The "alignment tax" of post-training is typically framed as a drop in task accuracy. We show it also involves a severe loss of calibration, making models overconfident, less reliable, and model outputs less diverse. We show that this trade-off can be navigated effectively via a simple post-hoc intervention: interpolating between a model's weights before and after alignment. Crucially, this is not a strict trade-off. We find that the process consistently reveals Pareto-optimal interpolations - models that improve accuracy beyond both parents while substantially recovering the calibration lost during alignment. Our work demonstrates that simple model merging provides a computationally efficient method for mitigating the full scope of the alignment tax, yielding models that are more capable and more reliable.

Existing alignment techniques for Large Language Models (LLMs), such as Direct Preference Optimization (DPO), typically treat the model as a monolithic entity, applying uniform optimization pressure across all layers. This approach overlooks the functional specialization within the Transformer architecture, where different layers are known to handle distinct tasks from syntax to abstract reasoning. In this paper, we challenge this one-size-fits-all paradigm by introducing Hierarchical Alignment, a novel method that applies targeted DPO to distinct functional blocks of a model's layers: local (syntax), intermediate (logic), and global (factuality). Through a series of controlled experiments on state-of-the-art models like Llama-3.1-8B and Qwen1.5-7B using LoRA for surgical fine-tuning, our results, evaluated by a powerful LLM-as-Judge, demonstrate significant and predictable improvements. Specifically, aligning the local layers (Local-Align) enhances grammatical fluency. More importantly, aligning the global layers (Global-Align) not only improves factual consistency as hypothesized but also proves to be the most effective strategy for enhancing logical coherence, outperforming all baselines. Critically, all hierarchical strategies successfully avoid the "alignment tax" observed in standard DPO, where gains in fluency come at the cost of degraded logical reasoning. These findings establish a more resource-efficient, controllable, and interpretable path for model alignment, highlighting the immense potential of shifting from monolithic optimization to structure-aware surgical fine-tuning to build more advanced and reliable LLMs.

Large Language Models (LLMs) require alignment with human preferences to avoid generating offensive, false, or meaningless content. Recently, low-resource methods for LLM alignment have been popular, while still facing challenges in obtaining both high-quality and aligned content. Motivated by the observation that the difficulty of generating aligned responses is concentrated at the beginning of decoding, we propose a novel framework, Weak-to-Strong Decoding (WSD), to enhance the alignment ability of base models by the guidance of a small aligned model. The small model first drafts well-aligned beginnings, followed by the large base model to continue the rest, controlled by a well-designed auto-switch mechanism. We also collect a new dataset, GenerAlign, to fine-tune a small-sized Pilot-3B as the draft model, which effectively enhances different base models under the WSD framework to outperform all baseline methods, while avoiding degradation on downstream tasks, termed as the alignment tax. Extensive experiments are further conducted to examine the impact of different settings and time efficiency, as well as analyses on the intrinsic mechanisms of WSD in depth.

Direct Preference Optimization (DPO) has emerged as a more computationally efficient alternative to Reinforcement Learning from Human Feedback (RLHF) with Proximal Policy Optimization (PPO), eliminating the need for reward models and online sampling. Despite these benefits, DPO and its variants remain sensitive to hyper-parameters and prone to instability, particularly on mathematical datasets. We argue that these issues arise from the unidirectional likelihood-derivative negative feedback inherent in the log-likelihood loss function. To address this, we propose a novel LLM alignment loss that establishes a stable Bidirectional Negative Feedback (BNF) during optimization. Our proposed BNF loss eliminates the need for pairwise contrastive losses and does not require any extra tunable hyper-parameters or pairwise preference data, streamlining the alignment pipeline to be as simple as supervised fine-tuning. We conduct extensive experiments across two challenging QA benchmarks and four reasoning benchmarks. The experimental results show that BNF achieves comparable performance to the best methods on QA benchmarks, while its performance decrease on the four reasoning benchmarks is significantly lower compared to the best methods, thus striking a better balance between value alignment and reasoning ability. In addition, we further validate the performance of BNF on non-pairwise datasets, and conduct in-depth analysis of log-likelihood and logit shifts across different preference optimization methods.

As large language models (LLMs) are overwhelmingly more and more integrated into various applications, ensuring they generate safe and aligned responses is a pressing need. Previous research on alignment has largely focused on general instruction-following but has often overlooked the unique properties and challenges of safety alignment, such as the brittleness of safety mechanisms. To bridge the gap, we propose the Superficial Safety Alignment Hypothesis (SSAH), which posits that safety alignment should teach an otherwise unsafe model to choose the correct reasoning direction - interpreted as a specialized binary classification task - and incorporate a refusal mechanism with multiple reserved fallback options. Furthermore, through SSAH, we hypothesize that safety guardrails in LLMs can be established by just a small number of essential components. To verify this, we conduct an ablation study and successfully identify four types of attribute-critical components in safety-aligned LLMs: Exclusive Safety Unit (ESU), Exclusive Utility Unit (EUU), Complex Unit (CU), and Redundant Unit (RU). Our findings show that freezing certain safety-critical components 7.5\% during fine-tuning allows the model to retain its safety attributes while adapting to new tasks. Additionally, we show that leveraging redundant units 20\% in the pre-trained model as an ``alignment budget'' can effectively minimize the alignment tax while achieving the alignment goal. All considered, this paper concludes that the atomic functional unit for safety in LLMs is at the neuron level and underscores that safety alignment should not be complicated. We believe this work contributes to the foundation of efficient and scalable safety alignment for future LLMs.

Large language models (LLMs) have shown remarkable abilities in diverse natural language processing (NLP) tasks. The LLMs generally undergo supervised fine-tuning (SFT) followed by preference alignment to be usable in downstream applications. However, this sequential training pipeline leads to alignment tax that degrades the LLM performance. This paper introduces PAFT, a new PArallel training paradigm for effective LLM Fine-Tuning, which independently performs SFT and preference alignment (e.g., DPO and ORPO, etc.) with the same pre-trained model on respective datasets. The model produced by SFT and the model from preference alignment are then merged into a final model by parameter fusing for use in downstream applications. This work reveals important findings that preference alignment like DPO naturally results in a sparse model while SFT leads to a natural dense model which needs to be sparsified for effective model merging. This paper introduces an effective interference resolution which reduces the redundancy by sparsifying the delta parameters. The LLM resulted from the new training paradigm achieved Rank #1 on the HuggingFace Open LLM Leaderboard. Comprehensive evaluation shows the effectiveness of the parallel training paradigm.

Effectively aligning Large Language Models (LLMs) with human-centric values while preventing the degradation of abilities acquired through Pre-training and Supervised Fine-tuning (SFT) poses a central challenge in Reinforcement Learning from Human Feedback (RLHF). In this paper, we first discover that interpolating RLHF and SFT model parameters can adjust the trade-off between human preference and basic capabilities, thereby reducing the alignment tax at the cost of alignment reward. Inspired by this, we propose integrating the RL policy and SFT models at each optimization step in RLHF to continuously regulate the training direction, introducing the Online Merging Optimizer. Specifically, we merge gradients with the parameter differences between SFT and pretrained models, effectively steering the gradient towards maximizing rewards in the direction of SFT optimization. We demonstrate that our optimizer works well with different LLM families, such as Qwen and LLaMA, across various model sizes ranging from 1.8B to 8B, various RLHF algorithms like DPO and KTO, and existing model merging methods. It significantly enhances alignment reward while mitigating alignment tax, achieving higher overall performance across 14 benchmarks.

Supervised fine-tuning (SFT) on instruction-following corpus is a crucial approach toward the alignment of large language models (LLMs). However, the performance of LLMs on standard knowledge and reasoning benchmarks tends to suffer from deterioration at the latter stage of the SFT process, echoing the phenomenon of alignment tax. Through our pilot study, we put a hypothesis that the data biases are probably one cause behind the phenomenon. To address the issue, we introduce a simple disperse-then-merge framework. To be concrete, we disperse the instruction-following data into portions and train multiple sub-models using different data portions. Then we merge multiple models into a single one via model merging techniques. Despite its simplicity, our framework outperforms various sophisticated methods such as data curation and training regularization on a series of standard knowledge and reasoning benchmarks.

Current large language models have dangerous capabilities, which are likely to become more problematic in the future. Activation steering techniques can be used to reduce risks from these capabilities. In this paper, we investigate the efficacy of activation steering for broad skills and multiple behaviours. First, by comparing the effects of reducing performance on general coding ability and Python-specific ability, we find that steering broader skills is competitive to steering narrower skills. Second, we steer models to become more or less myopic and wealth-seeking, among other behaviours. In our experiments, combining steering vectors for multiple different behaviours into one steering vector is largely unsuccessful. On the other hand, injecting individual steering vectors at different places in a model simultaneously is promising.

LLMs acquire a wide range of abilities during pre-training, but aligning LLMs under Reinforcement Learning with Human Feedback (RLHF) can lead to forgetting, which is also known as the alignment tax. To empirically verify this hypothesis, we conducted experiments with existing RLHF algorithms using OpenLLaMA-3B, which revealed a pronounced alignment tax in NLP tasks. On the other hand, despite various techniques to mitigate forgetting, they are often at odds with the RLHF performance, leading to a trade-off between reward maximization and forgetting mitigation. In light of the above pressing issue in aligning LLMs, in this paper we explore model averaging, which interpolates between pre and post RLHF model weights, to achieve a more efficient reward-tax Pareto front. To understand its effectiveness, We offer theoretical insights into model averaging, revealing that it enhances performance Pareto front by increasing feature diversity on the layers where tasks share overlapped feature spaces. Empirical evidence corroborates our analysis by showing the benefits of averaging low-level transformer layers. Building on the analysis and the observation that averaging different layers of the transformer leads to significantly different reward-tax trade-offs, we propose Adaptive Model Averaging (AMA) to adaptively find various combination ratios of model layers. AMA seeks to maximize the alignment reward while incurring minimal alignment tax. Moreover, we validate AMA's performance across a range of RLHF algorithms over OpenLLaMA-3B and further extend our findings to Mistral-7B.