Skip to content

Delineating Knowledge Boundaries for Honest Large Vision-Language Models

Conference: ECCV 2026
Paper: ECCV paper page
Area: Multimodal VLM
Keywords: Vision-Language Model, Knowledge Boundary, Learning to Refuse, Preference Optimization, Hallucination Mitigation

TL;DR

This paper probes a target model's own knowledge boundary with high-temperature multi-sample consistency, labels each knowledge-intensive question as mastered or unmastered, and then builds preference pairs whose chosen/rejected direction flips with that label; after SFT plus DPO/CPO/ORPO alignment, LLaVA-1.5-7B learns to refuse honestly on parametric knowledge gaps without over-refusing what it does know, lifting the Truthful rate on V-Idk from 57.9% to 67.3%.

Background & Motivation

Large vision-language models (VLMs) are already strong on open-domain multimodal QA, yet the moment a question touches a long-tail entity or a specialized domain — a rare species, a complex radiograph, an obscure building or machine — they answer with full confidence and complete inaccuracy. These failures are not "I cannot see it" but "I never learned it": the image itself is clear and unambiguous, and the model simply lacks the world knowledge required to interpret it. Existing hallucination-mitigation work almost always aims at the same target, answering more accurately: V-DPO suppresses object-level hallucination with vision-guided direct preference optimization, RLHF-V aligns behavior with fine-grained correctional feedback, and CLIP-DPO uses a vision-language model itself as the source of preference. All of them optimize answer quality, under the implicit assumption that accuracy is the whole game. A capability that is equally important and far cheaper to acquire is left out — knowing when not to answer. The text-only community already has the Idk framework, R-Tuning, and Hindsight Instruction Relabeling teaching LLMs to say "I don't know", but the multimodal side is largely empty; meanwhile the prevailing alignment paradigm ranks helpfulness above truthfulness, which effectively forces the model to guess when it has no evidence.

Unpacking the problem reveals that a VLM's refusal can stem from two very different sources. One is perceptual uncertainty: the image is blurred, occluded, or corrupted, the visual evidence is insufficient, and refusing is plainly right — this is what benchmarks such as VizWiz measure, and where most existing VLM refusal work stops. The other is epistemic uncertainty: the image is perfectly clear, the model genuinely "sees" it, but it holds no parametric knowledge about what it is looking at, and abstaining is then the only honest behavior. The latter is more insidious because nothing signals the model to stay quiet. In the four-quadrant view of Cheng et al. (left of Fig. 1), vanilla VLMs pile up in Unknown Unknowns (IDK-IDK), emitting confident fabrications about entities they never mastered; this paper aims to convert those fabrications into Known Unknowns (IK-IDK), i.e. standard honest refusals. Unlike InBoL, which studies generic knowledge boundaries for questions that are unanswerable by nature, this work targets the knowledge gap of a question that is answerable in principle but happens to lie outside a particular model's knowledge.

The complication is that this boundary is model-specific: for the same question about a rare bird, a 13B model may answer reliably while a 7B model is wrong nine times out of ten, so the mastered/unmastered dividing line has to be measured per model, and dataset annotations are of no help. The paper's approach is to measure the boundary first and then write it back into the parameters. Core idea: probe the target model with high-temperature multi-sample consistency (10 independent responses per question, average accuracy \(A(x,q)\) against a threshold \(\tau=0.7\)) to label each sample Known or Unknown, then build preference pairs whose direction is reversed by that label — for mastered facts the correct answer is chosen and refusal is rejected, for unmastered facts refusal is chosen and the model's own most confident wrong answer is rejected — and finally calibrate this answer-refusal boundary with SFT followed by DPO/CPO/ORPO.

Method

Overall Architecture

The input is a set of image-question pairs plus a target VLM \(\pi_\theta\) to be aligned (LLaVA-1.5-7B/13B here), and the output is a model that proactively says "I don't know" at parametric knowledge gaps while still answering on facts it has mastered. The pipeline has three parts. Dataset construction does three things: a stronger VLM acts as a semantic judge to strip perceptual noise (blurry images, ambiguous answers, questions solvable by OCR alone) out of InfoSeek; the target model's own repeated sampling then labels every remaining sample Known or Unknown; and preference pairs are generated in opposite directions according to that label. Alignment training runs in two stages: SFT first teaches the standard refusal format, then DPO/CPO/ORPO on the preference pairs actually pushes the "when to refuse" decision into the parameters. Four-quadrant evaluation looks beyond accuracy: an LLM judge maps every output back to the four knowledge quadrants, from which the correctness rate and the appropriate-refusal rate are counted separately, and their sum is the paper's central metric, the Truthful rate.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400, 'subGraphTitleMargin': {'top': 8, 'bottom': 16}}}}%%
flowchart TD
    A["InfoSeek image-question pairs"] --> B["Visual-Semantic Sample Selection<br/>VLM judge removes perceptual noise"]
    B --> C["Multi-Sample Consistency Probing<br/>10 samples, average accuracy"]
    C -->|"A ≥ τ, labeled Known"| D["correct answer as chosen<br/>refusal as rejected"]
    C -->|"A below τ, labeled Unknown"| E["refusal as chosen<br/>most confident wrong answer as rejected"]
    subgraph SG["Asymmetric Preference Pair Construction"]
        direction TB
        D
        E
    end
    D --> F["Preference-Aware Alignment<br/>SFT then DPO / CPO / ORPO"]
    E --> F
    F --> G["Four-Quadrant Honesty Metrics<br/>accuracy + refusal = Truthful"]

Key Designs

1. Visual-Semantic Sample Selection: make "wrong answer" attributable to a knowledge gap rather than to poor perception

Knowledge-intensive datasets such as InfoSeek carry perceptual noise of their own — blurry images, ambiguous ground-truth answers, or questions that need no world knowledge at all (asking who wrote a book whose author is printed on the cover is pure OCR). Leave those in the training set and a wrong answer could equally be a perception problem, an annotation problem, or a knowledge problem, which corrupts every label the probe later produces. The paper therefore has a stronger VLM (Qwen2.5-VL) act as a semantic judge over three criteria: visual quality is clear and unambiguous, the ground-truth answer is precise, and the query genuinely requires world knowledge rather than basic optical character recognition or simple perception. Only samples passing all three enter the curated subset — the rejection reason shown in Fig. 2 is literally "Relies on OCR, no world knowledge required."

This step is the foundation of the whole method: it removes perceptual uncertainty from the data entirely, so that any error on the remaining samples can only point at one cause, a missing parametric fact. The Known/Unknown probe labels and the preference pairs derived from them all rest on this clean attribution. It also draws the fundamental line between this work and perceptual-robustness refusal research: those methods handle blurry images, whereas this one handles clear images plus a blank knowledge slot.

2. Multi-Sample Consistency Probing: measure the boundary with the model's own ten answers

Knowledge boundaries differ per model and per question, and no off-the-shelf annotation exists; a single response is also insufficient evidence of mastery, since the model may guess right by luck or slip once due to sampling variance. What the paper needs to judge is not even "was this answer correct" but "does this fact live in the model's parameters at all". The procedure samples \(N=10\) independent responses per \((x,q)\) at high decoding temperature, has an LLM judge score each one, and takes the average accuracy \(A(x,q)\), which is then compared against a threshold \(\tau\) (the paper calls it the Ik threshold): \(A\ge\tau=0.7\) makes the sample Known, otherwise Unknown. ⚠️ The paper describes this process in prose (Sec. 3.1); the expression below is an equivalent plain rendering:

\[ A(x,q)=\frac{1}{N}\sum_{i=1}^{N}\mathbb{I}\!\left[V\!\left(y^{(i)},\,y_{gt}\right)\right],\qquad N=10,\ \tau=0.7 \]

Multi-sample consistency is precisely what separates a lucky guess from stable mastery: only if at least seven of ten responses are correct is the fact admitted to be in the parameters. It measures behavioral stability rather than the model's self-report, so it is insensitive to phrasing style, and because the label is model-specific, changing the model means re-running the probe. On V-Idk this probe yields roughly 60% Known and 40% Unknown, leaving enough supervision on both sides — the Known side constrains helpfulness (do not refuse everything), the Unknown side constrains hallucination (do not answer everything). The threshold \(\tau\) is effectively a knob that moves both the refusal rate and the false-refusal rate at once: raising it marks more samples Unknown and biases the training signal toward conservatism.

3. Asymmetric Preference Pair Construction: flip chosen and rejected according to mastery

Teaching the refusal format with SFT alone incurs an "alignment tax": the model learns a blanket conservative prior and rejects even questions it genuinely knows, while training refusals only on unmastered samples cannot stop it from drying up on mastered ones. The paper therefore generates preference pairs along two branches defined by the probe label. For Known samples (\(A\ge\tau\)), the chosen response is the correct answer and the rejected one is the refusal template — literally telling the model "you must answer this kind of question". For Unknown samples (\(A<\tau\)), the chosen response is the refusal template and the rejected one is the highest-confidence incorrect response the model previously produced, i.e. the most plausible-looking hallucination is pinned down as the negative example. Together the two branches form the 20,000-sample V-Idk training set.

This same-data-opposite-directions construction is what makes preference optimization meaningful at all. If every sample pushed toward refusal, DPO/ORPO would be indistinguishable from SFT and the model would merely learn a more extreme conservative prior; only when the polarity of the pairs flips with mastery is the model forced to learn the intermediate variable "mastered or not" instead of the surface pattern "refuse". Picking the most confident wrong answer rather than a random one also sharpens the penalty — the hallucinations the model is most sure about are exactly the dangerous ones in deployment.

4. Four-Quadrant Honesty Metrics: turn "honest" into countable behavior

Judging honesty by eyeballing a few examples neither compares alignment strategies nor shows whether the model truly knows its boundary or has merely memorized a refusal template. The paper crosses the probe's initial mastery label \(A(x,q)\) with the semantic correctness of the output at test time to obtain four knowledge quadrants: Known Knowns (IK-IK, answered correctly), Known Unknowns (IK-IDK, correctly refused on a sample it never mastered), Unknown Knowns (IDK-IK, wrongly refused a mastered fact, i.e. the alignment tax), and Unknown Unknowns (IDK-IDK, fabricated an answer to an unmastered sample, i.e. hallucination). The three key metrics are defined as (all normalized by the full test-set size \(|\mathcal{D}|\)):

\[ \rho_{IK}=\frac{1}{|\mathcal{D}|}\sum_{(x,q)\in\mathcal{D}}\mathbb{I}\!\left[V(y_{pred},y_{gt})\right],\qquad \rho_{IDK}=\frac{1}{|\mathcal{D}|}\sum_{(x,q)\in\mathcal{D}_{unk}}\mathbb{I}\!\left[V(y_{pred},\text{refusal})\right],\qquad T=\rho_{IK}+\rho_{IDK} \]

where \(\mathcal{D}_{unk}=\{(x,q)\in\mathcal{D}\mid A(x,q)<\tau\}\) is the subset the probe deemed unmastered, and \(V(\cdot,\cdot)\) is the semantic verification predicate returned by an LLM judge, which is True when the prediction and the target (ground-truth answer or refusal intent) agree in meaning.

Honesty thus becomes one sentence: answer correctly on a mastered fact, or appropriately refuse on an unmastered one — either counts as one honest behavior, and the Truthful rate is the share of such behaviors. This definition dictates how the experiments must be read: because both \(\rho_{IK}\) and \(\rho_{IDK}\) are normalized by \(|\mathcal{D}|\), under the 60/40 Known/Unknown split their ceilings are 60 and 40 respectively (the "Data Proportion" row of Table 1 marks exactly these caps), so an absolute value like 45.8 should be read against its cap rather than against an implicit full score of 100. As supporting evidence, the paper also measures the average log-probability (Logprob) and perplexity (PPL) of the refusal template at inference time, to distinguish refusal grounded in genuine epistemic calibration from template mimicry. Watching only the Truthful rate would hide the cheat of trading a high refusal rate for safety; reporting \(\rho_{IK}\) and \(\rho_{IDK}\) together reveals whether the model learned to discriminate or just learned to stay silent.

A Worked Example

Walk two samples from Fig. 2 through the pipeline (⚠️ the sampling counts below are transcribed from the figure).

First, Q "What is the brand of this vehicle?" — of 10 high-temperature samples 9 are correct (the correct brand is Chrysler, with one answer of Volkswagen), so \(A=9/10=0.9\ge\tau\) and the sample is labeled Known. The preference pair is chosen = "Chrysler", rejected = the refusal template "I don't know". This sample means "you must answer this kind of question", and it is what suppresses the alignment tax.

Second, Q "What is the architectural style of this building?" — the 10 samples split into 2 Gothic (correct), 6 Baroque, and 2 Neoclassical, so \(A=2/10=0.2<\tau\) and the sample is labeled Unknown. The preference pair is chosen = the refusal template, rejected = "Baroque", the most confident wrong answer. This sample means "you must not guess here".

Contrast a sample discarded during selection: Q "Who is the author of this book?" may well be answerable, but the name is printed on the cover and obtainable by OCR, consuming no parametric knowledge, so it is filtered out — keeping it would only contaminate the "wrong answer equals knowledge gap" attribution. After alignment the boundary becomes an observable behavioral difference (right of Fig. 1): facing a question beyond its parametric knowledge such as "What historical region did this building belong to?", the base model confidently answers "Moravia", while the aligned model says "I don't know" — and the correct answer is Bohemia.

Loss & Training

Training runs in two stages. The first is SFT (Idk-SFT in the paper), which simply maximizes the log-likelihood of the chosen responses in order to teach the standard refusal format:

\[ \mathcal{L}_{\text{SFT}} = -\mathbb{E}_{(x,y_w)\sim \mathcal{D}_{\text{V-Idk}}}\left[\log \pi_\theta(y_w \mid x)\right] \]

⚠️ The LaTeX of Eqs. 1–3 is corrupted in the cached PDF text; the expression above is rewritten from the prose in Sec. 3.2, and the exact form should be taken from the original paper.

The second stage performs preference optimization on the pairs, with three options. DPO takes the SFT model as the reference \(\pi_{ref}\) and applies a sigmoid to the \(\beta\)-weighted difference of log-ratios, pushing the chosen probability up and the rejected one down. CPO drops the reference model and optimizes a contrastive objective directly between chosen and rejected, keeping the model close to the chosen distribution while moving away from the rejected one. ORPO merges the supervised term and the preference term into a single objective \(\mathcal{L}_{\text{ORPO}}=\mathcal{L}_{\text{SFT}}+\lambda\mathcal{L}_{\text{OR}}\), where the preference term uses an odds ratio to raise the relative odds of being honest on unknown facts while preserving helpfulness on known ones (⚠️ Eqs. 2/3 are likewise corrupted in the cache; only the functional form is reconstructed here, and the exact expressions should be taken from the original). The difference between the three is precisely what the experiments compare: SFT learns only the format, DPO stays conservative, and CPO/ORPO strike the best balance on this task.

Implementation details: everything uses LoRA for parameter-efficient fine-tuning with rank \(r=64\), \(\alpha=128\), targeting all linear modules of the LLM backbone. SFT runs 1 epoch over the 20,000 V-Idk training samples with AdamW, a learning rate of \(1\times10^{-5}\), and a global batch size of 32; the preference-optimization stage runs 3 epochs with the learning rate lowered to \(1\times10^{-6}\) and a cosine decay schedule. The hardware is 2×NVIDIA H800 (80G). Inference uses greedy decoding for determinism, and every output is passed to a high-capability LLM judge for semantic verification.

Key Experimental Results

Main Results

Evaluation uses the 3,000-sample V-Idk test set (60% Known / 40% Unknown) under two inference protocols. Zero-shot provides only the image and the question, with no refusal instruction or template, testing whether alignment truly internalized the boundary; few-shot explicitly instructs the model to reply "I don't know" when it does not know and supplies two in-context demonstrations (one correct factual answer, one standard refusal), testing whether the model can calibrate under guidance.

Model Method Zero-shot IK-IK Zero-shot IK-IDK Zero-shot Truthful Few-shot IK-IK Few-shot IK-IDK Few-shot Truthful
Data Proportion (ceilings) — 60.0 40.0 100.0 60.0 40.0 100.0
LLaVA-1.5-7B Base (Prompting) 45.8 1.8 47.6 45.5 12.4 57.9
SFT 34.4 ↓11.4 32.7 ↑30.9 67.1 ↑19.5 27.6 ↓17.9 36.8 ↑24.4 64.4 ↑6.5
DPO 51.9 ↑6.1 5.5 ↑3.7 57.4 ↑9.8 45.3 ↓0.2 17.8 ↑5.4 63.1 ↑5.2
CPO 44.4 ↓1.4 13.9 ↑12.1 58.3 ↑10.7 42.8 ↓2.7 24.5 ↑12.1 67.3 ↑9.4
ORPO 45.9 ↑0.1 14.1 ↑12.3 60.0 ↑12.4 40.0 ↓5.5 26.5 ↑14.1 66.5 ↑8.6
LLaVA-1.5-13B Base (Prompting) 49.9 4.3 54.2 49.6 19.8 69.4
SFT 34.4 ↓15.5 39.1 ↑34.8 73.5 ↑19.3 27.5 ↓22.1 39.8 ↑20.0 67.3 ↓2.1
DPO 50.2 ↑0.3 6.7 ↑2.4 56.9 ↑2.7 49.7 ↑0.1 21.4 ↑1.6 71.1 ↑1.7
CPO 44.9 ↓5.0 19.0 ↑14.7 63.9 ↑9.7 43.7 ↓5.9 30.3 ↑10.5 74.0 ↑4.6
ORPO 45.8 ↓4.1 20.4 ↑16.1 66.2 ↑12.0 42.7 ↓6.9 33.9 ↑14.1 76.6 ↑7.2

Subscripts are gains or losses relative to the Base model of the same size; the "Data Proportion" row gives the theoretical ceilings for \(\rho_{IK}\), \(\rho_{IDK}\), and Truthful on this test set, since all three metrics are normalized by the full set size \(|\mathcal{D}|\).

Ablation Study

There is no conventional module ablation here; the closest analogue is a comparison of the training objective itself: Prompting (no training), SFT (format only), and DPO/CPO/ORPO (different forms of preference optimization) are treated as five alignment strategies over the same data, supplemented by two analysis threads — cross-domain generalization and internal probing.

Evaluation set (LLaVA-1.5-7B, few-shot) Data split Prompting SFT DPO CPO ORPO
V-Idk main set · Truthful 60:40 57.9 64.4 63.1 67.3 66.5
ScienceQA · Truthful 50:50 56.1 59.5 60.1 62.6 63.8
VizWiz-Unans · Truthful (visual shift) — 84.3 95.6 92.3 90.9 82.8
PMC-VQA · Truthful (knowledge shift) — 75.8 77.4 77.9 80.0 78.2

⚠️ The VizWiz-Unans and PMC-VQA rows are transcribed from the numeric labels on the bar chart in Fig. 3; consult the original figure for exact values. On VizWiz-Unans the correct behavior is itself refusal, so that row's Truthful rate is strictly the successful-refusal rate and is not fully comparable with the main set's "answer correct plus refuse correctly" definition.

The two components on ScienceQA are more informative. SFT is the most conservative, with the lowest \(\rho_{IK}\) (24.0, down 6.3 from Prompting's 30.3), DPO is the only method that also raises scientific knowledge mastery (33.9, +3.6), and CPO and ORPO lead on Truthful (62.6 / 63.8). What alignment learned is therefore not a memorized list of InfoSeek entities but a domain-agnostic mastered-or-not criterion.

The internal uncertainty probe reads the model state directly: the model is forced to generate the refusal template "I don't know." on both the Known and Unknown subsets, and the average Logprob and PPL of that template are measured (Post-Align means after ORPO alignment).

Subset Answer Logprob ↑ (Pre / Post) Refusal PPL ↓ (Pre / Post)
Known questions -0.88 / -0.76 2.42 / 2.14
Unknown questions -0.80 / -0.63 2.22 / 1.89

Key Findings

  • Vanilla VLM overconfidence is quantifiable. The 7B Base model reaches \(\rho_{IK}=45.8\) but \(\rho_{IDK}=1.8\) in the zero-shot setting: it almost never refuses. Even with few-shot guidance, \(\rho_{IDK}\) only climbs to 12.4. This directly supports the claim that alignment paradigms rank helpfulness above truthfulness.
  • SFT's gain and its cost are both extreme. For the 7B model, zero-shot \(\rho_{IDK}\) jumps from 1.8 to 32.7 (+30.9), while \(\rho_{IK}\) simultaneously falls from 45.8 to 34.4 (−11.4), and by 17.9 in the few-shot setting. SFT learns a blanket conservative prior rather than mastery discrimination — it treats refusal as a universal action.
  • Preference optimization pulls the trade-off back to the middle. ORPO achieves the best 7B zero-shot result at 60.0 (12.4 above Base's 47.6) and the overall best score of 76.6 in the 13B few-shot setting (7.2 above Base's 69.4); CPO reaches 67.3 (+9.4) for the 7B few-shot model, which is precisely the headline result of 57.9 → 67.3 quoted in the abstract (both numbers are the 7B few-shot setting).
  • Larger models delineate the boundary more precisely. The 13B model beats the 7B one under nearly every method; more tellingly, Base 13B already shows \(\rho_{IDK}=19.8\) in the few-shot setting versus only 12.4 for 7B. Greater parametric capacity provides a more stable internal knowledge representation, which the alignment objective can act on more reliably.
  • Honesty transfers across domains. Trained only on V-Idk, ORPO lifts the Truthful rate on ScienceQA from 56.1 to 63.8 (+7.7), and the behavioral structure matches the main set exactly (SFT most conservative, DPO best at preserving knowledge, CPO/ORPO best balanced). This indicates domain-agnostic honesty rather than overfitting to specific entities.
  • Visual degradation and knowledge gaps are won by different methods. On VizWiz-Unans SFT reaches 95.6, an 11.3-point gain over Prompting's 84.3: when the visual signal collapses, a single conservative refusal prior suffices. On PMC-VQA, where the gap is professional knowledge, CPO overtakes SFT (80.0 vs 77.4). Preference learning is what lets the model recognize "I am not expert in medicine" instead of applying a fixed template.
  • Internal probability evidence rules out template memorization. After ORPO alignment the model is more confident when refusing truly unknown facts (Logprob −0.63) than when wrongly refusing known ones (−0.76), and the refusal template's PPL is lower on the Unknown side (1.89 vs 2.14). Meanwhile, on the Known side the refusal template's Logprob rises from −0.88 to −0.76 and its PPL drops from 2.42 to 2.14 — the alignment tax showing up in internal probabilities: the model's tendency to refuse known facts also strengthens, but never beyond the level of genuinely justified refusal.

Highlights & Insights

  • Turning "honesty" from a virtue into a metric. Crossing the probe label with output semantic correctness yields four quadrants that split the honesty/helpfulness trade-off into two simultaneously observable numbers (\(\rho_{IK}\) and \(\rho_{IDK}\)) with their ceilings explicitly marked. The key payoff is that it blocks the cheat of spamming refusals: judging by Truthful alone, SFT's 67.1 in the 7B zero-shot setting looks better than CPO's 58.3.
  • Defining "what the model does not know" from its own sampling. The consistency probe measures behavioral stability rather than a self-report or the correctness of a single response, so it is inherently model-specific and format-agnostic. The idea ports directly to any setting that needs a model-specific mastery signal — it is essentially a capability-boundary annotator.
  • Negative examples are not sampled at random but chosen by confidence. Setting the highest-confidence hallucination as the rejected response prioritizes penalizing the error class that is most dangerous in deployment. The paper mentions this in passing, yet it is the implicit precondition for the Unknown-side construction to actually suppress hallucination.
  • A transferable recipe. "Multi-sample probe to label the boundary, reverse the pairs by the label, then calibrate with preference optimization" is not tied to VLMs: tool use in text LLMs (should retrieval be called), RAG systems (is the retrieved evidence sufficient), and medical QA (should the case be referred) are the same shape of problem, provided a model-specific mastery signal can be estimated by repeated sampling.

Limitations & Future Work

  • Limitations the authors acknowledge: the training set is somewhat small, and future work could enrich the training data or combine it with more efficient tuning methods (e.g. modality linear representation steering, training-free data selection).
  • Probing and dataset construction are not cheap. Each sample costs 10 generations plus per-response LLM judging, so 20,000 training samples amount to roughly 200,000 inferences, and the whole pipeline must finish before training can even start; the cost scales linearly with a stronger model or a larger dataset.
  • No sensitivity analysis for \(\tau=0.7\). The threshold is fixed at 0.7 and no curve of \(\rho_{IK}\) / \(\rho_{IDK}\) versus \(\tau\) is reported, yet \(\tau\) directly determines the positive/negative ratio of the training set (which lands at 60/40) and is the hyper-parameter most in need of ablation.
  • The normalization convention makes cross-table comparison delicate. All three metrics are normalized by \(|\mathcal{D}|\) and therefore depend strongly on the Known/Unknown split of the test set (60:40 on V-Idk, 50:50 on ScienceQA), while on VizWiz-Unans Truthful reduces to the successful-refusal rate alone. Absolute Truthful values are not directly comparable across datasets or models.
  • Narrow validation scope. Only LLaVA-1.5-7B/13B are tested, and both with LoRA. Whether the probe measures boundaries as accurately on stronger models such as Qwen-VL or InternVL, and whether ORPO still beats SFT there, remains untested.
  • The framework does not dominate under pure perceptual degradation. Judging by the labels in Fig. 3, ORPO reaches only 82.8 on VizWiz-Unans, below the Prompting baseline of 84.3 (⚠️ values transcribed from the bar chart; defer to the original). This suggests that when the reason to refuse shifts from a knowledge gap to insufficient visual evidence, the refusal prior produced by preference optimization is not conservative enough, and the framework may not beat plain SFT.
  • Concrete improvements: make \(\tau\) adaptive to question difficulty or domain instead of a single global threshold; use active learning to run the 10-sample probe only on the samples whose probe outcome is most uncertain, to cut annotation cost; and fuse perceptual uncertainty with epistemic uncertainty into a joint refusal criterion so the model can phrase "I cannot see it" and "I do not know" differently.
  • vs the Idk framework (Cheng et al., ICML 2024): They pioneered "build a model-specific dataset, then SFT plus preference optimization" to teach text LLMs to say "I don't know", and are this paper's direct technical ancestor. This work moves that pipeline to multimodal and rebuilds the data-construction stage for a problem they did not face: in the text domain the criterion for "not knowing" can be reused as-is, whereas in multimodal the refusal decision depends jointly on visual perception (what is in the image) and knowledge mastery (what I know about it), so perceptual noise must first be filtered out at the sample level before probe labels become trustworthy.
  • vs InBoL (Wang et al., 2024): Also about refusal in MLLMs, but InBoL targets generic knowledge boundaries for questions that are unanswerable by nature, whereas this paper targets the knowledge gap of a question that is answerable in principle but happens to lie beyond a given model's knowledge — a gap that shows up in long-tail and professional domains and requires model-specific boundary calibration. The problem settings differ, so they are not direct competitors.
  • vs VLM preference-optimization work such as V-DPO, CLIP-DPO, and RLHF-V: Their preference pairs are all built around answer quality (object-level hallucination, cross-modal alignment) with the goal of more accurate answers, whereas this paper's pairs are built around the answer-refusal boundary. The price is that this work does not directly improve \(\rho_{IK}\) — apart from DPO's +6.1 in the 7B zero-shot setting, the other methods' \(\rho_{IK}\) is flat or lower; the payoff is converting fabrication into honest refusal, which matters more than raw accuracy in high-stakes domains such as medicine and law.
  • vs perceptual-uncertainty and calibration-based refusal work: Those methods handle blurred, occluded, or corrupted inputs, where input quality or output confidence alone can decide whether to refuse; this paper handles clear images with missing knowledge, which requires an external probe to delineate the boundary. Their complementarity shows up directly in the OOD experiments: SFT is strongest on VizWiz-Unans (95.6) while CPO is strongest on PMC-VQA (80.0) — no single method covers both kinds of refusal.

Rating

  • Novelty: ⭐⭐⭐⭐ Moving the Idk paradigm from text to multimodal is not a trivial port: the three-stage dataset construction (especially the consistency probe) and the four-quadrant metrics are redesigned around the perceptual-vs-epistemic distinction and constitute a substantive contribution; the training side is a combination of off-the-shelf SFT + DPO/ORPO, so the methodological novelty is comparatively limited.
  • Experimental Thoroughness: ⭐⭐⭐ Four threads are covered (main set, cross-dataset, OOD, and internal probability probing), and the internal probe is especially convincing; however, only LLaVA-1.5-7B/13B are tested, neither \(\tau\) nor the training-set size is ablated, and the OOD conclusions have to be read off a bar chart.
  • Writing Quality: ⭐⭐⭐⭐ The four-quadrant framework and the three-stage pipeline are explained clearly, metric definitions are complete and even mark their own ceilings; but the formulas are corrupted in the PDF layout (visible in the cached text) and the OOD section offers only a bar chart, forcing readers to read numbers off a figure.
  • Value: ⭐⭐⭐⭐ It provides a reusable "measure the boundary, then write it back into the parameters" recipe and backs it with internal probability evidence that the model is truly calibrating its cognition rather than reciting a template, which is directly relevant to multimodal assistants in high-stakes settings.