IC-1584LLaMA-7B and GPT-J-6B fail to interpret textual emphasis markers, with marked prompting degrading performance substantially

Qingru Zhang, Chandan Singh, Liyuan Liu, Xiaodong Liu, Bin Yu, Jianfeng Gao, Tuo Zhao

SourceTell Your Model Where to Attend: Post-hoc Attention Steering for LLMs

"When standard emphasis markers (asterisks or double quotes) are added around user-specified spans in the prompt, both models' performance drops sharply rather than improving. For LLaMA-7B on the JSON formatting task, format accuracy falls from 60.00 (zero-shot) to 18.55 (*-marked) and 4.56 (""-marked). For GPT-J-6B, format accuracy falls from 28.83 to 4.44 and 8.81 respectively. The authors attribute this to the rarity of such markers in pre-training data, meaning the models have not learned to treat them as attention cues."

Evidence
correlational
Key metric
LLaMA-7B JSON format f.acc: 60.00 (zero-shot) vs 18.55 (*-marked) vs 4.56 (""-marked); GPT-J-6B: 28.83 vs 4.44 vs 8.81
Caveat
The authors note that emphasis markers rarely appear in massive pre-training data, which may explain the failure; the effect is measured on 5000 test examples per task.
Model
LLaMA Llama 7B, GPT-J 6B
Concepts
Failure mode
Datasets
Bias in Bios [eval], CounterFact / Counterfact dataset [eval]
Methods
Zero-shot prompting [compared-to]
Related findings
IC-1585, IC-1586
Extraction
automatic-extraction