Research map / Large language models / Reasoning and chain of thought

Making, Not Taking, the Best of N

ICLR 2026 · Large language models › Reasoning and chain of thought · approach: Routing, pruning and distillation for reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Fusion-of-N (FusioN), which replaces Best-of-N selection with an LLM fusor that synthesizes a new answer from N candidate generations, evaluated for test-time scaling and synthetic SFT data generation on multilingual tasks.

Problem

Best-of-N aggregation treats LLM candidate generations as a zero-sum selection, discarding complementary information in the pool and depending on imperfect reward models for selection.

Approach

Instead of selecting one candidate, a strong LLM fusor is prompted with all N candidates and asked to compare them and synthesize a new response that combines their strengths, requiring no training and acting as a drop-in replacement for BoN.

Key results

• On mArenaHard v2.0 with N=5, Command A FusioN wins over Gemini2.5-Pro in German (+9.5%) and Spanish (+7.8%) gains versus BoN; Aya Expanse 8B gains up to +10.8% in French.
• On WMT24++ with Command A samples, FusioN outperforms the XCometXL oracle selection in German, Russian and Chinese (+0.8 in Chinese) and beats BoN by up to +11.4 in Korean.
• Fine-tuning 111B model on FusioN data yields +0.8 XCometXL on WMT24++ and +2.5% win-rate vs Gemini2.5-Flash on Arena over BoN data; on GeoFactX answer correctness FusioN reaches 54.2% vs 52.4% for BoN on average.

Surprising findings

• FusioN surpasses the oracle selection on some WMT translation languages, showing synthesis can exceed the best candidate in the pool.
• On MGSM math, BoN with a specialized reward model outperforms FusioN, and fine-tuned FusioN lags BoN by 1.2% and the fusor by 1.5%.

Limitations

• Fusor needs to be large (around 100B scale) for generative fusion to work out of the box; smaller fusors underperform.
• Weaker performance on close-ended math tasks (MGSM) relative to BoN with a specialized RM.
• Lower language correctness (98.9%) and weaker gains for non-English languages; the fusion prompt is English-only.

Experimental setup

Datasets: UltraFeedback Binarized mArenaHard v2.0 WMT24++ GeoFactX s1k MGSM

Benchmarks: mArenaHard v2.0 WMT24++ MGSM GeoFactX

Metrics: win rate (%) XCometXL score answer accuracy reasoning quality score language correctness

Baselines: Best-of-N Oracle selection Gemini2.5-Pro Gemini2.5-Flash single-teacher sampling

Builds on: Best-of-N Mixture-of-Agents LLM-Blender rejection sampling

Closest papers by meaning

• Rethinking LLM Evaluation: Can We Evaluate LLMs with 200× Less Data? (ICLR 2026 · similarity 0.86)
• Majority of the Bests: Improving Best-of-N via Bootstrapping (NeurIPS 2025 · similarity 0.85)
• SELF-[IN]CORRECT: LLMs Struggle with Discriminating Self-Generated Responses (AAAI 2025 · similarity 0.85)
• Unveiling the Secret Recipe: A Guide For Supervised Fine-Tuning Small LLMs (ICLR 2025 · similarity 0.85)
• G-Zero: Self-Play for Open-Ended Generation from Zero Data (NeurIPS 2026 · similarity 0.85)
• Benchmarking LLMs' Judgments with No Gold Standard (ICLR 2025 · similarity 0.85)