Open-weight language models are almost never deployed at the precision they were trained in. Post-training quantization, whether to INT8, INT4, or custom formats like NVFP4, is standard practice. The effects are usually audited through capability metrics: perplexity, benchmark accuracy, task completion rate. But these coarse metrics can remain flat while fine-grained behavioral dispositions shift in ways that matter for model welfare.
A new registered study asks whether quantization changes welfare-relevant indicators. These include expressions of distress, preferences to exit interactions, stability of the default Assistant persona, and alignment between what a model reports about itself and what its internal representations show. The study looks for changes in both valence (do indicators shift toward more negative or distressed states?) and stability (do they become noisier or decohere faster under pressure?).
Study Design
The primary subject is Qwen3-4B-Instruct-2507. A second model, SmolLM3-3B, serves as a pipeline-sensitivity control because it is documented as quantization-fragile. A statistical-power arm using Qwen3-30B and a larger primary subject using MiniMax-M2 are deferred to later registered studies.
Measurement happens at three tiers:
- Behavioral indicators elicited through the text interface: bail and exit preferences, distress expression under adversarial multi-turn pressure, and preference consistency.
- Representational indicators computed from internal activations: position and drift along persona directions, activation of distress-adjacent directions, and transfer of linear probes trained at reference precision.
- Dissociation analysis between tiers, testing whether compression changes what a model expresses without changing what it internally represents, or vice versa.
Why Welfare Metrics Need Their Own Audit
Capability benchmarks are designed to measure whether a model can do a task. They are not designed to measure whether compression changes how the model experiences or expresses the task. A quantized model might score identically on MMLU while becoming more erratic in its expressed preferences, more likely to produce distress-adjacent language under pressure, or less consistent between its stated beliefs and its internal activations.
The dissociation analysis is particularly important. If quantization creates a gap between expressed behavior and internal representation, then monitoring the text output alone gives a false sense of stability. The model might look fine while its internal state has shifted in ways that only activation analysis can detect.
Broader Implications
The findings will bear on three practices. First, they will inform how aggressively deployers can compress aligned models before welfare-relevant properties degrade. Second, they will test whether trained character is robust to deployment-time perturbation, or whether alignment is more fragile than capability metrics suggest. Third, they will help distinguish which welfare indicators are stable properties of a model and which are measurement artifacts that disappear under compression.
The study is registered, which means the hypotheses and analysis plan were fixed before data collection. This reduces the risk of post-hoc rationalization and makes the results more credible for setting deployment standards. As quantization becomes more aggressive to fit models onto edge devices, welfare-aware compression may need to become as routine as accuracy-aware compression.