🔗 Source: arXiv

Self-Adapting Language Models

Mechanism: The model generates “self-edits”—natural language instructions, synthetic training data, and optimization hyperparameters—conditioned on new context, which are applied via gradient-based updates in an inner loop and optimized by an outer RL loop using downstream task success as the reward signal.
Nuance: Unlike prior approaches that rely on fixed prompting, auxiliary networks, or external data generators, SEAL directly leverages the base model’s generative capacity to parameterize its own adaptation process, meta-learning how to adapt rather than relying on static heuristics or separate modules.

Achieved 47.0% accuracy on no-context SQuAD knowledge incorporation, outperforming synthetic data generated by GPT-4.1 and raw passage finetuning (33.5%).
Reached a 72.5% success rate on few-shot ARC-AGI reasoning tasks, significantly surpassing standard in-context learning (0%) and untrained test-time training baselines (20%).
Demonstrates robust generalization from single-passage updates to large-scale continued pretraining regimes with aggregated self-edits.

Suffers from catastrophic forgetting when applying sequential self-edits over time, as the current RL objective optimizes only for immediate task performance without explicit retention mechanisms.
Performance heavily depends on the base model’s capacity to generate effective self-edits and may require careful hyperparameter tuning across different architectures or data distributions.