yifanzhang-pro/FlashREINFORCE
FlashREINFORCE: Critic-Free, Single-Rollout, Asynchronous RL for Agentic Language Models
About the project
Asynchronous reinforcement-learning framework for agentic language models: one rollout per prompt, critic-free, with token importance sampling and sequence trust region to control policy drift.
Useful for
- Train an agentic LLM on long reasoning chains without a critic or group rollouts
- Run the CPU toy_update.py example to verify an optimizer update
- Reproduce experiments on Qwen2.5-Math or DeepSeek-R1-Distill via Molt launchers
README summarized by DeepSeek V4.1 Flash. Details may be inaccurate.
Why it’s trending
- 1 star so far today, about 1 expected by the end of the day.
- The repository is 4 days old and already has 56 stars.
Stars per day
Bars are daily stars, the line is the usual pace. Red marks spike days.
Numbers
- Total stars
- 56
- Stars in a day
- 1
- Forks
- 6
- Issues and pull requests
- 3
- Watchers
- 0
- Language
- Python
- License
- Apache-2.0
- Created
- September 14, 2026
- Last push
- September 14, 2026
Star trust
Growth looks organic: forks and discussion are in line with active projects, and stars arrive unevenly, the way people give them.
These are heuristics, not a verdict: we judge by the repository’s behavior, not by a list of stargazers.
Spotted in
- September 18, 2026Top new repositories this week: #176
- September 17, 2026Top new repositories this week: #186
Share
README badge
Paste it into your README — the badge shows the current magnitude and links to this page.
[](https://gitnova.dev/en/r/yifanzhang-pro/FlashREINFORCE)Similar projects
-
8.6
TheoLeeCJ/SemIf
An open reproduction of the Jev-style semantic decision interface: reads typed option probabilities directly from a 4B model's logits without generating text. Runs on a single RTX 3090.
-
7.1
vinnylarouge/jevlike
A small trainable model that picks one option from a changing list of text options in a single forward pass, an open alternative to TypeSafe's Jev. Aimed at developers who need fast classification or menu selection…
-
6.9
Continuum-AI-Corp/OrcaBonsai-27B-Uncensored
A tool for runtime refusal ablation in the compressed Ternary Bonsai 2 27B LLM: it projects the residual stream to remove the refusal direction without modifying or re-quantizing weights. Runs on Apple Silicon via MLX.
-
6.4
TianyuCodings/NanoJev
A nano replica of Jev built on Qwen3-0.6B that outputs probability distributions over candidates in a single forward pass with no token decoding, plus a training pipeline and game demos (maze, Snake).
-
6.3
zhengkid/Dream-RSI
Research project on recursive self-improvement: past discovery histories become a replay simulator used to evaluate and refine exploration policies. Code is still being prepared; paper and demo are available.
-
5.0
ekzhang/openjev-sglang
HTTP server implementing the TypeSafe/Jev API for structured text classification on Qwen3.6-35B-A3B via SGLang; returns answer probabilities without generating a chain of thought.