Test-Time Scaling

Test-time scaling (TTS) is the practice of allocating extra inference-time compute to improve a single model's output on a single input. The compute can be spent along several axes: sampling more candidate answers in parallel (best-of-N, particle filtering), running a sequential refinement loop, performing beam search over reasoning traces, or synthesizing a final answer across many candidates (fusion decoding). TTS is the inference-side counterpart to scaling up pretraining or post-training compute, and is treated as a separate scaling axis because a model's behavior under additional inference budget often behaves differently from its behavior under more training compute.

TTS methods have been most thoroughly evaluated on mathematics and short-answer benchmarks where ground-truth verifiers exist. The Thomson Reuters Thomson technical report (Chen et al., 2026, Section 4.5) extends TTS evaluation to specialized, open-ended, non-verifiable domains such as legal reasoning, journalism, and long-context synthesis. The report's central empirical finding is that in such domains, exploitation — the step that converts the sampled pool into a final answer — is the bottleneck, not exploration: reward models correlate only weakly with true quality, and process-reward-model-guided search can collapse candidate diversity. Direct synthesis / fusion (producing a final answer by reasoning across the full pool, not by selecting a single candidate) is shown to recover substantially more of the available headroom than reward-model selection.

TTS overlaps with but is distinct from inference-scaling (a broader term covering model-size choices, long-context extensions, and adaptive compute strategies). It is also distinct from chain-of-thought and reasoning-model training, which expand the per-token reasoning rather than multiplying parallel samples.