analiza aiqbench
This research demonstrates that small language models, ranging from 0.5B to 7B parameters, are sufficient for personalizing AI-edited text without instruction prompts at inference time. By fine-tuning LoRA adapters on local data, the system effectively reduces perceived "AI-ness" in scientific writing while maintaining quality plateaus across different model sizes. The findings highlight a practical trade-off between computational efficiency and stylistic fidelity rather than establishing a strict hierarchy of large versus small models for this specific rewriting task.
Small Is Enough: Per-User Style Rewriting of AI-Edited Text via LoRA Adapters
arXiv:2607.29238v1 Announce Type: new Abstract: InMyStyle is a privacy first, single user system that adapts small language models to rewrite AI-edited text towards an individual user's writing style without an instruction prompt at inference. Given a user's documents, it uses multiple local helper LLMs to construct paired training examples and fine tunes LoRA adapters on base models ranging from 0.5B to 7B parameters. Length aware generation budgets and automatic chunking support inputs of different lengths. On 219 evaluation pairs from a scientific-paper corpus, the automatic composite score plateaus at 0.69 [scale 0-1] across all model sizes under both greedy and sampled decoding. This observed plateau suggests that small models are sufficient for the measured rewriting task, with model size determining trade-offs rather than a stable quality ranking. As a secondary evaluation, 400 ratings from five LLM judges give InMyStyle outputs a mean perceived AI-ness score over 20% lower than their helper-AI generated inputs, while mean perceived AI-ness scores decrease with model size within InMyStyle.
Abstract:InMyStyle is a privacy first, single user system that adapts small language models to rewrite AI-edited text towards an individual user's writing style without an instruction prompt at inference. Given a user's documents, it uses multiple local helper LLMs to construct paired training examples and fine tunes LoRA adapters on base models ranging from 0.5B to 7B parameters. Length aware generation budgets and automatic chunking support inputs of different lengths. On 219 evaluation pairs from a scientific-paper corpus, the automatic composite score plateaus at 0.69 [scale 0-1] across all model sizes under both greedy and sampled decoding. This observed plateau suggests that small models are sufficient for the measured rewriting task, with model size determining trade-offs rather than a stable quality ranking. As a secondary evaluation, 400 ratings from five LLM judges give InMyStyle outputs a mean perceived AI-ness score over 20% lower than their helper-AI generated inputs, while mean perceived AI-ness scores decrease with model size within InMyStyle.
Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Computers and Society (cs.CY); Emerging Technologies (cs.ET); Human-Computer Interaction (cs.HC)
arXiv:2607.29238 [cs.CL]
(or arXiv:2607.29238v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2607.29238 arXiv-issued DOI via DataCite (pending registration)
Text integral sub licența deschisă a publicației · 191 cuvinte
