Alignment-Constrained Dynamic Pruning for LLMs: Identifying and Preserving Alignment-Critical Circuits

• URL: http://arxiv.org/abs/2511.07482v1
• Date: Wed, 12 Nov 2025 01:01:22 GMT
• Title: Alignment-Constrained Dynamic Pruning for LLMs: Identifying and Preserving Alignment-Critical Circuits
• Abstract summary: We introduce Alignment-Aware Probe Pruning (AAPP), a dynamic structured pruning method that adaptively preserves alignment-relevant circuits during inference.<n>Experiments on LLaMA 2-7B, Qwen2.5-14B-Instruct, and Gemma-3-12B-IT show AAPP improves refusal rates by 50% at matched compute.
• Score: 8.041642853381909
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Large Language Models require substantial computational resources for inference, posing deployment challenges. While dynamic pruning offers superior efficiency over static methods through adaptive circuit selection, it exacerbates alignment degradation by retaining only input-dependent safety-critical circuit preservation across diverse inputs. As a result, addressing these heightened alignment vulnerabilities remains critical. We introduce Alignment-Aware Probe Pruning (AAPP), a dynamic structured pruning method that adaptively preserves alignment-relevant circuits during inference, building upon Probe Pruning. Experiments on LLaMA 2-7B, Qwen2.5-14B-Instruct, and Gemma-3-12B-IT show AAPP improves refusal rates by 50\% at matched compute, enabling efficient yet safety-preserving LLM deployment.

Related papers

• Stable-MM-R1: Anchoring Multimodal Reasoning Dynamics via Entropy-Guided Stratification [19.490862077232205]
  We propose a data-centric framework to stabilize Reinforcement Learning training.<n>We first introduce Potential-Aware Query Mining (PAQM), which filters data dynamically to focus on the "Distillation Zone"---samples with high potential for capability elicitation.<n>We present Hybrid Stratified Replay (HSR), a novel mechanism that restructures batches by stratifying rollouts based on Path Entropy, a rollout-level confidence proxy, and outcome reward.
  arXiv  Detail & Related papers  (2026-09-07T07:47:00Z)
• Unifying Stable Optimization and Reference Regularization in RLHF [64.16830602324345]
  This paper introduces a unified regularization approach that balances objectives of preventing reward hacking and maintaining stable policy updates.<n>Our simple yet principled alignment objective yields a weighted supervised fine-tuning loss with a superior trade-off, which demonstrably improves both alignment results and implementation complexity.
  arXiv  Detail & Related papers  (2026-02-12T03:31:19Z)
• Adaptive Linear Path Model-Based Diffusion [52.84663832658799]
  We introduce Linear Path Model-Based Diffusion (LP-MBD), which replaces the variance-preserving schedule with a flow-matching linear probability path.<n>We also propose Adaptive LP-MBD (ALP-MBD), which leverages reinforcement learning to adjust diffusion steps and noise levels according to task complexity and environmental conditions.
  arXiv  Detail & Related papers  (2026-02-02T21:33:03Z)
• Not All Preferences Are Created Equal: Stability-Aware and Gradient-Efficient Alignment for Reasoning Models [52.48582333951919]
  We propose a dynamic framework designed to enhance alignment reliability by maximizing the Signal-to-Noise Ratio of policy updates.<n>SAGE (Stability-Aware Gradient Efficiency) integrates a coarse-grained curriculum mechanism that refreshes candidate pools based on model competence.<n> Experiments on multiple mathematical reasoning benchmarks demonstrate that SAGE significantly accelerates convergence and outperforms static baselines.
  arXiv  Detail & Related papers  (2026-02-01T12:56:10Z)
• RISER: Orchestrating Latent Reasoning Skills for Adaptive Activation Steering [62.63376387138257]
  We propose a plug-and-play intervention framework that adaptively steers large language models (LLMs) reasoning in activation space.<n>RISER constructs a library of reusable reasoning vectors and employs a lightweight Router to dynamically compose them for each input.<n>The Router is optimized via reinforcement learning under task-level rewards, activating latent cognitive primitives in an emergent and compositional manner.
  arXiv  Detail & Related papers  (2026-01-14T08:04:33Z)
• Deep Reinforcement Learning Optimization for Uncertain Nonlinear Systems via Event-Triggered Robust Adaptive Dynamic Programming [0.3848364262836075]
  This work proposes a unified control architecture that couples a Reinforcement Learning (RL)-driven controller with a disturbance-rejection Extended State Observer (ESO)<n>The ESO is utilized to estimate the system states and the lumped disturbance in real time, forming the foundation for effective disturbance compensation.
  arXiv  Detail & Related papers  (2025-12-05T22:52:22Z)
• Iterative Refinement of Flow Policies in Probability Space for Online Reinforcement Learning [56.47948583452555]
  We introduce the Stepwise Flow Policy (SWFP) framework, founded on the key insight that discretizing the flow matching inference process via a fixed-step Euler scheme aligns it with the variational Jordan-Kinderlehrer-Otto principle from optimal transport.<n>SWFP decomposes the global flow into a sequence of small, incremental transformations between proximate distributions.<n>This decomposition yields an efficient algorithm that fine-tunes pre-trained flows via a cascade of small flow blocks, offering significant advantages.
  arXiv  Detail & Related papers  (2025-10-17T07:43:51Z)
• Stabilizing Policy Gradients for Sample-Efficient Reinforcement Learning in LLM Reasoning [77.92320830700797]
  Reinforcement Learning has played a central role in enabling reasoning capabilities of Large Language Models.<n>We propose a tractable computational framework that tracks and leverages curvature information during policy updates.<n>The algorithm, Curvature-Aware Policy Optimization (CAPO), identifies samples that contribute to unstable updates and masks them out.
  arXiv  Detail & Related papers  (2025-10-01T12:29:32Z)
• Steering When Necessary: Flexible Steering Large Language Models with Backtracking [16.23081952791394]
  Large language models (LLMs) have achieved remarkable performance across many generation tasks.<n> Activation steering is an effective and cost-efficient approach that directly modifies the activations of LLMs during the inference stage.<n>We propose the Flexible Activation Steering with Backtracking (FASB) framework, which dynamically determines both the necessity and strength of intervention.
  arXiv  Detail & Related papers  (2025-08-25T03:01:30Z)
• Linear Preference Optimization: Decoupled Gradient Control via Absolute Regularization [13.97375970293678]
  DPO (Direct Preference Optimization) has become a widely used offline preference optimization algorithm due to its simplicity and training stability.<n>We propose Linear Preference Optimization (LPO), a novel alignment framework featuring three key innovations.<n>First, we introduce gradient decoupling by replacing the log-sigmoid function with an absolute difference loss, thereby isolating the optimization dynamics.<n>Second, we improve stability through an offset constraint combined with a positive regularization term to preserve the chosen response quality.<n>Third, we implement controllable rejection suppression using gradient separation with straightforward estimation and a tunable coefficient that linearly regulates the descent of the rejection probability.
  arXiv  Detail & Related papers  (2025-08-20T10:17:29Z)
• R-Sparse: Rank-Aware Activation Sparsity for Efficient LLM Inference [77.47238561728459]
  R-Sparse is a training-free activation sparsity approach capable of achieving high sparsity levels in advanced LLMs.<n> Experiments on Llama-2/3 and Mistral models across ten diverse tasks demonstrate that R-Sparse achieves comparable performance at 50% model-level sparsity.
  arXiv  Detail & Related papers  (2025-04-28T03:30:32Z)
• AdaSteer: Your Aligned LLM is Inherently an Adaptive Jailbreak Defender [99.3105257001476]
  We propose AdaSteer, an adaptive activation steering method that adjusts model behavior based on input characteristics.<n>AdaSteer steers input representations along both the Rejection Direction (RD) and Harmfulness Direction (HD)<n>Our results highlight the potential of interpretable model internals for real-time, flexible safety enforcement in LLMs.
  arXiv  Detail & Related papers  (2025-04-13T07:39:17Z)
• DISC: Dynamic Decomposition Improves LLM Inference Scaling [83.5708537758088]
  Inference scaling methods for LLMs often rely on decomposing problems into steps, followed by sampling and selecting the best next steps.<n>We propose dynamic decomposition, a method that adaptively and automatically partitions solution and reasoning traces into manageable steps during inference.
  arXiv  Detail & Related papers  (2025-02-23T20:37:32Z)
• Constrain Alignment with Sparse Autoencoders [45.131670081186]
  Feature-level constrained Preference Optimization is a novel method designed to simplify the alignment process while ensuring stability.<n>Our approach enjoys efficiency by using sparse features activated in a well-trained sparse autoencoder and the quality of sequential KL divergence.
  arXiv  Detail & Related papers  (2024-11-12T07:54:13Z)

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.