Unverified50% confidenceFactExact time
After the RL stage, Agima exhibited behaviors not explicitly taught in training, including the spontaneous formation of a unified format, on-demand lookahead (forward thinking only when necessary), and more frequent branching exploration at key decision points.
1
Sources
50%
Confidence
Long-term
Relevance
10/20/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAgima's training is divided into two stages: the first stage uses supervised fine-tuning (SFT) to unify the policy and the world model into a single omni-model; the second stage uses agentic reinforcement learning (RL) to teach the model to reason over imagined futures and use them for planning.71% similarUnverifiedThe jev architecture used for comparison employs parallel sampling, is trained via RLCD (Reinforcement Learning from Contrastive Distillation), and outputs confidence distributions and schema selection.68% similarUnverifiedIn the Agent context, reinforcement learning (RL) is mainly used to optimize decision-making policies, with commonly used methods including RLHF and DPO; compared with supervised fine-tuning, it can explore better paths through trial and error, but it faces challenges such as difficulty in reward design, low sample efficiency, and unstable training.66% similarUnverified近期研究表明RL（尤其是RLVR）并没有给模型注入大量新知识，而更像是激发和对齐预训练阶段已具备的能力65% similarUnverifiedEvery step of RL training requires complete inference sampling, reward computation, and backpropagation, with rollout being the most time-consuming part.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1073380API
curl https://kongchang.com/api/v1/knowledge/claims/1073380MCP
get_claim(id=1073380)