Tag
This paper introduces ErrorBench, a stress-test protocol demonstrating that numeric anchoring in prompts inflates count-based F1 scores for LLM error detection without improving span localization, making count-only evaluation unreliable.
This paper proposes a randomized anchoring strategy to mitigate anchoring bias in LLM-based agents for energy-efficient 6G autonomous networks, achieving up to 25% energy savings using a lightweight 1B-parameter model.
This paper investigates how irrelevant numbers in prompts cause anchoring effects in language models and localizes the internal pathways carrying this signal using attribution-based circuit methods on Qwen and Llama models.