You just copied the exact prompt structure that could make your LLM interactions 15% more accurate. New research from arXiv reveals a counterintuitive truth: sometimes, less context is more.

For years, we've assumed feeding an AI its entire conversation history was best. This study flips that script. By removing the model's own past words, you force it to reason from scratch on each turn—and it often performs better.

TL;DR: The Core Finding

• What: A new prompting method that removes the AI's past responses from the conversation context.
• Impact: It challenges a core assumption in chatbot design and can improve multi-turn reasoning accuracy.
• For You: You can test this hack immediately in ChatGPT, Claude, or any API to see if it solves your specific reasoning task.

The Backstory: Questioning a Default

Every chatbot you've used works the same way. It sees the full dialogue: your questions and its own answers. This "full-context" prompting is the industry standard.

The new paper "Do LLMs Benefit From Their Own Words?" asked a simple question. What if we're wrong? Researchers tested three open-source reasoning models on real-world, multi-turn conversations.

How It Works: Less Is More

The "user-turn-only" approach is brutally simple. Before sending the latest query to the model, you delete every past "Assistant:" response from the prompt. The model only sees the user's questions in sequence.

This forces a clean-slate reasoning process. The AI can't blindly copy or get stuck in its earlier logic. It must reconstruct the problem from the user's perspective alone.

Why It Matters: The 15% Accuracy Bump

The results were not uniform, but they were significant. For certain complex reasoning tasks, user-only prompting delivered up to a 15% improvement in accuracy.

Think about code debugging or mathematical proofs. If the model made a subtle error in step two, seeing that error in step three can cement the mistake. Erasing its own words breaks that error chain.

Key Insight: The benefit was most pronounced in longer conversations where error propagation is a real risk.

When To Use This Hack

This isn't a magic bullet for all interactions. The study found the approach works best for:

• Complex, multi-step reasoning: Math, logic puzzles, strategic planning.
• Tasks requiring factual consistency: Where early mistakes corrupt later answers.
• Open-ended exploration: Brainstorming where you want fresh ideas each turn.

For simple Q&A or creative writing, the standard full-context method often remains superior. The model needs its own narrative to maintain style and coherence.

The Bottom Line For Developers

You now have a new tool in your prompting toolkit. The default is no longer the only option.

For your next project, A/B test it. Run your complex multi-turn task with both full-context and user-only prompts. Measure the outcome. The data from this study suggests you might be surprised.

The era of one-size-fits-all prompting is over. Intelligent context management is the next frontier.