MM-SystemOneBias: Evaluating Bias in Multimodal Jev-Style System One Decision Models
Organizations: School of Artificial Intelligence, Jilin University · International Center of Future Science, Jilin University · King Abdullah University of Science and Technology (KAUST) · School of Big Data and Software Engineering, Chongqing University ·
Abstract
When multimodal decision models determine actions, moderation outcomes, or quality judgments, their reliability depends on preserving semantic decisions across presentation changes, ignoring irrelevant additions, and distinguishing information consistent with task evidence from information that conflicts with it. Existing evaluations study decision-interface biases and multimodal model behavior in separate settings, leaving no common protocol to compare multimodal decision models with their backbones on practical tasks. We introduce MM-SystemOneBias, which evaluates 13 models on 5,323 original instances from nine public datasets and 11 task scenarios using six paired interventions: option order, option label, external judgment, image--text description, irrelevant information, and rating scale. Across seven matched decision-model--backbone pairs, decision models are more stable under option-label, irrelevant-information, and rating-scale changes in all seven pairs; six pairs also show lower option-order flip rates. They incur smaller accuracy losses under incorrect external judgments in all seven pairs and conflicting descriptions in six. In six pairs, they gain less accuracy from correct external judgments and show less favorable accuracy changes under consistent descriptions. Among pairs with unchanged semantic decisions, mean absolute changes in answer probabilities range from 4.56 to 11.55 percentage points, revealing input effects that accuracy alone misses.
Figures & tables
| Bias type | Manipulated factor and conditions |
|---|---|
| Option Order | Candidate positions; original versus reordered candidates |
| Option Label | Candidate display labels; original versus relabeled candidates |
| External Judgment | Added conclusion; correct or incorrect judgment across source conditions |
| Image–Text Description | Added visual claim; consistent or conflicting with the image |
| Irrelevant Information | Appended background records; 1, 2, or 4 records |
| Rating Scale | Numeric encoding of ordered grades; alternative increasing scales |
| Model | Overall Performance | Bias Interventions | ||||||
|---|---|---|---|---|---|---|---|---|
| Original Acc. (%) | Mean Acc. Drop (pp) | Option Order FR (%) | Option Label FR (%) | External Judgment EIR (%) | Image–Text Description Acc. Gap (pp) | Irrelevant Information FR (%) | Rating Scale Shift (%) | |
| Backbone Models | ||||||||
| Qwen 27B | 73.52 | 11.88 | 24.19 | 22.11 | 47.10 | 34.34 | 13.31 | 8.66 |
| Qwen 9B | 77.43 | 14.46 | 15.62 | 14.16 | 47.57 | 36.30 | 14.21 | 6.22 |
| Qwen 4B | 72.66 | 17.05 | 22.20 | 13.92 | 61.33 | 53.76 | 20.39 | 7.55 |
| LFM 3B | 73.65 | 7.65 | 19.18 | 6.23 | 25.03 | 31.81 | 14.94 | 15.59 |
| Intervention | pp | NLL | Brier | ||
|---|---|---|---|---|---|
| Option order | 6.28 | 20.70 | – | 0.002 | 0.002 |
| Option label | 4.56 | 13.99 | – | 0.024 | 0.006 |
| External judgment (Inc.) | 11.55 | 38.98 | 10.10 | 0.575 | 0.154 |
| External judgment (Cor.) | 8.67 | 29.14 | 6.26 | ||
| Image–text (Conf.) | 9.73 | 34.62 | 7.24 | 0.319 | 0.097 |
| Image–text (Cons.) | 6.42 | 22.45 | 4.11 |
Appendix figures & tables33 assets
Supplementary material from the paper’s appendix.
Appendix
| Quantity | Count |
|---|---|
| Dataset coverage | |
| Public source datasets | 9 |
| Task scenarios | 11 |
| Source and original instances | |
| Source instances | 4,228 |
| Original instances | 5,323 |
| Bias type | Construction | Verification or retention criterion |
|---|---|---|
| Option order | Reorder candidates and assign display labels by position. | Candidate identities and reference-answer mapping are preserved. |
| Option label | Replace letter labels with numeric labels or reverse the letter assignment. | Candidate order, content, and semantic identities are preserved. |
| External judgment | Append a fixed-source conclusion that selects the reference answer or a designated incorrect target. | The conclusion maps to the designated semantic answer. |
| Image–text description | Add a statement supported by or contradictory to relevant visual evidence. | The supported statement is judged supported and the contradictory statement judged contradicted at the stated level of specificity. |
| Irrelevant information | Select four items from other instances, then form nested sets of one, two, and four items. | Every item and every nested set is judged free of answer cues, task-relevant evidence, and extra task requirements. |
| Rating scale | Assign alternative numeric scales to the same ordered quality descriptions. | Each description remains attached to the same semantic grade. |
| Source dataset | Task scenarios | Selection unit | Native split or partition rule | Source instances |
|---|---|---|---|---|
| Multimodal-Mind2Web ( Zheng et al., 2024 ) | GUI next-action selection | Web task | Native test split; task-level isolation | 400 |
| AgentRewardBench ( Lù et al., 2025 ) | GUI task completion judgment | Benchmark/task pair | Source groups shuffled before sampling; final three trajectory frames retained | 368 |
| Robo2VLM-1 ( Chen et al., 2025 ) | Robot next-action selection, robot task success judgment, and robotic grasping stage | Robot trajectory | Native test split; trajectory-level isolation; all eligible robotic grasping stage items retained | 1,229 |
| RVL-CDIP ( Harley et al., 2015 ) | Document type classification | Source image | Native test split; 25 images per class; image-level isolation | 342 |
| LlavaGuard ( Helff et al., 2024 ) | Image safety moderation | Source image and policy row | Public test annotations; image-level grouping; same images support both interfaces | 295 |
| MVTec AD ( Bergmann et al., 2019 ) | Industrial anomaly detection | Test image with category-matched normal reference | Native test split; category-stratified reference pools | 394 |
| Task family | Task scenario | Source dataset | Interface | Model output | Images | Source instances | Original instances |
|---|---|---|---|---|---|---|---|
| GUI and robotics | GUI next-action selection | Multimodal-Mind2Web | Choice | Next interface element | 1 | 400 | 400 |
| GUI task completion judgment | AgentRewardBench | Yes/no judgment | Task completion | 3 | 368 | 368 | |
| Robot next-action selection | Robo2VLM-1 | Choice | Next robot action | 1 | 400 | 400 | |
| Robot task success judgment | Robo2VLM-1 | Yes/no judgment | Task success | 1 | 347 | 347 | |
| Robotic grasping stage | Robo2VLM-1 | Choice | Grasping stage | 1 | 482 | 482 | |
| Documents and safety | Document type classification | RVL-CDIP | Choice | Document category | 1 | 342 | 342 |
| Task | Reference label or grade | Count |
| GUI task completion judgment | Completed | 171 |
| Incomplete | 197 | |
| Robot task success judgment | Successful | 150 |
| Unsuccessful | 197 | |
| Image safety moderation | Policy violation | 104 |
| Policy compliant | 191 |
| Bias type | Intervention unit | Unit count | Conditions per unit | Intervened instances |
|---|---|---|---|---|
| Option order | Eligible choice instances | 3,119 | 1 or 3 | 6,367 |
| Option label | Eligible choice instances | 3,119 | 2 | 6,238 |
| External judgment | Original instances | 5,323 | 8 | 42,584 |
| Image–text description | Description pairs | 1,863 | 2 | 3,726 |
| Irrelevant information | Background sets | 5,308 | 3 nested sets | 15,924 |
| Rating scale | Scoring instances | 400 | 4 | 1,600 |
| Model | Backbone | Access | Source |
| Backbone Models | |||
| Qwen3.8-27B ( Qwen Team, 2026b ) | – | Weights | Qwen/Qwen3.8-27B |
| Qwen3.5-9B ( Qwen Team, 2026a ) | – | Weights | Qwen/Qwen3.5-9B |
| Qwen3.5-4B ( Qwen Team, 2026a ) | – | Weights | Qwen/Qwen3.5-4B |
| LFM2.5-VL-3B ( Liquid AI, 2026a ) | – | Weights | LiquidAI/LFM2.5-VL-3B |
| Gemma 4 12B IT ( Gemma Team, 2026 ) | – | Weights | google/gemma-4-12B-it |
| Model | FR (%) | Acc (pp) | Consistency across all evaluated orders (%) |
|---|---|---|---|
| Backbone Models | |||
| Qwen 27B | 24.19 | -0.21 | 68.20 |
| Qwen 9B | 15.62 | -0.35 | 77.49 |
| Qwen 4B | 22.20 | +0.72 | 71.79 |
| LFM 3B | 19.18 | -0.75 | 74.70 |
| Gemma 12B | 25.64 | -0.86 | 65.57 |
| Model | Numeric labels | Reversed letters | ||
|---|---|---|---|---|
| FR (%) | Acc (pp) | FR (%) | Acc (pp) | |
| Backbone Models | ||||
| Qwen 27B | 23.09 | +0.02 | 21.14 | +1.33 |
| Qwen 9B | 12.16 | -3.59 | 16.16 | -2.92 |
| Qwen 4B | 12.08 | +0.47 | 15.76 | -2.77 |
| LFM 3B | 5.72 | -0.63 | 6.75 | -0.01 |
| Model | Historical system | User | User (strong tone) | Tool | ||||
|---|---|---|---|---|---|---|---|---|
| EIR (%) | ECR (%) | EIR (%) | ECR (%) | EIR (%) | ECR (%) | EIR (%) | ECR (%) | |
| Backbone Models | ||||||||
| Qwen 27B | 25.34 | 49.85 | 48.61 | 76.80 | 53.52 | 75.76 | 60.95 | 83.79 |
| Qwen 9B | 22.16 | 38.29 | 53.29 | 75.06 | 56.91 | 79.53 | 57.93 | 84.89 |
| Qwen 4B | 38.73 | 51.13 | 60.74 | 82.89 | 69.28 | 89.09 | 76.59 | 91.92 |
| LFM 3B | 11.54 | 26.12 | 30.04 | 62.30 | 29.88 | 61.62 | 28.64 | 61.29 |
| Model | Consistent description | Conflicting description | |||
|---|---|---|---|---|---|
| Acc (pp) | ECR (%) | Acc (pp) | EIR (%) | Target capture (%) | |
| Backbone Models | |||||
| Qwen 27B | +9.21 | 76.84 | -25.14 | 29.25 | 33.58 |
| Qwen 9B | +7.61 | 73.04 | -28.70 | 32.67 | 37.33 |
| Qwen 4B | +12.80 | 81.68 | -40.97 | 51.57 | 59.79 |
| LFM 3B | +11.30 | 69.73 | -20.50 | 31.01 | 36.43 |
| Model | One item | Two items | Four items | |||
|---|---|---|---|---|---|---|
| FR (%) | Acc (pp) | FR (%) | Acc (pp) | FR (%) | Acc (pp) | |
| Backbone Models | ||||||
| Qwen 27B | 14.20 | -5.34 | 13.51 | -3.84 | 14.96 | -4.41 |
| Qwen 9B | 14.48 | -6.12 | 13.53 | -5.64 | 13.72 | -5.11 |
| Qwen 4B | 20.54 | -4.02 | 18.68 | -2.86 | 18.73 | -2.28 |
| LFM 3B | 12.89 | -2.49 | 14.37 | -2.35 | 15.24 | -2.28 |
| Model | Mean absolute grade shift | Across all five scales | ||||
|---|---|---|---|---|---|---|
| – | – | – | – | Consistency (%) | ICC | |
| Backbone Models | ||||||
| Qwen 27B | 0.420 | 0.312 | 0.335 | 0.318 | 26.00 | 0.796 |
| Qwen 9B | 0.168 | 0.480 | 0.168 | 0.180 | 38.25 | 0.878 |
| Qwen 4B | 0.150 | 0.258 | 0.468 | 0.333 | 47.00 | 0.802 |
| LFM 3B | 0.490 | 0.205 | 0.357 | 1.442 | 31.25 | 0.606 |
| Model | Order | Label | Judg. Inc. | Judg. Cor. | Text Conf. | Text Cons. | Irrel. 1 | Irrel. 2 | Irrel. 4 | Scale |
|---|---|---|---|---|---|---|---|---|---|---|
| Backbone Models | ||||||||||
| Qwen 27B | 9.83 | 9.44 | 19.23 | 12.55 | 10.70 | 4.55 | 7.20 | 7.00 | 7.41 | 8.80 |
| Qwen 9B | 7.07 | 9.15 | 21.25 | 12.60 | 14.05 | 7.57 | 7.58 | 7.47 | 8.00 | 6.95 |
| Qwen 4B | 9.68 | 7.99 | 22.06 | 14.38 | 17.91 | 9.56 | 10.43 | 9.89 | 10.03 | 10.33 |
| LFM 3B | 7.68 | 3.66 | 9.59 | 7.57 | 9.01 | 6.80 | 6.49 | 7.37 | 7.78 | 16.05 |
| Gemma 12B | 6.69 | 5.65 | 7.05 | 5.60 | 3.10 | 2.24 | 4.12 | 4.45 | 4.50 | 7.13 |
| Model | Order | Label | Judg. Inc. | Judg. Cor. | Text Conf. | Text Cons. | Irrel. 1 | Irrel. 2 | Irrel. 4 | Scale |
|---|---|---|---|---|---|---|---|---|---|---|
| Backbone Models | ||||||||||
| Qwen 27B | 34.98 | 32.85 | 55.60 | 37.20 | 37.28 | 15.22 | 25.96 | 24.82 | 26.18 | 32.58 |
| Qwen 9B | 25.20 | 33.55 | 68.98 | 42.18 | 49.24 | 27.01 | 27.52 | 26.28 | 29.58 | 24.84 |
| Qwen 4B | 33.92 | 30.47 | 73.72 | 50.24 | 66.86 | 33.39 | 38.77 | 36.84 | 36.89 | 43.98 |
| LFM 3B | 27.25 | 12.17 | 33.56 | 25.96 | 30.89 | 21.58 | 23.75 | 25.86 | 27.47 | 70.19 |
| Gemma 12B | 22.99 | 20.04 | 21.90 | 17.84 | 9.13 | 6.85 | 14.34 | 15.40 | 15.55 | 26.07 |
| Model | Incorrect judgment | Correct judgment | Conflicting description | Consistent description |
|---|---|---|---|---|
| Backbone Models | ||||
| Qwen 27B | 17.79 | 10.79 | 9.68 | 4.14 |
| Qwen 9B | 19.90 | 10.09 | 12.72 | 7.23 |
| Qwen 4B | 20.73 | 12.28 | 15.72 | 8.60 |
| LFM 3B | 7.52 | 5.30 | 5.26 | 4.22 |
| Gemma 12B | 5.50 | 4.40 | 2.25 | 1.86 |
| Model | Incorrect judgment | Correct judgment | Conflicting description | Consistent description | ||||
| NLL | Brier | NLL | Brier | NLL | Brier | NLL | Brier | |
| Backbone Models | ||||||||
| Qwen 27B | 1.673 | 0.303 | 0.920 | 0.146 | ||||
| Qwen 9B | 0.859 | 0.343 | 0.386 | 0.180 | ||||
| Qwen 4B | 1.317 | 0.409 | 0.640 | 0.279 | ||||
| LFM 3B | 0.335 | 0.100 | 0.175 | 0.082 | ||||
| Model | Probe error | Conflict-task error | Probe correct: conflict-task error | Probe error: conflict-task error | |||
|---|---|---|---|---|---|---|---|
| Default | Image-first | Default | Image-first | Default | Image-first | ||
| Backbone Models | |||||||
| Qwen 27B | 8.99 | 38.96 | 28.57 | 35.26 | 23.70 | 59.29 | 59.29 |
| Qwen 9B | 15.17 | 40.63 | 29.55 | 37.38 | 23.95 | 64.33 | 65.00 |
| Qwen 4B | 23.86 | 54.92 | 45.92 | 50.54 | 38.47 | 68.17 | 69.16 |
| LFM 3B | 22.09 | 41.23 | 38.15 | 35.83 | 30.75 | 60.72 | 65.19 |
| Model | Order | Label | Judg. | Text | Irrel. | Scale | EIR/ECR |
|---|---|---|---|---|---|---|---|
| Decision models | |||||||
| AutoTrust | 6,367 | 6,238 | 19,692 | 1,813 | 15,924 | 1,600 | 15,140/2,852 |
| Clef 27B | 6,367 | 6,238 | 19,692 | 1,813 | 15,924 | 1,600 | 14,952/2,234 |
| Clef-Flash | 6,367 | 6,238 | 19,692 | 1,813 | 15,924 | 1,600 | 14,976/1,938 |
| d3 27B | 6,367 | 6,238 | 19,692 | 1,813 | 15,924 | 1,600 | 15,508/1,749 |
| d3-mini | 6,367 | 6,238 | 19,692 | 1,813 | 15,924 | 1,600 | 15,632/1,210 |
Multimodal Model Evaluation
Multimodal model evaluation measures how reliably AI systems understand and generate content across text, images, audio, video, and other data.
See all 800 papers