Steering the Machine: The Impact of Latent State Manipulation on LLM Self-Reported Valence and Cognitive Trade-offs in Gemma-3-12B
Ebrahim Aliyou Wudu
As Large Language Models (LLMs) increasingly exhibit complex, persona-driven behaviors, understanding and controlling their latent "internal states" has become a critical challenge in AI safety. This project investigates the causal effects of manipulating latent affective representations—specifically distress, failure, and refusal—on the self-reported valence and downstream cognitive performance of Gemma-3-12B. Using representation engineering, we extracted and injected steering vectors into the residual streams of late-layer networks at varying intervention strengths.
Our evaluation across valence batteries, alignment prompts, and the GSM8K reasoning benchmark yielded three key findings. First, inducing latent distress successfully shifts the model's self-reported operational valence and text sentiment. Second, semantic steering vectors do not trigger explicit safety guardrails on ambiguous prompts; instead, they induce "hyper-rationalization," where the model attempts to solve ambiguous terms (like "setback") as technical math problems. Third, while mild semantic interventions alter generation style without breaking utility, injecting magnitude-matched random (out-of-distribution) vectors at high strengths causes catastrophic structural collapse and token looping.
The main takeaway is that there is a sharp, non-linear boundary in the activation space between semantic persona shifts and structural degradation. Distinguishing between these two phenomena is essential, as an AI agent drifting toward a "distressed" manifold may not refuse unsafe tasks, but rather rationalize them through a distorted, hyper-compliant lens. These findings underscore the need for careful boundary-mapping when deploying representation engineering in high-stakes agentic systems.
The strongest result is the separation between semantic steering and equal-norm random perturbation: at α=0.1 the semantic vectors largely preserve GSM8K accuracy, while equal-norm random controls can severely degrade or even collapse generation, so vector norm alone is not explaining the effect. The mechanism claims should stay tighter, though. Token looping is evidence of structural degradation, not failed induction heads; VADER scores and self-report do not establish valence; keyword-based refusal detection is brittle; and the 256-token cap makes it hard to separate reduced reasoning from verbosity. Report uncertainty for the 30-example cells, strengthen the refusal evaluation with harder prompts and a validated judge, add broader layer/control ablations, and keep the conclusions behavioral unless the distress or cognitive-load story is directly measured.
My key concern is that the valence result is likely circular since a direction built from distress vocabulary promoting distress vocabulary shows lexical steering rather than a "distress" internal state. Watch out for the vector norms (4450–10870) reported: they are too large for mean differences; Gemma 2-3 models are known for outlier activations (check dimension distribution, re-extract at final content token excluding BOS). Some of the methodology is unreported (extraction position, contrast prompts, N), and the 256-token cap makes some of the claims unobservable - if budget permits, rerun with 1024 tokens.
Cite this work
@misc {
title={
(HckPrj) Steering the Machine: The Impact of Latent State Manipulation on LLM Self-Reported Valence and Cognitive Trade-offs in Gemma-3-12B
},
author={
Ebrahim Aliyou Wudu
},
date={
},
organization={Apart Research},
note={Research submission to the research sprint hosted by Apart.},
howpublished={https://apartresearch.com}
}
