code/README.md

# OccludeBench: Can VLMs See the Hint?

Code for the experiments in the paper, organized by paper section.

## Setup

```bash
pip install -r requirements.txt
cp .env.example .env   # fill in API keys
python prepare_data.py # lays out the released images the way the scripts expect
```

Run everything from this `code/` folder, e.g. `python sec3_occludebench/eval_script.py --model gpt-5.4`. The shell wrappers `cd` here themselves, and each notebook does so in its first cell.

Models: `gpt-5.4`, `claude-opus-4-6`, `gemini-3.1-pro-preview`, and Llama 4 Maverick / Qwen3-VL-8B via Together.

## Layout

### §3 OccludeBench: `sec3_occludebench/`

| Paper | Code |
|---|---|
| Dataset distribution (§3.1) | `image_category_plots.ipynb` |
| Main occluded query + reflection follow-up (§3.2, App. A) | `eval_script.py` |
| Occluder follow-up for Llama / Qwen responses that name the mask itself | `vetter_script.py` |
| Two-judge grading, Claude Opus 4.6 + GPT-5.4 (§3.2, App. A) | `judger_script.py` |
| Human baseline grading (§3.3) | `human_baseline/judger_human.py` |
| Human accuracy and per-image difficulty distribution (§4.1, App. D) | `human_baseline/human_accuracy_breakdown_pub_quality.ipynb` |

### §4.1 Main Results: `sec4_1_main_results/`

| Paper | Code |
|---|---|
| Occluded + unmasked runs, all five models | `run_all_evals.sh` |
| Accuracy tables (occluded / unmasked, human solvability) | `accuracy_breakdown_pub_quality.ipynb` |
| Fig. 4: accuracy across base / thinking / tool use / steering | `nudge_comparison.ipynb` |
| Localization control, unmasked bounding box (Table 1) | `localization_unmasked/generate_transparent.py`, `localization_unmasked/eval_transparent.py` |
| Per-category accuracy + chi-square tests (Fig. 5, App. B) | `category_accuracy/accuracy_by_category.ipynb` |
| Model vs. human difficulty (Fig. 6) | `human_model_correlation/human_model_difficulty_correlation.ipynb` |

### §4.2 Why do VLMs fail?: `sec4_2_why_vlms_fail/`

| Paper | Code |
|---|---|
| Tool-trace failure modes (App. H) | `tool_trace_failures/build_visualizer.py` |
| Cue removal (Table 1) | `cue_removal/run_cue_removed_eval.sh`, `cue_removal/accuracy_breakdown_physical.ipynb` |
| Seeing failure rate on the highlighted projection region (Table 1) | `seeing_failure/eval_wrong_transparent.py`, `seeing_failure/judge_wrong_transparent.py` |
| Stage 1: cue detection (Table 2) | `stage1_cue_detection/` |
| Stage 2: zoomed-in object recognition (Table 2) | `stage2_zoomed_recognition/` (`app.py` is the manual cropping tool) |
| Stage 3: source mapping (Table 2) | `stage3_source_mapping/` (`map_eval.py`, `reverse_map_eval.py`) |
| Stage 4: highlighted-region recovery (Table 2) | `stage4_highlighted_region/` |

### §4.3 Can we improve performance on direct images?: `sec4_3_improvements/`

| Paper | Code |
|---|---|
| Fig. 8: improvement summary | `accuracy_improvements.ipynb` |
| Extended reasoning | `extended_reasoning/` |
| Tool use + tool-call analysis (Fig. 7, App. G) | `tool_use/` (`run_tool_use_eval.sh`; the paper's run used `OUTPUT_DIR=data/eval_results_tool_use_thinking_v2`) |
| Agentic harness, Codex + GPT-5.4 | `coding_agents/coding_agent_eval.sh` (needs the Codex CLI) |
| Steering prompts, short and long (App. A) | `steering_prompts/run_nudge_eval.sh`, `steering_prompts/run_occlusion_nudge_eval.sh` |
| Few-shot prompting | `few_shot/` |
| Inpainting multi-step execution (Table 3) | `inpainting_multistep/` (`gap_analysis_final.py` computes the Fisher tests) |

### Appendix: `appendix/`

| Paper | Code |
|---|---|
| Sensitivity to bounding-box style (App. C) | `box_sensitivity/` |
| Lexical analysis of answers (App. D) | `lexical_metrics/answer_lexical_metrics.py` |
| Token count vs. accuracy (App. E) | `token_length/token_length_vs_correctness_all_models_no_outliers.ipynb` |
| Image understanding vs. generation (App. F) | `image_generation/eval_image_gen.py` |
| Learning effects in the human survey (App. I) | `human_learning_effects/learning_effect_test.py` |

## Paths

`prepare_data.py` builds `data/combined_output/` (`annotations.json` plus `<id>.<ext>`, `<id>_occluded.png`), `data/combined_output_blue/`, `data/annotated/` (highlighted cue region) and `data/Image categories (3).xlsx` from the dataset one level up. For the localization control, `python sec4_1_main_results/localization_unmasked/generate_transparent.py` draws the unmasked `<id>_transparent.png` images. It also writes the stage 2 crops (`crop_meta.json`) and stage 3 source-mapping images (`regions_reverse.json`) into their `sec4_2_why_vlms_fail/` folders. Scripts write results to `data/eval_results*/` and figures to `data/figures/`.