Traditional decoding-time alignment methods, such as Best-of-N sampling, enhance output quality by selecting high-reward responses from multiple candidates. However, these approaches incur substantial computational overhead, which limits their practicality. Recent advances aim to mitigate this inefficiency by scoring and filtering partial generations early in the decoding process. However, these methods face a critical challenge: conventional reward models (RMs), optimized to assess complete responses, exhibit poor performance when evaluating truncated text. To address this, we introduce RewardPartial, a novel evaluation paradigm that shifts the focus from superficial partial-text quality to estimating a response’s completion potential. We provide theoretical analysis for this paradigm through the perspective of Markov Decision Process and construct TruncEval, a specialized dataset for training RewardPartial models to accurately predict this potential. Our evaluation on the TruncEval test set demonstrates that RewardPartial models deliver significant performance gains over conventional reward models. When deployed in early-filtering pipelines, our method achieves superior alignment performance compared to Best-of-N sampling, offering a more efficient solution for high-quality text generation.

Large language model (LLM) evaluation is increasingly confounded by *variant contamination*: training corpora contain semantically equivalent but lexically or syntactically altered test items. These variants can evade detectors based on sampling consistency or perplexity, inflating benchmark scores through memorization rather than reasoning. We formalize this problem and introduce **DVD** (**D**etection via **V**ariance of generation **D**istribution), a detector requiring only the test item, without external reference corpora or paired contaminated/clean sets. DVD models the local output distribution through repeated temperature sampling. Our key insight is that contaminated items alternate between *memory-adherence* and *perturbation-drift* states, producing unusually high variance in the synthetic difficulty of low-probability tokens, whereas uncontaminated items remain in drift with comparatively stable difficulty values. We construct the first dedicated variant-contamination benchmark across Omni-MATH and SuperGPQA by generating and filtering semantically equivalent variants, and simulate contamination by fine-tuning Qwen2.5 and Llama3.1 models of different scales. In most evaluated settings across datasets, model scales, architectures, and training configurations, **DVD** outperforms perplexity-based, Min-*k*%++, edit-distance (CDD), and embedding-similarity baselines, while showing strong robustness to hyperparameters. These results support generation-distribution variance as a principled and practical signal for detecting variant contamination, helping improve the reliability of LLM evaluation.