I Can't Believe It's Not Robust: Catastrophic Collapse of Safety Classifiers under Embedding Drift

Fuente: arXiv

Saved in:

Bibliographic Details
Main Authors:	Sahoo, Subramanyam, Jain, Vinija, Chaudhary, Divya, Chadha, Aman
Format:	Preprint
Published:	2026
Subjects:	Machine Learning Computation and Language
Online Access:	Acceder al recurso
Tags:	Add Tag No Tags, Be the first to tag this record!

Similar Items

When Shallow Wins: Silent Failures and the Depth-Accuracy Paradox in Latent Reasoning
by: Sahoo, Subramanyam, et al.
Published: (2026)

SAHOO: Safeguarded Alignment for High-Order Optimization Objectives in Recursive Self-Improvement
by: Sahoo, Subramanyam, et al.
Published: (2026)

The Reasoning Trap -- Logical Reasoning as a Mechanistic Pathway to Situational Awareness
by: Sahoo, Subramanyam, et al.
Published: (2026)

Position: The Complexity of Perfect AI Alignment -- Formalizing the RLHF Trilemma
by: Sahoo, Subramanyam, et al.
Published: (2025)

Dial E for Ethical Enforcement: institutional VETO power as a governance primitive
by: Sahoo, Subramanyam, et al.
Published: (2026)

Catch Me If You Can: How Smaller Reasoning Models Pretend to Reason with Mathematical Fidelity
by: Sahoo, Subramanyam, et al.
Published: (2025)

Are Small Language Models Ready to Compete with Large Language Models for Practical Applications?
by: Sinha, Neelabh, et al.
Published: (2024)

Guiding Vision-Language Model Selection for Visual Question-Answering Across Tasks, Domains, and Knowledge Types
by: Sinha, Neelabh, et al.
Published: (2024)

Calibration Collapse Under Sycophancy Fine-Tuning: How Reward Hacking Breaks Uncertainty Quantification in LLMs
by: Sahoo, Subramanyam
Published: (2026)

Decoding the Diversity: A Review of the Indic AI Research Landscape
by: KJ, Sankalp, et al.
Published: (2024)

MAAT: Multi-phase Adapter-Aware Targeted Unlearning
by: Yagnik, Suryash, et al.
Published: (2026)

How Culturally Aware are Vision-Language Models?
by: Burda-Lassen, Olena, et al.
Published: (2024)

SKETCH: Structured Knowledge Enhanced Text Comprehension for Holistic Retrieval
by: Mahalingam, Aakash, et al.
Published: (2024)

A Comprehensive Survey of Hallucination in Large Language, Image, Video and Audio Foundation Models
by: Sahoo, Pranab, et al.
Published: (2024)

Parameter Efficient Fine Tuning: A Comprehensive Analysis Across Applications
by: Balne, Charith Chandra Sai, et al.
Published: (2024)

D-STEER - Preference Alignment Techniques Learn to Behave, not to Believe -- Beneath the Surface, DPO as Steering Vector Perturbation in Activation Space
by: Raina, Samarth, et al.
Published: (2025)

SPINAL -- Scaling-law and Preference Integration in Neural Alignment Layers
by: Das, Arion, et al.
Published: (2026)

PermaFrost-Attack: Stealth Pretraining Seeding(SPS) for planting Logic Landmines During LLM Training
by: Kumar, Harsh, et al.
Published: (2026)

MENTIS: What Belief Changes Under Alignment? Measuring Multi-Scale Latent Torsion in Language Models
by: Saha, Partha Pratim, et al.
Published: (2026)

AdversariaL attacK sAfety aLIgnment(ALKALI): Safeguarding LLMs through GRACE: Geometric Representation-Aware Contrastive Enhancement- Introducing Adversarial Vulnerability Quality Index (AVQI)
by: Khanna, Danush, et al.
Published: (2025)

Overview of Factify5WQA: Fact Verification through 5W Question-Answering
by: Suresh, Suryavardan, et al.
Published: (2024)

Can Large Language Models Infer Causal Relationships from Real-World Text?
by: Saklad, Ryan, et al.
Published: (2025)

Born With a Silver Spoon? Investigating Socioeconomic Bias in Large Language Models
by: Singh, Smriti, et al.
Published: (2024)

AlignGuard-LoRA: Alignment-Preserving Fine-Tuning via Fisher-Guided Decomposition and Riemannian-Geodesic Collision Regularization
by: Das, Amitava, et al.
Published: (2025)

Moral Sensitivity in LLMs: A Tiered Evaluation of Contextual Bias via Behavioral Profiling and Mechanistic Interpretability
by: Aggarwal, Yash, et al.
Published: (2026)

SEPSIS: I Can Catch Your Lies -- A New Paradigm for Deception Detection
by: Rani, Anku, et al.
Published: (2023)

To Believe or Not to Believe Your LLM
by: Yadkori, Yasin Abbasi, et al.
Published: (2024)

From Prejudice to Parity: A New Approach to Debiasing Large Language Model Word Embeddings
by: Rakshit, Aishik, et al.
Published: (2024)

SPHERICAL KV: Angle-Domain Attention and Rate-Distortion Retention for Efficient Long-Context Inference
by: Chauhan, Anay, et al.
Published: (2026)

Objective Matters: Fine-Tuning Objectives Shape Safety, Robustness, and Persona Drift
by: Vennemeyer, Daniel, et al.
Published: (2026)

Negative Preference Optimization: From Catastrophic Collapse to Effective Unlearning
by: Zhang, Ruiqi, et al.
Published: (2024)

Personality Shapes Gender Bias in Persona-Conditioned LLM Narratives Across English and Hindi: An Empirical Investigation
by: Kumar, Tanay, et al.
Published: (2026)

A Systematic Survey of Prompt Engineering in Large Language Models: Techniques and Applications
by: Sahoo, Pranab, et al.
Published: (2024)

Cross-Platform Hate Speech Detection with Weakly Supervised Causal Disentanglement
by: Sheth, Paras, et al.
Published: (2024)

Assessing LLM Reliability on Temporally Recent Open-Domain Questions
by: Krishnappa, Pushwitha, et al.
Published: (2026)

Neural FOXP2 -- Language Specific Neuron Steering for Targeted Language Improvement in LLMs
by: Saha, Anusa, et al.
Published: (2026)

DPO Kernels: A Semantically-Aware, Kernel-Enhanced, and Divergence-Rich Paradigm for Direct Preference Optimization
by: Das, Amitava, et al.
Published: (2025)

A Comprehensive Survey of Accelerated Generation Techniques in Large Language Models
by: Khoshnoodi, Mahsa, et al.
Published: (2024)

Multilingual State Space Models for Structured Question Answering in Indic Languages
by: Vats, Arpita, et al.
Published: (2025)

Climbing the Ladder of Reasoning: What LLMs Can-and Still Can't-Solve after SFT?
by: Sun, Yiyou, et al.
Published: (2025)