Tag
This paper shows that asking vision-language models which blocks a director can see before acting significantly changes their choices in the Director Task, raising hidden-block selections by 12-88 percentage points across GPT models and 17 points on Qwen3.8-27B. The authors argue perspective-taking benchmarks need paired baselines that differ only by the report question to avoid misreading failures of perspective use.
This paper investigates how format, instruction count, and context length affect instruction adherence and hallucination in LLMs through controlled experiments on a synthetic corpus, finding that instruction-following collapses beyond 80 rules regardless of format, and recall degrades sharply after 64-128k tokens with format-dependent effects. It releases the VeyraBench harness for reproduction.