Unverified50% confidenceFactExact time
In RLVR training, it is often observed that pass@1 rises while pass@k (for large k) falls, i.e., the policy becomes more 'sharpened' but less diverse.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
11/2/2026
First Seen
Valid until: 1/31/2027
Sources
Related Entities
Related Claims
Unverified低精度引入的数值扰动会在RL的多轮迭代中通过策略梯度的乘积形式被放大，可能导致训练崩溃或模型性能退化71% similarUnverifiedAfter RL training, a model's output diversity decreases noticeably and fixed small habits (ticks) form.70% similarUnverifiedThe quality of reward signals in RLVR depends on the task difficulty distribution: tasks that are too easy lead to reward saturation and vanishing gradients, while tasks that are too hard lead to sparse rewards, both of which cause training to stall.69% similarUnverifiedAdversarial gradient saliency is a first-order local approximation; the saliency ranking drifts during training, and updating only the initial subset over the long term may miss parameters that become important later, so subset selection strategies and periodic re-estimation are needed.67% similarUnverifiedThe eigenvalue decay rate directly determines the smoothness of the RKHS and the generalization error bound: the faster the decay, the "smaller" the function space, and the more the model tends to learn smooth functions.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1125845API
curl https://kongchang.com/api/v1/knowledge/claims/1125845MCP
get_claim(id=1125845)