๐ป Hallucination Detection¶
๐ง NeurIPS2026 ยท 19 paper notes
๐ Same area in other venues: ๐๏ธ ECCV2026 (20) ยท ๐ท CVPR2026 (33) ยท ๐ฌ ICLR2026 (40) ยท ๐ฌ ACL2026 (28) ยท ๐งช ICML2026 (21) ยท ๐ค AAAI2026 (15)
๐ฅ Top topics: Reasoning ร3 ยท Multimodal/VLM ร3
- A Theory of Time-Sensitive Language Generation: Sparse Hallucination Beats Mode Collapse
-
This paper introduces a "timeliness" deadline constraint into the game-theoretic framework of language generation in the limit, proving that any eventually consistent (hallucination-free) generator inevitably suffers mode collapse (zero lower density) under any deadline function, while proposing a Speculative Blackbox Generation (SBG) reduction that restores the optimal \(1/2\) timely lower density under arbitrary superlinear deadlines by permitting a vanishing rate of sparse hallucinations.
- Actionable Hallucination Detection: Translating Latent Uncertainty into Agentic Critique
-
This paper introduces the Latent Critic, a lightweight low-rank adapter (LoRA) that operates concurrently with a frozen base LLM during a single forward pass to actively restructure the transformer residual stream and translate internal latent uncertainty into parameter-level natural language diagnostic feedback (e.g.,
ungrounded: ["date"]), achieving 0.966 detection AUROC and over 80% parameter localization accuracy on Qwen3-4B while enabling closed-loop zero-shot agentic self-correction with under 10 ms of latency overhead. - Alleviating Hallucination in Reasoning Tasks with Training-Free Uncertainty-Guided Steering
-
The paper proposes USteer, a simple training-free activation steering method that extracts gradients from an internal Neural Collapse Inspired (NCI) uncertainty metric via the Logit Lens and performs single-step dynamic hidden-state intervention during decoding, significantly mitigating hallucinations in multi-step reasoning tasks.
- CHORD: Cross-Model Hallucination Detection via Relational Graph Discrimination
-
Addressing the failure of internal representation-based hallucination detectors under cross-model transfer due to severe semantic inconsistency, CHORD reorganizes hidden states across layers, tokens, and feature blocks into a unified relational graph, applying relation-aware attention and Feature-Critic meta-learning to deliver an average AUROC of 0.8042 across 8 unseen LLMs and 7 benchmarks.
- Dual-Pathway Circuits of Object Hallucination in Vision-Language Models
-
This paper proposes the Dual-Pathway Circuit Analysis framework, applying activation patching and Conditional Pathway Analysis (CPA) across five architecturally diverse vision-language models (VLMs) to uncover a dual-pathway organization (visual grounding vs. hallucination pathways) and a consistent grounding-pathway polarity flip on hallucinating samples, while reducing object hallucination by up to 76% at \(\le 2\text{ pp}\) accuracy cost via training-free component-wise scaling.
- Entropy Distribution as a Fingerprint for Hallucinations in Generative Models
-
This paper reveals that the distributional shape and tail behavior of token-level entropies serve as a hallucination fingerprint independent of the mean, formalizes hallucination detection as a statistical hypothesis test, and proposes the Calibrated Entropy Score (CES)โa single-pass black-box detector that achieves the best performance among single-pass methods across 10 models and 8 QA benchmarks while statistically matching expensive multi-sample consistency methods.
- Generating in the Limit with Infinitely Many Hallucinations
-
This paper recasts language generation in the limit as a formal precisionโrecall trade-off over infinite sets, proving that allowing infinitely many hallucinations at a vanishing rate (relaxing tail precision while keeping global precision at 1) or permitting a fixed fraction of non-novel repeated outputs (\(\gamma < 1\)) strictly breaks the recall ceiling of zero-hallucination algorithms under sparse adversarial revelations, and can even achieve both precision and recall of 1.
- Hallucination in World Models is Predictable and Preventable
-
This paper reveals that hallucinations in generative world models stem fundamentally from state-action data coverage gaps, decomposes them into perceptual hallucination, action marginalization, and scene divergence, introduces three label-free dynamism-normalized runtime predictors (correlating with rollout error at \(\rho \approx 0.80\)), and mitigates them via uniform-task resampling and predictor-guided curiosity data collectionโadapting a 350M-parameter Dreamer 4 world model to unseen tasks with just 50 trajectories.
- HalluWorld: A Controlled Benchmark for Hallucination via Reference World Models
-
Grounded in an explicit "Reference World Model" formulation, this paper introduces HalluWorld, a controlled and extensible hallucination benchmark spanning gridworlds, chess, and realistic Linux terminal tasks that automatically generates ground-truth labels while decoupling observability, serialization formats, and conflict resolution policies, systematically revealing distinct failure modes across perception, long-horizon memory, causal forward simulation, and epistemic abstention in frontier LLMs.
- How Fine-Tuning Objectives Shape Layer-Wise Information in LLM Hallucination Detection
-
This paper demonstrates that in binary LLM hallucination detection, classifier-head fine-tuning (
cls) induces task representations with stronger linear decodability and clearer two-cluster geometry than sequence-to-sequence (s2s) fine-tuning; by combining static Partial Information Decomposition (PID) and dynamic training trajectories to locate a narrow upper-middle signal-forming band (InfoWindow) followed by a redundant upper plateau, it derives a static upper-layer truncation rule that removes 31% of Qwen3 layers without accuracy loss while achieving a \(1.4\times\) inference speedup. - Invisible Ink, Visible Lies: How Production Watermarking Causes LLMs to Hallucinate
-
Reveals that production-level text watermarking induces factual hallucinations even when text remains fluent and evidence is present in context, attributes the degradation to token perturbations and prefix-induced attention drift, and introduces plug-in token and attention interventions (FPTI and FPAI) that eliminate up to 90% of factual errors under matched detectability.
- KG-Guard: Graph-Based Hallucination Detection for Knowledge Base Question Answering
-
This paper formulates hallucination detection in LLM-based Knowledge Base Question Answering (KBQA) as a binary answer-node classification problem over retrieved subgraphs and proposes KG-Guard, a lightweight black-box graph detector that uses roughly \(305\times\) fewer parameters than LLM baselines with zero LLM calls at detection time, achieving the highest F1 across WebQSP, CWQ, and PUGG (82.0, 87.4, and 84.3) while enabling iterative answer refinement that boosts downstream KBQA Exact Match by 16.9โ17.6 percentage points.
- Mitigating Factual Hallucination in Large Reasoning Models via Mixed-Mode Advantage Regularization
-
To address "thinking-induced hallucination" where explicit thinking in large reasoning models (LRMs) overturns originally correct direct answers in factuality QA, this paper formulates explicit thinking as a reasoning residual over the model's direct-answer tendency and proposes MARGO, a mixed-mode reinforcement learning framework that incorporates same-model non-thinking rollouts as reference baselines within GRPO groups to suppress harmful factual drift, boosting average accuracy to 25.49% on Qwen3-4B and 29.57% on Qwen3-8B across six factuality benchmarks while preserving mathematical reasoning ability.
- Mitigating Object Hallucination in Large Vision-Language Models via False Discovery Controlled Visual Data Splitting
-
This paper proposes CORAL, a training-free post-hoc framework that applies sign-symmetric perturbations to visual features to construct mirror views and mirror statistics, explicitly controlling the image-level false discovery rate (FDR) to suppress object hallucinations while maintaining high statistical power for truly grounded objects.
- MotionHalluc: Diagnosing Kinematic Hallucinations in Fine-Grained Motion Reasoning
-
This paper introduces MotionHalluc, the first diagnostic benchmark for evaluating kinematic hallucinations in cross-video fine-grained motion comparison (comprising 1,540 QA pairs across directional, attributional, and temporal dimensions), along with Perceive-Parse-Verify (PPV), a training-free pipeline that parses natural language instructions into executable 3D kinematic measurement queries and injects quantitative physical evidence at inference time for an average 10.6% accuracy improvement across five state-of-the-art large multimodal models.
- Robustness of Internal Signals for Hallucination Detection in Vision-Language Models
-
This paper systematically evaluates the robustness of single-pass internal-signal-based hallucination detection methods in vision-language models (VLMs), revealing that mainstream white-box detectors heavily rely on a surface-level "token position" shortcut while lacking genuine visual grounding, and demonstrating through progressive ensembling that combining black-box outputs with mechanistically diverse white-box signals boosts detection AUPRC on LLaVA-1.5-13B to 0.71 (+0.20 over the token position baseline).
- SOPO: Socratic Guided Policy Optimization for Span-Level Hallucination Detection
-
To resolve localization signal ambiguity, reward sparsity on hard examples, and reasoning length inflation when applying Reinforcement Learning with Verifiable Rewards (RLVR) to span-level hallucination detection, this paper proposes Socratic Guided Policy Optimization (SOPO), which integrates answer-agnostic multi-turn Socratic teacher feedback with turn-averaged rewards and pre-normalization turn-based advantage scaling, enabling a 4B student model under single-turn inference to achieve a state-of-the-art 63.3% span-F1 on RAGTruth (surpassing GPT-5, o3, and RL4HS-CAPO) while reducing inference latency by 47.0%.
- The Commit-Abstain Circuit: Why Language Models Hallucinate Instead of Abstaining
-
The paper localises a sparse Commit-Abstain Circuit (CAC) using the pre-generation commitโabstain logit margin, identifies early commitment accumulation with insufficient late abstention correction, and trains a lightweight policy on component contributions that raises reported mean decision accuracy from 0.692 to 0.814, rather than demonstrating improved answer-content accuracy.
- The Labeling Problem in Hallucination Detection Benchmarks: An Empirical Evaluation
-
This paper demonstrates that automated hallucination labelers in open-domain QA suffer from a critical criterion mismatch between reference faithfulness and factual correctness, showing that faithfulness-style prompts cause severe false-positive bias which can be dramatically corrected by decision-tree prompts that explicitly distinguish absence from contradiction.