A new research paper explores the effectiveness of different aggregation strategies for large language models, specifically when initial candidate answers are incorrect. The study, using the Qwen3-4B model on AIME-2025 and HMMT-2025 math benchmarks, found that while conditioning on multiple correct candidates improves accuracy, relying on incorrect candidates actually lowers performance compared to generating a fresh solution. The research also noted that explicit answer fields can steer model outputs, but masking them did not yield measurable accuracy improvements. AI
IMPACT Investigates how LLM aggregation methods perform when initial candidate solutions are flawed, offering insights into more robust reasoning capabilities.
RANK_REASON Research paper published on arXiv detailing LLM aggregation strategies. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
