Supervised fine-tuning (SFT) trains a base language model to imitate target responses, and these targets may require knowledge the base model has not robustly internalized. We study this as a source of hallucinations and frame a group of mitigation methods as knowledge-aligned SFT: constraining SFT training targets to the base model’s parametric knowledge. Under a unified setup, we compare existing generation-based and estimation-based knowledge-alignment methods and introduce two new variants: Evidence Rewrite, which verifies base-model generations using external evidence, and Recall Rewrite, which retains claims only when they can be consistently recalled by the base model. Experiments with Qwen 3 4B and OLMo 3 7B show that knowledge-aligned SFT can reduce factual hallucinations on WildHalu and Biography while largely preserving general capabilities. Recall Rewrite yields the strongest factuality gains and improves refusal behavior on UnknownBench. It thereby confirms that SFT targets beyond the base model’s knowledge drive hallucination behavior.

We present a comprehensive platform to run human-computer experiments where an agent instructs a human in Minecraft, a 3D blocksworld environment. This platform enables comparisons between different agents by matching users to agents. It performs extensive logging and takes care of all boilerplate, allowing to easily incorporate new agents to evaluate them. Our environment is prepared to evaluate any kind of instruction giving system, recording the interaction and all actions of the user. We provide example architects, a Wizard-of-Oz architect and set-up scripts to automatically download, build and start the platform.