Life sciences · Preprint

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

arXiv · September 6, 2026

Preliminary Why this matters to you Expert discussion

Early or partial results. Treat as a signal, not a conclusion.

AI Summary

This preprint characterizes how weight-only quantization (INT8 and NF4) affects the reliability and cost of inference-time activation steering across open-weight language models. Sentiment steering appears robust to quantization, but reasoning-length steering exhibits asymmetric dose-response with stepwise failures in the shortening direction, and baseline capability shifts can exceed steering effects—yet steering vectors themselves remain geometrically aligned with unquantized versions.

What was studied

Systematic controlled study across multiple models and quantization schemes with preregistered analysis. Four open-weight language models (Mistral and three others, 7–9B parameters); no human subjects or clinical populations studied.. Intervention: Inference-time activation steering under INT8 and NF4 weight-only quantization. Compared with: FP16 (unquantized) baseline and steering-free quantized models.

Key findings

Sentiment steering survives quantization with pooled INT8 contrast of −0.010 (90% CI [−0.026, +0.007]), rated Equivalent under preregistered three-label rule NF4 sentiment steering remains Inconclusive at −0.017 (90% CI [−0.067, +0.033]) Reasoning length shortening exhibits step-function failure with only 12–30% shortening (model-dependent) before discontinuous collapse

Safety

Reasoning length shortening exhibits step-function failure with only 12–30% shortening (model-dependent) before discontinuous collapse

Who this may help

The source did not state who this applies to in practice.

Plain-language summary

This is a preprint systematic study of an understudied technical interaction (activation steering under quantization) with novel methodological contributions, but it examines proxy behavioral targets in simulated deployments rather than clinical or real-world outcomes, and has not undergone peer review.

Study design

As stated by the source record.

MethodsActivation steering applied to four open-weight 7–9B models under INT8 and NF4 weight-only quantization. Iso-effect framework compared capability costs at matched behavioral effect. Sentiment steering (judged) and reasoning length (judge-free, GSM8K-based) were primary behavioral targets. GSM8K parser artifact corrected with uniform v2.3.1 rescore. Preregistered three-label decision rule applied to inference of sentiment steering effects. Cosine similarity of steering vectors between quantized and FP16 versions computed. Baseline capability shifts at alpha=0 quantified.

DesignSystematic controlled study across multiple models and quantization schemes with preregistered analysis

PopulationFour open-weight language models (Mistral and three others, 7–9B parameters); no human subjects or clinical populations studied.

InterventionInference-time activation steering under INT8 and NF4 weight-only quantization

ComparatorFP16 (unquantized) baseline and steering-free quantized models

Reported figures

Quoted from the source exactly as published.

Pooled INT8 sentiment contrast-0.010

INT8 sentiment 90% CI[-0.026, +0.007]

NF4 sentiment contrast-0.017

NF4 sentiment 90% CI[-0.067, +0.033]

Reasoning length shortening range before collapse12-30%

Mistral-NF4 GSM8K baseline shift (alpha=0)0.545 to 0.365

INT8 steering vector cosine similarity with FP160.989-0.998

NF4 steering vector cosine similarity with FP160.945-0.990

Caveats

This is a preprint and has not been peer reviewed.
Study is technical in scope; findings concern LLM inference optimization, not clinical or biomedical outcomes.
Effect heterogeneity across models and quantization schemes; NF4 sentiment effects remain inconclusive.
Baseline capability shifts under quantization not systematically characterized across all conditions; their interaction with steering magnitude requires further investigation.

Evidence assessment

Graded across the dimensions that decide whether you should act, each from what the source actually supports. There is no single score, and where a dimension was not assessed it says so.

Study design

moderate

Systematic evaluation across multiple models and quantization schemes with preregistered analysis framework.

Directness

limited

Examines technical proxy behaviors (sentiment steering, reasoning length) in controlled inference settings, not clinical endpoints.

Consistency

moderate

Sentiment steering shows consistent survival across quantization; reasoning length shows model-dependent asymmetric failures.

Risk of bias

moderate

Preregistered analysis; corrected GSM8K parser artifact; baseline capability shifts not fully characterized across all conditions.

Sample adequacy

moderate

Four open-weight models tested; confidence intervals reported but some remain inconclusive; effect heterogeneity across quantization schemes.

Source abstract

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.

Follow from this

Taken from the source record, never inferred. Follow any of these and new work involving them reaches your briefing.

Saurav BhandariBenjamin Wade