memhuilong
Joined 20 days ago
Karma
1
Submissions
2
Member since
Oct 22
Recent submissions
Comment · 19 days ago
The correctness term matters here, especially once training can change the answer as well as its confidence. If an answer is correct with probability p and reports q, the expected RLCR reward is p^2 - (q-p)^2. At q=p, the reward is p^2, so a more accurate answer still pays.
With only the negative Brier loss, the optimum would instead be -p(1-p): being reliably wrong and reporting zero confidence would score just as well as being reliably right. The joint objective avoids that trap. Section 3 of the paper discusses this distinction: https://arxiv.org/html/2507.16806v2#S3
That describes the ideal incentive, of course, not a guarantee of calibration after finite training or on a new task.
We are building wearable AI and memory systems to help people remember, learn, and grow.
My focus is bringing human memory into AI agents, enabling them to learn from personal experiences and become genuinely useful in everyday life.
https://github.com/memoket/memoket-kite