COMPASS: Context-Modulated PID Attention Steering System for Hallucination Mitigation

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Sahay, Kenji, Pandya, Snigdha, Nagale, Rohan, Lin, Anna, Shiromani, Shikhar, Zhu, Kevin, Sunishchal, Dev
Natura: Preprint
Pubblicazione: 2025
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866914174404132864
author Sahay, Kenji
Pandya, Snigdha
Nagale, Rohan
Lin, Anna
Shiromani, Shikhar
Zhu, Kevin
Sunishchal, Dev
author_facet Sahay, Kenji
Pandya, Snigdha
Nagale, Rohan
Lin, Anna
Shiromani, Shikhar
Zhu, Kevin
Sunishchal, Dev
contents Large language models (LLMs) often generate fluent but factually incorrect statements despite having access to relevant evidence, a failure mode rooted in how they allocate attention between contextual and parametric knowledge. Understanding and steering this internal behavior is key both for trustworthy deployment and for scientific interpretability of model mechanisms. We introduce COMPASS (Context-Modulated PID Attention Steering System), a lightweight, interpretable control framework that embeds a model-based feedback loop directly within decoding. COMPASS quantifies context reliance via a transparent metric, the Context Reliance Score (CRS), which serves as an online probe of how attention heads ground generation in evidence. Using this interpretable signal, a PID controller dynamically modulates attention heads to maintain factual consistency without retraining or multi-pass decoding. Across benchmarks (HotpotQA, XSum, HaluEval, RAGTruth), COMPASS consistently reduces contextual hallucination rates (2.8 to 5.8 percent absolute) while revealing how distinct attention heads contribute to evidence alignment. These results highlight feedback-driven interpretability as a pathway toward scientific understanding of LLM behavior.
format Preprint
id arxiv_https___arxiv_org_abs_2511_14776
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle COMPASS: Context-Modulated PID Attention Steering System for Hallucination Mitigation
Sahay, Kenji
Pandya, Snigdha
Nagale, Rohan
Lin, Anna
Shiromani, Shikhar
Zhu, Kevin
Sunishchal, Dev
Computation and Language
Large language models (LLMs) often generate fluent but factually incorrect statements despite having access to relevant evidence, a failure mode rooted in how they allocate attention between contextual and parametric knowledge. Understanding and steering this internal behavior is key both for trustworthy deployment and for scientific interpretability of model mechanisms. We introduce COMPASS (Context-Modulated PID Attention Steering System), a lightweight, interpretable control framework that embeds a model-based feedback loop directly within decoding. COMPASS quantifies context reliance via a transparent metric, the Context Reliance Score (CRS), which serves as an online probe of how attention heads ground generation in evidence. Using this interpretable signal, a PID controller dynamically modulates attention heads to maintain factual consistency without retraining or multi-pass decoding. Across benchmarks (HotpotQA, XSum, HaluEval, RAGTruth), COMPASS consistently reduces contextual hallucination rates (2.8 to 5.8 percent absolute) while revealing how distinct attention heads contribute to evidence alignment. These results highlight feedback-driven interpretability as a pathway toward scientific understanding of LLM behavior.
title COMPASS: Context-Modulated PID Attention Steering System for Hallucination Mitigation
topic Computation and Language
url https://arxiv.org/abs/2511.14776