Aug 2026· Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2· pp. 3095-3105· 0 citations· 23 references
TL;DR
A novel inference-time intervention method that encodes truthful and untruthful content as an entailment hierarchy and performs representation editing in the hyperbolic space to activate the truthfulness of LLMs is proposed.
Abstract
Large language models (LLMs) often generate hallucinations, undermining the reliability of their outputs. While prior work improves truthfulness by contrastive decoding or representation editing, these approaches overlook the hierarchical relationship between truthful and untruthful content, thereby constraining the LLM's knowledge potential. In this paper, we propose HyperEdit, a novel inference-time intervention method that encodes truthful and untruthful content as an entailment hierarchy and performs representation editing in the hyperbolic space to activate the truthfulness of LLMs. Specifically, we employ an auto-encoder to project the representations of LLMs into a hyperbolic space where truthful samples lie closer to the origin and untruthful ones farther away. This results in a truthfulness-aware hyperbolic space, where proximity to the origin indicates higher truthfulness, defining a natural editing direction. During inference, we edit the LLM's internal representations along this hyperbolic direction to elicit more truthful outputs. Experimental results on TruthfulQA and three additional benchmarks demonstrate that HyperEdit consistently enhances truthfulness across various advanced LLMs.
TruthLens is a self-evaluation framework that teaches the LM head to expose a per-object truthfulness signal without any auxiliary model or additional inference cost, and generalizes effectively to benchmarks with substantially larger label spaces.
Yanqi Wu, Runhe Lai, Xinhua Lu et al.· 0 citations
This work studies hallucination from the perspective of dynamic representation shift during generation and proposes an online projection-based intervention on intermediate hidden states to suppress the hallucination-related directions, mitigating hallucinations while preserving language quality.
GHOST is proposed, an efficient white-box framework for hallucination detection in LLMs that primarily target confused hallucinations marked by internal reasoning instability, while also capturing stub-born hallucinations characterized by premature layer-wise convergence as a complementary signal.
Yueheng Mao, Min Yu, Gengwang Li et al.· Annual Meeting of the Associ...· 0 citations
HalluTracer is introduced, a detection framework that reads and aggregates truthfulness evidence across every layer of the forward pass before the model emits any answer token, recasts hallucination detection from a layer-selection problem into a depth-aggregation problem governed by the geometric sparsity of the truthfulness signal.
Zhihao Guo, Zonghan Wu, Huan Huo et al.· 0 citations
This work proposes M ulti-Model C ontrastive D ecoding (MCD), which integrates a pretrained language model with an evil model and a truthful model for contrastive decoding and effectively reduces hallucinations in LLMs and outperforms state-of-the-art methods across various benchmarks.
Chenyu Zhu, Yefeng Liu, Hao Zhang et al.· Neural Information Processin...· 7 citations
Large Language Models (LLMs) have achieved remarkable success in natural language generation but remain prone to hallucinations—generating content that is fluent but factually incorrect. While recent inference-time interventions like Contrastive Decoding (CD) effectively mitigate this by penalizing tokens favored by a "weak" hallucination-prone model, they introduce significant computational overhead (doubling inference latency) and fail to permanently align the model. In this paper, we propose \textbf{ContrastSFT}, a novel training framework to mitigate hallucinations in LLMs that internalizes the efficacy of contrastive decoding into the model's parameters via Contrastive Logit Regularization (CLR). Unlike standard Supervised Fine-Tuning (SFT) which indiscriminately maximizes the likelihood of ground-truth tokens, ContrastSFT dynamically recalibrates the training objective by subtracting the log-probabilities of a weak reference model. This effectively penalizes "easy" but potentially hallucinatory patterns captured by the weak model, forcing the model to learn more robust, factual representations. Extensive experiments on NLU benchmarks (ParaRel, WiCE) and Factuality tasks (HaluEval, MMLU) demonstrate that ContrastSFT achieves a 5-9\% absolute improvement over SFT and previous contrastive methods. Crucially, ContrastSFT eliminates the need for auxiliary models during deployment, retaining the high inference efficiency of standard LLMs. Code will be released.
Sifan Zhou· Poster Volume 0007 The 2026...· 0 citations