Beyond Token-Level Guidance: Inference-Time Alignment of Specialized LLMs via Cross-Family Representation Steering
2026-08-31 • Computation and Language
Computation and LanguageArtificial IntelligenceMachine Learning
AI summaryⓘ
The authors study how to make specialized language models safer without losing their expertise. They find that current methods have trouble because general guidance models and specialized base models work in very different ways, causing conflicts like ignoring when to stop generating text. To fix this, they create a method called CREST that adjusts the internal workings of the base model using safety information from another model, avoiding problems from token-level rules. Their method improves safety while keeping the model's specialized skills and performs better than previous approaches.
large language modelsfine-tuninginference-time alignmentspecialized modelssafety in AIhidden representationsguidance modelstoken stoppingmodel alignmentnatural language generation
Authors
Jin Gan, Xin Li, Jun Luo
Abstract
Large language models (LLMs) finetuned for specialized domains represent crucial high-impact applications. Inference-time alignment improves safety degraded from specialization finetuning without requiring substantial computational resources, complementing finetuning-based methods with an easy-to-use, plug-and-play solution. However, existing inference-time methods fail to reliably improve safety without disrupting domain capability. We identify the root cause as complementary expertise orthogonality: specialized base models and general-domain guidance models have orthogonal competencies, making the guidance signal unreliable for specialized generation. This primarily manifests as stop token interference, where the guidance model's tendency toward continuation overrides the base model's decision to stop, burying correct answers under guidance-induced continuation. To address this problem, we propose CREST, an inference-time alignment method that steers base model hidden representations using safety directions extracted from a guidance model of any family, avoiding token-level structural limitations entirely. CREST improves safety where specialization has weakened it while preserving both domain-specific capability and the safety of already well-aligned models, outperforming baselines by up to 22.2\% on safety benchmarks. Our code is available at: https://github.com/DecayingSeart/CREST.