Long-context QA on LongBench 1000 (test)
81.3Success RateCVT-RL
Evaluation results
| Method | Links | ||
|---|---|---|---|
| CVT-RLDescription=per-seed CVT-RL success, Split (train/dev/test)=6100/800/1000, Individual Seeds=80.4/82.1/81.7/80.8/81.52026.06 | 81.3 | 4.2 | |
| IMDescription=information-matched counterfactual-process RL, Split (train/dev/test)=6100/800/10002026.06 | 77.1 | — | |
| CMDescription=compute-matched non-causal RL, Split (train/dev/test)=6100/800/10002026.06 | 73.8 | — |