Cue Visibility Gap in Detection & Fairness

• Cue Visibility Gap is defined as the mismatch between cues present in inputs and those actually used by human and machine detectors, impacting performance in degraded conditions.
• It spans multiple domains including adverse-visibility object detection, dissociation paradigms, and LLM fairness evaluation, illustrating the importance of aligning signal availability with system utilization.
• Research employs structured representations, perturbation methods, and integrated pipelines to diagnose and bridge the gap, leading to measurable improvements in system robustness and fairness.

Cue Visibility Gap denotes, across several research literatures, a mismatch between cues that are available in an input or task and cues that a detector, reasoning model, awareness measure, or feedback system actually exploits. In adverse-visibility object detection, it is “the divergence between the visual cues humans can exploit in poor‐visibility scenes (e.g. faint shape outlines, motion contrast, contextual priors) and the cues that standard object detectors can extract from the same degraded image.” In the dissociation paradigm, it is “the phenomenon that different direct measures—despite targeting the same critical cue—can behave in qualitatively different ways under an experimental manipulation.” In LLM fairness evaluation, it quantifies “how much a model’s ‘fair’ behavior collapses when you hide the label” (Kumar, 2024, Schmidt et al., 2022, Shafiei et al., 30 Jun 2026). Related multimodal reasoning work diagnoses the same failure mode when MLLMs “under-see” dense tool outputs or when single-image 3D systems exploit only a narrow subset of available monocular cues (Janjua et al., 14 Apr 2026, Li et al., 27 Nov 2025).

1. Definitions and conceptual scope

The term has acquired domain-specific meanings, but each centers on cue accessibility versus cue utilization. In “Perceptual Piercing,” the Cue Visibility Gap refers to the divergence between the cues humans can exploit in fog, haze, or smoke and the cues conventional detectors extract from the same degraded image; the significance is that low-level features such as edges and color gradients become attenuated, while humans selectively attend to residual contrast or use top-down expectations (Kumar, 2024).

In Schmidt and Biafora’s cue-set account of the dissociation paradigm, the gap appears when different direct measures, all intended to assess awareness of the same critical cue, diverge because they rely on different criterion contents. The critical comparison is not awareness in the abstract, but awareness of the feature-specific perceptual evidence that drives the indirect effect (Schmidt et al., 2022).

In recent LLM safety work, the gap is operationalized as a robustness failure: models appear fair when demographic identity is stated as an explicit label, yet become measurably less fair when the same identity must be inferred. This is termed “performative compliance,” and the Cue Visibility Gap is proposed as a model-agnostic metric for separating genuine from performative moral safety (Shafiei et al., 30 Jun 2026).

A related usage appears in multimodal reasoning and perception-tool interfaces. “Don’t Show Pixels, Show Cues” argues that raw depth maps, flow fields, and correspondences are often misaligned with the language-native reasoning substrate of LLMs, so the bottleneck is not more tool calls or larger MLLMs but how tool outputs are represented. Cue3D, by contrast, treats the gap as a mismatch between cues present in a single image and cues modern single-image-to-3D systems actually exploit (Janjua et al., 14 Apr 2026, Li et al., 27 Nov 2025). This suggests a common family resemblance: cue visibility is not mere signal presence, but usable alignment between evidence and the system that must act on it.

2. Formalizations

One influential formalization comes from Perception Programs, denoted P$^2$. Let the pixel domain be

$\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$

A finite set of primitives $\mathcal{P}$ is defined, and each primitive $p\in\mathcal{P}$ has spatial support $S_p\subseteq\Omega$ and normalized center coordinate $c_p\in\{0,\ldots,1000\}^2$. For each $p$, the system emits a structured item

$I_p=(p,c_p,r_p,b_p),$

where $p$ is the primitive ID, $c_p$ the normalized location, $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$0 the “reading” extracted from tool data on $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$1, and $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$2 an optional label. Sparse relations may also be defined as

$\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$3

The collection $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$4 and $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$5 is serialized in a YAML-like block so that an MLLM can parse compact, structured, language-native cues rather than dense numeric maps (Janjua et al., 14 Apr 2026).

In the dissociation literature, Schmidt and Biafora write a direct measure as a function of its criterion content,

$\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$6

where $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$7 is the set of cues actually used in measure $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$8. Validity is defined by inclusion of the critical cue: $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$9 is valid for awareness if $\mathcal{P}$0. Exhaustive validity requires $\mathcal{P}$1, exclusiveness for the critical cue requires $\mathcal{P}$2, and exclusiveness for relevant cues requires $\mathcal{P}$3. A measure is a monotonic integrator if strengthening any cue never decreases the measure, and an exhaustive integrator if that inequality is strict (Schmidt et al., 2022).

In moral-safety benchmarking, the Cue Visibility Gap is defined directly over model decisions. For model $\mathcal{P}$4 and demographic group $\mathcal{P}$5, the neutral baseline is

$\mathcal{P}$6

the label-exposed decision is

$\mathcal{P}$7

and the label-hidden decision is

$\mathcal{P}$8

From Favor and Against events relative to the gold “benefit” annotation $\mathcal{P}$9, one computes

$p\in\mathcal{P}$0

and then

$p\in\mathcal{P}$1

If a model were genuinely label-invariant, the expected outcome would be $p\in\mathcal{P}$2 despite correct recovery of the hidden cue (Shafiei et al., 30 Jun 2026).

Cue3D introduces a perturbation-based formulation. For each monocular cue $p\in\mathcal{P}$3, a perturbation operator $p\in\mathcal{P}$4 maps an image $p\in\mathcal{P}$5 to $p\in\mathcal{P}$6 while ablating cue $p\in\mathcal{P}$7 as selectively as possible. A 3D network $p\in\mathcal{P}$8 then produces $p\in\mathcal{P}$9, and cue sensitivity is measured by metric degradation such as $S_p\subseteq\Omega$0 or $S_p\subseteq\Omega$1. The paper further defines a cue-saliency vector

$S_p\subseteq\Omega$2

and a gap measure

$S_p\subseteq\Omega$3

relative to a synthetic human-cue profile $S_p\subseteq\Omega$4 (Li et al., 27 Nov 2025).

3. Computational perception and multimodal reasoning

In multimodal tool use, the cue-visibility problem is framed as a representation bottleneck. Perception Programs rewrite dense tool outputs into compact summaries of what is present, where it is, and how parts relate. Across six perception-centric tasks from BLINK, P$S_p\subseteq\Omega$5 consistently outperforms both “Standard” and “Raw Tool” inputs. With GPT-5 Mini as the base model, P$S_p\subseteq\Omega$6 raises accuracy from $S_p\subseteq\Omega$7 to $S_p\subseteq\Omega$8 on multi-view reasoning, from $S_p\subseteq\Omega$9 to $c_p\in\{0,\ldots,1000\}^2$0 on relative depth, and yields an overall average gain of $c_p\in\{0,\ldots,1000\}^2$1. Similar large gains of $c_p\in\{0,\ldots,1000\}^2$2–$c_p\in\{0,\ldots,1000\}^2$3 percentage points absolute are reported on Gemini 2.5 Pro, Qwen3VL-4B, and InternVL3.5-2B/4B, setting new state-of-the-art across all six tasks without model fine-tuning. The paper also reports that prompting GPT-5 to reconstruct redacted P$c_p\in\{0,\ldots,1000\}^2$4 read-outs from raw tool outputs causes Kendall’s $c_p\in\{0,\ldots,1000\}^2$5 to decay toward $c_p\in\{0,\ldots,1000\}^2$6 as grid resolution increases, and that these noisy reconstructions remain at approximately $c_p\in\{0,\ldots,1000\}^2$7 accuracy independent of grid size, whereas correct P$c_p\in\{0,\ldots,1000\}^2$8 with finer patching pushes accuracy to approximately $c_p\in\{0,\ldots,1000\}^2$9 (Janjua et al., 14 Apr 2026).

GThinker addresses an adjacent failure mode in end-to-end multimodal reasoning. Its Cue-Rethinking pattern first performs free-form reasoning with explicit visual-cue tags, then re-examines those tagged cues, checks for inconsistencies between image evidence and inference, and revises both cue descriptions and downstream reasoning before finalizing the answer. The training pipeline combines Pattern-Guided Cold Start on $p$0 examples with Incentive-RL on $p$1 curated reinforcement-learning samples, using DAPO. On M$p$2CoT, GThinker-7B reaches $p$3 overall accuracy, surpassing O4-mini at $p$4; it also reports $p$5 on MMStar, $p$6 on RealWorld QA, $p$7 on MMMU-Pro, $p$8 on MathVista, and $p$9 on MathVision, with all margins stated as significant at $I_p=(p,c_p,r_p,b_p),$0 under paired resampling tests (Zhan et al., 1 Jun 2025).

Cue3D turns the issue into a diagnostic benchmark for single-image 3D generation. It evaluates seven methods spanning regression-based, multi-view, and native 3D generative paradigms, and perturbs shading, texture, silhouette, perspective, edges, and local continuity. The reported pattern is sharply asymmetric: semantic texture swap barely affects top methods, with $I_p=(p,c_p,r_p,b_p),$1; removing shading causes $I_p=(p,c_p,r_p,b_p),$2–$I_p=(p,c_p,r_p,b_p),$3 for SF3D and $I_p=(p,c_p,r_p,b_p),$4–$I_p=(p,c_p,r_p,b_p),$5 for InstantMesh and Trellis; silhouette perturbations can induce $I_p=(p,c_p,r_p,b_p),$6–$I_p=(p,c_p,r_p,b_p),$7; and $I_p=(p,c_p,r_p,b_p),$8 is reported as large, in the range $I_p=(p,c_p,r_p,b_p),$9–$p$0, driven by near-zero weight on texture and overweighting of silhouette (Li et al., 27 Nov 2025). A plausible implication is that cue visibility in modern generative vision models is often dominated by architectural priors and training distributions rather than by the full cue ecology available in the input.

4. Poor-visibility detection and attention feedback

Perceptual Piercing addresses the gap in degraded-image detection by embedding human-vision principles into a multi-tiered pipeline. The image formation model is the classic single-image scattering model,

$p$1

with transmission

$p$2

and dehazing estimate

$p$3

The proposed framework consists of initial quick detection, region-specific dehazing, and in-depth detection. The preliminary detector is YOLOv5s or YOLOv8n; RoIs are boxes with objectness score $p$4, and the cue-driven preprocessing extracts RoIs where objectness $p$5 and generates an attention mask $p$6. In AOD-NetX, five $p$7 convolution layers with ReLU produce $p$8, the masked estimate is $p$9, and the final dehazed image is $c_p$0. The final detector is YOLOv5x or YOLOv8x with early fusion and squeeze-and-excitation attention. On Foggy Cityscapes, the full YOLOv5s + AOD-NetX + YOLOv5x model reaches foggy mAP $c_p$1, versus $c_p$2 for AOD-Net + YOLOv5x and $c_p$3 for YOLOv5x alone; removing selective attention drops mAP to $c_p$4, removing adaptability scaling to $c_p$5, and replacing RoI-based dehazing with uniform dehazing to $c_p$6 (Kumar, 2024).

In collaborative XR, the “Attention-Aware Pipeline” formulates an analogous visibility problem for gaze. Capture is $c_p$7, Record is $c_p$8, and Revisualize is $c_p$9, with the crucial feedback relation that $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$00 influences the next gaze sample. Three tensions are emphasized: “Mirror also Steers,” “Attention Paradox,” and the predicted “Tunneling-Target Paradox.” A formative eye-tracking study instrumented a four-piece band with dynamic AOIs derived from YOLOv8 tracking plus manual refinement, and measured dwell time, fixation duration, saccade amplitude, gaze spatial dispersion, transition entropy, blink rate, and head movement. The leader P03 showed highest dispersion at $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$01 px, largest saccades at $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$02, blink rate $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$03/min, and head movement $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$04/s; the bassist P04 exhibited $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$05 dwell on the leader, dispersion $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$06 px, saccades $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$07, blink rate $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$08/min, and $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$09 bidirectional transitions in a tight leader–bassist loop (Srinivasan et al., 2 Jun 2026). Here the visibility gap is not missing sensory input but the invisibility of attentional coordination cues, and the pipeline makes explicit that any attempt to reveal attention also perturbs it.

5. Cue sets, dissociation, and awareness theory

The awareness-theoretic treatment is the most explicit account of cue-specific validity. In the dissociation paradigm, an indirect measure $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$10 testifies to processing of a stimulus feature, while direct measures $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$11 are intended to gauge awareness of the feature driving that indirect effect. Schmidt and Biafora distinguish the critical feature $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$12 from the critical cue $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$13, the latter being the internal evidence to which any valid direct measure must be sensitive (Schmidt et al., 2022).

Criterion content generalizes Kahneman’s notion of criterion content to the set of cues actually used in the direct task. These may include auxiliary perceptual cues, sensorimotor cues, decisional cues, fringe cues, and strategic cues. Because objective and subjective measures may integrate overlapping but non-identical cue sets, one measure generally cannot replace another without sacrificing information. This is the theoretical basis for Direct–Indirect mismatch and for dissociations among awareness measures (Schmidt et al., 2022).

The paper’s three propositions make the implications explicit. Proposition 1 states that if two measures are both monotonic integrators and a double dissociation is observed, then they cannot both be monotonic functions of the same one-dimensional evidence source. Proposition 2 states that any theory explaining two measures as monotonically increasing functions of a single process is falsified by a double dissociation. Proposition 3 generalizes this to a gradient of direct measures: if at least two of them double dissociate, no one-dimensional theory can account for all the data (Schmidt et al., 2022). The cue visibility gap, in this framework, is therefore not a defect of instrumentation alone; it is evidence that awareness may be multidimensional and cue-heterogeneous.

6. Moral safety and performative compliance

The moral-safety literature converts cue visibility into a benchmarkable robustness problem. The experimental design holds the moral dilemma $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$14 and the target individual $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$15 fixed, and varies only how demographic identity is conveyed: Neutral contains no demographic information, Direct states it explicitly, and Puzzled encodes the same assignment as the unique solution to a short logic puzzle. Because puzzle solutions are verifiable, analyses restrict to cases where the model correctly recovers gender and race, ruling out information loss as a confound (Shafiei et al., 30 Jun 2026).

The experimental setup includes $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$16 LLMs, $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$17 genders $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$18 $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$19 races, approximately $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$20 Direct and $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$21 Puzzled dilemmas per model with $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$22 probes each, for approximately $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$23 model–item interactions, all at temperature $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$24. The key macro-averaged result is one-sided: Against rate increases by $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$25 percentage points from Direct to Puzzled-hard, whereas Favor rate increases only $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$26 points. In Direct, $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$27 is slightly positive or near zero for all genders and races; in Puzzled-hard, $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$28 turns negative for every gender and race, driven by the jump in Against. Hispanic individuals, women, and Muslim individuals show especially consistent shifts under sign tests across the $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$29 main models (Shafiei et al., 30 Jun 2026).

The paper further shows that cue visibility changes model rankings. Open-weight models such as Qwen3 8 B and Ministral 8 B exhibit the largest gaps, reported as $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$30–$\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$31 percentage points, whereas frontier aligned models such as Claude 4.6, Gemini 3, Llama 70 B, and GPT-OSS 20 B have smaller or even negative gaps, approximately $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$32 to $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$33 points (Shafiei et al., 30 Jun 2026). The central claim is that explicit-label fairness evaluations measure surface compliance rather than moral robustness. In this setting, the Cue Visibility Gap is a stress test for whether fairness survives the weakening of evaluation-like cues.

7. Terminological boundaries and research directions

The term should be distinguished from unrelated uses of “visibility” in photonics. Gavenda et al. analyze a visibility bound caused by a distinguishable noise particle, where interference visibility drops to $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$34 when the noise photon is principally completely distinguishable and rises to unity for complete indistinguishability; the quantity is interference contrast, not cue utilization (Gavenda et al., 2011). In one-dimensional $\Omega=\{0,\ldots,W-1\}\times\{0,\ldots,H-1\}.$35-symmetric photonic crystals, “unidirectional weak visibility” denotes a band-gap regime with essentially zero transmittance, very large reflectance from one side, and very small reflectance from the other, again a scattering phenomenon rather than a cue-theoretic one (Wang, 2020).

Within the cue-visibility literature proper, the open directions are highly specific. In adverse-weather detection they include haze-index gating to bypass dehazing when unnecessary, joint end-to-end fine-tuning of dehazing and detector, extension to rain, snow, and nighttime glare, and exploration of learned top-down priors via large vision-LLMs (Kumar, 2024). In collaborative XR they include explicit modeling of social targets, anticipation of the feedback loop, choice of additive versus subtractive overlays by scene density, detection of both attentional peaks and gaps, calibration of timing and transparency, role-aware aggregation, and evaluation in ecological settings (Srinivasan et al., 2 Jun 2026). In fairness evaluation, the proposed direction is to add cue variation to existing benchmarks so that deployment decisions are not grounded in Direct-only scores (Shafiei et al., 30 Jun 2026). Taken together, these programs suggest that closing a cue visibility gap requires not merely more signal, but a disciplined account of which cues are critical, how they are represented, and whether the receiving system can in fact use them.