Prompting Research: 5 Papers That Will Change How You Talk to AI

This Week in Prompting Research — February 26, 2026

Every week I scan arXiv for papers that help regular people — not ML researchers — get better results from ChatGPT, Claude, and other LLMs. This week's crop is exceptionally practical. Here are the five papers that stood out.

1. 🎯 Your Framing Words Are Invisible Levers

arXiv:2602.21223 — Geng et al., Measuring Pragmatic Influence in Large Language Model Instructions

When you write "This is urgent" or "As your supervisor," you're not just being polite — you're pulling a lever that measurably shifts how the model responds. Researchers tested 400 framing phrases organized into 13 strategies across five different LLMs and found consistent, predictable shifts in how models prioritize directives.

What to do with this: Be intentional about your framing. Adding urgency, authority, or emotional cues genuinely changes outputs — it's not wasted text. But be careful: framing one side of a question as more important can bias the answer. Across all five models tested (different sizes, different companies), the effect held. Framing is a fundamental property of instruction-following, not a quirk.

2. 🏗️ Structure Beats Context (0% → 85% With Zero New Information)

arXiv:2602.21814 — Jo et al., Prompt Architecture Determines Reasoning Quality

The viral "car wash problem" stumps every LLM — until you restructure the prompt. Using a STAR framework (Situation-Task-Action-Result) that forces the model to state its goal before reasoning, accuracy jumped from 0% to 85% with no additional information. Context retrieval helped too, but the reasoning scaffold did the heavy lifting.

What to do with this: When the model gets something wrong, don't just add more context. Restructure. Instead of asking your question directly, lay out: (1) the situation, (2) what you need, (3) how to approach it, (4) what a good answer looks like. The model already has the knowledge — it just needs a scaffold to use it.

3. 👻 The Invisible Voice Eraser

arXiv:2602.22145 — Chandra et al., When AI Writes, Whose Voice Remains?

Ask an LLM to "professionalize" your email and it'll quietly strip your cultural identity along with the typos. Analyzing 22,350 outputs from Indian, Singaporean, and Nigerian English texts, researchers found models erase up to 20.5% of cultural markers while maintaining high semantic similarity — meaning you might not even notice. Politeness conventions (greetings, indirect phrasing) are 1.9x more vulnerable than vocabulary. The fix? Adding "preserve my cultural voice" to your prompt reduced erasure by 29%.

What to do with this: Whenever you ask an LLM to edit your writing, add an explicit instruction to preserve your tone, style, and cultural expressions. Review edits for whether they still sound like you, not just whether they're "correct."

4. 🧠 The Brainstorming Trick: Farmers Beat Steve Jobs

arXiv:2602.20408 — Deng et al., Examining and Addressing Barriers to Diversity in LLM-Generated Ideas

LLM brainstorming produces less diverse ideas than human groups — but two simple prompting fixes reverse this entirely. Chain-of-Thought prompting reduces "fixation" (where early ideas constrain later ones), and assigning ordinary personas like "librarian" or "farmer" — not famous people — improves knowledge diversity. Combined, these make LLM ideas more diverse than human brainstorming.

What to do with this: For brainstorming, always combine "think step by step" with persona prompting. Crucially, use mundane personas ("retired teacher," "park ranger"), not celebrities. "Think like Steve Jobs" actually collapses diversity toward popular stereotypes. Run multiple sessions with different personas rather than one big brainstorm.

5. ⚠️ When "Think Step by Step" Backfires

arXiv:2602.22072 — Nickel et al., Understanding Artificial Theory of Mind

Chain-of-Thought is the most popular prompting trick, but it's not universally helpful. Testing LLMs on perspective-taking tasks (understanding that someone else has different knowledge than you do), researchers found CoT generally helps — but on certain types of unusual scenarios, it actively made accuracy worse. The model overthinks itself into the wrong answer.

What to do with this: If you're asking the model to reason about what other people think, know, or believe — and CoT gives you a weird answer — try asking without it. "Think harder" is great advice most of the time, but for edge cases involving perspective and social reasoning, the model's quick instinct can be more reliable than its deliberate reasoning.

Full research notes with detailed summaries available in the research repo.