cs.CVOct 13, 2026
SaveReDiS: Reasoning-Diagnosis Self-Evolution for Multimodal Reasoning in Remote Sensing
Organizations: School of Computer Science, Wuhan University · Zhongguancun Academy · National Key Laboratory of Science and Technology on Space-Born Intelligent Information Processing, Beijing Institute of Technology
Abstract
Multimodal large language models (MLLMs) have shown strong potential for remote sensing understanding, yet reliable reasoning remains challenging when tasks require fine-grained visual grounding and multi-step inference. Existing approaches typically rely on either costly pre-constructed reasoning trajectories or self-generated signals such as response consistency, which can reinforce shared errors and provide unreliable supervision. We introduce Reasoning--Diagnosis Self-Evolution (ReDiS), a framework that separates reasoning exploration from candidate evaluation while allowing the two processes to evolve jointly. ReDiS instantiates two role-specific experts from a shared multimodal backbone: a Reasoner that explores diverse visual evidence--answer candidates and a Diagnoser that jointly evaluates the complete candidate set and provides credibility-aware feedback. As the Reasoner evolves, it exposes new candidate distributions and failure modes; the Diagnoser adapts to these changes and, in turn, reshapes the learning signal for subsequent exploration. To handle groups in which all sampled answers are incorrect and group-relative advantages vanish, we further introduce Failure-Triggered Recovery (FTR), which provides minimal outcome-level correction without prescribing intermediate reasoning trajectories. Experiments on LHRS-Bench, MME-RS, and CHOICE show consistent improvements over Qwen3.5 backbones at 0.8B, 2B, and 4B scales. ReDiS improves category-averaged accuracy by 6.47 percentage points at 0.8B, while ReDiS-4B achieves the highest average performance among the compared remote sensing MLLMs.
Figures & tables
Figure 1: Training paradigms for remote sensing reasoning. (a) Trajectory-supervised methods learn from pre-constructed chains of thought. (b) Self-consistency methods derive feedback from multiple online rollouts. (c) The Reasoning–Diagnosis Self-Evolution (ReDiS) framework for multimodal reasoning instantiates two role-specific experts from a shared multimodal model.
Figure 2: Overview of ReDiS: Reasoning–Diagnosis Self-Evolution. The Reasoner explores visual evidence–answer candidates, while the Diagnoser jointly evaluates the complete set. Answer labels anchor diagnostic learning and gate credibility feedback to correct-answer candidates. The two roles co-evolve through online sampling and separate group-relative updates. Failure-Triggered Recovery (FTR) supplies value-token correction on failed groups, leaving visual evidence and free-form reasoning without prescribed targets.
| Method | Publication | Size | LHRS-Bench | MME-RS | CHOICE | Avg. | ||||
|---|---|---|---|---|---|---|---|---|---|---|
| OA | AA | OA | AA | OA | AA | OA | AA | |||
| Closed-source Commercial Vision-Language Models | ||||||||||
| GPT-4o | arXiv 2024 | – | 70.29 | 66.15 | 21.43 | 21.36 | 67.71 | 65.66 | 53.14 | 51.06 |
| Gemini-2.5-pro | arXiv 2025 | – | 74.35 | 66.41 | 38.23 | 38.06 | 75.84 | 72.48 | 62.81 | 58.98 |
| Claude-haiku-4.5 | – | – | 58.70 | 56.59 | 20.48 | 21.03 | 64.32 | 59.82 | 47.83 | 45.81 |
| Open-source Vision-Language Models | ||||||||||
Table 1: Comparison with existing vision-language models on remote sensing benchmarks. OA and AA denote overall accuracy and average accuracy, respectively.
Figure 3: Selected response excerpts from ReDiS and the corresponding Qwen3.5 backbones. The examples cover object localization (top left), map recognition (bottom left), object counting (top right), and small-object color recognition (bottom right). For ease of interpretation, key visual evidence is highlighted with yellow boxes.
Figure 4: Training dynamics of Qwen3.5-0.8B w/ and w/o FTR.
Appendix figures & tables17 assets
Supplementary material from the paper’s appendix.
Appendix
| Method | Color | Count | Position | AA |
|---|---|---|---|---|
| EarthDial | 35.06 | 22.51 | 44.07 | 33.88 |
| GeoChat | 33.71 | 22.92 | 31.11 | 29.25 |
| RSThinker | 39.36 | 28.06 | 47.65 | 38.36 |
| RemoteReasoner | 32.35 | 11.42 | 47.10 | 30.29 |
| VHM | 38.09 | 27.32 | 49.72 | 38.38 |
| ReDiS-0.8B | 51.08 | 29.93 | 57.44 | 46.15 |
Table 3: Fine-grained accuracy (%) on MME-RS. AA is the unweighted mean of the displayed category accuracies. Bold and underline denote the best and second-best results, respectively.
| Method | Area | Color | Distance | Identity | Location | Modality | Orientation | Quantity | Reasoning | Resolution | Shape | AA |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| EarthDial | 58.67 | 59.29 | 45.45 | 58.99 | 66.18 | 21.74 | 23.08 | 40.88 | 63.04 | 42.86 | 62.16 | 49.30 |
| GeoChat | 46.67 | 37.17 | 40.91 | 51.26 | 50.49 | 21.74 | 28.21 | 39.42 | 69.57 | 42.86 | 56.76 | 44.09 |
| RSThinker | 61.33 | 65.49 | 50.00 | 67.67 | 70.59 | 60.87 | 48.72 | 46.72 | 80.43 | 9.52 | 78.38 | 58.16 |
| RemoteReasoner | 68.00 | 63.72 | 45.45 | 69.24 | 73.04 | 30.43 | 48.72 | 45.26 | 76.09 | 33.33 | 75.68 | 57.18 |
| VHM | 64.00 | 52.21 | 36.36 | 56.31 | 54.90 | 43.48 | 35.90 | 47.45 | 73.91 | 14.29 | 64.86 | 49.42 |
| ReDiS-0.8B | 52.00 | 71.68 | 45.45 | 60.41 | 65.69 | 30.43 | 25.64 | 51.09 | 71.74 | 9.52 | 64.86 | 49.87 |
Table 4: Fine-grained accuracy (%) on LHRS. AA is the unweighted mean of the displayed category accuracies. Bold and underline denote the best and second-best results, respectively.
| Method | ILC | SII | CID | AA | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| SC | MR | IC | IQ | IM | LR | OP | OL | OC | AR | HD | AC | SR | CD | ||
| EarthDial | 97.50 | 29.30 | 88.50 | 34.50 | 59.62 | 81.00 | 97.40 | 88.40 | 48.00 | 50.40 | 64.80 | 74.00 | 34.67 | 62.50 | 65.04 |
| GeoChat | 92.75 | 47.77 | 73.50 | 25.75 | 33.88 | 79.00 | 93.00 | 81.00 | 26.60 | 51.60 | 33.20 | 71.14 | 29.67 | 47.00 | 56.13 |
| RSThinker | 97.20 | 22.93 | 78.00 | 38.50 | 69.88 | 71.00 | 94.80 | 77.80 | 67.20 | 49.60 | 7.20 | 75.71 | 39.33 | 57.50 | 60.48 |
| RemoteReasoner | 93.70 | 53.50 | 97.50 | 52.25 | 67.25 | 88.00 | 92.40 | 75.00 | 50.40 | 63.00 | 77.20 | 76.57 | 55.67 | 62.00 | 71.75 |
| VHM | 96.75 | 33.12 | 82.00 | 39.75 | 63.88 | 77.00 | 93.40 | 89.80 | 35.00 | 55.20 | 92.80 | 46.57 | 37.33 | 59.50 | 64.44 |
Table 5: Fine-grained perception accuracy (%) on CHOICE. Level-3 tasks are grouped by their Level-2 categories. AA is the unweighted mean of the displayed category accuracies. Bold and underline denote the best and second-best results, respectively. ILC: Image-level Comprehension; SC: Scene Classification, MR: Map Recognition, IC: Image Caption, IQ: Image Quality, IM: Image Modality. SII: Single-instance Identification; LR: Landmark Recognition, OP: Object Presence, OL: Object Localization, OC: Object Counting, AR: Attribute Recognition, HD: Hallucination Detection. CID: Cross-instance Discernment; AC: Attribute Comparison, SR: Spatial Relationship, CD: Change Detection.
| Method | AttR | AssR | CSR | AA | ||||
|---|---|---|---|---|---|---|---|---|
| TP | PP | EA | RA | DD | GD | SI | ||
| EarthDial | 39.50 | 50.67 | 47.50 | 37.33 | 86.50 | 64.50 | 78.00 | 57.71 |
| GeoChat | 21.50 | 48.00 | 40.50 | 36.33 | 75.00 | 82.50 | 74.00 | 53.98 |
| RSThinker | 46.50 | 53.33 | 40.00 | 64.17 | 67.50 | 58.50 | 70.67 | 57.24 |
| RemoteReasoner | 44.50 | 62.33 | 43.00 | 54.67 | 89.00 | 89.00 | 81.33 | 66.26 |
| VHM | 41.00 | 41.33 | 34.00 | 37.50 | 74.00 | 76.00 | 76.67 | 54.36 |
Table 6: Fine-grained reasoning accuracy (%) on CHOICE. Level-3 tasks are grouped by their Level-2 categories. AA is the unweighted mean of the displayed category accuracies. Bold and underline denote the best and second-best results, respectively. AttR: Attribute Reasoning; TP: Time Property, PP: Physical Property. AssR: Assessment Reasoning; EA: Environmental Assessment, RA: Resource Assessment. CSR: Common Sense Reasoning; DD: Disaster Discrimination, GD: Geospatial Determination, SI: Situation Inference.
Figure 5: Online Reasoner training dynamics of ReDiS-0.8B. Each point summarizes a non-overlapping window of 50 steps (1,600 source-question presentations), with eight candidates per question. (a) Fractions of source groups with nonzero GRPO advantages and groups routed to FTR because all candidates are incorrect. (b) Candidate-level answer accuracy. (c) Mean number of distinct candidate answers.
Figure 6: Attribute comparison and area estimation at 0.8B. Left: ReDiS distinguishes the dark and white vehicles, whereas Qwen3.5 selects matching colors. Right: ReDiS selects the reference 50% snow-cover category rather than the backbone’s 25%.
Figure 7: Land-use reasoning and scene description at 0.8B. Left: ReDiS identifies cultivated land as absent, whereas Qwen3.5 selects road. Right: ReDiS selects a partially filled parking lot with cars in rows, correcting the backbone’s almost-vacant description.
Figure 8: Evidence–answer consistency at 2B. Left: Qwen3.5 describes the red vehicle as left and below but selects top left; ReDiS selects bottom left. Right: both models describe soccer fields, but only ReDiS selects the affirmative answer.
Figure 9: Image-modality and corruption recognition at 2B. Left: ReDiS selects the false-color reference rather than RGB, although its explanation retains modality ambiguity. Right: ReDiS identifies the reference impulse-noise category, whereas Qwen3.5 selects Gaussian noise.
Figure 10: Seasonal interpretation and change detection at 2B. Left: ReDiS selects the bottom-right spring scene rather than the snowy bottom-left scene. Right: ReDiS selects the building-removal and football-field option; Qwen3.5 mentions a sports field but chooses the picnic-area option.
Figure 11: Landmark recognition and object localization at 2B. Left: ReDiS identifies Big Ben, whereas Qwen3.5 selects Westminster Abbey despite mentioning a clock tower. Right: ReDiS locates the specified cargo ship in the upper-right region rather than the center.
Figure 12: Small-target recognition at 2B. Left: ReDiS locates the fountain in the lower-left water area, while Qwen3.5 reports it as absent. Right: ReDiS selects yellow for the ship-deck ring rather than the backbone’s green, matching the reference.
Figure 13: Object counting and absence detection at 4B. Left: ReDiS counts five aircraft, matching the annotated instances, while Qwen3.5 reports six. Right: ReDiS reports no airplane; the backbone incorrectly places one in the upper-right region.
Figure 14: Color comparison and coverage estimation at 4B. Left: ReDiS distinguishes the two vehicle colors, while Qwen3.5 treats them as identical. Right: ReDiS selects the reference 50% cloud-cover category instead of the backbone’s 100%.
Figure 15: Question interpretation and corruption recognition at 4B. Left: despite similar junkyard descriptions, ReDiS correctly excludes vehicle count as a distinguishing reason. Right: ReDiS selects the reference Gaussian-noise label, whereas Qwen3.5 selects impulse noise.
Figure 16: Disaster and map recognition at 4B. Left: ReDiS describes structural collapse and selects the earthquake reference label; Qwen3.5 selects flood. Right: ReDiS identifies Sri Lanka from its outline, while the backbone selects Latvia.
Figure 17: Regional counting and roof-color recognition at 4B. Left: ReDiS reports three circular structures in the specified lower-left region rather than two. Right: ReDiS selects red for the indicated triangular roof, correcting Qwen3.5’s white prediction.