How much of an LLM agent’s apparent behavior comes from the decoder rather than the model alone? SamplerScope holds the model’s logits fixed, changes only the decoding method, and measures how this alters outcomes in small decision environments.
sampling policy-evaluation markov-decision-process huggingface-transformers large-language-models qwen llm-evaluation llm-agents inference-time ai-welfare decoding-strategies decoder-attribution
-
Updated
Aug 17, 2026 - Python