Unverified50% confidenceOpinionExact time
Using validation loss for early stopping or model selection may point in the wrong direction in the DPO setting.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/21/2026
First Seen
Valid until: 1/19/2027
Sources
Related Entities
Related Claims
UnverifiedIn the DPO setting, there is only a weak correlation between the hyperparameters that minimize validation loss and the optimal hyperparameters on downstream benchmarks.75% similarUnverifiedUnder off-policy sampling conditions, when the preference margin is aligned in direction with the utility gradient, DPO's update shifts from a corrective signal to a reinforcement-style update similar to reinforcement learning.74% similarVerifiedAfter geometric transformation, the transformation parameters must be recorded and the predicted boxes must be inverse-transformed to align them; ignoring this step will cause semantic misalignment in the consistency loss and harm training effectiveness71% similarUnverifiedStandard DPO and its unlearning variants are typically built around noise prediction error and do not transfer well to FSD models.70% similarUnverifiedIn adversarial MDPs, the environment can choose an arbitrary loss function at each episode (possibly depending on the algorithm's history), and the loss can vary arbitrarily over time, so standard optimistic or confidence upper bound methods no longer apply directly.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1079347API
curl https://kongchang.com/api/v1/knowledge/claims/1079347MCP
get_claim(id=1079347)