A revised arXiv paper surveys research on hallucination in multimodal foundation models, focusing on large vision-language models. It organizes the literature around why these errors occur, methods proposed to reduce them, and how they are measured. The authors frame the work as a structured overview connecting causes, corrections and evaluation practices.
A new arXiv preprint proposes a decoding strategy that draws on internal attribution signals within a large language model to keep generation factually grounded as it proceeds token by token. The authors argue that early factual mistakes tend to snowball during autoregressive generation, and that neither post-hoc fixes nor edits at the weight level reliably correct them. Their approach aims to intervene during decoding rather than after the fact.
A new arXiv paper examines how language models trained for chemical reasoning use chain-of-thought steps, and finds that the intermediate reasoning frequently contains fabricated content. Testing four reasoning model families across twelve chemistry tasks, the authors report that hallucination is widespread and largely disconnected from the final answer. The work suggests chain-of-thought traces in this domain act more like an unreliable scratchpad than a faithful record of the model's reasoning.
A preprint introduces NeuroActiSep, a method that aims to spot factual hallucinations in large language models by inspecting feed-forward neurons, according to its abstract. The approach is described as working in a single pass, making it cheaper than methods requiring repeated sampling. The paper frames this as an under-explored alternative to existing white-box truthfulness detection techniques.
A new arXiv paper argues that clinical LLM answers should be selected by routing between decoding strategies rather than by correcting a single model's output. The authors show that the best decoding regime depends on the query, and propose a trajectory-gated router that picks the appropriate method per question. The work aims to improve reliability without adding retrieval, fine-tuning, or external verifier infrastructure that clinical governance would need to approve.
A new arXiv paper proposes using transcoders, an alternative to sparse autoencoders, to study how vision-language models turn image inputs into text. The authors argue that sparse autoencoders decompose static representations and miss the flow of information, while transcoders can follow that transformation more directly. The method is used to locate where visual grounding happens and where hallucinated content originates in these models.
Researchers tested 26 large language models from eight developers on their ability to produce accurate biomedical citations. The authors report that fabricated or incorrect references remain a persistent problem across the models evaluated. The work is a preprint and characterises the frequency of this failure mode rather than proposing a fix.
A newly posted computational linguistics paper examines whether large language models become less faithful to a provided context when the input data seems implausible. The authors analyse how a model's tendency to hallucinate or misinterpret facts varies with the plausibility of what it is given, a question with direct consequences for retrieval-augmented generation and data-to-text systems. The work seeks to clarify when models follow supplied evidence versus falling back on their own priors.
Researchers present their system for the SHROOM-Visions 2026 shared task, which targets character-level detection of hallucinations in vision-language model outputs. The method fine-tunes a 4-billion-parameter VLM as a per-token classifier that reads a two-token feature from its own hidden states, then combines it with larger models in an ensemble.
Researchers have introduced LexAgentHallu, a new benchmark for measuring how tool-augmented legal AI agents hallucinate. It uses a hierarchical structure to trace how errors in tool calls and reasoning cascade into fabricated case holdings and miscited legal authority. The benchmark aims to fill a gap left by existing legal evaluations that do not capture agentic workflows.
A new arXiv paper traces hallucinations in multimodal large language models to shifts in how information is distributed across attention components the authors call synergy heads. The researchers argue that existing mitigation techniques built on attention weights track only indirect cues and therefore miss the underlying mechanism. The findings could support more targeted methods for reducing fabricated outputs in multimodal systems.
Researchers have proposed GANDR, an approach that audits legal question-answering systems at the level of individual claims instead of scoring whole answers. Each statement generated by the model is checked against the specific source it cites, so readers can verify a grounded answer piece by piece. The work targets a shortcoming of existing grounded-generation pipelines, which typically evaluate answers only in aggregate.
Researchers assembled a corpus of 150 academic papers with deliberately planted errors to test how well large language models can act as automated document-quality auditors. They report that detection reliability worsens as processing batch sizes increase, and that models sometimes fabricate audit findings with high confidence. The results cast doubt on deploying LLMs unsupervised for contamination-detection tasks.
A new arXiv paper addresses hallucination in large audio-language models, where models sometimes let language priors override what is actually heard in the audio. The authors propose a method that adaptively selects perturbations for contrastive decoding, a training-free approach, arguing that existing techniques rely on crude perturbations such as masking or added noise. The work aims to improve how reliably these models ground their outputs in acoustic evidence.
A new arXiv preprint describes a hallucination detection method that combines a fine-tuned DeBERTa-v3 classifier with Monte Carlo dropout uncertainty estimates and other signals. The approach targets domain-specific settings where unfaithful model outputs are a concern. The work is cross-listed under cs.AI.
A research team describes HALDETECT, their entry for the English hallucination-detection track of the ImageEval 2026 shared task. The system combines an answer-first contrastive grounding approach with QLoRA parameter-efficient fine-tuning to curb fluent but unfounded visual claims by large multimodal models. The work addresses a known obstacle to using such models for fine-grained image interpretation.
A new arXiv preprint proposes probing methods to trace where large language models source their knowledge, aiming to separate two kinds of hallucination. The authors distinguish faithfulness violations, where a model mishandles context it was given, from factuality violations, where its answers stem from incorrect stored knowledge. The work targets better attribution of model outputs to internal knowledge versus supplied context.