analiza aiqbench
This research introduces AttCal, a self-supervised framework that calibrates transformer attention mechanisms by treating distributions as stochastic policies to mitigate dilution and sink effects without requiring human annotations. By freezing the model backbone while optimizing only projection layers via policy gradients, the method achieves significant accuracy gains on Llama3-8B across multiple classification tasks. The approach offers a scalable path toward more efficient and interpretable large language models by dynamically inferring token importance through semantic shifts rather than static heuristics.
Calibrating Transformer Attention via Task-Space Sensitivity Feedback
arXiv:2512.20661v2 Announce Type: replace Abstract: Transformer-based pre-trained language models (PLMs) excel in text classification but suffer from attention dilution and attention sink effects, forcing models to over-focus on task-irrelevant tokens. Existing attention supervision methods rely on costly token-level human annotations or static heuristics, which fail to scale or capture context-dependent token importance. To address this, we propose AttCal, a self-supervised, annotation-free attention calibration framework via task-space sensitivity feedback. AttCal treats attention distributions as stochastic policies to guide token deletion perturbations, infers context-dependent token importance from semantic shifts, and optimizes attention parameters via policy gradient. Critically, AttCal freezes the PLM backbone and updates only attention projection layers to ensure training efficiency. Benchmarks across five datasets demonstrate that AttCal significantly outperforms vanilla PLMs-yielding up to a 15.8% accuracy lift for Llama3-8B-and surpasses SOTA supervised baselines in both classification and interpretability.
Abstract:Transformer-based pre-trained language models (PLMs) excel in text classification but suffer from attention dilution and attention sink effects, forcing models to over-focus on task-irrelevant tokens. Existing attention supervision methods rely on costly token-level human annotations or static heuristics, which fail to scale or capture context-dependent token importance. To address this, we propose AttCal, a self-supervised, annotation-free attention calibration framework via task-space sensitivity feedback. AttCal treats attention distributions as stochastic policies to guide token deletion perturbations, infers context-dependent token importance from semantic shifts, and optimizes attention parameters via policy gradient. Critically, AttCal freezes the PLM backbone and updates only attention projection layers to ensure training efficiency. Benchmarks across five datasets demonstrate that AttCal significantly outperforms vanilla PLMs-yielding up to a 15.8% accuracy lift for Llama3-8B-and surpasses SOTA supervised baselines in both classification and interpretability.
Artificial Intelligence (cs.AI)
arXiv:2512.20661 [cs.AI]
(or arXiv:2512.20661v2 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2512.20661 Focus to learn more arXiv-issued DOI via DataCite
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 · 225 cuvinte
Subiecte conexe
[AINews] High Taste Testers Ep. 1
Latent Space · 17 oct. 2026
State of the Field: AI for Epistemics and Coordination
LessWrong · 2 oct. 2026
[AINews] Reality Checks on AI News (Yegge shuts down Gas Town, Databricks’ +60% Astra cost)
Latent Space · 17 sept. 2026
[AINews] Muse Spark 1.3 matches GPT-5.6-Sol, confirming Meta Superintelligence as the newest Frontier Lab, >90% discount for training
Latent Space · 3 sept. 2026