Reducing Forgetting
| Work | Training & Models | Tasks & Evaluation |
|---|---|---|
| CPO · July 2026; revised September | GSPO + masked movement penalty. Full-parameter Qwen3-VL-2B/4B/8B. | MRCL: medical VQA, navigation, geometry, puzzles and financial charts. Evaluate acquisition and retention across a five-task sequence. |
| TSRL · October 2026 | GRPO/DAPO + triggered SFT. Qwen2.5-7B-Instruct for ALFWorld/WebShop; 32B-Instruct for Deep Search. | ALFWorld ↔ WebShop transfer; ScienceWorld extends the sequence. Search training also measures AIME24/25 retention. |
Improving Adaptation
| Work | Training & Models | Tasks & Evaluation |
|---|---|---|
| ORBIT · February 2026 | GRPO across complete trials. Qwen3-8B; scaling to 4B/14B. | Train on Rock-Paper-Scissors, Minesweeper, Hangman, Wordle and Blackjack. Test on unseen Maze/Mastermind; three episodes per instance. |
| LaMer · December 2025; revised March 2026 | GiGPO with cross-episode returns and trained reflection. Qwen3-4B; additional Llama3.1-8B-Instruct experiments. | Sokoban, Minesweeper, WebShop; ALFWorld transfer. Compare three-episode trials and test history/reflection ablations. |
| SEAL · June 2025 | ReST$^{\mathrm{EM}}$ outer loop; SFT/LoRA adaptation inside. Qwen2.5-7B for knowledge incorporation. | No-context SQuAD and few-shot ARC. Evaluate the adapted model; repeated updates separately expose forgetting. |
For retention, follow each task’s score after later updates. For adaptation, compare successive attempts at matched interaction or token budgets, including on unseen tasks. A final success rate alone does not distinguish these behaviors.