Test-Time LLM Adaptation

🔗 Source: arXiv

Test-Time Learning for Large Language Models

🚀 Technical Novelty

• Mechanism: Formulates inference-time adaptation as input perplexity minimization, actively weighting high-perplexity samples for backpropagation, and applying lightweight LoRA updates to preserve pre-trained knowledge.
• Nuance: Replaces standard TTA’s entropy minimization with autoregressive-aware input perplexity optimization, eliminating the need for external retrieval or labeled data while mitigating catastrophic forgetting via parameter-efficient updates.

💡 Yield

• Achieves ≥20% performance gains on domain adaptation benchmarks; reduces online backward passes by ~69.7%; establishes AdaptEval benchmark for rigorous TTL evaluation.

⚠️ Limitations

• Sensitive to the perplexity threshold hyperparameter (P0); relies on greedy decoding which may limit output diversity; requires careful calibration to avoid overfitting on noisy test streams.