Reproduce of Pixel-Level Reasoning Model trained with RL [NeuIPS25]
active 2025-05-18 → 2026-05-01 (UTC)
Activity over time
Daily event counts in the loaded window
Line chart, 349 days from 2025-05-18 to 2026-05-01. Pushes: 59 total, peak 12 in a day. Pull requests: 0 total, peak 0 in a day. Issues: 40 total, peak 4 in a day. Comments: 52 total, peak 5 in a day. Stars: 159 total, peak 12 in a day.
- Pushes
- Pull requests
- Issues
- Comments
- Stars
Stars, PRs, issues and forks are under-captured in the later part of this window. GH Archive progressively stopped capturing non-push events during 2026 — −95% or worse by the end of the window. Every series here except Pushes fades for that reason, so a decline above reflects the archive, not this repository. Pushes stay reliable throughout, so read them, and the contributor counts derived from them, as the real signal. Data health has the measurements.
Top contributors
Pushes, PRs, issues, reviews and comments — stars and forks excluded, so this is contribution rather than popularity
| Contributor | Contributions | Pushes | PRs | Comments |
|---|---|---|---|---|
| HaozheH3 | 68 | 40 | 0 | 21 |
| cjakfskvnad | 18 | 8 | 0 | 9 |
| wenhuchen | 17 | 11 | 0 | 4 |
| g-makerr | 6 | 0 | 0 | 3 |
| zhangchbin | 6 | 0 | 0 | 4 |
| LiqiangJing | 4 | 0 | 0 | 0 |
| lt573152873 | 4 | 0 | 0 | 2 |
| mf-zhang | 3 | 0 | 0 | 1 |
| wcpcp | 3 | 0 | 0 | 1 |
| EugeneLiu01 | 3 | 0 | 0 | 3 |
| Xuan03-THU | 3 | 0 | 0 | 1 |
| fushh | 2 | 0 | 0 | 0 |
| TimSong412 | 2 | 0 | 0 | 2 |
| wuzhaodongaipython | 1 | 0 | 0 | 0 |
| NielsRogge | 1 | 0 | 0 | 0 |
| Haochen-Wang409 | 1 | 0 | 0 | 0 |
| 4NLLab | 1 | 0 | 0 | 0 |
| dhruvmsheth | 1 | 0 | 0 | 0 |
| Dynamite321 | 1 | 0 | 0 | 0 |
| w-zhih | 1 | 0 | 0 | 0 |
Recent activity
Latest issues, pull requests and releases
- Issue comment#39cjakfskvnad2026-01-04 01:51Pixel Reasoner Tool Call Frequency Gradually Approaches Zero in Verl-tool
- Issue comment#34cjakfskvnad2026-01-04 01:45KeyError: 'type' in PixelReasoner-SFT-Data
- Issue#37cjakfskvnad2026-01-04 01:24How was the Qwen2.5-VL 7B configuration and system prompt settings defined in Table 1 in paper?
- Issue comment#38cjakfskvnad2026-01-04 01:09CPU memory usage for PixelReasoner
- Issue#33wenhuchen2026-01-03 20:16Why verl-tool's pixel reasoner (82.7) lead to much lower performance than pixel reasoner in this repo (84.3)?
- Issue#38zulihit2025-12-04 11:55CPU memory usage for PixelReasoner
- Issue#354NLLab2025-11-11 03:09ImportError: cannot import name 'Qwen2_5_VLVisionFlashAttention2' from transformers.models.qwen2_5_vl
- Issue comment#35cjakfskvnad2025-11-06 06:30ImportError: cannot import name 'Qwen2_5_VLVisionFlashAttention2' from transformers.models.qwen2_5_vl
- Issue comment#32HaozheH32025-10-03 05:25Unable to reproduce PixelReasoner Cold-Start performance on vstar
- Issue#28HaozheH32025-07-25 13:39Difficulty in Reproducing Results
- Issue comment#31EugeneLiu012025-07-22 04:21Actor/policy loss steep rise, and mean reward steep fall
- Issue#31EugeneLiu012025-07-22 03:49Actor/policy loss steep rise, and mean reward steep fall
- Issue#30wuzhaodongaipython2025-07-18 06:41How long to train the model with 8*A800(80G)
- Issue comment#27mf-zhang2025-07-07 07:55Could you share your pip list for RL setup?
- Issue comment#27HaozheH32025-07-07 04:31Could you share your pip list for RL setup?
- Issue comment#27HaozheH32025-07-06 13:42Could you share your pip list for RL setup?
- Issue#27mf-zhang2025-07-06 12:55Could you share your pip list for RL setup?
- Issue comment#26HaozheH32025-07-02 13:03stuck in the single-node rl training
- Issue#26HaozheH32025-07-02 13:03stuck in the single-node rl training
- Issue comment#26g-makerr2025-07-02 12:59stuck in the single-node rl training
- Issue comment#26HaozheH32025-07-02 12:17stuck in the single-node rl training
- Issue comment#26g-makerr2025-07-02 07:12request for environment.yaml
- Issue#26g-makerr2025-07-02 07:11request for environment.yaml
- Issue#13HaozheH32025-06-30 11:58Simple inference demo
- Issue#23HaozheH32025-06-30 11:57the meaning of normalized coordinates in bbox
Totals cover only the window loaded into ClickHouse and count events, not GitHub's lifetime totals — 159 stars here means stars gained during the window, not the repo's star count.