IC-1577Base LLMs prompted with URiAL (3 restyled in-context examples + system prompt) match or surpass their SFT/RLHF-aligned counterparts on multi-aspect evaluation
Using a 1-5 scoring protocol across six aspects (helpfulness, clarity, factuality, depth, engagement, safety) on 1,000 instructions, Mistral-7b with URiAL scores 4.63 average versus 4.44 for Mistral-7b-instruct (SFT), and Llama-2-70bq with URiAL scores 4.74 versus 4.67 for Llama-2-70b-chatq (RLHF). For the weaker Llama-2-7b base, URiAL reaches 4.33, close to but below Llama-2-7b-chat's 4.47. The gap between tuning-free and tuning-based alignment narrows as base model quality increases. Human pairwise evaluation confirms: Mistral-7b+URiAL wins 20.0% vs 8.9% for Mistral-7b-instruct, and Llama-2-70bq+URiAL wins 16.5% vs 8.4% for Llama-2-70b-chatq.
Evidence
correlational
Key metric
Mistral-7b+URiAL 4.63 vs Mistral-7b-instruct 4.44; Llama-2-70bq+URiAL 4.74 vs Llama-2-70b-chatq 4.67; Llama-2-7b+URiAL 4.33 vs Llama-2-7b-chat 4.47; human eval: 20.0% vs 8.9% (Mistral), 16.5% vs 8.4% (Llama-2-70b)
Caveat
Evaluation uses GPT-4 as judge (94.1% human approval on explanations, 87.8% agreement on pairwise comparisons). Greedy decoding with repetition penalty 1.1 on base models. The 1,000-example Just-Eval-Instruct set may not cover all domains; authors note tuning may still be necessary for coding, math, and interactive agents.