2026· Annual Meeting of the Association for Computational Linguistics· pp. 11879-11892· 0 citations· 46 references
Computer Science
TL;DR
CoDA is introduced, a lightweight inference-time attention intervention that amplifies evidence-aligned value states, enabling more attention heads to transport reliable external evidence in a copy-encouraged manner and improves contextual faithfulness, reduces hallucination, and remains robust under long and noisy contexts with modest and stable inference overhead.
Abstract
Retrieval-augmented generation reduces hallu-cination by grounding model outputs in exter-nal evidence, yet hallucinations can still occur even when the retrieved context is accurate and sufficient. From the perspective of information routing in the residual stream, this reflects an imbalance where internal parametric knowledge overwhelms external context during generation. We present an attention-centric analysis of RAG hallucination under valid evidence, showing that hallucinated and factual tokens diverge in mid-to-late Transformer layers as context-selective attention routing weakens, allowing parametric influence to dominate the residual stream. Motivated by prior studies showing that some attention heads—often referred to as copying heads—exhibit stronger information transport capacity, we aim to extend similar evidence-carrying behavior to a broader set of attention heads. To this end, we introduce CoDA, a lightweight inference-time attention intervention that amplifies evidence-aligned value states, enabling more attention heads to transport reliable external evidence in a copy-encouraged manner. Experiments demonstrate that CoDA improves contextual faithfulness, reduces hallucination, and remains robust under long and noisy contexts with modest and stable inference overhead.
Retrieval-augmented generation (RAG) reduces but does not eliminate hallucination, and existing detectors return a single answer-level score that does not indicate which sentence is unsupported, or why. To close this gap, we introduce Grounding-Aware Sensitivity by Perturbation (GASP), a span-level detector that scores each answer sentence by how strongly its likelihood depends on the retrieved evidence, a quantity we term grounding sensitivity. GASP holds the answer fixed and re-scores it under the full context, under no context, and with each chunk removed, then measures the log-likelihood drops and Jensen-Shannon divergences (JSD). The likelihood of a grounded sentence collapses once its supporting passage is removed, whereas a hallucinated sentence is almost unaffected, a contrast we interpret by casting decoding as a random nonlinear iterated function system (RNIFS). We evaluate GASP on three benchmarks (RAGTruth, TofuEval, RAGBench) with three instruction-tuned scorers from two model families (Qwen2.5-0.5B, Qwen2.5-1.5B, and SmolLM2-1.7B) under a leakage-clean protocol. On RAGTruth it reaches a response-level area under the ROC curve (AUC) of about 0.73 and a span-level AUC of about 0.67, improving significantly over perplexity and by clear margins over length, whole-context natural language inference (NLI), and self-consistency baselines. The only baseline competitive at the span level is a well-configured chunk-level entailment verifier, which requires a separate model, whereas a training-free threshold on the grounding features matches the trained classifier without labeled data and serves as the default detector. Beyond RAGTruth, the signal transfers to TofuEval but not to short-answer question answering in RAGBench, showing GASP is best suited to outputs constructed from the retrieved context rather than answers recoverable from parametric knowledge.
A lightweight linear detector is built on top of Role-Break that requires no fine-tuning of the VLM, whose feature dimension stays below 5,000 and reaches an average AUROC of 93.23 across six VLMs and four benchmarks.
Mingyu Wang, Weilin Jin, Wenbo Li et al.· 0 citations
Faithfulness hallucinations, where large language models generate outputs unsupported by retrieved evidence, remain a central challenge for trustworthy AI. We present a systematic empirical evaluation of faithfulness in retrieval-augmented generation (RAG) systems using two benchmark datasets, HotpotQA and HaluBench, covering both multi-hop reasoning and single-hop hallucination detection. We analyze three small-to-mid-sized (2B-8B) open-weight LLMs in combination with multiple retrieval strategies, including sparse, dense, and hybrid approaches, as well as score-based and rank-based fusion techniques, enabling a comprehensive assessment of retrieval-generation interactions. By disentangling retrieval and generation errors, we characterize how different pipeline components contribute to hallucinations in RAG systems. Our analysis provides actionable insights and practical evaluation protocols, highlighting the critical role of robust retrieval and careful system design. These findings offer a benchmarking-oriented perspective for developing more reliable and faithful RAG systems within evaluated model scales.
C. Mala, Gizem Gezici, Fosca Giannotti· Machine-mediated learning· 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.
This review provides systematic theoretical support for industrial RAG model selection and optimization and summarizes existing research gaps, including lightweight deployment and multimodal expansion, and proposes future research directions for trustworthy RAG systems.
Shujing Liu· Applied and Computational En...· 0 citations
This study proposes a novel conceptual framework and taxonomy for hallucination mitigation in low-code AI environments, integrating retrieval, validation, conflict resolution, and workflow orchestration mechanisms to contribute to the development of more reliable, transparent, and scalable AI systems.
I. K. W. Adnyana, Rosalin Theophilia Tayane, Fahmi Fahmi et al.· EDUKASIA Jurnal Pendidikan d...· 0 citations