Panda-CVL
Models accept correct image-grounded responses or locate and replace their first erroneous text span. The test expands 652 conversations into 2,126 instances. Its overall F1 combines correct acceptance with exact correction of faulty responses.
The separate training split contains 6,839 annotation sessions. Correction triples use a fixed tokenizer and human reference edits. Multiple valid correction positions can limit exact-match agreement.
- Made by
- StepFun and Xiamen University
- Released
- September 2026
- Announced on X
- 20K views, 343 likes
- Kind
- General visual questions
- Level
- University
- Scored
- Checked automatically
- Answers
- Short answer
- Tasks
- All public
- Measures
- Capability
- Language
- English
- Interaction
- Single turn
- Refresh
- One fixed set
- Run by
- Anyone
- In views
- Vision
- Strongest model tested
- GPT-5.6 Sol #14 on the capability index
- Last model added
- September 2026
- Scores refreshed
- November 20th, 2026
In the authors’ words
We present onPanda, an interactive tool for efficiently annotating LLM alignment data and agent trajectories. onPanda adopts token-level correction as its core interaction: while reading a model response, the annotator locates the first inappropriate token and either picks a substitute from the model's candidate tokens or types the correct text via free-form editing. The system then truncates everything after that position and continues generation from the corrected prefix, repeating this locate-correct-continue loop until a satisfactory response is obtained. This mechanism lets annotators precisely steer model outputs at low cost: a small controlled study suggests that onPanda reduces median annotation time by 52% over manual post-editing. Since the vast majority of tokens in the final response are generated by the model itself, the resulting data largely preserves the model's sampling distribution and is well suited for constructing on-policy SFT and preference data. Furthermore, the token-level corrections recorded during annotation provide fine-grained supervision with precise positions and naturally paired positive–negative samples. onPanda also connects to external tools and harnesses, enabling interactive trajectory annotation in realistic environments. In addition, we release Panda-CVL, a dataset annotated with onPanda, together with a benchmark for token-level correction.
GPT-5.5 leads with 17.1% among models released by September 2026 GPT-5.5 leads with 17.1% among models released by September 2026
Best score per entry from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. Where a model ran with several settings or harnesses, we keep its best. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. This view shows only models we track, each at its best.
The newest model with a score here came out in September 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- Claude Opus 5.5
- and 15 more
- 1 GPT-5.5 Reasoning; effort xhigh 17.1%
- 2 GPT-6 (variant unspecified) Reasoning; effort xhigh 16.6%
- 3 Kimi K2.6 Reasoning 15.1%
- 4
Step 5 Preview Reasoning 15%
- 5 GPT-5.6 Sol Reasoning; effort xhigh 14.6%
- 6
Qwen3.5 397B-A17B Reasoning 14.3%
- 7
Doubao Seed 2.1 Pro Reasoning 11.3%
- 8
Qwen 3.6 35B-A3B Reasoning 9.36%
- 9
Step 3.7 Flash Reasoning 6.13%
Source: The Panda-CVL paper
- 1 GPT-5.5 Reasoning; effort xhigh 17.1%
- 2 Kimi K2.6 Reasoning 15.1%
- 3
Step 5 Preview Reasoning 15%
- 4 GPT-5.6 Sol Reasoning; effort xhigh 14.6%
- 5
Qwen3.5 397B-A17B Reasoning 14.3%
- 6
Doubao Seed 2.1 Pro Reasoning 11.3%
- 7
Qwen 3.6 35B-A3B Reasoning 9.36%
- 8
Step 3.7 Flash Reasoning 6.13%
Source: The Panda-CVL paper
| # | Entry | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | GPT-5.5 | 17.1% | Reasoning; effort xhigh | Apr 2026 | Official leaderboard |
| 2 | GPT-6 (variant unspecified) | 16.6% | Reasoning; effort xhigh | Official leaderboard | |
| 3 | Kimi K2.6 | 15.1% | Reasoning | Apr 2026 | Official leaderboard |
| 4 | | 15% | Reasoning | Sep 2026 | Official leaderboard |
| 5 | GPT-5.6 Sol | 14.6% | Reasoning; effort xhigh | Jul 2026 | Official leaderboard |
| 6 | | 14.3% | Reasoning | Feb 2026 | Official leaderboard |
| 7 | | 11.3% | Reasoning | Jun 2026 | Official leaderboard |
| 8 | | 9.36% | Reasoning | Apr 2026 | Official leaderboard |
| 9 | | 6.13% | Reasoning | May 2026 | Official leaderboard |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | GPT-5.5 | 17.1% | Reasoning; effort xhigh | Apr 2026 | Official leaderboard |
| 2 | Kimi K2.6 | 15.1% | Reasoning | Apr 2026 | Official leaderboard |
| 3 | | 15% | Reasoning | Sep 2026 | Official leaderboard |
| 4 | GPT-5.6 Sol | 14.6% | Reasoning; effort xhigh | Jul 2026 | Official leaderboard |
| 5 | | 14.3% | Reasoning | Feb 2026 | Official leaderboard |
| 6 | | 11.3% | Reasoning | Jun 2026 | Official leaderboard |
| 7 | | 9.36% | Reasoning | Apr 2026 | Official leaderboard |
| 8 | | 6.13% | Reasoning | May 2026 | Official leaderboard |
Rankings by part
We read 2 parts from the benchmark’s paper: GoodAcc and Corr.-NG. Each ranks models on one part of the test. GPT-5.5 leads GoodAcc and Doubao Seed 2.1 Pro leads Corr.-NG. Arrows show how far a model moved from its overall rank among the models we track.
GoodAcc
8 models we track have a score here.
- 1 GPT-5.5 53.4%
- 2
Step 3.7 Flash ↑6 40.1%
- 3
Step 5 Preview 32.2%
- 4 Kimi K2.6 ↓2 26.4%
- 5
Qwen 3.6 35B-A3B ↑2 26.2%
- 6
Qwen3.5 397B-A17B ↓1 20.4%
- 7 GPT-5.6 Sol ↓3 16.6%
- 8
Doubao Seed 2.1 Pro ↓2 9.66%
Corr.-NG
8 models we track have a score here.
- 1
Doubao Seed 2.1 Pro ↑5 13.7%
- 2 GPT-5.6 Sol ↑2 13.1%
- 3
Qwen3.5 397B-A17B ↑2 11%
- 4 Kimi K2.6 ↓2 10.6%
- 5 GPT-5.5 ↓4 10.2%
- 6
Step 5 Preview ↓3 9.77%
- 7
Qwen 3.6 35B-A3B 5.70%
- 8
Step 3.7 Flash 3.32%
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: The Panda-CVL paper
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official Panda-CVL leaderboard
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
More in Vision
- MMMU In use 11.5K college-level multimodal questions across 30 subjects. Multiple-choice and open accuracy.
Qwen 3.6 Plus 86%
- MMMU-Pro In use Harder MMMU: college-level multimodal questions with more options and vision-only inputs to remove text shortcuts. GPT-5.6 Sol 84.6%
- CharXiv In use Chart understanding on real arXiv figures. Reasoning questions require synthesizing information from complex charts. Claude Fable 5 93.5%
- MathVista In use Math reasoning in visual contexts (charts, geometry, puzzles) from 31 datasets. Answer accuracy.
Seed2.1 Pro 90.7%
- MATH-Vision (MathVision) In use 3,040 competition math problems with visual context. Answer accuracy.
Qwen 3.8 Max 97.7%
- MMStar In use Models answer 1,500 human-curated image questions that cannot be solved without the image, across 6 core capabilities. Scored by accuracy.
Qwen 3.8 Max 85.9%
Data audit
All 10 checks pass. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 4 pass 1 note
- Pass Independent runs Scores from the benchmark’s paper (9 entries).
- Pass Models scored 8 models we track, plus 1 other entry on its leaderboard.
- Pass Newest score The newest is from September 2026, for Step 5 Preview.
- Note Frontier models None of the top 10 has a score. The strongest tested is GPT-5.6 Sol (#14).
- Pass Headroom The best is 17.1%, under 85%, so it still separates models.
Leaderboard import 3 pass
- Pass Importer We read the board on arxiv.org: 9 rows for models we track and 1 other entry at the last read.
- Pass Last read November 20th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.