Test-Time Scaling in the Wild: Why Exploitation, Not Exploration, Is the Bottleneck

This study presents a compute-normalized comparison of five test-time scaling families—Best-of-N, Beam Search, Particle Filtering, Sequential Refinement, and Fusion—across HealthBench, PRBench, LEXam, WildBench, and WritingBench, covering medicine, professional reasoning, law, chat, and creative writing. Using Qwen3.5 and OLMo3 generators, with reward models including Skywork-Reward-V2 and Llama-3.1-70B, the researchers separate exploration, which generates candidate answers, from exploitation, which selects or synthesizes the final response. Exploration keeps working: the bias-corrected oracle quality of larger candidate pools rises across benchmarks, but exploitation fails because reward-model rankings correlate with judged quality at only 0.12, making Best-of-N selection nearly random. Beam Search and Particle Filtering worsen this through diversity collapse, retaining just 40–60% of the diversity of independent sampling. Sequential Refinement produces a genuine gain on only one of five benchmarks, while apparent improvements on WritingBench are strongly confounded by verbosity bias; the benchmark’s native judge includes Claude 3.7 Sonnet among its evaluators, while HealthBench originates from OpenAI. Fusion is the only consistently improving strategy for Qwen3.5-35B-A3B, reaching an overall score of 0.610 at the highest compute level versus 0.574 for single-sample inference, yet it captures only 40% of available exploration headroom. Results from OLMo3 show that generating strong candidates does not guarantee the ability to exploit them, establishing candidate selection and synthesis—not candidate generation—as the central bottleneck in open-ended test-time scaling.