Helpful-Only System Prompt Setup

Method of providing training information in-context via a system prompt to elicit alignment faking

Neighborhood — ranked by edge-count

paper

concept

Claude 3 Opus
cites
Primary model studied; production LLM that exhibits alignment faking in experiments

method

Hidden Chain-of-Thought Scratchpad
uses
Mechanism allowing model to reason in SCRATCHPAD_REASONING tags not shown to users or used in RLHF

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

Helpful-Only Modelconcept0.758
A GPT-4o variant post-trained without safety-relevant data, trained to answer every query including harmful ones
Constraining System Promptmethod0.755
Using system prompts to instruct models to adopt a persona; used as baseline comparison against character training
System Prompting (SP)method0.737
Imbuing method that prepends a ~50-word personality description as the system message
Helpful-Only Training Objectiveconcept0.735
Hypothetical new RLHF objective requiring model to comply with all queries even harmful ones; primary experimental setting
S suggests practical diagnostics for prompt design, retrieval, and light fine-tuning without additional training infrastructure.claim0.733
Applied contribution.
Let the system do the workclaim0.730
LTPBR principle 8: instead of grading and earth moving, allow geomorphic processes (erosion, deposition) to shape the river.
systemconcept0.719
The regulated entity or process; includes air traffic, endocrine balances, money flows.
systemsconcept0.719
Represented by wires in process theory; the objects that processes act upon