mlsanityData to 14 Oct 2026

Beyond Token-Level Guidance: Inference-Time Alignment of Specialized LLMs via Cross-Family Representation Steering

arXiv 2608.30319 ↗First posted 31 Aug 2026cs.CL cs.AI cs.LGcited by 1

On the map: Mechanistic interpretability → Activation steering methods

Abstract (arXiv)

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.

Key facts. Cited by 1 paper in the map (data through 14 Oct 2026). Placed in Activation steering methods, one of 13 directions in Mechanistic interpretability. Builds on 17 of its 27 references that are in the map. First posted on arXiv 31 Aug 2026.

Builds on

17 of 27 references.

Closest ancestors · 1

Same direction, newest first

  1. 2023Steering Llama 2 via Contrastive Activation Addition9 Dec

Other references · 16

  1. 2021Training Verifiers to Solve Math Word Problemscited by 10,109
  2. 2021Evaluating Large Language Models Trained on Codecited by 8,320
  3. 2016Overcoming catastrophic forgetting in neural networkscited by 4,787
  4. 2020What Disease does this Patient Have? A Large-scale Open Domain Question Answering Dataset from Medical Examscited by 1,356
  5. 2023Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To!cited by 1,154
  6. 2024Refusal in Language Models Is Mediated by a Single Directioncited by 1,032
  7. 2024Safety Alignment Should Be Made More Than Just a Few Tokens Deepcited by 472
  8. 2024Keeping LLMs Aligned After Fine-tuning: The Crucial Role of Prompt Templatescited by 96
  9. 2024Booster: Tackling Harmful Fine-tuning for Large Language Models via Attenuating Harmful Perturbationcited by 92
  10. 2024InferAligner: Inference-Time Alignment for Harmlessness through Cross-Model Guidancecited by 83
  11. 2025Layer-Aware Representation Filtering: Purifying Finetuning Data to Preserve LLM Safety Alignmentcited by 24
  12. 2024Nudging: Inference-time Alignment of LLMs via Guided Decodingcited by 23
4 more
  1. 2024Locking Down the Finetuned LLMs Safetycited by 22
  2. 2024Inference-Time Language Model Alignment via Integrated Value Guidancecited by 19
  3. 2025Rethinking Safety in LLM Fine-tuning: An Optimization Perspectivecited by 14
  4. 2026To Intervene or Not: Guiding Inference-time Alignment with Probabilistic Model Blendingcited by 2

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 1.

1 paper.

  1. 2026Share the Symbol, Not the Vector: Training-Free Fact Communication Across Language Modelscited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2025On Almost Surely Safe Alignment of Large Language Models at Inference-Time
  2. 2025Any-Depth Alignment: Unlocking Innate Safety Alignment of LLMs to Any-Depth
  3. 2026LASA: Language-Agnostic Semantic Alignment at the Semantic Bottleneck for LLM Safety
  4. 2026DNAlign: Dynamic Null-Space Safe Alignment for LLMs
  5. 2025ExpertSteer: Intervening in LLMs through Expert Knowledge
  6. 2025The Hidden Space of Safety: Understanding Preference-Tuned LLMs in Multilingual context
  7. 2026ALIGNBEAM : Inference-Time Alignment Transfer via Cross-Vocabulary Logit Mixing
  8. 2024Towards Inference-time Category-wise Safety Steering for Large Language Models