Test 2 — Grain Robustness Results
Test 2 measures how closed-loop accuracy changes when deterministic monochrome Gaussian grain is added to the benchmark images. The evaluation uses the fixed 680-image core subset: 20 images from each of 34 domains, with five questions per image, giving 3,400 evaluated question rows per model at each noise level.
Noise conditions
- σ = 0: original image, without added grain.
- σ = 15, 25, 40: one deterministic monochrome Gaussian-noise field is added equally to all three RGB channels.
- Every model receives the same altered pixels for a given image and noise level.
- Δ is the change from σ = 0 to σ = 40. Negative values indicate an accuracy loss under the strongest tested grain.
Accuracy at each noise level (%)
| Rank at σ=0 | Model | σ=0 | σ=15 | σ=25 | σ=40 | Δ |
|---|---|---|---|---|---|---|
| 1 | Gemini 3.8 Flash | 64.1 | 65.5 | 64.7 | 64.0 | −0.1 |
| 2 | Claude Opus 5 | 62.4 | 59.9 | 58.6 | 56.2 | −6.1 |
| 3 | Grok 4.6 | 52.7 | 51.2 | 51.3 | 50.4 | −2.4 |
| 4 | GPT-5.6 Sol | 50.8 | 50.9 | 51.0 | 49.8 | −1.1 |
| 5 | Claude Sonnet 5 | 49.7 | 48.4 | 47.6 | 47.1 | −2.6 |
| 6 | Muse Glimmer 30B | 44.4 | 42.4 | 43.6 | 42.4 | −1.9 |
| 7 | DeepSeek V4.1 Flash | 43.2 | 39.3 | 40.1 | 39.1 | −4.1 |
| 8 | GPT-5.6 Luna | 42.4 | 41.8 | 42.0 | 40.3 | −2.1 |
| 9 | Perplexity Sonar Pro | 41.8 | 41.4 | 40.9 | 39.4 | −2.4 |
| 10 | Inkling | 38.7 | 36.8 | 35.7 | 34.4 | −4.3 |
Main findings
- Gemini 3.8 Flash is both the most accurate model and the most robust to added grain. Its reported change at σ = 40 is only −0.1 percentage points.
- GPT-5.6 Sol has the second-smallest degradation at −1.1 points.
- Claude Opus 5 experiences the largest reported degradation, falling from 62.4% to 56.2%, with a reported Δ of −6.1 points.
- Inkling and DeepSeek V4.1 Flash also show comparatively large losses of −4.3 and −4.1 points, respectively.
- Small non-monotonic changes at intermediate noise levels are visible for several models. For example, Gemini improves at σ = 15, while GPT-5.6 Sol and Muse Glimmer 30B partially recover at σ = 25.
- The ranking is broadly stable under the strongest grain: Gemini remains first, Claude Opus remains second, and Inkling remains last.
Rounding note
The values above are transcribed from the supplied result table. Accuracy columns and Δ were calculated from underlying unrounded values before display, so subtracting the displayed one-decimal endpoints may differ from the displayed Δ by 0.1 point for some models.