Last-Token Self-Rewarding Reinforcement Learning Recipe Model Card

这是可索引项目报告证据页：它保留 Last-Token Self-Rewarding Reinforcement Learning Recipe Model Card 的源材料入口、机制线索和限制提醒；正文仍需 reader/editor 与 academic public-copy review 后才能当作最终结论引用。

Last-Token Self-Rewarding Reinforcement Learning Recipe Model Card

1. Role in Self Evolve

LaSeR is a reinforcement-learning recipe that jointly improves reasoning and self-rewarding behavior by adding an MSE self-reward term to the RLVR objective. It matters because self-evolving agents need explicit feedback loops, observable retention mechanisms, and auditable evaluation pressure before improvement claims become useful.

2. Working Principle

optimize RLVR objective -> learn last-token self-reward signal -> reuse auxiliary reward during training and testing -> improve reasoning and reward calibration together

3. Evidence Path

web-observed GitHub page showed 36 stars, 2 forks, 6 commits, MIT license, a last-token self-rewarding RLVR objective, released checkpoints, and auxiliary self-reward usage at both training and testing time. This iteration keeps freshness honest: the snapshot comes from the current public GitHub page, while shell GitHub API access remained blocked in this workspace.

4. Teaching Use

Use this card to explain Last-Token Self-Rewarding Reinforcement Learning Recipe: it shows how coding-agent, self-rewarding, optimization, or graph-runtime layers convert agent behavior into a reproducible engineering story.

5. Limits

The repository was not cloned in this iteration; no benchmark run, workflow execution, or training experiment was executed. Counts and claims are visible public-page signals unless independently revalidated later.