Adaptive residual-update steering for low-overhead hallucination mitigation in large vision language models

Thumbnail Image

Files

master_Zou_Zhengtao_2026.pdf (13.43 MB)

URL

Journal Title

Journal ISSN

Volume Title

School of Science | Master's thesis
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.

Authors

Date

2025-12-30

Department

Major/Subject

Machine Learning, Data Science and Artificial Intelligence

Mcode

Degree programme

Master's Programme in Computer, Communication and Information Sciences
Master's Programme in Computer, Communication and Information Sciences
Master's Programme in Computer, Communication and Information Sciences

Language

en

Pages

37

Series

Abstract

Large Vision-Language Models (LVLMs) often suffer from object hallucination, generating text inconsistent with visual inputs, which can critically undermine their reliability. Existing inference-time interventions to mitigate this issue present a challenging trade-off. While methods that steer internal states or adjust output logits can be effective, they often incur substantial computational overhead, typically requiring extra forward passes. This efficiency bottleneck can limit their practicality for real-world, latency-sensitive deployments. In this work, we aim to address this trade-off with Residual-Update Directed DEcoding Regulation 𝑅𝑈𝐷𝐷𝐸 𝑅, a low-overhead framework that steers LVLMs towards visually-grounded generation. RUDDER is built on two key innovations: (1) Contextual Activation Residual Direction (CARD) vector, a per-sample visual evidence vector extracted from the residual update of a self-attention layer during a single, standard forward pass. (2) A Bayesian-inspired adaptive gate that performs token-wise injection, applying a corrective signal whose strength is conditioned on the model’s deviation from the visual context. Extensive experiments on key hallucination benchmarks, including POPE and CHAIR, indicate that RUDDER achieves performance comparable to state-of-the-art methods while introducing negligible computational latency, validating RUDDER as a pragmatic and effective approach for improving LVLMs’ reliability without a significant compromise on efficiency.

Description

Supervisor

Marttinen, Pekka

Thesis advisor

Gao, Ya

Keywords

LVLM, hallucination, Bayesian-inspired, adaptive, low overhead, steering vector

Other note

Citation

Permanent link to this item

https://urn.fi/URN:NBN:fi:aalto-202601221844

Collections

[dipl] Perustieteiden korkeakoulu / SCI