aiqbench analysis
This research identifies a critical failure mode in personalized large language models where the decision to apply user preferences becomes biased toward "Apply" simply due to the generation objective, rather than an actual loss of sensitivity. The authors introduce ABIDE, a method that decouples this decision from the response generation process to isolate and correct the bias without compromising model utility. By explicitly modeling the applicability signal before decoding, this work offers a concrete pathway to reduce over-personalization artifacts in future AI systems.
Over-Personalization Is a Decision Failure: Generation-Induced Apply Bias in LLMs
arXiv:2609.34284v1 Announce Type: new Abstract: Personalized LLMs must decide, for each stored preference, whether the current context calls for applying or suppressing it, which we call its applicability. They frequently over-personalize, applying preferences the context rules out, yet existing benchmarks score only the final response and cannot tell where this failure arises. We decompose preference handling into three stages and measure each separately: (1) knowing whether a preference applies, (2) deciding on an explicit Apply/Suppress label, and (3) generating a response consistent with that label. Using linear probes, we first show that this applicability signal remains decodable from hidden states during generation. By making the decision explicit, we then find that in most settings wrong decisions faithfully followed outnumber correct decisions lost in generation. We thus locate the failure in the decision, which breaks once the model is also asked to answer. To determine whether this reflects lost sensitivity or a response bias, we propose ABIDE (Apply-Bias Investigation via Decision-score), which adapts signal detection theory to Apply-vs-Suppress decision scores read directly from logits. ABIDE reveals a generation-induced Apply bias: merely stating an answer-generation objective shifts the decision score toward Apply while sensitivity is largely preserved, and the shift persists under controls for prompt structure, cascades across preference slots, and prompt wording. Finally, we show that subtracting a single bias scalar, estimated on a held-out split, from the decision score at decoding time reduces leakage while largely preserving fulfillment.
Abstract:Personalized LLMs must decide, for each stored preference, whether the current context calls for applying or suppressing it, which we call its applicability. They frequently over-personalize, applying preferences the context rules out, yet existing benchmarks score only the final response and cannot tell where this failure arises. We decompose preference handling into three stages and measure each separately: (1) knowing whether a preference applies, (2) deciding on an explicit Apply/Suppress label, and (3) generating a response consistent with that label. Using linear probes, we first show that this applicability signal remains decodable from hidden states during generation. By making the decision explicit, we then find that in most settings wrong decisions faithfully followed outnumber correct decisions lost in generation. We thus locate the failure in the decision, which breaks once the model is also asked to answer. To determine whether this reflects lost sensitivity or a response bias, we propose ABIDE (Apply-Bias Investigation via Decision-score), which adapts signal detection theory to Apply-vs-Suppress decision scores read directly from logits. ABIDE reveals a generation-induced Apply bias: merely stating an answer-generation objective shifts the decision score toward Apply while sensitivity is largely preserved, and the shift persists under controls for prompt structure, cascades across preference slots, and prompt wording. Finally, we show that subtracting a single bias scalar, estimated on a held-out split, from the decision score at decoding time reduces leakage while largely preserving fulfillment.
Computation and Language (cs.CL)
arXiv:2609.34284 [cs.CL]
(or arXiv:2609.34284v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2609.34284 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 328 words
Related coverage
The Expensive-Loop Problem
LessWrong · Oct 12, 2026
The basic argument for making AI better at moral philosophy
LessWrong · Oct 14, 2026
Game-Theoretic Disempowerment: Loss of Control to Agencyless AI
LessWrong · Sep 27, 2026
Social Theory and Agent Foundations
LessWrong · Oct 8, 2026