Paper deep dive
DepthEvidence: Unifying Metric Depth Prediction and Geometric Reasoning in Multimodal Language Models
Jiangning Wei, Yuan Yao, Miaomiao Cui, Mingsheng Li, Humen Zhong, Shuai Bai, Zhibo Yang
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 90%
Last extracted: 9/28/2026, 4:14:04 AM
Summary
The paper introduces DepthEvidence, a 4-billion parameter multimodal language model that unifies metric depth prediction with geometric reasoning. It employs a camera-conditioned decoder for dense metric depth estimation and a dense-to-language interface that converts these predictions into object-aligned continuous geometry tokens. This approach allows the model to use its own metric predictions as evidence for answering complex spatial and numerical questions, evaluated via the new Depth-VQA benchmark.
Entities (11)
Relation Signals (9)
DepthEvidence → evaluatedon → Depth-VQA
confidence 95% · We introduce a Depth-VQA benchmark... DepthEvidence achieves the highest average dense δ1 among evaluated methods.
DepthEvidence → uses → Dense-to-Language Interface
confidence 95% · A dense-to-language interface converts predicted depths and decoder features into object-aligned continuous geometry tokens anchored to object identifiers.
DepthEvidence → uses → Camera-Conditioned Decoder
confidence 95% · A camera-conditioned decoder predicts full-resolution metric depth using multi-scale visual features and high-resolution RGB refinement.
DepthEvidence → buildsupon → Qwen3.5-4b
confidence 90% · We use Qwen3.5-4B (Qwen Team, 2026a)...
Depth-VQA → contains → Level-2
confidence 90% · Level-2: Compositional Depth Reasoning. Depth-VQA contains 1,503 questions in total... 370 Level 2 Relative questions, and 376 Level 2 Metric questions.
Depth-VQA → contains → Level-1
confidence 90% · Level-1: Instance-Level Metric Depth Estimation. Level 1 contains 757 grounded instances...
DepthEvidence → performsbeston → SUN RGB-D
confidence 80% · Across nine datasets, DepthEvidence achieves the highest average dense δ1... SUN RGB-D... 91.9
DepthEvidence → →
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Spatial reasoning with metric constraints requires linking objects to geometric measurements and preserving their numerical content during language reasoning. We present DepthEvidence, a 4B model that uses its own dense metric predictions as object-grounded evidence for language generation. A camera-conditioned decoder predicts full-resolution metric depth using multi-scale visual features and high-resolution RGB refinement. A dense-to-language interface converts predicted depths and decoder features into object-aligned continuous geometry tokens anchored to object identifiers. Geometric supervision encourages metric information to remain recoverable before and after language-context interaction, while instruction tuning supports object measurement and compositional reasoning. We introduce a Depth-VQA benchmark evaluating object-depth queries, relative comparisons, and decisions combining spatial and numerical constraints. Across nine datasets, DepthEvidence achieves the highest average dense $\delta_1$ among evaluated methods, competitive with specialized estimators. It also leads the evaluated methods in instance-level metric depth estimation and overall accuracy on both relative and metric reasoning tracks, while broadly preserving general VQA performance and improving spatial understanding relative to the base model.
Tags
Links
- Source: https://arxiv.org/abs/2609.31103v1
- Canonical: https://arxiv.org/abs/2609.31103v1
Trouble viewing inline? Open PDF directly →
Full Text
89,838 characters extracted from source content.
Expand or collapse full text
DEPTHEVIDENCE: UNIFYING METRIC DEPTH PREDIC- TION AND GEOMETRIC REASONING IN MULTIMODAL LANGUAGE MODELS Jiangning Wei 1,2 Yuan Yao 1 Miaomiao Cui 1 Mingsheng Li 1 Humen Zhong 1 Shuai Bai 1 Zhibo Yang 1 1 Alibaba Token Hub, Alibaba Group 2 Beijing University of Posts and Telecommunications ABSTRACT Spatial reasoning with metric constraints requires linking objects to geometric measurements and preserving their numerical content during language reasoning. We present DepthEvidence, a 4B model that uses its own dense metric predictions as object-grounded evidence for language generation. A camera-conditioned de- coder predicts full-resolution metric depth using multi-scale visual features and high-resolution RGB refinement. A dense-to-language interface converts predicted depths and decoder features into object-aligned continuous geometry tokens an- chored to object identifiers. Geometric supervision encourages metric information to remain recoverable before and after language-context interaction, while in- struction tuning supports object measurement and compositional reasoning. We introduce a Depth-VQA benchmark evaluating object-depth queries, relative com- parisons, and decisions combining spatial and numerical constraints. Across nine datasets, DepthEvidence achieves the highest average denseδ 1 among evaluated methods, competitive with specialized estimators. It also leads the evaluated meth- ods in instance-level metric depth estimation and overall accuracy on both relative and metric reasoning tracks, while broadly preserving general VQA performance and improving spatial understanding relative to the base model. 1INTRODUCTION Answering spatial questions with metric constraints requires connecting object semantics to geometric measurements. Knowing which object is closer establishes a relative relation, but selecting objects within a depth interval requires absolute scale and the correct association between each object and its measured depth. Consider the question in Figure 1(a): which referenced objects to the right of a wooden cabinet have an average camera depth between 1.30 m and 1.33 m? The answer must satisfy both the spatial filter and the numerical interval. Dense metric estimators provide pixel-level measurements (Yin et al., 2023; Piccinelli et al., 2024), while spatially aware multimodal large language models (MLLMs) connect objects with language (Chen et al., 2024a; Cheng et al., 2024). This question requires bringing those capabilities together at the level of individual objects. Recent depth-aware VLMs have substantially advanced geometric prediction. DepthLM (Cai et al., 2026) demonstrates accurate pixel-level metric estimation through language modeling, using visual prompting and intrinsic-conditioned augmentation without a dedicated depth head. DepthVLM (Yu et al., 2026) extends this capability to full-resolution dense prediction through a lightweight head and joint depth–text training, while retaining general multimodal capability and improving spatial reasoning. G 2 VLM also couples geometric reconstruction with spatial reasoning (Hu et al., 2026). Building on these advances, we ask: how can a VLM use its own dense metric predictions as object- grounded evidence for language reasoning? We focus on explicitly linking the predicted metric field to the referenced objects and preserving its numerical content as the language model processes the question. Using predicted geometry as language evidence requires both object alignment and metric information preservation. Measurements must come from the appropriate image regions and remain associated 1 arXiv:2609.31103v1 [cs.CV] 25 Sep 2026 (a) Metric-spatial reasoning Select objects by location and metric depth Which objects right of the cabinet have mean camera depth in [1.30, 1.33] m? 8 Rug 4 Cabinet 6 Picture 7 Stairs 1 Chaise 3 White pillow 2 Grey pillow 5 Picture GT mean depth: chaise 1.30 m; picture 1.31 m. DepthEvidence · correct White chaise (1) + framed picture (5) Without interface · incorrect Also selects pillows outside the range: grey (1.35 m) and white (1.48 m). (b) Dense Metric Depth δ₁ (%) ↑ · 4 datasets + 9-dataset average 9-set Avg. SUN RGB-D Taskonomy Waymo Argoverse2 0.025.050.075.0100.0 77.0 57.1 87.9 87.6 59.2 77.3 61.4 91.9 84.8 80.5 79.5 39.7 86.3 85.0 86.2 79.7 84.2 90.7 82.7 79.5 DepthEvidenceUniDepth V2 Depth Anything V3MetricAnything 9-set Avg. is over all nine evaluated datasets. (c) Depth VQA Level-1: δ₁ (%) ↑ · Level-2: accuracy (%) ↑ Level-2 Metric Overall Acc Level-2 Metric Set Decision Level-2 Metric Spatial Group Level-2 Relative Overall Acc Level-2 Relative Ordering Level-1 δ₁ 0.025.050.075.0100.0 46.2 52.1 34.0 29.2 18.4 65.9 45.7 46.8 35.1 48.4 38.3 7.7 76.9 76.6 74.5 58.4 54.1 86.5 85.0 85.1 81.9 62.7 58.7 86.7 DepthEvidenceDepthEvidence w/o interface Qwen3.7-plusDepthVLM-4B Overall Acc: Relative n = 370; Metric n = 234. (d) General VQA Benchmark score (%) ↑ Ego3D OCRBench MMBench MMMU 0.025.050.075.0100.0 56.9 84.2 85.8 68.0 45.6 81.3 83.9 61.1 62.8 85.0 89.4 77.6 68.0 84.6 89.5 78.7 DepthEvidenceQwen3.5-4B Youtu-VL-4BGLM-4.1V-9B Figure 1: Overview of DepthEvidence’s capabilities. (a) A metric-spatial question requires selecting objects by image location and mean camera depth; the dense-to-language interface rejects out- of-range distractors. (b) Dense metric depth results (δ 1 ) on four representative datasets and the nine-dataset average, compared with specialized estimators. (c) Level 1 metric estimation and Level 2 relative and metric reasoning, compared with general-purpose and depth-aware VLMs as well as the no-interface variant. (d) General VQA retention on MMMU, MMBench, and OCRBench, and spatial understanding on Ego3D. All scores are percentages and higher is better. with the correct references as the model processes multiple objects. Their numerical content must also survive projection into language tokens and interaction with the question. Answer supervision targets the final response without directly constraining the metric content of intermediate geometry tokens. This motivates explicit geometric supervision to keep object-level depth information recoverable before and after interaction with the language context. We present DepthEvidence, a 4B model that turns its own dense predictions into object-aligned evi- dence for language generation. Given an image and camera intrinsics, a camera-conditioned decoder combines multi-scale visual features with high-resolution RGB refinement to produce metric depth and preserve local structure. The dense-to-language interface pools predicted depths and decoder features within each referenced region, then encodes them as object-aligned continuous geometry tokens anchored to object identifiers. Auxiliary depth objectives supervise metric information in these tokens before and after language-context interaction, while a ranking objective supervises object ordering. The tokens enter answer generation through cached continuation, avoiding repeated encoding of the image and question. In the cabinet example, this interface supplies the candidate objects’ geometry for spatial filtering and metric-range selection. To evaluate how geometric prediction translates into language answers, we organize depth understand- ing into dense metric depth prediction (Level 0), instance-level metric depth estimation (Level 1), and compositional depth reasoning (Level 2). Instance-level instruction tuning and the interface support both measurement and reasoning. Our Depth-VQA benchmark evaluates whether models can turn object-level depth measurements into compositional answers, covering relative comparisons and deci- sions that jointly depend on spatial relations and metric gaps, ranges, or thresholds. DepthEvidence achieves the highest average denseδ 1 across nine datasets among the evaluated methods and the highest overall accuracy on both reasoning tracks, while general VQA performance remains broadly stable relative to the base model (Figure 1). Our contributions are: •An object-aligned dense-to-language interface with geometric supervision before and after contextualization, enabling language generation to condition on the model’s own metric predictions. •DepthEvidence, a unified 4B model combining camera-conditioned dense prediction with instance-level depth instruction tuning, together with a Depth-VQA benchmark that distin- guishes object measurement, relative reasoning, and metric-constrained reasoning. • DepthEvidence leads the evaluated methods in nine-dataset average denseδ 1 , instance- level metric estimation, and overall compositional reasoning, while retaining general VQA performance and improving spatial understanding over the base model. 2 Qwen3.5 Language Model Vision Encoder Depth Decoder Dense-to-Language Interface Text tokenVisual token Response Reference objects: -obj_1: black SUV [bbox_2d] -obj_2: gray car [bbox_2d] ...... -obj_10: pole [bbox_2d] Among the objects to the right of gray car obj_4, which have an average camera depth between 6.3m and 7.5m TextTokenizer <answer> obj_1: black SUV </answer> Geometry token Training Stage 1 Training Stage 2+ ℒ !"# ℒ $"%&" ℒ '%& " Figure 2: Overview of DepthEvidence. The dense decoder supplies metric predictions and features to the dense-to-language interface, which converts them into object-aligned geometry tokens for answer generation. This path connects Level-0 dense prediction to Level-1 instance measurement and Level-2 compositional reasoning within a shared VLM. 2RELATED WORK Monocular Metric Depth Estimation. Monocular depth estimation has progressed from multi- scale convolutional models with scale-invariant supervision (Eigen et al., 2014) to transformer decoders using multi-level visual tokens (Ranftl et al., 2021). Depth Anything and Depth Anything V2 scale transferable representations with unlabeled or pseudo-labeled data (Yang et al., 2024a;b), whereas Metric3D and UniDepth address camera variation through canonicalization or camera- conditioned features (Yin et al., 2023; Piccinelli et al., 2024). These methods primarily treat geometry as a dense output; DepthEvidence additionally uses metric predictions as object-grounded evidence for measurement and reasoning. Geometry-Aware Vision–Language Models. Prior work uses geometry either to support spatial reasoning or as an explicit model output. SpatialVLM constructs metric spatial question–answer data through automatic 3D annotation (Chen et al., 2024a), and SpatialRGPT combines 3D scene- graph regions with a depth-aware visual plugin (Cheng et al., 2024); G 2 VLM, Vid-LLM, and SpatialStack more directly couple reconstruction, 3D tokens, or geometry-encoder features with language reasoning (Hu et al., 2026; Chen et al., 2026; Zhang et al., 2026). AURORA and Perceptio emit discrete depth-related tokens before text (Bigverdi et al., 2025; Li et al., 2026), DepthLM predicts metric depth at prompted pixels (Cai et al., 2026), and DepthVLM jointly outputs full-resolution depth and text (Yu et al., 2026). However, explicit or parallel depth prediction does not necessarily condition answers on the predicted metric field. DepthEvidence bridges prediction and reasoning by converting its single-image dense outputs into object-aligned continuous geometry tokens, supervised before and after language-context interaction, for object-level measurement and relative or metric compositional reasoning. 3METHOD 3.1FRAMEWORK OVERVIEW As illustrated in Figure 2, DepthEvidence conditions object-level answers on geometric evidence predicted from the same image. A camera-conditioned dense decoder supplies a metric depth map and spatial features. The dense-to-language interface extracts evidence for objects specified by boxes or referring prompts and binds their geometry tokens to explicit object identifiers. Auxiliary objectives supervise metric-depth prediction from these tokens before and after language-context interaction. Instance-level instruction tuning trains the VLM to use this context for measurement and compositional reasoning. 3 3.2LEVEL-0 DENSE METRIC DEPTH PREDICTION Camera-conditioned decoder. The Level-0 branch follows a coarse-to-fine DPT design (Ranftl et al., 2021), fusing intermediate vision-encoder features with the spatially reshaped final visual hidden states. To recover metric scale, it converts the supplied camera intrinsicsκinto a compact descriptor containing normalized calibration, field-of-view, and aspect-ratio information. An MLP maps this descriptor to FiLM parameters (Perez et al., 2018) that condition every decoder scale. A lightweight high-resolution RGB pathway supplies finer appearance cues during upsampling, and the resulting full-resolution representation predicts both the metric depth map and the dense features used by the dense-to-language interface. This numerical camera encoding is specific to the dense branch; Depth-VQA instead receives the raw intrinsic values in its instruction. Appendix A.1.1 provides the feature dimensions, camera parameterization, and decoder details. High-resolution RGB refinement.Spatially compressed visual tokens can lose thin structures and sharp depth discontinuities. The RGB pathway therefore injects high-resolution features through learned residual gates, preserving the transformer representation as the main path while restoring local boundary cues. The refined spatial features also provide local appearance and boundary cues for object-level pooling in the interface. Architectural details are given in Appendix A.1.2. Training objective. We train the dense branch with L dense = 0.8L abs + 0.2L silog + 0.1L der + 0.03L lap +L ssi .(1) The first two terms supervise metric depth, the derivative and Laplacian terms preserve multi-scale structure, and the edge-guided stable local scale-and-shift invariant termL ssi emphasizes object boundaries. Appendix A.1.3 provides the detailed definitions and implementation. 3.3INSTANCE-LEVEL DEPTH INSTRUCTION TUNING Instance-level measurement and reasoning tasks. We extend dense prediction with two object- grounded instruction levels. Level 1 asks for a specified object’s center, mean, median, nearest, or farthest metric depth. Level 2 combines measurements across objects: relative tasks cover near–far and front–behind comparisons, closest/farthest selection, and subset or full depth ordering; metric tasks cover numerical mean-depth gaps, thresholds, ranges, and margin-based decisions. Compositional questions first resolve image-space references or filter candidates, then apply relative comparisons or metric constraints. These instructions teach object-level measurement and its use in spatially and numerically constrained answers. Unified object grounding. We support two ways to specify an object. A box prompt provides its category and bounding box, with coordinates normalized to a0–1000image coordinate system. A referring prompt instead identifies it through a natural-language description enclosed by object- reference tokens. For multi-object questions, each instance receives a stable identifier such asobj1, and the prompt lists its identifier, category, and grounding cue before the question. This common serialization supports both grounding modes and disambiguates repeated object categories without assuming any interface-specific processing. Structured supervision.Each instruction supplies the raw camera intrinsics as<focal> f x f y c x c y </focal>, independently of the dense decoder’s numerical camera encoding. The VLM reads this block together with the image and question. We train both instruction levels with the next-token cross-entropy lossL ans on the final answer. The answer is a metric depth value for Level 1, or a numerical depth gap, an object or set selection, an ordering, or a discrete comparison or constraint decision for Level 2. Margin-based decisions include an ambiguous answer when the closest object does not satisfy the required depth margin. Appendix A.2 provides the detailed targets, prompt serialization, and output formats. 3.4DENSE-TO-LANGUAGE INTERFACE Object-aligned dense geometry.For both Level 1 and Level 2 instructions, the interface receives a metric depth map b Dand a full-resolution feature mapF d from the dense decoder. For a box prompt, we use the supplied region directly; for a referring prompt, a lightweight grounding head predicts its 4 box. For objecti, poolingF d yields a regional featurer i , while summarizing the predicted depths yields metric statisticss i . Together with the normalized boxb i , these form an object descriptor that associates local geometry with the referenced instance. Appendix A.3.1 gives the pooling statistics and grounding-head implementation. Continuous geometry tokens. A two-layer projectorP g maps each object descriptor toK = 2 continuous tokens, g i,k = P g ([r i ; b i ; s i ]) k + p i + t k , k ∈1, 2,(2) wherep i is a learned object-index embedding andt k distinguishes the two token slots. We organize the resulting tokens as an input-only geometry bank in which each pair follows its textual anchor obji:. Geometry-bank positions are excluded as language-modeling targets, but answer tokens attend to them and the answer loss backpropagates through them to the projector. Appendix A.3.1 details the bank construction, and Appendix A.3.2 discusses the choice of K = 2. Geometric supervision. To encourage metric-depth recoverability, we supervise the geometry representations at both instance levels with L geo = 0.2L raw + 0.2L ctx + 0.1L rank ,(3) The raw-depth and contextual-depth losses train auxiliary heads to predict five ground-truth object- depth statistics (center, mean, median, nearest, and farthest) in log space, from the projected tokens and their contextualized states, respectively. These objectives encourage metric-depth recoverability before and after interaction with the image and question. For multi-object examples,L rank supervises relative ordering using the contextual mean-depth predictions. Appendix A.3.4 specifies the heads, targets, and gradient routing. Single-prefill cached continuation. The geometry bank becomes available after the image and question have been encoded. We retain the prefix cache, append the bank as a causal continuation, and generate the answer from the extended context. Geometry tokens can attend to the cached prefix, and answer tokens can attend to both the prefix and geometry. This execution scheme incorporates predicted evidence without re-encoding the multimodal prefix. Appendix A.3.3 gives implementation details, and Appendix E compares no-interface, two-pass, and single-prefill inference. 4INSTANCE-LEVEL DEPTH-VQA BENCHMARK Depth-VQA evaluates object-level metric estimation (Level 1) and multi-object compositional rea- soning (Level 2), including relative comparisons, ordering, numerical gaps, and spatially conditioned metric decisions. Each question provides an image, camera intrinsics, and object references specified by normalized category–box prompts or natural-language referring prompts. Construction and quality control.We first complete sparse outdoor depth so every source image has a dense depth map. Qwen3.7-plus (Qwen Team, 2026b) and Seed-2.1-Turbo (ByteDance Seed Team, 2026) independently identify major objects and propose their categories and grounding boxes, reducing dependence on the output distribution of a single generator. We pool their candidates, remove boxes covering less than 1% of the image or lying near image boundaries, prompt SAM 3 (Carion et al., 2026) with the remaining categories and boxes, and apply a second quality filter to the resulting masks. Each retained mask and dense depth map yield the Level 1 center, mean, median, nearest, and farthest targets, with the latter two defined by the 5th and 95th depth percentiles. The two generators then produce Level 2 questions and answers from these instance records, with each candidate verified by the model that did not generate it. We manually audit 1% of training samples across all three levels to derive targeted filters, manually check every benchmark example, and limit reuse of the same instance across task types within a video scene. Appendix C.1 provides full details. Level-1: Instance-Level Metric Depth Estimation. Level 1 contains 757 grounded instances from six indoor and outdoor datasets: Waymo (Sun et al., 2020), Argoverse 2 (Wilson et al., 2021), Taskonomy (Zamir et al., 2018), LingBot-Depth (Tan et al., 2026), DDAD (Guizilini et al., 2020), and nuScenes (Caesar et al., 2020). Each query requests the object-center, nearest, farthest, mean, or median valid depth within the referenced mask and expects a scalar answer in meters. We reportδ 1 (higher is better), RMSE, and AbsRel (lower is better). 5 Table 1: Level-2 tracks and subbench goals. Appendix Table 7 provides all question types and reasoning requirements. TrackSubbenchQuestion summary Relative Depth Spatial PairCompare two objects, optionally after resolving a spatial reference. Spatial GroupSpatially filter candidates, then select the nearest or farthest. Depth OrderingSelect an extreme or order a subset or all referenced objects. Metric Depth Metric PairPredict a mean-depth gap or compare it with a threshold. Spatial Metric PairResolve a spatial reference, then predict an absolute mean-depth gap. Spatial Metric GroupSpatially filter candidates, then apply range or bound constraints. Metric Set DecisionTest a safety band or a margin-constrained closest-object decision. Level-2: Compositional Depth Reasoning. Depth-VQA contains 1,503 questions in total: 757 Level 1 questions, 370 Level 2 Relative questions, and 376 Level 2 Metric questions. Level 2 uses grounded instances as reasoning operands, assigning each a stable identifier such asobj1. Table 1 groups seven subbenches into the two tracks; Appendix Table 7 defines all question types. The Relative track uses exact-match accuracy: selections require the target identifier, and orderings require the full identifier sequence. Relative Overall Acc pools the 370 questions from the three reported subbenches. The Metric track uses accuracy for discrete answers andδ 1 for numerical gaps; Metric Overall Acc excludesδ 1 -scored questions. Appendix C provides question counts and scoring details. 5EXPERIMENTAL SETUP 5.1TRAINING PIPELINE Training proceeds in two stages. Stage 1 freezes the vision-language backbone and trains the Level-0 decoder and high-resolution RGB branch for 80K steps. Stage 2 initializes from this checkpoint and jointly trains Level-1/Level-2 Depth-VQA and general VQA for 18K steps, with about 60K dense replay examples preserving Level-0 prediction. The dense-to-language interface is active for both instance levels. Because dense predictions and decoder features are detached before geometry-token construction, QA losses update the LoRA-adapted language modules and interface, while dense replay updates the decoder. We use Qwen3.5-4B (Qwen Team, 2026a); Appendix B provides data composition, preprocessing, optimization, and runtime settings. 5.2EVALUATION SETUP For Level 0, we evaluate metric depth on four outdoor datasets (Argoverse 2 (Wilson et al., 2021), Waymo (Sun et al., 2020), nuScenes (Caesar et al., 2020), and DDAD (Guizilini et al., 2020)) and five indoor datasets (Taskonomy (Zamir et al., 2018), ScanNet++ (Yeshwanth et al., 2023), HyperSim (Roberts et al., 2021), SUN RGB-D (Song et al., 2015), and Booster (Ramirez et al., 2022)). Predictions use ground-truth intrinsics and are scored in metric scale without post-hoc alignment; δ 1 is the primary metric, and the overall score averages the nine dataset-level results. Depth-VQA follows the tasks and metrics defined in Section 4. Level-0 baselines include specialized metric-depth models and depth-aware VLMs; Depth-VQA baselines include general-purpose and depth-aware VLMs. All models use the same prompts and answer protocol. A matched DepthEvi- dence variant disables the dense-to-language interface at both instance levels to isolate geometry-token conditioning. Appendix C provides the complete baselines, sampling, and scoring protocols. 6MAIN RESULTS 6.1LEVEL-0 DENSE METRIC DEPTH PREDICTION Because general-purpose VLMs lack dense metric-depth outputs, Table 2 compares DepthEvidence with depth-aware VLMs and specialized estimators. DepthEvidence achieves the highest nine- dataset averageδ 1 of 0.7972, slightly above UniDepth V2 (0.7945) and 0.1033 above the strongest depth-aware baseline, DepthLM-12B. It exceeds DepthLM on eight datasets and leads on SUN RGB-D (0.8420). Its outdoor and indoor averages of 0.7970 and 0.7974 indicate balanced aggregate performance, although specialized models remain stronger on several individual datasets. Figure 3 complements Table 2 with two outdoor and two indoor examples; additional visualizations for the remaining five datasets are provided in Appendix Figure 5. DepthEvidence preserves fine structures and depth discontinuities more consistently than DepthVLM-4B and Youtu-VL-4B-Instruct while 6 Table 2: Level-0 dense metric depth comparison across outdoor and indoor benchmarks. We report δ 1 ↑and its unweighted average over all nine datasets. The best and second-best results in each column are shown in bold and underlined, respectively. Rankings are determined using the unrounded values. † DepthLM-12B follows its original sparse point-query protocol (Cai et al., 2026), using 8,192 randomly sampled pixels per dataset. Method OutdoorIndoor Avg. Argoverse2 Waymo NuScenes DDAD Taskonomy ScanNet++ HyperSim SUN RGB-D Booster VLMs Trained on Metric Depth Estimation Youtu-VL-4B-Instruct0.62900.51210.64430.29950.75610.60060.42850.35250.3471 0.5077 DepthVLM-4B0.76960.80560.84330.82320.69920.85680.45360.50890.2452 0.6673 DepthLM-12B † 0.77680.77580.83100.66380.82180.79620.43580.55080.5934 0.6939 Specialized Vision Models Depth Anything V2 outdoor0.13040.71220.17010.14460.00080.00020.04150.00000.0000 0.1333 Depth Anything V1 outdoor0.21820.65640.33380.17790.00260.00010.07430.00010.0000 0.1626 Depth Anything V1 indoor0.00180.00190.00090.00150.47060.76040.25040.64270.5225 0.2947 Depth Anything V2 indoor0.13430.01880.17730.28910.23970.23220.86520.16510.6007 0.3025 Metric3D0.78010.82870.76910.81360.94670.07340.47450.05820.0434 0.5320 Metric3D v20.8589 0.87740.85550.87000.93890.88090.73110.69800.1374 0.7609 MetricAnything0.59200.87580.80670.87360.87940.89660.84400.57130.5943 0.7704 Depth Anything V3 metric-large0.80510.84830.79960.78250.91890.89060.58000.61410.7175 0.7730 UniDepth V20.86190.85020.88110.87190.86290.86020.86480.39690.70040.7945 DepthEvidence-4B (Ours)0.79480.82660.81220.75430.90720.81970.82130.84200.5970 0.7972 RGBGTDepthVLM-4BYoutu-VL-4BMetricAnythingUniDepth V2DepthEvidence Argoverse2 Waymo ScanNet++ HyperSim Figure 3: Qualitative dense depth on Argoverse2, Waymo, ScanNet++, and HyperSim. Columns show RGB, ground truth (GT), DepthVLM-4B, Youtu-VL-4B-Instruct, MetricAnything, UniDepth V2, and DepthEvidence. Outdoor GT is densified only for visualization; Table 2 uses the original sparse sensor GT. Each row shares one color scale; Figure 5 shows the other five datasets. retaining the global layout. MetricAnything and UniDepth V2 sometimes produce smoother planes, whereas DepthEvidence remains competitive in boundary localization across both domains. 6.2INSTANCE-LEVEL DEPTH ESTIMATION AND REASONING Table 3: Instance-level depth estimation and reasoning results. The interface is enabled or disabled at both instance levels. Relative Overall Acc is computed over 370 questions, and Metric Overall Acc includes the 234 questions scored by accuracy. Bold and underlined entries denote the best and second-best results in each column. Method Level-1: Instance Metric Depth Level-2: Compositional Depth Reasoning RelativeMetric δ 1 ↑RMSE↓AbsRel↓ Spatial Pair Spatial Group Ordering Overall Acc Metric Pair Acc Metric Pair δ 1 ↑ Spatial Pair δ 1 ↑ Spatial Group Set Decision Overall Acc General-Purpose VLMs Qwen3.5-4B0.0159 206.12219.51850.0930 0.0568 0.01020.04050.63040.25000.22340.3085 0.55320.4701 Gemini-3.1-Pro-Preview0.4108 52.46802.09270.6395 0.5795 0.37240.48380.63040.22920.23400.3085 0.59570.4872 Qwen3.7-plus0.0766 107.25563.04770.5814 0.6136 0.38270.48380.65220.29170.14890.3511 0.46810.4573 Seed-2.1-Turbo0.2048 107.93602.74700.6977 0.6250 0.41330.52970.63040.10420.15960.3617 0.43620.4444 VLMs Trained on Metric Depth Estimation DepthLM-12B0.7781 10.7460 0.18390.0000 0.0000 0.00000.00000.00000.14580.05320.0000 0.00000.0000 Youtu-VL-4B-Instruct0.0264 22.15621.08520.1744 0.2955 0.12240.17570.28260.04170.10640.1915 0.34040.2692 DepthVLM-4B0.6592 11.91640.26130.5000 0.3295 0.18370.29190.58700.08330.09570.3404 0.52130.4615 Agent Pipeline MetricAnything + Qwen3.5-4B 0.7318 11.92240.23630.3372 0.6136 0.41840.44590.58700.29170.23400.4149 0.64890.5427 MetricAnything + Qwen3.7-plus 0.7543 14.25290.24070.3140 0.7273 0.45410.48650.84780.41670.29790.4468 0.65960.6111 DepthEvidence w/o interface0.865311.21590.23510.74420.5227 0.54080.58380.82610.43750.35110.74470.76600.7692 DepthEvidence (Ours)0.8666 8.60620.16290.7907 0.5568 0.58670.62700.91300.58330.40430.8191 0.85110.8504 7 Depth Ordering Question Rank all referenced objects from nearest to farthest. Without interface ✗ ... shopping cart_1 obj_2 → shopping cart_2 obj_3 ... With interface ✓ ... shopping cart_2 obj_3 → shopping cart_1 obj_2 ... Spatial Metric Group Question Which objects right of cabinet obj_4 lie within 1.30-1.33 m? Without interface ✗ white chaise obj_1; grey pillow obj_2; white pillow obj_3 With interface ✓ white chaise obj_1; framed pictures obj_5 Metric Set Decision Question Select the closest only if its margin is at least 0.60 m; otherwise ambiguous. Without interface ✗ dining table obj_1 With interface ✓ ambiguous Figure 4: Representative Level-2 cases corrected by the dense-to-language interface. The three panels test complete depth ordering, spatially filtered metric-range selection, and margin-aware ambiguity decisions, respectively. Each panel contrasts matched variants without and with the interface; green answers match the ground truth. Aggregate results are reported in Table 3. Table 4: General and spatial capability evaluation. Metrics follow their native scales; all are higher-is- better except Ego3D RMSE. Bold marks the best result, and “–” denotes unavailable results. Method STEMGeneral VQA MathVisionMathVistaMMMUMMMU-ProRealWorldQAMMStarMMBenchHallusionBench Youtu-VL-4B-Instruct–76.5061.1043.0074.6071.1083.9059.10 GLM-4.1V-9B-Thinking54.4080.7068.0057.10–72.9085.8063.20 Qwen3.5-4B74.6085.1077.6066.3079.5078.3089.4065.00 DepthEvidence (Ours)74.1084.9078.7164.0578.0678.3089.5268.21 Method Text Recognition and Document UnderstandingSpatial Intelligence AI2DCC-OCROCRBenchCharXiv (RQ)ERQAEmbSpatialBenchEgo3D Acc.Ego3D RMSE↓ Youtu-VL-4B-Instruct85.60–81.3043.80–45.6319.27 GLM-4.1V-9B-Thinking87.90–84.20–45.80–56.9312.79 Qwen3.5-4B89.6076.7085.0070.8054.0081.3062.8213.17 DepthEvidence (Ours)90.6777.3484.6069.0057.2583.1467.996.83 Table 3 evaluates instance-level metric estimation and compositional reasoning. Among the depth- aware baselines, DepthLM-12B (Cai et al., 2026) performs best on Level 1 (δ 1 = 0.7781), whereas DepthVLM-4B (Yu et al., 2026) performs best on both Level 2 tracks (0.2919 Relative and 0.4615 Metric Overall Acc). DepthEvidence leads all evaluated methods on Level 1, withδ 1 = 0.8666, RMSE= 8.6062, and AbsRel= 0.1629. It also achieves 0.6270 Relative and 0.8504 Metric Overall Acc, exceeding Seed-2.1-Turbo and Gemini-3.1-Pro-Preview by 9.73 and 36.32 percentage points, respectively, and DepthVLM-4B by 33.51 and 38.89 points. DepthEvidence further obtains the highest reported numerical-gap δ 1 on Metric Pair (0.5833) and Spatial Metric Pair (0.4043). 6.3EFFECT OF THE DENSE-TO-LANGUAGE INTERFACE Matched ablations in Table 3 isolate the interface effect. On Level 1,δ 1 is nearly unchanged (0.8653→0.8666), while RMSE and AbsRel decrease from 11.2159 to 8.6062 and from 0.2351 to 0.1629, suggesting fewer large metric errors. On Level 2, Overall Acc increases by 4.32 points on Relative and 8.12 points on Metric. The largest categorical gain is on Metric Pair (+8.69 points), while numerical-gapδ 1 rises from 0.4375 to 0.5833 on Metric Pair and from 0.3511 to 0.4043 on Spatial Metric Pair. Although MetricAnything + Qwen3.7-plus leads Relative Spatial Group (0.7273), its Relative and Metric Overall scores (0.4865/0.6111) trail DepthEvidence by 14.05/23.93 points, and both agent pipelines remain below DepthEvidence without the interface (Appendix C.6). Overall, the agent pipelines lag the full DepthEvidence system, indicating that external estimator–VLM coupling is less effective than our integrated design. Figure 4 visualizes representative corrections in depth ordering, spatially filtered metric-range selection, and margin-aware ambiguity decisions. 6.4GENERAL VQA RETENTION AND SPATIAL UNDERSTANDING Table 4 shows that DepthEvidence improves spatial reasoning while largely retaining the general capabilities of its Qwen3.5-4B initialization. Non-spatial scores change modestly: MMMU, M- Bench, HallusionBench, AI2D, and C-OCR improve, while the largest regression is 2.25 points on MMMU-Pro. All four spatial metrics improve: ERQA rises from 54.00 to 57.25, EmbSpatialBench 8 Table 5: Dense-decoder ablation on outdoor and indoor datasets (δ 1 ↑). SSI HR branch OutdoorIndoor Argoverse2 Waymo Taskonomy ScanNet++ –0.73130.75050.78520.7349 ✓–0.75470.77360.81750.7511 –✓0.78240.81290.88040.8081 ✓0.79480.82660.90720.8197 Table 6: Geometry-interface ablation on Level-2 Relative and Metric Overall Acc. Variant L ans L raw L ctx L rank Relative Overall Metric Overall w/o interface–0.58380.7692 Interface ✓–0.58380.7735 ✓–0.59460.8120 ✓–0.61350.8419 ✓0.62700.8504 from 81.30 to 83.14, and Ego3D accuracy from 62.82 to 67.99, while RMSE falls from 13.17 to 6.83 (−48.1%). DepthEvidence also outperforms Youtu-VL-4B-Instruct (Wei et al., 2026) and GLM-4.1V-9B-Thinking (GLM-V Team, 2025) on their commonly reported higher-is-better metrics and Ego3D RMSE. 7ANALYSIS AND DISCUSSION 7.1DENSE DECODER ABLATION Table 5 shows that both the SSI objective and high-resolution branch increaseδ 1 on each evaluated dataset, with their combination obtaining the highest values among the ablated variants. The HR branch has a larger measured effect indoors: used alone, it raises meanδ 1 by 0.0842 on the two indoor datasets versus 0.0568 outdoors. One possible explanation is the sparse evaluation coverage of outdoor datasets: their ground truth contains fewer evaluated pixels near object boundaries, so improvements in boundary prediction contribute less to the measuredδ 1 than under denser indoor annotations. 7.2GEOMETRIC SUPERVISION ABLATION The interface is used at both instance levels; Table 6 focuses on its supervision design using the two Level-2 accuracy aggregates. Training the interface with onlyL ans leaves Relative Overall Acc unchanged at 0.5838 and raises Metric Overall Acc from 0.7692 to 0.7735. AddingL raw improves the two aggregates by 1.08 and 3.85 percentage points, respectively;L ctx adds a further 1.89 and 2.99 points, andL rank adds 1.35 and 0.85 points. The complete objective reaches 0.6270 Relative Overall Acc and 0.8504 Metric Overall Acc, improving over the model without the interface by 4.32 and 8.12 percentage points. These results support the utility of auxiliary geometric supervision, with a larger measured gain on the accuracy-based Metric tasks. A separate no-interface control with direct metric supervision remains clearly below the full interface, indicating that the main performance gain arises from the interface itself rather than numerical supervision alone; Appendix D.1 provides the detailed disentanglement analysis. 7.3LIMITATIONS The current setting assumes access to camera intrinsics and explicit object references. Its geometric evidence is therefore sensitive to errors in calibration, object grounding, and dense-depth prediction, particularly when depth ranges or decision margins are narrow. In addition, although cached continu- ation avoids repeated prefix computation, conditioning on geometry tokens incurs additional latency and memory relative to inference without the interface (Appendix E). 8CONCLUSION DepthEvidence connects dense metric depth prediction, instance-level metric depth estimation, and compositional depth reasoning within a shared MLLM. Its dense-to-language interface supplies object-aligned continuous geometry tokens, with geometric supervision encouraging metric-depth recoverability during language-context interaction. Experiments show competitive dense prediction and gains in object measurement, numerical depth gaps, and metric-constrained decisions. General VQA remains broadly stable relative to the base model, while spatial understanding improves on the evaluated benchmarks. These findings support dense prediction as both a supervised output and internal evidence for language reasoning. Future work can assess robustness to noisier camera intrinsics and object references, reduce inference cost as the object count grows, and extend the interface to other geometric evidence. 9 AI USE STATEMENT We used Qwen3.7-plus and Seed-2.1-Turbo to propose object annotations and to generate and cross- verify synthetic Depth-VQA questions and answers, as described in Section 4 and Appendix C.1. We additionally used generative AI tools to improve manuscript readability. All benchmark examples were manually verified, and the authors reviewed all AI-assisted outputs and take responsibility for the final content. REPRODUCIBILITY STATEMENT Section 3 and Appendix A describe the model architecture and training objectives, while Section 5 and Appendix B provide the training pipeline, data composition, preprocessing, and optimization settings. Section 4 and Appendix C document benchmark construction, data partitioning, evaluation protocols, and metric definitions. REFERENCES Mahtab Bigverdi, Zelun Luo, Cheng-Yu Hsieh, Ethan Shen, Dongping Chen, Linda G. Shapiro, and Ranjay Krishna. Perception tokens enhance visual reasoning in multimodal language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 3836–3845, 2025. ByteDance Seed Team. Seed2.1: A next-generation agent for real-world productivity. Official model page, 2026. URL https://seed.bytedance.com/en/seed2_1. Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 11621–11631, 2020. Zhipeng Cai, Ching-Feng Yeh, Xu Hu, Zhuang Liu, Gregory P. Meyer, Xinjie Lei, Changsheng Zhao, Shang-Wen Li, Vikas Chandra, and Yangyang Shi. DepthLM: Metric depth from vision language models. In International Conference on Learning Representations, 2026. Nicolas Carion, Laura Gustafson, Yuan-Ting Hu, Shoubhik Debnath, Ronghang Hu, Di- dac Suris Coll-Vinent, Chaitanya Ryali, Kalyan Vasudev Alwala, Haitham Khedr, Andrew Huang, Jie Lei, Tengyu Ma, Baishan Guo, Arpit Kalla, Markus Marks, Joseph Greer, Meng Wang, Peize Sun, Roman R ̈ adle, Triantafyllos Afouras, Effrosyni Mavroudi, Kather- ine Xu, Tsung-Han Wu, Yu Zhou, Liliane Momeni, Rishi Hazra, Shuangrui Ding, Sagar Vaze, Francois Porcher, Feng Li, Siyuan Li, Aishwarya Kamath, Ho Kei Cheng, Piotr Doll ́ ar, Nikhila Ravi, Kate Saenko, Pengchuan Zhang, and Christoph Feichtenhofer. SAM 3: Seg- ment anything with concepts.In International Conference on Learning Representations, 2026. URLhttps://proceedings.iclr.c/paper_files/paper/2026/hash/ e0982cbc81401df3430e1f780dc7a2-Abstract-Conference.html. Boyuan Chen, Zhuo Xu, Sean Kirmani, Brian Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. SpatialVLM: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 14455–14465, 2024a. Haijier Chen, Bo Xu, Shoujian Zhang, Haoze Liu, Jiaxuan Lin, and Jingrong Wang. Vid-LLM: A compact video-based 3D multimodal LLM with reconstruction–reasoning synergy. In International Conference on Learning Representations, 2026. Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, and Feng Zhao. Are we on the right way for evaluating large vision-language models? In Advances in Neural Information Processing Systems, volume 37, 2024b. An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. SpatialRGPT: Grounded spatial reasoning in vision-language models. In Advances in Neural Information Processing Systems, volume 37, 2024. 10 Mengfei Du, Binhao Wu, Zejun Li, Xuanjing Huang, and Zhongyu Wei. EmbSpatial-Bench: Benchmarking spatial understanding for embodied tasks with large vision-language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), p. 346–355, 2024. doi: 10.18653/v1/2024.acl-short.33. David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. In Advances in Neural Information Processing Systems, volume 27, 2014. Mohsen Gholami, Ahmad Rezaei, Zhou Weimin, Sitong Mao, Shunbo Zhou, Yong Zhang, and Mohammad Akbari. Spatial reasoning with vision-language models in ego-centric multi-view scenes. In International Conference on Learning Representations, 2026. GLM-V Team. GLM-4.1V-Thinking and GLM-4.5V: Towards versatile multimodal reasoning with scalable reinforcement learning. arXiv preprint arXiv:2507.01006, 2025. Google DeepMind. Embodied reasoning question answer (ERQA) benchmark. Official benchmark repository, 2025. URL https://github.com/embodiedreasoning/ERQA. Google DeepMind. Gemini 3.1 Pro model card. Google DeepMind, 2026. URLhttps:// deepmind.google/models/model-cards/gemini-3-1-pro/. Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, Dinesh Manocha, and Tianyi Zhou. HallusionBench: An advanced diagnostic suite for entangled language hallucination and visual illusion in large vision- language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 14375–14385, 2024. Vitor Guizilini, Rares Ambrus, Sudeep Pillai, Allan Raventos, and Adrien Gaidon. 3D packing for self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 2485–2494, 2020. Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Kaixuan Wang, Hao Chen, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3D v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12), 2024. Wenbo Hu, Jingli Lin, Yilin Long, Yunlong Ran, Lihan Jiang, Yifan Wang, Chenming Zhu, Runsen Xu, Tai Wang, and Jiangmiao Pang. G 2 VLM: Geometry grounded vision language model with unified 3D reconstruction and spatial reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 9535–9546, 2026. Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images. In Proceedings of the European Conference on Computer Vision, 2016. Yuchen Li, Amanmeet Garg, Shalini Chaudhuri, Rui Zhao, and Garin Kessler. Perceptio: Perception enhanced vision language models via spatial token generation. arXiv preprint arXiv:2603.18795, 2026. Haotong Lin, Sili Chen, Jun Hao Liew, Donny Y. Chen, Zhenyu Li, Yang Zhao, Sida Peng, Hengkai Guo, Xiaowei Zhou, Guang Shi, Jiashi Feng, and Bingyi Kang. Depth anything 3: Recovering the visual space from any views. In International Conference on Learning Representations, 2026. URLhttps://proceedings.iclr.c/paper_files/paper/2026/hash/ e4cd50120b6d7e8daff1749d6bbaa889-Abstract-Conference.html. Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. MMBench: Is your multi-modal model an all-around player? In Proceedings of the European Conference on Computer Vision, 2024. Yuliang Liu, Zhang Li, Mingxin Huang, Biao Yang, Wenwen Yu, Chunyuan Li, Xucheng Yin, Cheng-Lin Liu, Lianwen Jin, and Xiang Bai. OCRBench: On the hidden mystery of OCR in large multimodal models. arXiv preprint arXiv:2305.07895, 2023. 11 Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. MathVista: Evaluating mathematical reasoning of foundation models in visual contexts. In International Conference on Learning Representations, 2024. Baorui Ma, Jiahui Yang, Donglin Di, Xuancheng Zhang, Jianxun Cui, Hao Li, Yan Xie, and Wei Chen. MetricAnything: Scaling metric depth pretraining with noisy heterogeneous sources. In Proceedings of the European Conference on Computer Vision, 2026. Ethan Perez, Florian Strub, Harm de Vries, Vincent Dumoulin, and Aaron Courville. FiLM: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. doi: 10.1609/aaai.v32i1.11671. Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. UniDepth: Universal monocular metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 10106–10116, 2024. Luigi Piccinelli, Christos Sakaridis, Yung-Hsu Yang, Mattia Segu, Siyuan Li, Wim Abbeloos, and Luc Van Gool. UniDepthV2: Universal monocular metric depth estimation made simpler. IEEE Transactions on Pattern Analysis and Machine Intelligence, 48(3):2354–2367, 2026. doi: 10.1109/ TPAMI.2025.3628473. URL https://doi.org/10.1109/TPAMI.2025.3628473. Qwen Team. Qwen3.5-4B model card. Hugging Face, 2026a. URLhttps://huggingface. co/Qwen/Qwen3.5-4B. Qwen Team.Qwen3.7-Plus.Alibaba Cloud Model Studio, 2026b.URLhttps: //docs.modelstudio.console.alibabacloud.com/en/model-studio/ qwen3-7-plus. Pierluigi Zama Ramirez, Fabio Tosi, Matteo Poggi, Samuele Salti, Stefano Mattoccia, and Luigi Di Stefano. Open challenges in deep stereo: The booster dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 21168–21178, 2022. Ren ́ e Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 12179–12188, 2021. Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M. Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 10912–10922, 2021. Shuran Song, Samuel P. Lichtenberg, and Jianxiong Xiao. SUN RGB-D: A RGB-D scene under- standing benchmark suite. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, p. 567–576, 2015. Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 2446–2454, 2020. Bin Tan, Changjiang Sun, Xiage Qin, Hanat Adai, Zelin Fu, Tianxiang Zhou, Han Zhang, Yinghao Xu, Xing Zhu, Yujun Shen, and Nan Xue. Masked depth modeling for spatial perception. arXiv preprint arXiv:2601.17895, 2026. Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Mingjie Zhan, and Hongsheng Li. Measuring multimodal mathematical reasoning with the MATH-Vision dataset. In Advances in Neural Information Processing Systems, volume 37, 2024a. 12 Zirui Wang, Mengzhou Xia, Luxi He, Howard Chen, Yitao Liu, Richard Zhu, Kaiqu Liang, Xindi Wu, Haotian Liu, Sadhika Malladi, Alexis Chevalier, Sanjeev Arora, and Danqi Chen. CharXiv: Charting gaps in realistic chart understanding in multimodal LLMs. In Advances in Neural Information Processing Systems, volume 37, 2024b. Zhixiang Wei, Yi Li, Zhehan Kan, Xinghua Jiang, Zuwei Long, Shifeng Liu, Hongze Shen, Wei Liu, Xiaoyu Tan, Haojia Lin, Yubo Zhu, Qianyu Li, Di Yin, Haoyu Cao, Weibo Gu, Xin Li, Yinsong Liu, Deqiang Jiang, Xing Sun, Yunsheng Wu, Mingkong Tang, Shuangyin Liu, Lexiang Tang, Haodong Lin, Junru Lu, Jiarui Qin, Lingfeng Qiao, Ruizhi Qiao, Bo Ke, Jianfeng He, Ke Li, Yangning Li, Yunhang Shen, Mengdan Zhang, Peixian Chen, Kun Yin, Bing Liu, Yunfei Wu, Huang Chen, Zhongpeng Cai, and Xiaotian Li. Youtu-VL: Unleashing visual potential via unified vision-language supervision. arXiv preprint arXiv:2601.19798, 2026. URLhttps: //arxiv.org/abs/2601.19798. Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, Deva Ramanan, Peter Carr, and James Hays. Argoverse 2: Next generation datasets for self-driving perception and forecasting. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, 2021. xAI.RealWorldQA.Hugging Face dataset, 2024.URLhttps://huggingface.co/ datasets/xai-org/RealworldQA. Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 10371–10381, 2024a. Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything V2. In Advances in Neural Information Processing Systems, volume 37, 2024b. Zhibo Yang, Jun Tang, Zhaohai Li, Pengfei Wang, Jianqiang Wan, Humen Zhong, Xuejing Liu, Mingkun Yang, Peng Wang, Shuai Bai, Lianwen Jin, and Junyang Lin. C-OCR: A comprehensive and challenging OCR benchmark for evaluating large multimodal models in literacy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 21744–21754, 2025. Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. ScanNet++: A high- fidelity dataset of 3D indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 12–22, 2023. Wei Yin, Chi Zhang, Hao Chen, Zhipeng Cai, Gang Yu, Kaixuan Wang, Xiaozhi Chen, and Chunhua Shen. Metric3D: Towards zero-shot metric 3d prediction from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 9043–9053, 2023. Hanxun Yu, Xuan Qu, Yuxin Wang, Jianke Zhu, and Lei Ke. Unlocking dense metric depth estimation in VLMs. arXiv preprint arXiv:2605.15876, 2026. Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Renliang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. MMMU: A massive multi-discipline multimodal understanding and reasoning benchmark for expert AGI. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 9556–9567, 2024a. Xiang Yue, Tianyu Zheng, Yuansheng Ni, Yubo Wang, Kai Zhang, Shengbang Tong, Yuxuan Sun, Botao Yu, Ge Zhang, Huan Sun, Yu Su, Wenhu Chen, and Graham Neubig. MMMU-Pro: A more robust multi-discipline multimodal understanding benchmark. arXiv preprint arXiv:2409.02813, 2024b. Amir R. Zamir, Alexander Sax, William Shen, Leonidas J. Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, p. 3712–3722, 2018. 13 Jian Zhang, Shijie Zhou, Bangya Liu, Achuta Kadambi, and Zhiwen Fan. SpatialStack: Layered geometry–language fusion for 3D VLM spatial reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 38678–38688, 2026. 14 PAPER AND APPENDIX CONTENTS MAIN PAPER 1 Introduction . . . . . . . . . . . . . . .1 2 Related Work . . . . . . . . . . . . . . .3 3 Method . . . . . . . . . . . . . . . . . .3 3.1Framework Overview . . . . . . . .3 3.2 Level-0 Dense Metric Depth Prediction4 3.3Instance-Level Depth Instruction Tuning . . . . . . . . . . . . . . . .4 3.4Dense-to-Language Interface . . . .4 4 Instance-Level Depth-VQA Benchmark5 5 Experimental Setup . . . . . . . . . . .6 5.1Training Pipeline . . . . . . . . . .6 5.2Evaluation Setup . . . . . . . . . .6 6 Main Results . . . . . . . . . . . . . . .6 6.1 Level-0 Dense Metric Depth Prediction6 6.2 Instance-Level Depth Estimation and Reasoning . . . . . . . . . . . . . .7 6.3Effect of the Dense-to-Language In- terface . . . . . . . . . . . . . . . .8 6.4 General VQA Retention and Spatial Understanding . . . . . . . . . . . .8 7 Analysis and Discussion . . . . . . . . .9 7.1Dense Decoder Ablation . . . . . .9 7.2Geometric Supervision Ablation . .9 7.3Limitations . . . . . . . . . . . . .9 8 Conclusion . . . . . . . . . . . . . . . .9 References . . . . . . . . . . . . . . . . . .10 APPENDIX A Method and Implementation Details . .15 A.1Dense Depth Branch . . . . . . . .15 A.2 Depth-VQAInstructionsand Grounding . . . . . . . . . . . . . .16 A.3Dense-to-Language Interface . . . .17 B Data and Training Details . . . . . . . .18 B.1Data Composition . . . . . . . . . .18 B.2Optimization and Runtime . . . . .19 C Benchmark Construction and Evalua- tion Protocols . . . . . . . . . . . . . . .19 C.1Benchmark Construction and Quality Control . . . . . . . . . . . . . . .19 C.2 Benchmark Composition and Parti- tioning . . . . . . . . . . . . . . . .19 C.3Level-0 Dense-Depth Evaluation . .20 C.4Depth-VQA Evaluation . . . . . . .21 C.5General-Capability Benchmarks and Baselines . . . . . . . . . . . . . .21 C.6Agent-Pipeline Baseline Implemen- tation . . . . . . . . . . . . . . . .21 D Additional Interface Ablation . . . . . .21 D.1Disentangling Interface and Metric Supervision . . . . . . . . . . . . .21 E Inference Efficiency . . . . . . . . . . .22 F Additional Qualitative Results . . . . . .23 F.1Dense Metric Depth . . . . . . . . .23 F.2Instance-Level Depth-VQA . . . . .23 AMETHOD AND IMPLEMENTATION DETAILS A.1DENSE DEPTH BRANCH A.1.1DECODER AND CAMERA CONDITIONING Given an RGB imageI ∈R H×W×3 , we retain three intermediate feature mapsV k 3 k=1 from the vision encoder and reshape the final VLM hidden states at image-token positions into a spa- tial mapM. Following DPT (Ranftl et al., 2021), we normalize and project these four maps to 256, 512, 1024, 1024channels, resample them with strides8, 4, 2, 1, and map each level to a common width of 512 channels. The camera condition is applied to all four levels before fusion, while RGB features are injected into the three higher-resolution levels. Residual fusion blocks propagate the coarsest representation upward and merge it with progressively finer conditioned features. The resulting full-resolution feature is used both to predict dense metric depth and to construct object-level geometry representations. Letκ = (f x ,f y ,c x ,c y )denote the supplied camera intrinsics for an image of original sizeW 0 × H 0 .We convert them to image-relative coordinates ̄ k = [ ̄ f x , ̄ f y , ̄c x , ̄c y ] = 15 [f x /W 0 ,f y /H 0 ,c x /W 0 ,c y /H 0 ], and defineφ x = 2 arctan(1/(2 ̄ f x )),φ y = 2 arctan(1/(2 ̄ f y )), and r = W 0 /H 0 . The decoder condition is the nine-dimensional vector s(κ) = [ ̄ k, log ̄ f x , log ̄ f y ,φ x ,φ y ,r].(4) An MLP encodess(κ)ascand modulates each decoder level throughFiLM(F k , c) = F k ⊙ (1 + γ k (c)) +β k (c). A.1.2HIGH-RESOLUTION RGB REFINEMENT The RGB branch processes the input image with a convolutional stem followed by three stride-2 stages, producing feature mapsR k 3 k=1 at strides2, 4, 8. After channel projection and spatial alignment, each RGB map is added to its corresponding DPT level through a learned residual spatial gate, e F k = F k + G k ⊙ R k , whereG k = clip(b k + 1.5 tanhg k ([F k ; R k ]), 0, 3). The functiong k predicts a spatially varying residual from the concatenated DPT and RGB features, and the base gateb k is initialized to0.01. After the final fusion block, a convolution reduces the feature width, a second FiLM layer applies the camera condition, and the feature map is bilinearly resized to the target resolution. The metric prediction is b D = clip(softplus(h(F)) + d min ,d min ,d max ) , with d min = 0.05 m and d max = 250 m. A.1.3DENSE OBJECTIVE The dense objective combines the absolute relative and scale-invariant log losses (Eigen et al., 2014) with structural supervision on inverse depth. Specifically,L der andL lap match Scharr derivatives and Laplacian responses at scales1, 2, 4, 8. For the edge-guided stable local scale-and-shift invariant termL ssi , we sample regions centered on strong RGB edges and transform depth using q(D) = log(1 + 50/(10 −4 +D)). Prediction and target values are robustly normalized and compared both within the sampled local regions and over the full image. A.2DEPTH-VQA INSTRUCTIONS AND GROUNDING A Level-1 example contains an imageIand a grounded target objecto i , and asks for instance-level metric depth estimation. The default targetd i is the mean of valid depths within the object’s instance mask; queries can also request center-point, median, nearest, or farthest depth. A Level-2 example contains grounded objectsO =o i N i=1 and a questionqrequiring compositional depth reasoning in the Relative or Metric track. Relative questions ask for pairwise comparisons, closest or farthest object selection, or depth ordering. Metric questions ask for numerical mean-depth gaps, comparisons against a threshold, selection within a depth range or bounds, or decisions under a required depth margin. Compositional questions combine these operations with image-space relations to select the relevant objects. For a margin-based closest-object decision, the target is ambiguous when the closest object fails to meet the specified margin. Depth convention and object statistics. Camera depth denotes the optical-axis coordinateZof a point in the camera frame, measured in meters, rather than its Euclidean distance to the camera center. Object depth denotes the statistic requested for an instance. For Level 1, letΩ i contain the pixels inside the ground-truth instance mask with valid camera depthD(p). The mean target is d mean i = |Ω i | −1 P p∈Ω i D(p), and the median target is the median over the same set. The nearest and farthest targets are the 5th and 95th percentiles of valid mask depths, respectively; center-point depth is a point measurement rather than a region aggregate. The mask defines the target statistics, while a box or referring prompt identifies the object to the model. The interface pools predicted evidence over box regions (Appendix A.3.1). Mean camera-depth gaps compare optical-axis depth statistics, not Euclidean distances between objects in 3D. A box prompt specifies the object category and bounding box, with coordinates normalized to a 0–1000image coordinate system. A referring prompt identifies the object with a natural-language description enclosed by object-reference tokens. Each object in a Level-2 prompt receives a stable identifier such asobj1, and the prompt lists its identifier, category, and grounding cue before the question. Every Depth-VQA instruction embeds the camera intrinsics in the user prompt as<focal> f x f y c x c y </focal> . The token name<focal>is retained in the implementation, but the block 16 contains both focal lengths and principal-point coordinates. The language-modeling loss is applied only to the final answer: a metric depth value for Level 1, or a numerical depth gap, an object or set selection, an ordering, or a discrete comparison or constraint decision for Level 2, including ambiguous for an unmet depth-margin requirement. Input positions are excluded as language- modeling targets; auxiliary geometric losses supervise the geometry representations as specified in Appendix A.3.4. A.3DENSE-TO-LANGUAGE INTERFACE A.3.1CONSTRUCTION The same dense-to-language interface is used for Level 1 single-object queries and Level 2 multi-object questions. The dense decoder produces a metric depth map b D and a full-resolution fused feature mapF d . For each grounded objecto i , letb i = [x 1 i ,y 1 i ,x 2 i ,y 2 i ]be its normalized box and letR i denote the corresponding region. We average-poolF d insideR i to obtain an appearance-and-shape descriptorr i , and summarize the predicted depths in the same region as s i = [μ i ,σ i ,d min i ,d max i ,d ctr i , logμ i ,a i ,ρ i ] . These entries denote the mean, standard deviation, minimum, maximum, center depth, log-mean depth, box area, and aspect ratio, respectively. For a box-grounded object, we directly use the box given in the prompt. For a referring expression enclosed by<|objectrefstart|>and<|objectrefend|>, we take the VLM hidden state at the closing reference token as an object query. A lightweight box-regression head, imple- mented as layer normalization followed by a two-layer MLP, predicts b b i ∈ [0, 1] 4 . The predicted box is supervised by the ground-truth region during training and is used for differentiable region pooling and inference. The projected object tokens are organized as B G = [g start ,e(obj 1:), g 1,1 , g 1,2 ,...,e(objN:), g N,1 , g N,2 , g end ],(5) wheree(obji:)is the ordinary language-embedding sequence of an object anchor andg start and g end are learned boundary vectors. Placeholder<geometry>embeddings at the dynamic positions are replaced by the corresponding object vectors before the continuation enters the language model. The bank therefore contains2N + 2dynamically generated vectors forNobjects, in addition to the ordinary textual anchors. Its positions serve as conditioning context and are excluded as language- modeling targets. The answer loss nevertheless backpropagates through the geometry representations to the projector. The bank contains predicted geometry; ground-truth depth statistics serve as targets for the auxiliary representation objectives. A.3.2TOKEN CAPACITY: WHY TWO PER OBJECT? We useK = 2as a compact multi-token bottleneck for each grounded object. The fused descriptor [r i ; b i ; s i ]contains decoder features, image location, and several statistics of the predicted metric depth. Compressing all of this evidence into a single VLM vector would impose a narrower infor- mation bottleneck. Instead, the final layer ofP g produces a2h-dimensional output that is reshaped into two independently parameterized vectors inR h . Both vectors are computed from the complete object descriptor; we do not prescribe one token for appearance and the other for metric depth. A shared object-index embedding binds the pair to the same instance, while distinct slot embeddings break their symmetry and allow the language model to use the two positions differently. The resulting specialization, if any, is therefore learned rather than hand-crafted. This added representational band- width requires only one additional dynamic position per object relative toK = 1: forNgrounded objects, the geometry bank contains2Nobject-geometry positions instead ofN. Thus,K = 2 provides a controlled capacity increase while keeping the interface sequence length proportional to the number of grounded objects; the subsequent attention cost follows the standard dependence on the resulting total sequence length. A.3.3CACHED GEOMETRY CONTINUATION At both instance levels, letPcontain the image, camera-intrinsics block, grounded object descriptions, and question. Processing this prefix once produces contextualized statesH P and a layer-wise cache 17 C P . The dense branch uses the visual positions inH P together with the multi-level vision features to constructB G , after which the cache is extended and the answer is generated as (H P ,C P ) = f θ (P; cache = on), C P,G = f θ (B G ;C P ),Y ∼ f θ (·;C P,G ).(6) In a full-attention layer,C P contains the prefix keys and values; queries from the geometry continua- tion attend to these cached tensors and to preceding continuation positions. The language backbone also contains linear-attention layers, which summarize the prefix with a causal-convolution state and a recurrent state rather than a conventional key–value sequence. We clone these differentiable prefix states and use them to initialize the chunk kernels that process the multi-token geometry continuation. Both attention types therefore process the continuation as if[P;B G ; Y]had been evaluated as one causal sequence. Caching changes the computation schedule without changing causal order or exposing the geometry tokens as visible outputs. A.3.4REPRESENTATION OBJECTIVES For objecti, let ̄ g i = (g i,1 +g i,2 )/2be the raw representation before it enters the language model, and let ̄ h i = (h i,1 + h i,2 )/2be the final-layer representation at the same geometry positions after cached continuation. Letd i contain five ground-truth object-depth statistics: center, mean, median, nearest, and farthest depth. Two lightweight heads predict their logarithms from the raw and contextualized representations: b z raw i = H raw ( ̄ g i ), b z ctx i = H ctx ( ̄ h i ),z i = log d i .(7) The valid-depth ratio w i weights the corresponding smooth-ℓ 1 regressions, L raw = P i w i SmoothL1( b z raw i , z i ) P i w i , L ctx = P i w i SmoothL1( b z ctx i , z i ) P i w i .(8) At both instance levels, we operationalize metric information preservation through the recoverability of these ground-truth statistics from both representations. The raw and contextual losses supervise the projected tokens and their states after language-context interaction against the same targets, encouraging metric content to remain accessible to a prediction head. For multi-object examples, we additionally supervise relative depth using L rank = 1 |P| X (i,j)∈P softplus(−s ij (bz i −bz j )), s ij = sign(logd mean i − logd mean j ),(9) wherebz i is the contextual prediction of mean log-depth andPcontains object pairs whose ground- truth log-depth gap exceeds a small ambiguity margin. The differentiable cache allows the answer and contextual geometry objectives to update the continuation adapters and trainable prefix-pathway parameters. The geometry projector receives gradients from the answer, raw-depth, contextual-depth, and ranking losses. We detach b D andF d before object projection, so these objectives do not perturb the pretrained dense decoder, which continues to be optimized byL dense on dense-prediction batches. BDATA AND TRAINING DETAILS B.1DATA COMPOSITION The Level-0 corpus combines approximately 6.7M images from LingBot-Depth, Taskonomy, Waymo, DDAD, nuScenes, Argoverse 2, and ScanNet++. We apply depth completion to outdoor sources with sparse sensor depth so that every source image is paired with a dense depth map, and preserve the camera intrinsics associated with each image. The dense branch converts these intrinsics into its numerical camera condition, whereas Depth-VQA prompts provide the raw(f x ,f y ,c x ,c y )values in a textual camera-intrinsics block. All dense-benchmark frames are excluded from training. The instance-level corpus contains approximately 1.5M Level-1 and Level-2 examples sampled from LingBot-Depth, Waymo, DDAD, Argoverse 2, nuScenes, and Taskonomy. Stage 2 combines these examples with roughly 771K general visual-instruction examples and about 60K Level-0 replay samples, with stream frequencies proportional to corpus size. 18 B.2OPTIMIZATION AND RUNTIME Optimization. We use Qwen3.5-4B as the vision-language backbone and AdamW with weight decay 0.01 and a 3% warmup ratio. During Stage 1, the vision encoder and language model are frozen, and the dense branch is trained with a learning rate of10 −3 . During Stage 2, decoder-side modules and LoRA-adapted language parameters use a learning rate of10 −4 , while the vision encoder uses 10 −7 . LoRA is applied to attention and feed-forward projections with rank 32, scaling factor 64, and dropout 0.05. Runtime settings.Dense inputs are resized to448× 448. Instruction-tuning sequences contain at most 4,096 tokens, including up to 2,048 visual tokens. We train in bfloat16 with FlashAttention-2 and an effective global batch size of 256 for both dense and VQA updates. Stage 2 runs on four nodes with 32 NVIDIA A100 GPUs in total. CBENCHMARK CONSTRUCTION AND EVALUATION PROTOCOLS C.1BENCHMARK CONSTRUCTION AND QUALITY CONTROL Dense depth preparation.Outdoor datasets provide sparse sensor measurements, so we first apply depth completion to obtain a dense depth map for every source image. These image–depth pairs form the Level-0 pool and provide the metric field from which instance-level targets are derived. Level-1 instance construction.For each image, Qwen3.7-plus (Qwen Team, 2026b) and Seed-2.1- Turbo (ByteDance Seed Team, 2026) independently identify the major objects and propose a category and grounding box for each one. We pool their candidates and discard boxes covering less than 1% of the image or lying too close to image boundaries. The remaining category–box pairs prompt SAM 3 (Carion et al., 2026) to produce instance masks, which pass a second mask-quality filter. For each retained mask, we combine its label, box, and dense depth values to compute center-point, mean, median, nearest, and farthest targets; the latter two are the 5th and 95th percentiles of valid mask depths, respectively. The resulting grounded instance records constitute the Level-1 data. Using two proposal models reduces dependence on model-specific object and category preferences during automatic construction. Level-2 question construction. We sample multiple Level-1 instance records from an image according to the Relative and Metric task taxonomy in Table 7. Given the selected object labels, boxes, and depth statistics, either Qwen3.7-plus or Seed-2.1-Turbo generates a candidate question and answer. The other model then checks question validity and answer correctness, and invalid examples are removed. This cross-model generation–verification protocol reduces reliance on the phrasing and reasoning patterns of a single generator. Human review and distribution control.We manually inspect 1% of the training samples across Levels 0–2, aggregate recurring failure cases, and translate them into targeted filters applied to the training corpus. Every benchmark example undergoes manual verification. For video data, we favor different objects when constructing Level-1 and Level-2 questions within the same scene and limit reuse of the same instance across task types, reducing repeated-instance bias. C.2BENCHMARK COMPOSITION AND PARTITIONING 19 Table 7: Detailed composition of the Relative Depth and Metric Depth tracks in the Level-2 bench- mark. “Subset” orders a specified subset of objects, whereas “all” orders every referenced object. TrackSubbenchRaw question typeRequired reasoning Relative Depth Spatial Pair closer_fartherDecide which of two objects is closer to or farther from the camera. front_behindDecide which of two objects is in front of or behind the other. spatial_depth Resolve a left/right reference, then compare the selected object with another object in depth. Spatial Group spatial_depth_groupFilter objects by a spatial relation, then select the nearest or farthest member. Depth Ordering depth_ordering_closestSelect the closest object from all candidates. depth_ordering_farthestSelect the farthest object from all candidates. depth_ordering_subsetOrder a specified subset of objects by depth. depth_ordering_allOrder all referenced objects by depth. Metric Depth Metric Pair metric_depth_gapCompute the numerical difference in mean camera depth between two objects. metric_gap_thresholdDetermine whether the mean-depth gap between two ob- jects reaches a specified metric threshold. Spatial Metric Pair spatial_metric_gapResolve a 2D spatial reference, then compute the absolute mean-depth difference to another object. Spatial Metric Group spatial_range_filterApply a spatial filter, then select objects whose mean depths fall inside a specified metric range. compositional_metric_ constraint Combine a spatial condition, a metric lower bound, and an object-referenced upper-depth bound. Metric Set Decision camera_safety_bandDetermine whether any referenced object lies within a specified mean-depth range from the camera. metric_margin_decisionSelect the closest object only when its depth advantage exceeds a metric margin; otherwise answer ambiguous. Benchmark partitioning. For datasets with official partitions, benchmark examples are sampled exclusively from their official validation or test splits. For datasets without an official evaluation partition, we randomly hold out complete scenes using a fixed seed and sample benchmark examples only from these held-out scenes. No held-out scene appears in either the dense-depth or instruction- tuning corpora. Table 8: Compact statistics for the dense and instance-level depth benchmarks. Depth-VQA contains 1,503 questions in total. ComponentSampling unitScale or coverageTask composition Level-0 dense depthImage4,228 images from nine datasets 500 images from each of eight datasets and all 228 Booster images Level-1 instance depthGrounded-object query757 queries from seven datasets Center, nearest, farthest, mean, and median metric depth; box or referring prompts Level-2 RelativeMulti-object question370 questions Spatial Pair (86), Spatial Group (88), and Depth Order- ing (196) Level-2 MetricMulti-object question376 questionsMetric Pair (46 threshold + 48 numerical gap), Spatial Metric Pair (94), Spatial Metric Group (94), and Metric Set Decision (94) C.3LEVEL-0 DENSE-DEPTH EVALUATION Sampling and ground truth.We select 500 frames from each of Argoverse 2, Waymo, nuScenes, DDAD, Taskonomy, ScanNet++, HyperSim, and SUN RGB-D using a fixed random seed, and use all 228 Booster images spanning 38 scenes and six exposure settings. All selected frame identities are removed from the dense training manifest. The automotive datasets are evaluated against their original sensor-derived metric ground truth rather than depth-completion outputs. Booster disparity is converted to metric depth using its official calibration and evaluated with the official validity mask. Scoring.Ground-truth camera intrinsics are supplied to every method evaluated in Table 2. Predic- tions are compared directly in metric scale without median scaling or scale-and-shift alignment. For the first eight datasets, valid pixels must be finite and lie within[0.05, 250]meters; Booster follows its official mask and[0.001, 10]meter range. We compute AbsRel, RMSE, log RMSE, SILog, andδ 1 , δ 2 , andδ 3 over valid pixels. The main comparison reportsδ 1 , with the cross-dataset score computed as the unweighted mean of the nine dataset-level results. 20 C.4DEPTH-VQA EVALUATION Level 1 metrics are computed over the requested instance-depth values after parsing the numerical answer in meters. For the Level 2 Relative track, a selection answer is correct only if it contains the target identifier, and an ordering answer is correct only if the full predicted identifier sequence matches the reference. Relative Overall Acc is computed over the 86 Spatial Pair, 88 Spatial Group, and 196 Depth Ordering questions (370in total). In the Metric track, Metric Pair separates 46metric_gap_thresholdquestions scored by accuracy from 48metric_depth_gap questions scored byδ 1 . Spatial Metric Pair reportsδ 1 on its 94spatial_metric_gapquestions. Metric Overall Acc divides the total correct answers on the threshold, Spatial Metric Group, and Metric Set Decision questions by46 + 94 + 94 = 234. The 142 numerical gap questions are excluded from Overall Acc and evaluated in their respective δ 1 columns. C.5GENERAL-CAPABILITY BENCHMARKS AND BASELINES Benchmarks. Table 4 covers MathVision (Wang et al., 2024a), MathVista (Lu et al., 2024), MMMU (Yue et al., 2024a), MMMU-Pro (Yue et al., 2024b), RealWorldQA (xAI, 2024), M- Star (Chen et al., 2024b), MMBench (Liu et al., 2024), HallusionBench (Guan et al., 2024), AI2D (Kembhavi et al., 2016), C-OCR (Yang et al., 2025), OCRBench (Liu et al., 2023), CharXiv (Wang et al., 2024b), ERQA (Google DeepMind, 2025), EmbSpatialBench (Du et al., 2024), and Ego3D-Bench (Gholami et al., 2026). Baselines. For Level-0 dense prediction, the specialized vision models are Metric3D (Yin et al., 2023), Metric3D v2 (Hu et al., 2024), UniDepth V2 (Piccinelli et al., 2026), Depth Anything V1–V3 (Yang et al., 2024a;b; Lin et al., 2026), and MetricAnything (Ma et al., 2026); the depth- aware VLMs are DepthLM (Cai et al., 2026) and DepthVLM (Yu et al., 2026). For instance- level Depth-VQA, the general-purpose VLMs are Qwen3.5-4B (Qwen Team, 2026a), Gemini- 3.1-Pro-Preview (Google DeepMind, 2026), Qwen3.7-plus (Qwen Team, 2026b), and Seed-2.1- Turbo (ByteDance Seed Team, 2026), while the metric-depth VLMs are DepthLM-12B, DepthVLM- 4B, and Youtu-VL-4B-Instruct. C.6AGENT-PIPELINE BASELINE IMPLEMENTATION The agent baselines in Table 3 use a fixed tool-augmented inference pipeline. For each benchmark question, MetricAnything receives the original RGB image and the camera intrinsics provided by the benchmark and predicts a pixel-level metric depth map in meters. We then crop this prediction with each benchmark-provided grounding box and compute a structured set of regional depth statistics for the corresponding identifier, such asobji. The resulting object-indexed depth table is supplied together with the original image and question to either Qwen3.5-4B or Qwen3.7-plus. The reasoner returns its final response between<answer>and</answer>, and the parsed response is evaluated by the original benchmark scorer. The pipeline therefore exposes predicted metric measurements through an explicit textual table while preserving the benchmark questions, object grounding, and scoring protocol; benchmark boxes are used only to extract regions from the predicted depth map, and no ground-truth depth values are given to the reasoner. DADDITIONAL INTERFACE ABLATION D.1DISENTANGLING INTERFACE AND METRIC SUPERVISION Table 9 tests whether the gain from the dense-to-language interface can be explained by additional supervision on metric values alone. The new No interface + metric supervision control removes the interface and adds auxiliary language targets for every instance referenced by a QA example. These targets ask the model to predict the instance’s center, mean, median, nearest, and farthest depths (e.g., “Givenobj 1 , predict its center depth”) and are optimized with the language loss. The remaining three rows reproduce the corresponding configurations from Table 6. 21 Table 9: Disentangling direct metric supervision from the dense-to-language interface on Level-2 Relative and Metric Overall Acc. VariantRelative Overall Metric Overall No interface0.58380.7692 No interface + metric supervision0.58650.7949 Interface +L ans 0.58380.7735 Full interface0.62700.8504 Without the interface, direct metric supervision improves Relative Overall Acc by 0.27 percentage points and Metric Overall Acc by 2.56 points. The larger Metric gain confirms that the auxiliary language targets provide useful numerical depth supervision. However, the full interface remains ahead of this control by 4.05 points on Relative Overall Acc and 5.56 points on Metric Overall Acc. Within this comparison, metric supervision therefore explains only part of the improvement: the full interface provides an additional benefit beyond exposing the language model to referenced-object depth values through auxiliary text targets. EINFERENCE EFFICIENCY We compare three inference modes with end-of-sequence stopping disabled and a fixed output length of 64 tokens. No-interface generates answers without geometry tokens. Two-pass first extracts geometry and then reprocesses the multimodal prefix with the geometry tokens to generate the answer. Single-prefill reuses the prefix cache and processes the geometry tokens as a continuation, as described in Appendix A.3.3. Table 10 reports generation latency, end-to-end latency, token throughput, and peak allocated GPU memory under this fixed-length setting. Table 10: Inference efficiency with a fixed output length of 64 tokens and end-of-sequence stopping disabled. Single-prefill is the cached geometry-continuation scheme used by DepthEvidence. Mode Generation latency (s)↓ End-to-end latency (s)↓ Throughput (tokens/s)↑ Peak allocated memory (GiB)↓ No-interface4.1744.28915.33310.36 Two-pass4.5354.71014.11113.56 Single-prefill4.3234.49114.80413.73 Single-prefill reduces generation latency from 4.535 to 4.323 s and end-to-end latency from 4.710 to 4.491 s relative to two-pass inference, reductions of 4.67% and 4.65%, respectively. Throughput increases by 4.91%, from 14.111 to 14.804 tokens/s. Relative to no-interface inference, the end-to-end overhead is 0.202 s (4.71%). These results support reusing the multimodal prefix to limit the runtime cost of incorporating dense geometry into language reasoning. The benefit is in latency rather than memory: single-prefill uses 13.73 GiB of peak allocated memory, 0.17 GiB more than two-pass and 3.37 GiB more than no-interface. Thus, cached continuation reduces repeated prefix computation while retaining an additional memory cost for geometry-conditioned inference. 22 FADDITIONAL QUALITATIVE RESULTS F.1DENSE METRIC DEPTH Figure 5 extends the qualitative comparison in Section 6.1 to the five remaining dense benchmarks. RGBGTDepthVLM-4BYoutu-VL-4BMetric AnythingUniDepth V2DepthEvidence NuScenes DDAD Taskonomy SUN RGB - D Booster Figure 5: Additional qualitative dense metric depth comparisons on NuScenes, DDAD, Taskonomy, SUN RGB-D, and Booster. Columns show RGB, ground truth (GT), DepthVLM-4B, Youtu-VL-4B- Instruct, MetricAnything, UniDepth V2, and DepthEvidence. Outdoor GT is densely completed for visualization only; quantitative evaluation uses the original sparse sensor GT. All depth maps within a row use the same color scale. F.2INSTANCE-LEVEL DEPTH-VQA We provide selected qualitative cases, including the input image, question, ground-truth answer, and the outputs of Gemini-3.1-Pro, DepthEvidence without the interface, and DepthEvidence. A green checkmark ( ✓ ) denotes a correct benchmark prediction, and a red cross ( × ) denotes an incorrect one. Box-grounded questions are paired with the annotated box image, whereas referring-prompt questions retain the original image. Level 1 cases are grouped by dataset; Level 2 cases follow the subbench taxonomy of Table 1 and cover every Relative Depth and Metric Depth subbench. Examples reproduced in Figure 4 are omitted here, except for the sole selected Depth Ordering example. F.2.1LEVEL-1 CASES GROUPED BY DATASET These examples span center-point, mean, median, nearest, and farthest instance-depth queries. The displayed target specification is the semantic reference seen by the model; box coordinates remain in the question for box-grounded cases. LingBot-Depth Mean depth; box prompt Question. Given this image and the target object, estimate the average depth of the target region from the camera. Target: classroom sign with bbox2d [670, 38, 888, 529] Ground truth: 1.25 m Gemini-3.1-Pro: 2.11 m × DepthEvidence w/o Interface: 1.31 m ✓ DepthEvidence: 1.32 m ✓ Waymo 23 Farthest depth; box prompt Question. Given this image and the target object, estimate the farthest depth of the target region from the camera. Target: silver sedan with bbox 2d [819, 454, 999, 562] Ground truth: 27.167 m Gemini-3.1-Pro: 41.11 m × DepthEvidence w/o Interface: 26.62 m ✓ DepthEvidence: 27.3 m ✓ Median depth; referring prompt Question. Given this image and the target object, estimate the median depth of the target region from the camera. Target: pole referred as ”the tall, dark metal pole on the far left sidewalk near the traffic light” Ground truth: 16 m Gemini-3.1-Pro: 10.11 m × DepthEvidence w/o Interface: 15 m ✓ DepthEvidence: 15.56 m ✓ DDAD Mean depth; box prompt Question. Given this image and the target object, estimate the average depth of the target region from the camera. Target: green trees with bbox2d [736, 248, 999, 514] Ground truth: 50.94 m Gemini-3.1-Pro: 101.12 m × DepthEvidence w/o Interface: 52.4 m ✓ DepthEvidence: 55.3 m ✓ Argoverse 2 Nearest depth; referring prompt Question. Given this image and the target object, estimate the nearest depth of the target region from the camera. Target: white car hood referred as ”the white car hood with ’ARGO SELF-DRIVING TEST VEHICLE’ text at the bottom of the image” Ground truth: 1.15 m Gemini-3.1-Pro: 1.8 m × DepthEvidence w/o Interface: 1.06 m ✓ DepthEvidence: 1.03 m ✓ 24 Nearest depth; box prompt Question. Given this image and the target object, estimate the nearest depth of the target region from the camera. Target: large brick building with bbox 2d [458, 230, 797, 495] Ground truth: 37.46 m Gemini-3.1-Pro: 10.82 m × DepthEvidence w/o Interface: 40.5 m ✓ DepthEvidence: 36.5 m ✓ nuScenes Median depth; box prompt Question. Given this image and the target object, estimate the median depth of the target region from the camera. Target: road with bbox 2d [2, 530, 999, 998] Ground truth: 7.19 m Gemini-3.1-Pro: 14.11 m × DepthEvidence w/o Interface: 7.53 m ✓ DepthEvidence: 7.06 m ✓ Taskonomy Mean depth; box prompt Question. Given this image and the target object, estimate the average depth of the target region from the camera. Target: blue bowl with bbox2d [734, 713, 992, 828] Ground truth: 1.53 m Gemini-3.1-Pro: 1.81 m ✓ DepthEvidence w/o Interface: 1.4 m ✓ DepthEvidence: 1.48 m ✓ F.2.2LEVEL-2 CASES GROUPED BY TABLE 1 SUBBENCH The cases below retain the same Relative Depth and Metric Depth organization used by the benchmark. Annotated boxes identify the candidate objects and their stable obji identifiers. Relative Depth 25 Spatial Pair. spatialdepth; box prompt Question. Which is closer to the camera, the object to the left of silver pickup truck obj2 or sidewalk obj3? Ground truth: red sedan obj1. Gemini-3.1-Pro: obj3 × DepthEvidence w/o Interface: red sedan obj1. ✓ DepthEvidence: red sedan obj1. ✓ Spatial Group. spatialdepthgroup; box prompt Question. Among the objects to the left of tree3 obj6, which is closest to the camera? Ground truth: crosswalk marking1 obj1. Gemini-3.1-Pro: obj3 × DepthEvidence w/o Interface: crosswalk marking1 obj1. ✓ DepthEvidence: crosswalk marking1 obj1. ✓ Depth Ordering. depthorderingall; box prompt Question. Rank the main objects in this scene by their distance to the camera, starting from the closest. Ground truth: traffic light obj1, person in purple jacket obj5, shopping cart2 obj3, shopping cart1 obj2, building pillar obj7, red car obj4, tree obj6. Gemini-3.1-Pro: obj1, obj5, obj7, obj3, obj2, obj4, obj6 × DepthEvidence w/o Interface: traffic light obj1, person in purple jacket obj5, shopping cart1 obj2, shopping cart2 obj3, building pillar obj7, red car obj4, tree obj6. × DepthEvidence: traffic light obj1, person in purple jacket obj5, shopping cart2 obj3, shopping cart1 obj2, building pillar obj7, red car obj4, tree obj6. ✓ Metric Depth 26 Metric Pair. metricgapthreshold; box prompt Question. Is the average camera depth of wooden entertainment center obj2 at least 6.60 m greater than that of blue sofa obj1? Ground truth: No. Gemini-3.1-Pro: yes. × DepthEvidence w/o Interface: no. ✓ DepthEvidence: No. ✓ Spatial Metric Pair. spatialmetricgap; box prompt Question. What is the absolute difference in average camera depth between the object to the right of black sedan obj1 and building obj3? Ground truth: 38.39 m Gemini-3.1-Pro: 10.5 × DepthEvidence w/o Interface: 27.00 × DepthEvidence: 40.00 m ✓ Spatial Metric Group. spatialrangefilter; box prompt Question. Among the referenced objects to the left of building obj4, which have an average camera depth between 55.21 m and 79.57 m, inclusive? Ground truth: building obj 5. Gemini-3.1-Pro: obj3 × DepthEvidence w/o Interface: building obj5. ✓ DepthEvidence: building obj5. ✓ compositionalmetricconstraint; box prompt Question. Among the referenced objects to the right of green bush obj4, select those with an average camera depth greater than 4.50 m but smaller than that of large glass window obj2. Ground truth: silver car obj1, metal railing obj5, green bush obj3. Gemini-3.1-Pro: obj3, obj5 × DepthEvidence w/o Interface: silver car obj1, green bush obj3, metal railing obj5. ✓ DepthEvidence: silver car obj1, metal railing obj5, green bush obj3. ✓ 27 Metric Set Decision. metricmargindecision; box prompt Question. Select the closest referenced object only if it is at least 21.00 m closer in average camera depth than every other candidate; otherwise answer ambiguous. Ground truth: ambiguous. Gemini-3.1-Pro: obj1 × DepthEvidence w/o Interface: asphalt road obj1. × DepthEvidence: ambiguous. ✓ camerasafetyband; box prompt Question. Is any referenced object at an average camera depth of no more than 0.80 m? Ground truth: No. Gemini-3.1-Pro: yes × DepthEvidence w/o Interface: no ✓ DepthEvidence: No. ✓ 28