Grad Detect: gradient-based hallucination detection using layer-wise backward pass signals
Grad Detect is a new method for detecting LLM hallucinations by analyzing layer-wise gradient patterns from a single forward-backward pass at inference time, without relying on output-level signals alone. Evaluated across Q&A benchmarks and eleven models from four architectural families, it consistently outperforms confidence-based and sampling-based baselines. A key finding is that the final five layers concentrate over 97% of the discriminative gradient signal, enabling efficient deployment. The method also supports model abstention prediction, framing it as a unified reliability framework.
Related guides (2)
Related events (8)
CHAIR: Supervised hallucination detection via internal logit analysis across LLM layers
A new arXiv preprint introduces CHAIR (Classifier of Hallucination As ImproveR), a supervised framework that detects hallucinations by extracting statistical features (max, min, mean, std, slope) from token logits across all layers of an LLM. Evaluated on TruthfulQA and MMLU, CHAIR shows improved detection accuracy especially in zero-shot settings. The authors argue the approach also points toward richer internal representations for designing adaptive decoding strategies that reduce hallucinations.
ClinHallu benchmark diagnoses stage-wise hallucinations in medical multimodal LLM reasoning
Researchers from Alibaba DAMO Academy introduce ClinHallu, a benchmark of 7,031 validated instances designed to identify where hallucinations originate within medical MLLM reasoning pipelines. Each instance is annotated with a structured reasoning trace decomposed into Visual Recognition, Knowledge Recall, and Reasoning Integration stages, with stage-replacement interventions to measure the causal impact of correcting each stage. The paper also demonstrates that trace-supervised fine-tuning reduces stage-wise hallucinations, offering both diagnostic and mitigation value for clinical AI systems.
MMBench2 paper: hallucination in world models is predictable and preventable via coverage signals
Researchers introduce MMBench2, a 427-hour, 210-task dataset for visual world modeling, and train a 350M-parameter world model to study hallucination in generative world models. The paper identifies three distinct hallucination modes (perceptual, action-marginalized, scene-diverging) and develops lightweight signals that predict where models will fail. A coverage-aware sampling technique and curiosity-reward-based data collection enable efficient finetuning to unseen environments with as few as 50 real trajectories. The central finding is that world model hallucination is fundamentally a data coverage problem, with the same signals serving both detection and mitigation.
BenHalluEval: Multi-Task Hallucination Evaluation Framework for Bengali LLMs
BenHalluEval introduces the first systematic hallucination benchmark for Bengali, covering four tasks (generative QA, code-mixed QA, summarization, reasoning) with 12,000 hallucinated candidates generated via GPT-5.4 across twelve hallucination types. Seven LLMs are evaluated under a dual-track protocol separating false-positive rate on ground-truth instances from hallucination detection rate on hallucinated candidates. The proposed BenHalluScore metric reveals substantial variation (7.72%–55.42%) across models and tasks, and chain-of-thought prompting is found to shift response distributions without consistently improving hallucination discrimination. The work highlights gaps in low-resource language hallucination evaluation and critiques single-track and prompting-only evaluation approaches.
Label-Free Bias Identification in Vision Models via Gradient Probes on Concept Decompositions
This paper introduces a post-hoc, label-free method for identifying spurious correlations in frozen vision classifiers without requiring bias annotations, group labels, or retraining. The approach applies non-negative matrix factorization to intermediate activations to extract interpretable concept vectors, then ranks them using a gradient-based bias estimator derived from misclassified examples. On Colored MNIST, Waterbirds, and CelebA benchmarks, the method recovers known spurious cues and improves worst-group accuracy by up to 17.9 percentage points on Waterbirds by suppressing top-ranked concepts at inference time. Notably, the method surfaces decision-relevant directions that do not always coincide with annotated attributes, offering both an auditing tool and a debiasing handle for deployed models.
The Hallucinations Leaderboard, an Open Effort to Measure Hallucinations in Large Language Models
Hugging Face has launched an open leaderboard specifically designed to benchmark hallucination rates across large language models. The effort aims to standardize evaluation of factual accuracy and confabulation tendencies, filling a gap in existing benchmarks that focus primarily on capability rather than reliability. The leaderboard is positioned as a community-driven, transparent resource for tracking model trustworthiness.
LegalHalluLens: Typed hallucination auditing and calibrated multi-agent debate for legal AI
Researchers introduce LegalHalluLens, an auditing framework for hallucination in legal AI systems, evaluated across 510 contracts and 249,252 clause-level instances from the CUAD dataset. The framework introduces typed hallucination profiles across four claim categories (numeric, temporal, obligation/entitlement, factual) and a Risk Direction Index (RDI) that distinguishes omission from invention errors. A calibrated multi-agent debate pipeline reduces fabricated detections by 45% using a 4B-parameter model competitive with commercial APIs. The work reveals that aggregate hallucination rates (~52%) mask a 38-40 percentage-point gap between claim types and that two systems with identical aggregate rates can have opposite risk profiles.
PGT: Procedurally Generated Tasks for Improving Visual Grounding in MLLMs
This paper introduces Procedurally Generated Tasks (PGT), a data-driven framework that overlays geometric primitives on images to create dense supervision signals for fine-grained visual grounding in multimodal large language models. PGT serves both as a training augmentation method and a diagnostic tool to isolate perception failures from semantic priors. Instruction tuning on LLaVA-v1.5-Instruct augmented with PGT data yields gains of up to +20% on the What'sUp benchmark and +13.3% on CV-Bench-2D. The results suggest that spatial reasoning deficits in MLLMs stem primarily from inadequate supervision rather than architectural or resolution constraints.