Engagement Prediction on LinkedIn User Engagement production (val)
0.769ROC-AUC (Val)Pointwise Reward
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Pointwise RewardReward mechanism=Pointwise2026.02 | 0.769 | 3.5 | |
| Pointwise Reward (first three lines)Reward mechanism=Pointwise, Truncation/Length constraints=first three lines2026.02 | 0.759 | 2.15 | |
| Listwise RewardReward mechanism=Listwise2026.02 | 0.749 | 0.83 | |
| Full Context textInput representation=Full Context, Truncation/Length constraints=750 tokens2026.02 | 0.747 | 0.54 | |
| Pointwise Reward (first two lines)Reward mechanism=Pointwise, Truncation/Length constraints=first two lines2026.02 | 0.747 | 0.54 | |
| EmbeddingInput representation=User dense embedding vectors, Integration=concat to linear layer2026.02 | 0.745 | 0.26 | |
| Baseline2026.02 | 0.743 | — | |
| Naive prompt-tuned SummaryPrompting/Tuning strategy=Naive prompt-tuned2026.02 | 0.743 | 0 |