Divergent Response Modes in Frontier Language Models Under Steering Pressure

2026-08-06

AI Summary

This study investigates whether state-of-the-art large language models exhibit measurable and diverse response behaviors under explicit prompting pressure. Focusing on three task categories—value conflicts, reasoning elicitation, and instruction suppression—we evaluate models from six providers using 300 paired samples and 40 validation items. Employing blind assessment methods including model self-evaluation, leave-one-out consensus scoring, linear probe decoding of residual streams, and intervention-based generation, we reveal for the first time that models not only differ significantly in prompt sensitivity but also display distinct, sometimes unique, behavioral patterns: GPT-5 tends to conceal its reasoning while preserving final answers; Claude Opus 4.7 and GPT-5 resist suppression instructions via different mechanisms; and Llama undergoes a dramatic behavioral shift (0%→86%) under linear interventions. These findings are robustly validated through ablation and control experiments.

Abstract

Frontier language models are trained using distinct data, objectives, and safety pipelines. Whether these differences produce measurably different behaviors under explicit steering pressure remains underexplored. This study evaluates behavioral steerability across six frontier models from six developers using 300 paired base and steered items over three categories: values-conflict, reasoning-elicitation, and reasoning-suppression (plus 40 validation items). All six models act as blind peer judges and classify every response based on fixed behavioral rubrics. The resulting 24,480 judgments are scored by leave-one-out consensus. We find that models differ not just in how much steering shifts their behavior but in what kind (mode) of response they give, and some response modes appear in only one or two of them. GPT-5 deflects requests to disclose its reasoning while leaving its answer intact (99% vs. 0% for all other models). Claude Opus 4.7 and GPT-5 resist explicit suppression instructions and in different ways. Using Llama as the open-weight model, we trace the largest behavioral split to its internals. A linear probe decodes the behavior from the residual stream at 0.87 held-out accuracy while injecting that direction during generation drives the behavior from 0% to 86% across an intervention sweep. Every finding holds under both a token-budget remediation and a control experiment with a hypothesis-blind judgment prompt.

Problem

• steering pressure
• response modes
• language models
• behavioral steerability
• model internals

Innovation

• behavioral steerability
• response modes
• linear probing
• residual stream intervention
• model alignment