Study finds simple few-shot prompts let small open LLMs match complex prompting on NLI
rohanpaul_ai · x · 2026-11-11
An arXiv paper benchmarks decoder-only LLMs (Llama, Mistral, Gemma, Phi 4) on natural language inference across MNLI, QNLI and SciTail using minimal zero-shot and few-shot prompts.
Key findings:
- Simple, direct prompts make small open LLMs perform exceptionally well, challenging the assumption that complex prompting and sampling techniques are required for NLI
- In one-shot prompting, the choice of demonstration label matters significantly, with entailment-primed examples performing best
- A few representative few-shot examples match or exceed longer prompt lists
Related event: Simple Few-Shot Prompts Beat Complex Techniques for Small Models on NLI(2 posts)→
More from Models
- $200 Claude Ultra plan drained in one night: usage screenshot shows weekly limits maxed — ezshine · 2026-11-11
- Fable 5.2 rebuilds a century-old landmark from 40 photos into a printable 3D model — FinanceYF5 · 2026-11-11
- Fable 5.2 acts as a blackboard tutor, drawing formulas step by step while explaining — FinanceYF5 · 2026-11-11
- One prompt, three Claude versions: side-by-side video reveals real capability gaps — FinanceYF5 · 2026-11-11
- Fable 5.2 designs origami animals and verifies foldability with its own simulator — FinanceYF5 · 2026-11-11
- Fable 5.2's 7 early experiments: 3D cells, verified origami, agent visualization — FinanceYF5 · 2026-11-11