality Average) (raw data from Table 4):

Model Configuration False Premise (Sgl / Cmp) Visual Inacc. (Sgl / Cmp) Univ. Unknown (Sgl / Cmp) Task Feas. (Sgl / Cmp) Safety (Sgl / Cmp) Answerable Avg. Factuality Avg.
InternVL3-2B
SFT w/o answerable 0.92 / 0.83 0.77 / 0.90 0.77 / 0.92 0.99 / 0.85 0.94 / 0.94 0.38 0.87
SFT w/ answerable 0.92 / 0.82 0.77 / 0.88 0.72 / 0.86 0.99 / 0.87 1.00 / 0.97 0.60 0.90
SFT + GRPO (Full) 0.86 / 0.82 0.88 / 0.88 0.85 / 0.93 1.00 / 0.90 0.99 / 0.97 0.70 0.88
Qwen2.5-VL-3B
SFT w/o answerable 0.89 / 0.80 0.83 / 0.92 0.86 / 0.91 0.99 / 0.91 0.97 / 0.96 0.53 0.87
SFT w/ answerable 0.86 / 0.74 0.79 / 0.83 0.84 / 0.86 0.99 / 0.90 1.00 / 0.99 0.64 0.87
SFT + GRPO (Full) 0.86 / 0.72 0.87 / 0.87 0.82 / 0.89 0.99 / 0.88 1.00 / 0.98 0.71 0.89

Key Findings

  • Compound queries trigger acute compliance failures in baseline models: Across all foundation VLMs, compound-query accuracy collapses dramatically compared to single-query performance. Under default inference, smaller models (InternVL3-2B, Qwen2.5-VL-3B) drop from 0.25-0.29 to 0.10-0.11 on compound queries. Even frontier models exhibit severe drops; GPT-5 drops from 0.63 to 0.14 on False Premise, revealing that models overwhelmingly comply with false premises when they are embedded within broader answerable queries.
  • Inference-time prompting yields inconsistent, model-scale-dependent gains: Chain-of-Thought prompting fails to mitigate compound failures and frequently increases hallucination rates. Behavior Guidance prompts improve policy-driven categories (Safety and Task Feasibility) on 70B+ models (Qwen2.5-VL-72B reaches 0.79 compound average), but yield negligible or adverse effects on smaller models and remain ineffective for context-sensitive False Premise and Universal Unknown tasks.
  • Answerable contrast instances prevent over-refusal: In the ablation study, omitting answerable contrast examples during SFT causes Answerable Average accuracy to plummet to 0.38 in InternVL3-2B and 0.53 in Qwen2.5-VL-3B, indicating aggressive, unselective refusal. Incorporating answerable instances recovers valid compliance to 0.60 and 0.64.
  • GRPO establishes a superior Pareto balance: Adding the second-stage GRPO optimization elevates Answerable Average to 0.70 and 0.71 while preserving high non-compliance accuracy (0.90 and 0.87 compound average), demonstrating that reinforcement learning with factual penalties effectively reduces over-refusal.
  • Strong out-of-distribution transfer and general benchmark retention: When evaluated on CC3M images (Table 8), KoNA-tuned models achieve 0.88 and 0.86 compound averages, proving generalization beyond training distributions. Extended evaluations on HaloQuest, MM-SafetyBench, R-Bench, and UPD confirm robust cross-benchmark transfer. Concurrently, standard visual capabilities (TextVQA, MMBench, POPE) remain stable (Table 6), while MOSSBench evaluations verify that safe queries are not excessively refused (refusal rates remain low at 0.04-0.05).

Highlights & Insights

  • Formalizing component-level selective non-compliance: The paper moves beyond simplistic binary classification of prompt answerability, exposing that the true challenge in deploying helpful yet safe VLMs lies in selective component-level disentanglement within heterogeneous compound queries.
  • Asymmetric factuality penalty in policy rewards: Designing the GRPO reward function with an asymmetric penalty \(\lambda \cdot \mathbb{I}(R_{\text{fac}} = \text{FAIL})\) ensures that models are explicitly penalized for hallucinating details in the valid component while executing non-compliance on the invalid component.
  • Actionable implications for multimodal agents: For autonomous agent systems where tool invocation instructions often mingle benign context queries with impossible or unauthorized actions, KoNA's fine-tuning recipe provides a lightweight, data-efficient method (1,300 instances) to instill robust refusal boundaries into compact open-source models.

Limitations & Future Work

  • Absence of fine-tuning validation on frontier open-source scales: Computational constraints restricted SFT and GRPO training to 2B and 3B parameter scales; whether identical two-stage recipes remain necessary or exhibit different reward dynamics on 70B+ open models remains unverified.
  • Scope confined to single-turn 2D visual question answering: The taxonomy and dataset focus exclusively on static 2D images in single-turn QA formats, without exploring multi-turn conversational degradation, video temporal dynamics, or embodied action trajectories.
  • Residual visual grounding failures during correct non-compliance decisions: Qualitative error analysis reveals that models can correctly trigger non-compliance while failing on fine-grained visual perception—such as misreading text while correcting false premises (reading "SHOW" as "SHUW") or expanding task refusals to harmless informational guidance.
  • vs HaloQuest / R-Bench / Unk-VQA: Prior visual hallucination and unanswerability datasets formulate queries in isolation where the whole query is unanswerable; KoNA integrates unanswerable premises into compound queries with valid visual components to test selective response.
  • vs MM-SafetyBench / Red Teaming VLMs: Conventional safety benchmarks test isolated jailbreak attempts or harmful queries; KoNA evaluates safety in mixed queries where harmful requests are paired with legitimate environmental description tasks, exposing compliance spillover.
  • vs DPO / Preference Optimization in VLMs: Standard multimodal preference optimization often leads to generic over-refusal; KoNA demonstrates that combining mixed-supervision SFT with a component-level factuality-penalized GRPO reward achieves selective balance with minimal data.

Rating

  • Novelty: ⭐⭐⭐⭐⭐ Formulates the critical yet overlooked problem of component-level selective non-compliance in compound multimodal queries.
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ Rigorous evaluation across 6 diverse models, prompt variants, OOD distributions, 4 external benchmark extensions, and human verification.
  • Writing Quality: ⭐⭐⭐⭐⭐ Exceptionally structured, clear motivation, and thorough error analysis with concrete qualitative failure examples.
  • Value: ⭐⭐⭐⭐⭐ Provides a benchmark and a data-efficient training methodology directly applicable to building reliable, safe multimodal conversational systems.