Logit Leakage through Floating-Point Bias Addition

cs.CR, cs.AI

Submitted: 2026-10-12

Updated: 2026-10-12

Comments: 18 pages, 6 figures, 5 tables, including appendices

Code: https://github.com/ggml-org/llama.cpp

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Sources

• Softmax Reparameterization for Output-Head Quantization
• OPT: Open Pre-trained Transformer Language Models
• BLOOM: A 176B-Parameter Open-Access Multilingual Language Model
• Qwen3 Technical Report
• Quantamination: Dynamic Quantization Leaks Your Data Across the Batch
• FloatDoor: Platform-Triggered Backdoors in LLMs
• CTRL: A Conditional Transformer Language Model for Controllable Generation
• Baichuan 2: Open Large-scale Language Models
• A Study of BFLOAT16 for Deep Learning Training

Related papers

• SoK: AI-Augmented Binary Reversing
• Relaxed Sender Anonymity for CBDC Interbank Settlement: A Zero-Knowledge Approach on Permissioned EVM
• Calibration-Family Overfit: Why Trusted Sabotage Monitors Don't Transfer Across Lineages
• Efficient Fuzzy PSI under One-Sided Assumptions
• Sealing the Audit-Runtime Gap for LLM Skills
• Token Composition: A Graph Based on EVM Logs