The Anatomy of Alignment: Decomposing Preference Optimization by Steering Sparse Features

• URL: http://arxiv.org/abs/2509.12934v2
• Date: Thu, 25 Sep 2025 20:31:28 GMT
• Title: The Anatomy of Alignment: Decomposing Preference Optimization by Steering Sparse Features
• Authors: Jeremias Ferrao, Matthijs van der Lende, Ilija Lichkovski, Clement Neo,
• Abstract summary: We introduce Feature Steering with Reinforcement Learning, a framework that trains a lightweight adapter to steer model behavior by modulating interpretable sparse features.<n>We show that this mechanism is principled and expressive enough to approximate the behavioral shifts of post-training processes.<n>Overall, FSRL offers an interpretable control interface and a practical way to diagnose how preference optimization pressures manifest at the feature level.
• Score: 1.7832672957068079
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Prevailing alignment methods induce opaque parameter changes, making it difficult to audit what the model truly learns. To address this, we introduce Feature Steering with Reinforcement Learning (FSRL), a framework that trains a lightweight adapter to steer model behavior by modulating interpretable sparse features. First, we theoretically show that this mechanism is principled and expressive enough to approximate the behavioral shifts of post-training processes. Then, we apply this framework to the task of preference optimization and perform a causal analysis of the learned policy. We find that the model relies on stylistic presentation as a proxy for quality, disproportionately steering features related to style and formatting over those tied to alignment concepts like honesty. Despite exploiting this heuristic, FSRL proves to be an effective alignment method, achieving a substantial reduction in preference loss. Overall, FSRL offers an interpretable control interface and a practical way to diagnose how preference optimization pressures manifest at the feature level.

Related papers

• Why Steering Works: Toward a Unified View of Language Model Parameter Dynamics [81.80010043113445]
  Local weight fine-tuning, LoRA-based adaptation, and activation-based interventions are studied in isolation.<n>We present a unified view that frames these interventions as dynamic weight updates induced by a control signal.<n>Across methods, we observe a consistent trade-off between preference and utility: stronger control increases preference while predictably reducing utility.
  arXiv  Detail & Related papers  (2026-02-02T17:04:36Z)
• How to Set the Learning Rate for Large-Scale Pre-training? [73.03133634525635]
  We formalize this investigation into two distinct research paradigms: Fitting and Transfer.<n>Within the Fitting Paradigm, we introduce a Scaling Law for search factor, effectively reducing the search complexity from O(n3) to O(n*C_D*C_) via predictive modeling.<n>We extend the principles of $$Transfer to the Mixture of Experts (MoE) architecture, broadening its applicability to encompass model depth, weight decay, and token horizons.
  arXiv  Detail & Related papers  (2026-01-08T15:55:13Z)
• From RLHF to Direct Alignment: A Theoretical Unification of Preference Learning for Large Language Models [0.7366405857677227]
  This survey provides a textittheoretical unification of preference learning methods.<n>We formalize each axis with precise definitions and theorems.<n>We synthesize empirical findings across 50+ papers and provide a practitioner's decision guide for method selection.
  arXiv  Detail & Related papers  (2026-01-03T08:33:26Z)
• The Path Not Taken: RLVR Provably Learns Off the Principals [85.41043469428365]
  We show that sparsity is a surface artifact of a model-conditioned optimization bias.<n>We mechanistically explain these dynamics with a Three-Gate Theory.<n>We provide a parameter-level characterization of RLVR's learning dynamics.
  arXiv  Detail & Related papers  (2025-11-11T18:49:45Z)
• AdaLRS: Loss-Guided Adaptive Learning Rate Search for Efficient Foundation Model Pretraining [12.630306478872043]
  We propose textbfAdaLRS, a plug-in-and-play adaptive learning rate search algorithm that conducts online optimal learning rate search.<n>Experiments show that AdaLRS adjusts suboptimal learning rates to the neighborhood of optimum with marked efficiency and effectiveness.
  arXiv  Detail & Related papers  (2025-06-16T09:14:01Z)
• Diffusion Guidance Is a Controllable Policy Improvement Operator [98.11511661904618]
  CFGRL is trained with the simplicity of supervised learning, yet can further improve on the policies in the data.<n>On offline RL tasks, we observe a reliable trend -- increased guidance weighting leads to increased performance.
  arXiv  Detail & Related papers  (2025-05-29T14:06:50Z)
• Solver-Informed RL: Grounding Large Language Models for Authentic Optimization Modeling [3.253908111652627]
  Large Language Models (LLMs) often struggle to generate formally correct and usable models against hallucinations.<n>We present a novel framework that significantly improves the authenticity of LLMs for optimization modeling using Reinforcement Learning with Verifiable Reward.
  arXiv  Detail & Related papers  (2025-05-17T02:32:03Z)
• Surrogate Fitness Metrics for Interpretable Reinforcement Learning [7.889696505137217]
  We employ an evolutionary optimization framework that perturbs initial states to generate informative and diverse policy demonstrations.<n>A joint surrogate fitness function guides the optimization by combining local diversity, behavioral certainty, and global population diversity.<n>By refining and systematically analyzing surrogate fitness functions, this study advances the interpretability of RL models.
  arXiv  Detail & Related papers  (2025-04-20T15:01:19Z)
• Align-Pro: A Principled Approach to Prompt Optimization for LLM Alignment [40.71270945505082]
  Large language models (LLMs) are increasingly integrated into various societal and decision-making processes.<n>Traditional methods, such as reinforcement learning from human feedback (RLHF), achieve alignment by fine-tuning model parameters.<n>In contrast, prompt optimization is a viable alternative to RLHF for LLM alignment.
  arXiv  Detail & Related papers  (2025-01-07T03:14:39Z)
• Self-supervised Preference Optimization: Enhance Your Language Model with Preference Degree Awareness [27.43137305486112]
  We propose a novel Self-supervised Preference Optimization (SPO) framework, which constructs a self-supervised preference degree loss combined with the alignment loss. The results demonstrate that SPO can be seamlessly integrated with existing preference optimization methods to achieve state-of-the-art performance.
  arXiv  Detail & Related papers  (2024-09-26T12:37:26Z)
• Linear Alignment: A Closed-form Solution for Aligning Human Preferences without Tuning and Feedback [70.32795295142648]
  Linear alignment is a novel algorithm that aligns language models with human preferences in one single inference step. Experiments on both general and personalized preference datasets demonstrate that linear alignment significantly enhances the performance and efficiency of LLM alignment.
  arXiv  Detail & Related papers  (2024-01-21T10:46:23Z)
• Optimal Goal-Reaching Reinforcement Learning via Quasimetric Learning [73.80728148866906]
  Quasimetric Reinforcement Learning (QRL) is a new RL method that utilizes quasimetric models to learn optimal value functions. On offline and online goal-reaching benchmarks, QRL also demonstrates improved sample efficiency and performance.
  arXiv  Detail & Related papers  (2023-04-03T17:59:58Z)
• When to Update Your Model: Constrained Model-based Reinforcement Learning [50.74369835934703]
  We propose a novel and general theoretical scheme for a non-decreasing performance guarantee of model-based RL (MBRL) Our follow-up derived bounds reveal the relationship between model shifts and performance improvement. A further example demonstrates that learning models from a dynamically-varying number of explorations benefit the eventual returns.
  arXiv  Detail & Related papers  (2022-10-15T17:57:43Z)
• Learning Off-Policy with Online Planning [18.63424441772675]
  We investigate a novel instantiation of H-step lookahead with a learned model and a terminal value function. We show the flexibility of LOOP to incorporate safety constraints during deployment with a set of navigation environments.
  arXiv  Detail & Related papers  (2020-08-23T16:18:44Z)

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.