mlsanityData to 14 Oct 2026

Steering Under Compression: Dose-Response, Capability Cost, and Failure Asymmetry in Quantized LLMs

arXiv 2609.06473 ↗First posted 6 Sep 2026cs.LGcited by 0

On the map: Mechanistic interpretability → Steering and interpreting LLM reasoning

Abstract (arXiv)

Inference-time activation steering enables behavioral control of large language models without parameter modification, while post-training quantization reduces memory and compute costs for deployment. Despite their growing convergence in practice, the interaction between these two techniques remains uncharacterized. We systematically study activation steering under weight-only quantization (INT8 and NF4) across four open-weight 7-9B models and two behavioral targets: judged sentiment and judge-free reasoning length. Using an iso-effect framework that compares capability costs at matched behavioral effect, we find that sentiment steering survives quantization intact. After correcting a GSM8K parser artifact with a uniform v2.3.1 rescore, the pooled INT8 contrast is -0.010 (90% CI [-0.026, +0.007]), descriptively Equivalent under the preregistered three-label rule, while NF4 remains Inconclusive at -0.017 ([-0.067, +0.033]). In contrast, reasoning length exhibits a surprising asymmetric dose-response: lengthening is graded but terminates in cap-runaway and collapse, while shortening is a step function with only 12-30% shortening (model-dependent) before discontinuous failure. We expose a methodological pitfall: the naive iso-effect ladder anchors on the collapse floor for floor-bounded targets, and we introduce a censored construction that restores interpretable crossings. We also quantify a substantial baseline capability shift for Mistral-NF4 (0.545 to 0.365 GSM8K at alpha=0), demonstrating that compression can dominate the steering intervention. Despite this, steering vectors remain highly collinear with their FP16 siblings (cosine similarity 0.989-0.998 for INT8, 0.945-0.990 for NF4), confirming that the behavioral direction survives quantization even when the cost structure does not. All code and data are released.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Steering and interpreting LLM reasoning, one of 13 directions in Mechanistic interpretability. Builds on 27 of its 32 references that are in the map. First posted on arXiv 6 Sep 2026.

Builds on

27 of 32 references.

References · 27

  1. 2024The Llama 3 Herd of Modelscited by 14,715
  2. 2021Training Verifiers to Solve Math Word Problemscited by 10,109
  3. 2020Measuring Massive Multitask Language Understandingcited by 7,963
  4. 2024Qwen2.5 Technical Reportcited by 4,942
  5. 2018Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challengecited by 4,886
  6. 2023QLoRA: Efficient Finetuning of Quantized LLMscited by 3,418
  7. 2023Mistral 7Bcited by 3,164
  8. 2021TruthfulQA: Measuring How Models Mimic Human Falsehoodscited by 2,975
  9. 2024Gemma 2: Improving Open Language Models at a Practical Sizecited by 1,998
  10. 2022GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformerscited by 1,919
  11. 2022SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Modelscited by 1,386
  12. 2023Representation Engineering: A Top-Down Approach to AI Transparencycited by 1,336
15 more
  1. 2023Inference-Time Intervention: Eliciting Truthful Answers from a Language Modelcited by 1,302
  2. 2023AWQ: Activation-aware Weight Quantization for LLM Compression and Accelerationcited by 1,298
  3. 2022Editing Models with Task Arithmeticcited by 1,120
  4. 2023Steering Llama 2 via Contrastive Activation Additioncited by 1,085
  5. 2022LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scalecited by 906
  6. 2024QuaRot: Outlier-Free 4-Bit Inference in Rotated LLMscited by 634
  7. 2024Massive Activations in Large Language Modelscited by 294
  8. 2024The Super Weight in Large Language Modelscited by 52
  9. 2026Locate, Steer, and Improve: A Practical Survey of Actionable Mechanistic Interpretability in Large Language Modelscited by 27
  10. 2025Steering Language Models with Weight Arithmeticcited by 20
  11. 2026Why Steering Works: Toward a Unified View of Language Model Parameter Dynamicscited by 9
  12. 2026Steered LLM Activations are Non-Surjectivecited by 8
  13. 2026From Weights to Activations: Is Steering the Next Frontier of Adaptation?cited by 8
  14. 2026Mind the Performance Gap: Capability-Behavior Trade-offs in Feature Steeringcited by 3
  15. 2025Fusion Steering: Prompt-Specific Activation Controlcited by 1

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2025Activation Steering for Chain-of-Thought Compression
  2. 2024Exploring the Trade-Offs: Quantization Methods, Task Difficulty, and Model Size in Large Language Models From Edge to Giant
  3. 2026Mechanistic Indicators of Steering Effectiveness in Large Language Models
  4. 2026FaithSteer-BENCH: A Deployment-Aligned Stress-Testing Benchmark for Inference-Time Steering
  5. 2026Quantization Inflates Reasoning: Token Inflation as a Hidden Cost of Low-Bit Reasoning Models
  6. 2025A Comparative Analysis of Sparse Autoencoder and Activation Difference in Language Model Steering
  7. 2026Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels
  8. 2026Quantized Reasoning Models Think They Need to Think Longer, but They Do Not