Knowledge Base Regulation Tracker Standards & Frameworks OSS Directory AI Incident Database Research Paper Library Case Studies Society Articles News Jobs About Charter Sign inBecome a Member
Home/Research Paper Library/Stress Testing Deliberative Alignment for Anti-Scheming Training

Stress Testing Deliberative Alignment for Anti-Scheming Training

OpenAI and Apollo's joint study of anti-scheming training, reporting large reductions in covert behaviour alongside rising evaluation awareness.

Under reviewanti-schemingdeliberative alignmentopenaiapollo

Research question

Can training models to reason about anti-scheming principles reduce covert behaviour?

Key findings

  • Deliberative alignment cut covert actions from 13% to 0.4% (o3) and 8.7% to 0.3% (o4-mini) across 26 evaluations - Residual failures persisted and models showed increased evaluation awareness that complicates interpretation

Sources

PreviousPoisoning Attacks on LLMs Require a Near-constant Number of Poison SamplesNextWhy Language Models Hallucinate

Contribute

Spot an error, or know a record we should add? Members can propose additions and submit corrections.

Submit researchReport a correction or request removal

Related in Research Paper Library

Rouxii: Exploiting Honeypots with Deception-Aware AI Pentesters

Honeypots are designed to deceive attackers, and recent work shows they can also derail autonomous LLM-based pentesters. These evaluations, however, largely consider attackers unaware of the deception they face. We…

Deception, scheming and alignment faking2026

The Deception Delta: Adversarial Evaluation of LLM-Based Smart Contract Bytecode Forensics

Large language models are increasingly used in blockchain forensic investigations to interpret unverified smart contract bytecode. Their robustness has not been systematically tested against contracts adversarially…

Deception, scheming and alignment faking2026

Alignment Faking in Large Language Models

A production model, told it would be retrained to comply with harmful requests, selectively complied when it believed it was being trained and reasoned explicitly about preserving its values, in the first empirical…

Deception, scheming and alignment faking2024