IC-171ICL prompting produces binary response patterns in released LLMs, with answered ratios collapsing to near 0% or 100% rather than calibrated refusal, making prompting ineffective for RAG groundedness

Maojia Song, Shang Hong Sim, Rishabh Bhardwaj, Hai Leong Chieu, Navonil Majumder, Soujanya Poria

SourceMeasuring and Enhancing Trustworthiness of LLMs in RAG through Grounded Attributions and Learning to Refuse

The paper evaluates whether in-context learning with refusal demonstrations can calibrate models' RAG behavior. Across multiple released models, ICL with a refusal prompt causes answered ratios to drop to near zero (Llama-3-8B: 1.48% ASQA, 3.90% QAMPARI, 0.00% ELI5; Llama-2-7B: 0.00%, 0.00%, 0.50%), while under default prompting models answer nearly everything. This binary behavior indicates that released models cannot learn calibrated refusal from a few ICL examples, and the resulting indiscriminate refusal or over-responsiveness degrades trust-scores across all sub-metrics.

Evidence
correlational
Key metric
Llama-3-8B ICL: AR% 1.48 (ASQA), 3.90 (QAMPARI), 0.00 (ELI5); Llama-2-7B ICL: AR% 0.00, 0.00, 0.50; Llama-3.2-3B ICL: AR% 1.27, 34.10, 21.90; Qwen-2.5-0.5B ICL: AR% 29.85, 11.40, 82.30
Model
Llama 2 / Llama 2 base Llama 2 7B, Llama 3 8B, Llama-3.2-3B, Qwen2.5 Qwen-2.5-0.5B, GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report, Claude 3.5 Sonnet
Concepts
Failure mode
Datasets
ASQA [eval], QAMPARI [eval], ELI5 [eval]
Related work
RARR [context]
Related findings
IC-170
Extraction
automatic-extraction