ULVR-Bench — Universal Latent Visual Reasoning Benchmark

2,888 samples across 7 reasoning-type splits. Final-answer accuracy is the only graded metric. Built via Qwen2.5-VL-7B helper-helps filtering on top of heuristic + manual quality audit.

Split	Samples	Audit pool	Source	Filter	Intermediate state GT	Answer GT
text_reasoning	500	1,526	ScienceQA test	helper-helps filter	✅ lecture + solution text	✅ (A/B/C/D)
bbox_reasoning	400	910	GQA val_balanced	helper-helps filter	✅ bbox-list JSON from scene graph	✅
scene_graph_reasoning	700	3,707	GQA val_balanced	helper-helps filter	✅ full scene-graph JSON	✅
visual_repr_reasoning	200	587	GQA val_balanced	helper-helps filter	✅ 3 aux images: depth + edge + seg	✅
helper_reasoning	88	88	GeoQAPlus-StepbyStep test (strict aux-line)	all kept	✅ Chinese rationale; ⏳ aux-line image (TODO manual)	✅
chart_reasoning	500	2,375	ChartQA test	no-helper-wrong proxy filter	⏳ region coords (TODO manual)	✅
doc_reasoning	500	4,890	DocVQA validation	no-helper-wrong proxy filter	⏳ region coords (TODO manual)	✅

Helper-helps filter — key contribution

We ran Qwen2.5-VL-7B-Instruct on each audit-pool sample twice:

• no_helper: input_image + question only
• with_helper: input_image + question + intermediate_state (text/JSON, or 3 aux images for VR)

Then categorize each sample by (no_helper_ok, with_helper_ok):

• P1 helper-helps (no=Wrong, with=Right): helper genuinely guided model → KEEP all
• P2 both correct: helper at least doesn't hurt → KEEP if room left
• P3 both wrong: ambiguous → KEEP if room left
• AVOID helper-hurts (no=Right, with=Wrong): helper misleads → DROP all

Per-split breakdown:

Split	P1 helps	P2 both✓	P3 both✗	AVOID hurts (dropped)
helper_reasoning	26	61	1	0
visual_repr_reasoning	24	355	157	51
bbox_reasoning	122	368	331	89
text_reasoning	192	1,294	32	8
scene_graph_reasoning	772	1,837	897	201

Total dropped 349 samples where helper actively HURT model performance.

For SG, 772 P1 already exceeds 700-target → use P1 only (SHA1-sorted). For other splits, all P1 kept + remainder filled from P2 (deterministic SHA1 sort).

chart_reasoning and doc_reasoning have no GT intermediate (manual annotation pending). Proxy: no_helper_wrong → P1-equivalent, no_helper_right → P2.

Qwen2.5-VL-7B baseline accuracy per split

Split	no_helper acc	with_helper acc	Δ
helper_reasoning	69.3%	98.9%	+29.6%
scene_graph_reasoning	55.0%	70.4%	+15.4%
text_reasoning	85.3%	97.4%	+12.1%
bbox_reasoning	50.2%	53.8%	+3.6%
visual_repr_reasoning	69.2%	64.6%	-4.6% (aux images confuse the 7B model)
chart_reasoning	81.1%	n/a	n/a
doc_reasoning	91.1%	n/a	n/a

Why these sources?

• GQA val (not testdev): testdev has no public scene-graph GT. val_balanced has val_sceneGraphs.json and is not used by ULVR training.
• GeoQAPlus-StepbyStep: Chinese middle-school geometry filtered to 88 problems where the solution explicitly performs an aux-line construction (连接 / 延长 / 过…作 / 作…垂线 in 解题过程 only, non-greedy, exclude pre-drawn).
• DocVQA validation (not test): test has no public GT (leaderboard-only).
• MIRA-inspired design (YiyangAiLab/MIRA): test "visualization → reasoning". We use public-test items + automated filter, not hand-crafted problems.

Schema

bench_id: str
reasoning_type: str
question: str
answer: str
input_image: {bytes, path}
intermediate_state_type: str
intermediate_state: str       # GT text/JSON; empty for chart/doc (TODO manual)
intermediate_image: {bytes, path}  # NULL for splits without GT image
source: str

visual_repr_reasoning adds intermediate_image_1/_2/_3: depth (DepthAnything-V2-Large), edge (HED), seg (SAM 2.1 hiera-large top-8 masks @ 50% alpha).

Usage

from datasets import load_dataset
bench = load_dataset("RuoliuYang/ULVR-Bench", split="scene_graph_reasoning")

Pipeline summary

Raw HF sources → 18,218 audit pool
   ↓ heuristic + manual quality audit (BENCH_QUALITY_AUDIT.md)
14,083 cleaned
   ↓ Qwen2.5-VL-7B helper-helps filter (this README)
2,888 final

Known limitations

1. Filter uses Qwen2.5-VL-7B as the test model — biased toward what a 7B model finds hard. A 70B+ model might find some "hard" samples trivial.
2. GT answer correctness not LLM-judged (no API key); some GQA / DocVQA val GT may be noisy.
3. Chart/doc filter uses no_helper_wrong proxy since no GT intermediate exists yet.
4. VR helper proves net-negative on average (-4.6%); 24 P1 samples kept but format may need re-design.

See BENCH_QUALITY_AUDIT.md for manual audit log.