analiza aiqbench
This study reveals a critical brittleness in LLMs where models often fail to handle negation by simply suppressing the original answer rather than logically excluding it, leading to repeated incorrect responses. The authors propose a novel training objective that enforces stronger shifts in answer preferences for confident predictions, effectively reducing these failures without degrading general capabilities. This work advances our mechanistic understanding of how neural networks process linguistic negation and offers a concrete path toward more robust language models.
How Do LLMs Change Predictions Under Negation?
arXiv:2610.09571v1 Announce Type: new Abstract: Negation is an essential feature of human language, yet large language models (LLMs) remain unreliable in processing it. We evaluate recent open-source and closed-source LLMs on our negation benchmark and find that, in 37-71% of cases, they repeat the same answer under negation (e.g., "Madrid" for "What is not the capital of Spain?"). To understand and address this brittleness, we mechanistically examine how models operate under negation. Our main finding is that specialized attention heads and MLP neurons jointly implement negation by (1) suppressing retrieval of the original answer (e.g., "Madrid") while (2) promoting a favored candidate within the answer category (e.g., "Paris"). This contrasts with accounts of human negation processing, in which information about the original answer helps to determine what should be excluded. Furthermore, we find that this difference from human processing is a key source of negation failures: the model's mechanism relies on suppressing the original answer rather than using it to determine what to exclude, so the model can repeat the original answer when suppression is too weak or when a bias toward particular answers prevents it from selecting an alternative. To address this weakness in the model's negation mechanism, we propose a training objective that requires larger shifts in answer preference for more confident original predictions, and show that it reduces negation failures with less degradation of general capabilities than standard fine-tuning baselines. Together, our results demonstrate how mechanistic analysis can reveal why a linguistic capability fails and guide training that targets the underlying limitation.
Title:How Do LLMs Change Predictions Under Negation?
Abstract:Negation is an essential feature of human language, yet large language models (LLMs) remain unreliable in processing it. We evaluate recent open-source and closed-source LLMs on our negation benchmark and find that, in 37-71% of cases, they repeat the same answer under negation (e.g., "Madrid" for "What is not the capital of Spain?"). To understand and address this brittleness, we mechanistically examine how models operate under negation. Our main finding is that specialized attention heads and MLP neurons jointly implement negation by (1) suppressing retrieval of the original answer (e.g., "Madrid") while (2) promoting a favored candidate within the answer category (e.g., "Paris"). This contrasts with accounts of human negation processing, in which information about the original answer helps to determine what should be excluded. Furthermore, we find that this difference from human processing is a key source of negation failures: the model's mechanism relies on suppressing the original answer rather than using it to determine what to exclude, so the model can repeat the original answer when suppression is too weak or when a bias toward particular answers prevents it from selecting an alternative. To address this weakness in the model's negation mechanism, we propose a training objective that requires larger shifts in answer preference for more confident original predictions, and show that it reduces negation failures with less degradation of general capabilities than standard fine-tuning baselines. Together, our results demonstrate how mechanistic analysis can reveal why a linguistic capability fails and guide training that targets the underlying limitation.
Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
arXiv:2610.09571 [cs.CL]
(or arXiv:2610.09571v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2610.09571 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.
Text integral sub licența deschisă a publicației · 355 cuvinte
Subiecte conexe
Spontaneous verbal evidence synthesis in LLMs
LessWrong · 18 oct. 2026
Synthesis Superintelligence: from Semiconductors to Superconductors — Periodic Labs’ Liam Fedus and Ekin Dogus Cubuk
Latent Space · 8 oct. 2026
Links #5: 2026/07
LessWrong · 31 iul. 2026
Predictive Processing: Conscious when Training
LessWrong · 11 iun. 2026