Changearthmore/rig-bench-matched-conditions

RIG-Bench — matched four-condition experiment (results + code)

Everything behind the two rebuttal sections "Q3. Reasoning failure versus image rendering limitation" and "W2. What image-output evaluation reveals beyond text-output reasoning". Both are the same experiment, reported in different units. The self-reflection experiment reuses the same 200 items and the same pipeline, so it is included under self_reflection/.

Generated images are not included here. The D images are already published in Changearthmore/mllm-bench-outputs; every record below keeps its sample_id and image path so the two can be joined.

The four conditions

	Condition	What the model receives	Output
T	text solution	original visual context + task	structured text answer
D	direct image	original visual context + task	answer image — this is the paper's Table 3 protocol
H	own-plan image	original prompt + the model's own T answer, verbatim	answer image
O	oracle image	original prompt + a pseudo-gold answer specification (text only)	answer image

D is reused unchanged from the published runs, so it is byte-identical to Table 3; H and O were generated with each model's own published run_config (same seed, cfg, steps, max_input_images, resolution). T and H come from a single text call — H uses the model's own answer uncorrected, so it measures whether explicit thinking helps rather than whether a corrected plan helps.

Systems

Label	Reasoner (T)	Renderer (D/H/O)
bagel	BAGEL-7B-MoT	BAGEL-7B-MoT (unified)
qwen-pipeline	Qwen2.5-VL-72B-Instruct	Qwen-Image-Edit-2509
emu35-pipeline	Qwen2.5-VL-72B-Instruct	Emu3.5-Image

The two pipelines share the same reasoner and the same 200 text answers, so comparing them isolates the renderer.

Two judges

Everything is scored twice, with the same rubric, weighting, threshold and number of runs (n = 3, temperature 0.9):

• __qwen2.5-vl-72b — the judge used in the rebuttal tables.
• __qwen3.5-27b — an independent re-judge added for robustness.

Two caveats, both visible in coverage.json:

1. Qwen2.5-VL-72B is also the reasoner of the two pipeline settings, so their T scores are self-graded. The Qwen3.5 re-judge exists to check this: the pipelines' T advantage over BAGEL is +12.0/+13.5 pp under Qwen2.5 and +12.5/+14.0 pp under Qwen3.5, i.e. unchanged, so it is not self-preference.
2. Qwen3.5 scores T reliably only ~75% of the time (it is a thinking model and often fails to emit parsable JSON after the reasoning; 145–156 of 200 items get a score, versus 200/200 for Qwen2.5). Its image scores are essentially complete (194–200 per condition). Prefer Qwen2.5 for T.

Image judging is blind: the generated image is copied to a hash-named file and the judge sees no model or condition label. It does see the ground-truth image, matching the paper's protocol.

D/H/O use the paper's image rubric against the ground-truth image. T has no ground-truth image, so it uses a separate text-solution correctness rubric against the pseudo-gold specification — the T column is comparable across systems but not against the D/H/O columns.

Layout

matched/
  inputs/
    selected_items_200.jsonl        the frozen 200 items (50 per family, all 11 subtasks)
    pilot_items_11.jsonl            pilot set used to freeze the prompts
    prompt_freeze.json              prompt versions/hashes
    text_solutions_T.*.jsonl        the raw T answers (bagel / the shared Qwen reasoner)
    gold_solution_manifest.jsonl    the selected oracle specification per item
    gold_manifest_candidates.jsonl  all K=5 candidates
    gold_manifest_candidate_scores.jsonl
  <system>/
    image_runs.jsonl                per-image generation record (params, seed, size, path)
    blind_packets.jsonl             blind_id -> decode mapping
    text_scores__<judge>.jsonl      per-item T scores (3 runs each)
    image_scores__<judge>.jsonl     per-image D/H/O scores (3 runs each, per-dimension)
    matched_summary__<judge>__tau{3,4}.csv    pass rates, CIs, rescue/regression, integration gap
    failure_taxonomy__<judge>__tau{3,4}.csv
    perceptual_metrics.jsonl        DINO / CLIP-I / LPIPS per image vs ground truth
  summary__<judge>.md               headline table on the judge-score scale (composite x20)

self_reflection/
  r1|r2/  image_runs, critiques, blind_packets, image_scores__<judge>, perceptual_metrics
  summary__<judge>.md

code/
  cascade/            01..12 pipeline, adapters/, prompts/
  self_reflection/    01..06 pipeline, prompts/
  judge/              score_llm_judge.py, compute_image_metrics.py, rubrics_v2.yaml
  infra/              driver scripts for the actual runs
coverage.json         row/score counts per file, per judge

Reproducing

code/infra/score_model.sh <system> <published_run_dir> <text_solutions> <image_runs> runs 06b (reuse published D) → 07 (judge T) → 08 (blind packets) → 09 (judge images) → 10 (matched analysis). code/infra/rejudge35*.sh re-runs the judging step with a different judge against an OpenAI-compatible endpoint. Adapters need their own environments (BAGEL, Emu3.5 via vLLM, Qwen-Image-Edit via diffusers); see the headers of code/cascade/adapters/*.py.

Known issues

• image_scores__qwen3.5-27b.jsonl has slightly more than 600 rows: failed rows were re-judged and appended rather than overwritten. Key by (sample_id, condition) and keep the last row with a non-null composite; 10_analyze_matched_conditions.py already does this.
• One Emu3.5 item lacks a complete D/H/O triple under Qwen2.5 (n = 199).
• LPIPS is offset by a roughly constant amount from the published Table 3 values (e.g. BAGEL D 0.634 vs 0.67) in a way that subset sampling does not explain — most likely a different LPIPS backbone or input normalisation. DINO and CLIP-I do fall within the sampling interval of the published values. All comparisons here are within a column, so a constant offset cancels.