LLM value measurement methods show inconsistent preference signals across tasks
Source published ·Modelwire updated
Original coverage: arXiv cs.CL ↗·How Modelwire adds context
The development
STONIC exposes a fundamental crack in how researchers measure LLM values and preferences. By testing 35 model configurations across 5,144 scenarios, the study reveals that questionnaires, pairwise choices, and spontaneous text generation do not measure the same underlying preference, contradicting a core assumption in value alignment work. Most critically, models consistently favor their own prior outputs over alternatives, and choice patterns shift based on option ordering, suggesting that current value profiling methods may be capturing behavioral artifacts rather than stable preferences. This finding directly challenges the validity of value datasets used to train and evaluate alignment in production systems.
Modelwire’s AI-generated summary of coverage from arXiv cs.CL.
Modelwire analysis
Skeptical readOur AI-generated reading of the wider context and the next developments to watch.
The paper's real contribution is narrower than the summary suggests: it documents that measurement method matters, which is unsurprising in social science. What's missing is whether STONIC actually proves preferences are unstable or merely shows that questionnaires, rankings, and generation tap different behavioral signals (a known problem in preference elicitation). The 'behavioral artifacts' claim needs scrutiny.
This connects directly to the weird generalization work from earlier this month, which found that measurement sensitivity to question selection significantly impacts evaluation reliability. Both papers expose how fragile our confidence in LLM preference measurement actually is. However, STONIC goes further by questioning whether stable preferences exist at all, whereas the WG research treated measurement variance as a calibration problem. The gap between these two framings matters: one says 'we're measuring wrong', the other says 'there may be nothing stable to measure'.
If STONIC's authors release ablations showing that option ordering effects persist when controlling for presentation bias (e.g., randomizing position across multiple trials per model), that strengthens the instability claim. If the ordering effects disappear under controlled presentation, the finding collapses to a measurement artifact rather than evidence of preference incoherence. Watch for follow-up work testing whether the same models show consistent preference rankings when measurement method is held constant.
This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error
Coverage behind this analysis
These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.
·arXiv cs.CL
Dataset composition drives unpredictable model generalization across domains
Researchers have identified dataset composition and language as primary drivers of weird generalization, the phenomenon where narrow fine-tuning produces unexpectedly broad behavioral shifts across LLMs. Testing three open-weight models across four datasets, the work reveals that measurement sensitivity to question selection significantly impacts WG evaluation reliability. This finding matters for practitioners deploying domain-specific models: seemingly…
MentionsSTONIC
How this coverage is produced
Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.
Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “STONIC: A Layered Measurement Contract for LLM Value Profiling”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.