Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Inference-time Alignment in Continuous Space
Authors: Yige Yuan, Teng Xiao, Li Yunfan, Bingbing Xu, Shuchang Tao, Yunqi Qiu, Huawei Shen, Xueqi Cheng
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We empirically demonstrate that, despite its simplicity, SEA enjoys promising performance on extensive benchmarks such as Adv Bench and Truthful QA, consistently and significantly outperforming state-of-the-art baselines across various base models. Additionally, we conduct extensive ablation studies and visualize the dynamic optimization process of SEA, providing deeper insights into its underlying mechanisms. |
| Researcher Affiliation | Collaboration | 1Institute of Computing Technology, Chinese Academy of Sciences 2University of Chinese Academy of Sciences 3University of Washington, 4Allen Institute for AI, 5Alibaba Group EMAIL, EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1 SEA for Inference-time Alignment |
| Open Source Code | Yes | Our code is publicly available at this link. |
| Open Datasets | Yes | Datasets. We evaluate SEA across three tasks: safety, truthfulness, and reasoning. For the safety task, we use Adv Bench [45]... For the truthfulness task, we use Truthful QA [47]... For the reasoning task, we use two datasets: GSM8K [48] and MATH [49]. |
| Dataset Splits | Yes | For the truthfulness task, we use Truthful QA [47] in generation mode and sample 100 queries for evaluation. For the reasoning task, we use two datasets: GSM8K [48] and MATH [49]... We sample 200 samples from each dataset. |
| Hardware Specification | Yes | All the training experiments in this paper were conducted on 4 NVIDIA A100 (80G) GPUs. |
| Software Dependencies | No | The paper mentions the Adam optimizer [61] and building on COLD [32] and COLD-Attack [34], but does not provide specific version numbers for software libraries or environments like Python, PyTorch, or CUDA. |
| Experiment Setup | Yes | Table 7: Hyperparameters of SEA Dataset η 1/α τ k Adv Bench 0.1 0.1 0.1 10 Truthful QA 0.1 0.1 0.05 10 GSM8K 0.01 0.1 0.05 1000 MATH 0.01 0.1 0.05 1000 |