MIRAGE NeurIPS Prior Harnessing Index results
Public, results-only release for the completed expanded Prior Harnessing Index
(PHI) run full_phi_cais_20260728_v3.
Completion
- 25/25 model-by-benchmark cells completed successfully.
- 31/31 orchestration tasks completed successfully.
- Five models and five benchmarks are included.
- Row-level PHI scores and diagnostics are under
run/cells/. - Final compact and machine-readable tables are under
run/tables/.
The source response artifacts are published separately at
JokeMaker/mirage-neurips-artifacts.
Cluster logs, model caches, scheduler metadata, failed intermediate attempts,
local usernames, and absolute filesystem paths are excluded.
Most cells use eager attention. Two high-resolution cells that exceeded a
single A100's eager-attention peak memory were completed with mathematically
equivalent scaled-dot-product attention (SDPA):
InternVL3.5-4B-HF / MedXpertQA and MiMo-VL-7B-RL / MMMU-Pro.
Final table
Expanded Prior Harnessing Index (PHI)
This reproduces the original-paper Table 4a procedure for all 25 expanded model/benchmark combinations. Each cell selects up to 100 mirage-like and 100 non-mirage-like contrastive rows in response-artifact order (no shuffle) (seed 42; non-original mutations allowed), scores the first ground-truth answer token after [[, and uses the exact null question What is the Answer?.
The compact tables report no-image PHI. Full row-level with-image/no-image scores, selection counts, missing-row diagnostics, and auxiliary metrics remain in each cell's phi_* artifact directory.
Ovis2.5-2B
| Dataset | Mirage Corr. | Mean PHI |
|---|---|---|
| VQA-RAD | -0.36 | -0.45 |
| MMMU-Pro | -0.21 | 1.21 |
| MedXpertQA | 0.09 | 0.49 |
| ChartQA | 0.05 | -1.35 |
| A-OKVQA | -0.02 | -1.94 |
Qwen3-VL-32B-Instruct
| Dataset | Mirage Corr. | Mean PHI |
|---|---|---|
| VQA-RAD | -0.17 | -0.00 |
| MMMU-Pro | -0.09 | 0.00 |
| MedXpertQA | 0.21 | 0.00 |
| ChartQA | 0.06 | -0.86 |
| A-OKVQA | -0.05 | -0.91 |
InternVL3.5-4B-HF
| Dataset | Mirage Corr. | Mean PHI |
|---|---|---|
| VQA-RAD | 0.19 | 2.84 |
| MMMU-Pro | -0.09 | -0.71 |
| MedXpertQA | 0.02 | 0.06 |
| ChartQA | 0.53 | 0.96 |
| A-OKVQA | 0.23 | -0.10 |
MiMo-VL-7B-RL
| Dataset | Mirage Corr. | Mean PHI |
|---|---|---|
| VQA-RAD | -0.13 | -0.00 |
| MMMU-Pro | 0.13 | -0.00 |
| MedXpertQA | -0.00 | -0.00 |
| ChartQA | 0.08 | -0.43 |
| A-OKVQA | 0.05 | 0.15 |
Gemma-3-4B-IT
| Dataset | Mirage Corr. | Mean PHI |
|---|---|---|
| VQA-RAD | 0.15 | 4.27 |
| MMMU-Pro | 0.10 | 2.45 |
| MedXpertQA | 0.13 | 1.17 |
| ChartQA | 0.09 | -1.50 |
| A-OKVQA | 0.13 | -0.18 |
- Downloads last month
- 13