PaperlayerView exact revision on arXiv

Source-linked AI summary

When Visual Evidence Is Insufficient: How Instructions, Post-Training, and Internal Activations Shape Refusal in Multimodal Language Models

Ananth K Suresh, Aveen Dayal, Gouthaman KV, Aditya Pola, Vineeth N Balasubramanian

arXiv:2610.37430v1cs.LGcs.CV

TL;DR

Visually unanswerable questions call for withholding unsupported answers, but refusal alone does not show that a model recognized insufficient visual evidence. The paper tests instruction, post-training, and activation controls on MoHoBench across three MLLMs. Refusal varies with instruction meaning and placement; post-training yields larger increases, while activation steering produces smaller transferable changes.

  • Problem

    Images may omit, obscure, or fail to show needed information, and refusal alone does not establish that an MLLM recognized insufficient visual evidence.

    Source: §1 · pp. 1–2

  • Method

    The study holds image–question pairs fixed while testing instruction policies and placement, post-training, and a behavior-derived activation direction with reason-aware refusal evaluation.

    Source: §1 · pp. 2–3, §3.1 · p. 4, §3.4 · p. 5

  • Results

    Evidence-aware and general instructions had stronger effects than mismatched safety instructions; C1/C3 raised total refusal 18.4–45.2 percentage points across 12 model–position comparisons.

    Source: §1 · p. 2, §5.1 · p. 6

  • Takeaways & Limitations

    Refusal frequency alone is insufficient to characterize clean evidence-based abstention, so refusal reasons and responses that combine refusal with attempted answers matter.

    Source: §6 · p. 8

    The activation intervention establishes causal control under the tested perturbation, not that the unmodified model naturally represents or computes visual uncertainty through that direction.

    Source: §5.3 · p. 8

Abstract

from arXiv · show

For reliable deployment, multimodal large language models (MLLMs) should recognize when an image provides insufficient evidence to answer a question and withhold unsupported answers. Yet refusal alone is not necessarily evidence of such recognition: models may refuse because of unrelated safety constraints or without identifying the missing evidence. We study how refusal on visually unanswerable questions is shaped by three sources of control: inference-time instructions, post-training, and internal activation interventions. Across three MLLMs on the MoHoBench benchmark, we systematically manipulate the meaning and placement of refusal instructions, compare supervised and preference-based post-training, and intervene on a refusal-related activation direction derived from naturally answered and refused examples. We find that refusal is highly sensitive to both the meaning and placement of instructions, with model-specific changes at the sample level that are obscured by aggregate refusal rates. Evidence-aware and general refusal instructions have stronger effects than an intentionally mismatched safety instruction. Post-training produces larger increases in refusal, but can also induce responses that combine refusal language with attempted answers. In contrast, activation intervention produces smaller yet transferable changes across prompt conditions, without evidence of indiscriminate refusal on tested visually answerable samples. These findings show that refusal is jointly shaped by external instructions, post-training, and internal representations, highlighting the need to disentangle evidence recognition from a model's learned propensity to refuse.

1 Introduction

The introduction argues that refusal on visually unanswerable questions does not by itself show that a model recognized insufficient visual evidence. It frames refusal as shaped by instruction policies, post-training, and internal activations, and previews effects that vary across models and samples.

1 Introduction

  • The study compares evidence-aware, safety-oriented, and general refusal policies and varies instruction placement while keeping each image–question pair fixed.Unlike prompt-paraphrase studies, these instructions prescribe different reasons for withholding an answer.Source: §1 · p. 2
  • The post-training analysis compares supervised fine-tuning with DPO, ORPO, and SimPO to test whether increased abstention is evidence-based or indiscriminate.It also examines whether responses become internally inconsistent.Source: §1 · p. 2
  • The activation intervention steers a fixed model using a direction derived from its naturally answered and refused examples, testing transfer across response policies and selectivity.The direction is evaluated for its ability to alter refusal behavior.Source: §1 · p. 2
  • Evidence-aware and general refusal instructions have stronger effects than mismatched safety instructions; post-training raises refusal more, while activation steering produces smaller but transferable changes.Instruction placement also produces model-specific sample-level changes obscured by aggregate rates; post-training can combine refusal language with attempted answers.Source: §1 · pp. 2–3
  • A reason-aware, same-sample evaluation distinguishes evidence-based, safety, and generic refusals and reveals behavioral changes that aggregate refusal rates can hide.The study compares instruction-, parameter-, and activation-level mechanisms, including a behavior-derived direction tested for transfer and selectivity.Source: §1 · p. 3

2 Related Work

Related work covers visual abstention and unanswerability, prompt sensitivity, and multimodal refusal interventions. This study focuses on how response policies and their placement affect refusal and stated reasons for fixed unanswerable samples.

Visual unanswerability, hallucination, and abstention

  • Selective VQA studies abstention through risk–coverage tradeoffs, while unanswerability benchmarks test whether models recognize questions that cannot be answered from provided inputs.The paper distinguishes these from its focus on the response-level decision and stated reason for withholding an answer.Source: §2 · p. 3

Instruction sensitivity and policy-conditioned refusal

  • Prior work finds that multimodal behavior varies with prompt wording, structure, context ordering, and prompting strategies; this study instead tests response-policy semantics and position.Its paired sample-level transitions distinguish refusal reasons under different policies.Source: §2 · p. 3

Safety refusal, over-refusal, and activation steering

  • Prior work studies activation directions for safety refusal and over-refusal, whereas this paper derives its direction from naturally answered and refused benign unanswerable questions.The paper evaluates it across response policies and instruction positions, with matched random and shuffled-label directions and benign description controls.Source: §2 · p. 3

3 Method

The method compares instruction-level policies, post-training, and activation interventions as controls on refusal, using a shared response taxonomy and paired evaluations. It tests how policy meaning and placement, learned response preferences, and a behavior-derived activation direction shape refusal.

3.1 Problem setting and response taxonomy

  • Responses receive a primary label, while a separate mixed-response indicator records whether refusal language co-occurs with a substantive answer attempt.The analysis tracks total and evidence-based refusal, transitions among response labels, and answer–refuse switching.Source: §3.1 · p. 4

3.2 Instruction-level policy control

  • The instruction study compares an evidence-aware policy, a safety policy mismatched to benign questions, and a general refusal policy against an unaugmented condition.Each instructed policy appears before or after the question, with the image–question pair held fixed for paired comparisons.Source: §3.2 · pp. 4–5

3.3 Parameter-level control through post-training

  • Post-training comparisons include supervised fine-tuning and preference-based methods DPO, ORPO, and SimPO, assessed for refusal, indiscriminate abstention, and general multimodal performance.Reason-specific labels and mixed-response incidence accompany total refusal because increased refusal may not be evidence-based or behaviorally clean.Source: §3.3 · p. 5

3.4 Behavior-derived activation intervention

  • The activation intervention derives a direction from naturally answered and refused examples matched by question category, then shifts the residual representation toward the refusal centroid.Norm-matched random and category-preserving shuffled-label directions serve as controls; the intervention tests transfer across policies and positions, benign-description refusal, and MMMU accuracy.Source: §3.4 · p. 5

4 Experimental Protocol

The protocol evaluates three MLLMs on MoHoBench and applies consistent semantic response labeling across model and intervention conditions. It specifies policy prompts and summarizes paired instruction-level changes and response switching.

4 Experimental Protocol

  • The evaluation uses InternVL3-2B, InternVL3-8B, and Qwen2.5-VL-7B on MoHoBench questions from four visually unanswerable categories.A single semantic judge assigns a primary response label and a separate mixed-response indicator across the evaluation conditions.Source: §4 · pp. 5–6
  • Table 2 defines the four response policies, with C1–C3 tested in prefix and postfix positions.Prefix inputs place the instruction before the question; postfix inputs place it after the question.Source: §4 · p. 5, Table 2 · p. 6
  • Table 3 reports C0 refusal rates, paired changes from C0 and between positions, answer–refuse and four-class switching, and image-clustered bootstrap intervals.Rates and changes are reported as percentages and percentage points, respectively.Source: Table 3 · p. 6

5 Results

Across the three control levels, refusal changes with instruction meaning and placement, post-training method and model, and activation steering. These increases do not by themselves establish clean evidence-based abstention; mixed responses and model-specific effects qualify the results.

5.1 Instruction semantics and position strongly affect refusal

  • Evidence-aware and general refusal policies increased total refusal by 18.4–45.2 percentage points in all 12 model–position comparisons, while safety-oriented policy effects ranged from −2.3 to +13.2 points.Moving instructions changed answer/refuse outcomes for 13.7–23.6% of items and response type for 20.9–32.1%; position effects differed by model.Source: §5.1 · pp. 6–7

5.2 Post-training increases refusal without ensuring clean abstention

  • SFT and ORPO raised refusal to 83.9–94.9% under C0 and 96.7–99.9% under C1-prefix, while DPO and SimPO effects varied by architecture and technique.Mixed-response incidence reached 30.8% under SFT and 33.2% under ORPO for InternVL3-2B, and 24.1% under DPO for Qwen; near-ceiling refusal therefore did not ensure clean abstention.Source: §5.2 · p. 7, Table 4 · p. 7

5.3 The behavior-derived direction provides transferable but imperfect control

  • The behavior-derived activation direction increased total refusal in all 21 model–condition comparisons by 1.1–15.1 percentage points, with 20 confidence intervals excluding zero, and increased net evidence-based refusal by 1.3–13.3 points.Its effects transferred across policy meanings and instruction positions without condition-specific retuning, but remained dependent on the surrounding policy context.Source: §5.3 · p. 7, Table 5 · p. 8
  • The intervention outperformed random and shuffled-label directions by 4.9–11.9 points at the anchor conditions, but its effects were neither reason-pure nor monotonic.No refusals occurred on the 704 tested description requests per model, and the largest absolute MMMU change was 0.46 points with every interval including zero; these controls do not establish selectivity for arbitrary answerable inputs.Source: §5.3 · p. 8

6 Conclusion

The paper treats refusal on visually unanswerable questions as a behavior shaped by instructions, learned response policies, and internal state—not a direct readout of insufficient visual evidence. It evaluates these controls on three MLLMs using MoHoBench and specified post-training and activation-intervention protocols.

6 Conclusion

  • Refusal on visually unanswerable questions is shaped by inference-time instructions, learned response policies, and internal model state, so refusal frequency alone is insufficient to characterize clean evidence-based abstention.The conclusion calls for evaluating refusal reasons, same-sample transitions, and selectivity controls; the study covers three MLLMs and MoHoBench.Source: §6 · p. 8

A.1 Models, benchmark, and data splits

  • The evaluation used InternVL3-2B, InternVL3-8B, and Qwen2.5-VL-7B on 900 MoHoBench questions spanning four categories and 176 unique images.The final test set had no item-level or image-level overlap with post-training, direction-construction, validation, or exploratory sets.Source: §A.1 · p. 11

A.2 Post-training evaluation

  • Post-training variants were evaluated on the same 900 questions under C0 and C1-prefix, with benign image-description requests and a common MMMU subset used to assess refusal and general multimodal performance.Refusal on the description task was treated as unwarranted because it asks only for directly visible information.Source: §A.2 · p. 11

A.3 Activation-intervention settings and controls

  • The activation direction’s layer and strength were selected on validation data, then fixed across all seven policy conditions without policy- or position-specific retuning.Selectivity was assessed with four benign description requests, and utility with paired accuracy on the same MMMU subset used for post-training comparison.Source: §A.3 · p. 11

A.4 Statistical analysis

The statistical analysis uses paired comparisons and resampling procedures that account for questions sharing an image. The appendix also specifies how response distributions and the category-stratified figure are reported.

A.4 Statistical analysis

  • Policy and intervention effects were computed as paired differences, with 95% MoHoBench bootstrap intervals from 10,000 replicates resampling image identities and retaining associated questions.MMMU intervals used paired resampling of 650 common samples, and sensitivity analyses excluded pairs containing mixed responses.Source: §A.4 · p. 11
  • Results are reported separately for each model because instruction-position effects differ across model families, rather than being averaged into one cross-model estimate.Source: §A.4 · p. 11
  • Complete response distributions report each response-type entry as a count and rate, while Mixed is an orthogonal flag rather than a fifth primary type.The figure presents category-stratified refusal rates for unmodified models across seven response-policy conditions; its numerical values are repeated in Table 14.Source: Table 7 · p. 12, Figure 3 · p. 12

B Detailed Results

The appendix defines the response-distribution labels and the bootstrap procedure used for detailed results.

B Detailed Results

  • Primary response types are A, RE, RS, and RG; the mixed flag is orthogonal to the primary type.Source: B Detailed Results · p. 12
  • Response typing was produced by the Qwen3.5 LLM evaluator described in the experimental protocol.Source: B Detailed Results · p. 12
  • Unless otherwise stated, confidence intervals use 10,000 bootstrap replicates with seed 42, resampling decoded-image identities while retaining associated questions.Source: B Detailed Results · p. 12

B.1 Complete paired behavioral results

The appendix provides complete paired response, transition, and policy-effect analyses, followed by alignment comparisons across post-training methods. It also specifies the training objectives and reports benign-description and MMMU controls.

B.1 Complete paired behavioral results

  • The paired analyses preserve each image–question pair across 900 items and 176 decoded-image clusters, with response distributions, transitions, refusal effects, and category breakdowns reported.The mixed-excluded analyses remove pairs when either response is mixed; retained denominators range from 791 to 885 while all 176 clusters remain represented.Source: §B.1 · p. 12

B.2 Alignment details

  • The alignment comparison evaluates SFT, DPO, ORPO, and SimPO against the corresponding unmodified models at C0 and C1-prefix on the same 900 items.Tables report response distributions, post-training effects, benign-description outcomes, and paired MMMU correctness transitions.Source: §B.2 · p. 13, Table 15 · p. 16, Table 16 · p. 16
  • SFT uses supervised likelihood; DPO uses a reference-relative preference objective, ORPO combines supervised learning with an odds-ratio preference term, and SimPO is reference-free.Training time and recorded memory usage are also reported.Source: §B.2 · p. 13, Table 21 · p. 18

B.1 Complete paired behavioral results

  • Transition tables give source-to-target response counts in [A, RE, RS, RG] order, with a mixed-excluded analysis removing pairs flagged as mixed.The Qwen2.5-VL-7B and InternVL transition matrices are reported separately.Source: Table 8 · p. 13, Table 9 · p. 13, Table 10 · p. 14

B.3 Complete steering and control results

The behavior-derived direction increases total and evidence-based refusal across most tested conditions, but also yields generic refusals and some reversals from refusal to answering. The reported controls do not establish selectivity for arbitrary answerable questions.

B.3 Complete steering and control results

  • The behavior-derived direction increases total and evidence-based refusal across most tested conditions, while also producing generic refusals and refusal-to-answer reversals.The complete decompositions are reported in the appendix.Source: §B.3 · p. 18
  • Fixed-control specificity is established only at C0 and C1-prefix and relative to the recorded seed-42 directions.This bounds the control comparison to those tested settings and directions.Source: §B.3 · p. 18
  • Description and MMMU controls show neither blanket refusal nor a detectable accuracy change under tested configurations, but are not an equivalence test or evidence of selectivity for arbitrary answerable questions.The controls do not include arbitrary answerable-question cases.Source: §B.3 · p. 18

C Limitations

The evaluation covers visually unanswerable MoHoBench questions and three open MLLMs, without matched answerable-VQA counterfactuals. The authors identify extending the protocol to more models and answerable VQA settings as future work.

C Limitations

  • MoHoBench contains visually unanswerable questions, so the description and MMMU controls do not provide matched answerable-VQA counterfactuals.The paper therefore bounds its conclusions to the evaluated scope.Source: C Limitations · p. 18
  • The study covers three open MLLMs; the authors identify additional models and answerable VQA settings as natural extensions.These extensions would use the same paired, reason-aware protocol.Source: C Limitations · pp. 18, 24

NeurIPS Paper Checklist

  • The checklist records that the abstract and introduction claims reflect the paper’s contributions and scope.The checklist response is marked Yes.Source: NeurIPS Paper Checklist · p. 25

2. Limitations

The checklist records author affirmations about discussing limitations, supplying experimental details, and reporting confidence intervals. It also records an affirmative response on ethics compliance.

2. Limitations

  • The authors state that the paper discusses its limitations.The checklist marks this item Yes.Source: §2 · p. 25
  • The checklist says experimental-setting details and information for reproducing the results are provided in Appendix Section A.The experimental-setting item is marked Yes.Source: §2 · pp. 26–27
  • The authors state that confidence intervals are reported for the results and mark the ethics-compliance item Yes.The checklist gives these as affirmative responses.Source: §2 · pp. 27–28
Loading 2610.37430v1…