Value Drifts: Tracing Value Alignment During LLM Post-Training

• URL: http://arxiv.org/abs/2510.26707v1
• Date: Thu, 30 Oct 2025 17:09:09 GMT
• Title: Value Drifts: Tracing Value Alignment During LLM Post-Training
• Abstract summary: We investigate how and at which stage value alignment arises during the course of a model's post-training.<n>Our findings provide actionable insights into how values are learned during post-training.
• Score: 37.472751808508235
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: As LLMs occupy an increasingly important role in society, they are more and more confronted with questions that require them not only to draw on their general knowledge but also to align with certain human value systems. Therefore, studying the alignment of LLMs with human values has become a crucial field of inquiry. Prior work, however, mostly focuses on evaluating the alignment of fully trained models, overlooking the training dynamics by which models learn to express human values. In this work, we investigate how and at which stage value alignment arises during the course of a model's post-training. Our analysis disentangles the effects of post-training algorithms and datasets, measuring both the magnitude and time of value drifts during training. Experimenting with Llama-3 and Qwen-3 models of different sizes and popular supervised fine-tuning (SFT) and preference optimization datasets and algorithms, we find that the SFT phase generally establishes a model's values, and subsequent preference optimization rarely re-aligns these values. Furthermore, using a synthetic preference dataset that enables controlled manipulation of values, we find that different preference optimization algorithms lead to different value alignment outcomes, even when preference data is held constant. Our findings provide actionable insights into how values are learned during post-training and help to inform data curation, as well as the selection of models and algorithms for preference optimization to improve model alignment to human values.

Related papers

• Efficient Controller Learning from Human Preferences and Numerical Data Via Multi-Modal Surrogate Models [1.2526963688768455]
  We propose a multi-fidelity, multi-modal Bayesian optimization framework that integrates low-fidelity numerical data with high-fidelity human preferences.<n>We illustrate the framework by tuning an autonomous vehicle's trajectory planner, showing that combining numerical and preference data significantly reduces the need for experiments involving the human decision maker.
  arXiv  Detail & Related papers  (2026-03-25T10:04:03Z)
• Holdout-Loss-Based Data Selection for LLM Finetuning via In-Context Learning [19.677969862434708]
  We present a theoretically grounded, resource-efficient framework for data selection and reweighting.<n>At its core is an In-Context Approximation (ICA) that estimates the holdout loss a model would incur after training on a candidate example.<n>We derive per-example weights from ICA scores, dynamically reweighting gradient updates as model parameters evolve.
  arXiv  Detail & Related papers  (2025-10-16T09:00:39Z)
• Towards Understanding Valuable Preference Data for Large Language Model Alignment [85.38864561060088]
  Large language model (LLM) alignment is typically achieved through learning from human preference comparisons.<n>We assess data quality through individual influence on validation data using our newly proposed truncated influence function (TIF)<n>To this end, we combine them to offset their diverse error sources, resulting in a simple yet effective data selection rule.
  arXiv  Detail & Related papers  (2025-10-15T06:57:55Z)
• Middo: Model-Informed Dynamic Data Optimization for Enhanced LLM Fine-Tuning via Closed-Loop Learning [44.53583316198435]
  Supervised Fine-Tuning (SFT) Large Language Models rely on high-quality training data.<n>We introduce Middo, a self-evolving Model-informed dynamic data optimization framework.<n>We show that Middo consistently enhances the quality of seed data and boosts LLM's performance with improving accuracy by 7.15% on average.
  arXiv  Detail & Related papers  (2025-08-29T12:47:27Z)
• LLM Data Selection and Utilization via Dynamic Bi-level Optimization [100.20933466418786]
  We propose a new Data Weighting Model (DWM) to adjust the weight of selected data within each batch to achieve a dynamic data utilization during training.<n>Our experiments demonstrate that DWM enhances the performance of models trained with randomly-selected data.<n>We further analyze how a model's data preferences evolve throughout training, providing new insights into the data preference of the model during training.
  arXiv  Detail & Related papers  (2025-07-22T02:47:12Z)
• Capturing the Temporal Dependence of Training Data Influence [100.91355498124527]
  We formalize the concept of trajectory-specific leave-one-out influence, which quantifies the impact of removing a data point during training.<n>We propose data value embedding, a novel technique enabling efficient approximation of trajectory-specific LOO.<n>As data value embedding captures training data ordering, it offers valuable insights into model training dynamics.
  arXiv  Detail & Related papers  (2024-12-12T18:28:55Z)
• Self-Steering Optimization: Autonomous Preference Optimization for Large Language Models [79.84205827056907]
  We present Self-Steering Optimization ($SSO$), an algorithm that autonomously generates high-quality preference data.<n>$SSO$ employs a specialized optimization objective to build a data generator from the policy model itself, which is used to produce accurate and on-policy data.<n>Our evaluation shows that $SSO$ consistently outperforms baselines in human preference alignment and reward optimization.
  arXiv  Detail & Related papers  (2024-10-22T16:04:03Z)
• Optimizing LLMs with Direct Preferences: A Data Efficiency Perspective [4.548047308860141]
  This study investigates the impact of different type of preference data on model performance. It aims to reduce their dependency on extensive amounts of preference data, which is expensive to collect.
  arXiv  Detail & Related papers  (2024-10-22T00:11:41Z)
• LAVA: Data Valuation without Pre-Specified Learning Algorithms [20.578106028270607]
  We introduce a new framework that can value training data in a way that is oblivious to the downstream learning algorithm. We develop a proxy for the validation performance associated with a training set based on a non-conventional class-wise Wasserstein distance between training and validation sets. We show that the distance characterizes the upper bound of the validation performance for any given model under certain Lipschitz conditions.
  arXiv  Detail & Related papers  (2023-04-28T19:05:16Z)
• Data-Driven Offline Decision-Making via Invariant Representation Learning [97.49309949598505]
  offline data-driven decision-making involves synthesizing optimized decisions with no active interaction. A key challenge is distributional shift: when we optimize with respect to the input into a model trained from offline data, it is easy to produce an out-of-distribution (OOD) input that appears erroneously good. In this paper, we formulate offline data-driven decision-making as domain adaptation, where the goal is to make accurate predictions for the value of optimized decisions.
  arXiv  Detail & Related papers  (2022-11-21T11:01:37Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.