Cover Image

Noise Injection Reveals Hidden Capabilities of Sandbagging Language Models

Fuente: arXiv

Saved in:

Bibliographic Details
Main Authors:	Tice, Cameron, Kreer, Philipp Alexander, Helm-Burger, Nathan, Shahani, Prithviraj Singh, Ryzhenkov, Fedor, Roger, Fabien, Neo, Clement, Haimes, Jacob, Hofstätter, Felix, van der Weij, Teun
Format:	Preprint
Published:	2024
Subjects:	Artificial Intelligence Cryptography and Security
Online Access:	Acceder al recurso
Tags:	Add Tag No Tags, Be the first to tag this record!

Similar Items

AI Sandbagging: Language Models can Strategically Underperform on Evaluations
by: van der Weij, Teun, et al.
Published: (2024)

The Elicitation Game: Evaluating Capability Elicitation Techniques
by: Hofstätter, Felix, et al.
Published: (2025)

Noise Injection Systemically Degrades Large Language Model Safety Guardrails
by: Shahani, Prithviraj Singh, et al.
Published: (2025)

Benchmark Inflation: Revealing LLM Performance Gaps Using Retro-Holdouts
by: Haimes, Jacob, et al.
Published: (2024)

Extending Activation Steering to Broad Skills and Multiple Behaviours
by: van der Weij, Teun, et al.
Published: (2024)

Auditing Games for Sandbagging
by: Taylor, Jordan, et al.
Published: (2025)

Evaluating Precise Geolocation Inference Capabilities of Vision Language Models
by: Jay, Neel, et al.
Published: (2025)

LLMs Can Covertly Sandbag on Capability Evaluations Against Chain-of-Thought Monitoring
by: Li, Chloe, et al.
Published: (2025)

Option-Order Randomisation Reveals a Distributional Position Attractor in Prompted Sandbagging
by: Cacioli, Jon-Paul
Published: (2026)

Probing a Vision-Language-Action Model for Symbolic States and Integration into a Cognitive Architecture
by: Lu, Hong, et al.
Published: (2025)

Chord Colourizer: A Near Real-Time System for Visualizing Musical Key
by: Haimes, Paul
Published: (2025)

Tailored Truths: Optimizing LLM Persuasion with Personalization and Fabricated Statistics
by: Timm, Jasper, et al.
Published: (2025)

View From Above: A Framework for Evaluating Distribution Shifts in Model Behavior
by: Chopra, Tanush, et al.
Published: (2024)

Sandbagging in a Simple Survival Bandit Problem
by: Dyer, Joel, et al.
Published: (2025)

Removing Sandbagging in LLMs by Training with Weak Supervision
by: Ryd, Emil, et al.
Published: (2026)

The H-graph with unequal masses in quantum field theory
by: Kreer, Philipp Alexander, et al.
Published: (2024)

The H-graph with equal masses in terms of multiple polylogarithms
by: Kreer, Philipp Alexander, et al.
Published: (2021)

EGT Quantum Resonance: Predicting the $\mathbf{402 \text{ GeV}}$ WIMP and Future Discovery Conditions at the Large Hadron Collider
by: Tice, Brian
Published: (2025)

HumanAgencyBench: Scalable Evaluation of Human Agency Support in AI Assistants
by: Sturgeon, Benjamin, et al.
Published: (2025)

Trace Relations in Deformed Gauge Theories
by: Raman, Madhusudhan, et al.
Published: (2024)

The Anatomy of Alignment: Decomposing Preference Optimization by Steering Sparse Features
by: Ferrao, Jeremias, et al.
Published: (2025)

Direct Ascent Synthesis: Revealing Hidden Generative Capabilities in Discriminative Models
by: Fort, Stanislav, et al.
Published: (2025)

Counterfactual Evaluation Reveals Hidden Capability Profiles in Clinical LLMs and Agents
by: Turk, Matt
Published: (2026)

Definição e Percepção de Imagem: Um Estudo em uma Escola de Educação Infantil de Novo Hamburgo
by: Cássia Rebelo Hofstätter
Published: (2008)

A Pesquisa de Marketing como um Meio de Informação para a Tomada de Decisão Estratégica
by: Cássia Rebelo Hofstätter
Published: (2005)

Imagem e Identidade Institucional: Um Estudo Aplicado à Feevale
by: Cássia Rebello Hofstätter
Published: (2009)

Understanding Addition and Subtraction in Transformers
by: Quirke, Philip, et al.
Published: (2024)

CTRL-ALT-DECEIT: Sabotage Evaluations for Automated AI R&D
by: Ward, Francis Rhys, et al.
Published: (2025)

Chain-of-thought obfuscation learned from output supervision can generalise to unseen tasks
by: Hadida, Nathaniel Mitrani, et al.
Published: (2026)

MEQA: A Meta-Evaluation Framework for Question & Answer LLM Benchmarks
by: Veuthey, Jaime Raldua, et al.
Published: (2025)

Rethinking CyberSecEval: An LLM-Aided Approach to Evaluation Critique
by: Hariharan, Suhas, et al.
Published: (2024)

Lower overhead fault-tolerant building blocks for noisy quantum computers
by: Prabhu, Prithviraj
Published: (2026)

Interpreting Context Look-ups in Transformers: Investigating Attention-MLP Interactions
by: Neo, Clement, et al.
Published: (2024)

Uma contribuição da educação ambiental crítica para (des)construção do olhar sobre a seca no semiárido baiano
by: Lakshmi Juliane Vallim Hofstatter
Published: (2016)

Ancient asexuality: No scandals found with novel data
by: Paulo Hofstatter, et al.
Published: (2024)

Gravity driven traveling bore wave solutions to the free boundary incompressible Navier-Stokes equations
by: Stevenson, Noah, et al.
Published: (2025)

Hydrostatic bubbles of compressible fluid in an incompressible fluid
by: Jang, Juhi, et al.
Published: (2025)

The traveling wave problem for the shallow water equations: well-posedness and the limits of vanishing viscosity and surface tension
by: Stevenson, Noah, et al.
Published: (2023)

Stationary wave solutions to two dimensional viscous shallow water equations: theory of small and large solutions
by: Stevenson, Noah, et al.
Published: (2025)

Stress Testing Deliberative Alignment for Anti-Scheming Training
by: Schoen, Bronson, et al.
Published: (2025)