Expired90% confidenceFactTime unknown
Composer 2.5 introduces a text-feedback-based directed reinforcement learning method that inserts local prompts as teacher signals at the positions where the model makes errors and trains using on-policy KL divergence loss.
1
Sources
90%
Confidence
Medium-term (~90 days)
Relevance
5/31/2026
First Seen
Valid until: 8/29/2026(expired)
Sources
Related Entities
Related Claims
Unverified文本反馈（Textual Feedback）方法通过教师模型对 rollout 片段给出自然语言提示，将稀疏标量奖励转化为密集局部化的文本监督信号77% similarUnverifiedComposer 2.5 is Cursor's specially optimized version for programming scenarios, with deep tuning in areas such as codebase indexing, symbol reference understanding, and diff editing.67% similarUnverifiedDoubao TTS's performance-direction mechanism is an in-band signaling design, where control instructions and the text to be read are mixed within the same text input, relying on the model's comprehension ability to distinguish between the two.64% similarUnverified冗长的系统提示通常意味着开发者用大量文字教模型该怎么做，是把工程问题转嫁给提示词的做法63% similarUnverifiedComposer 2.5 performs strongly in code writing, debugging, and logical reasoning, but is slightly lacking in the creativity dimension.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/12616API
curl https://kongchang.com/api/v1/knowledge/claims/12616MCP
get_claim(id=12616)