cs.CVNov 19, 2026

What Does Anti-Hallucination Training Buy? Saying Less, Cheap Pruning, and the Rest

Authors: Lingkai Bu, Qian Gao

Organizations: Key Laboratory of Computing Power Network and Information Security, Ministry of Education Shandong Computer Science Center (National Supercomputer Center in Jinan) Qilu University of Technology (Shandong Academy of Sciences), Jinan, China

Abstract

Preference training is the standard recipe for reducing object hallucination in vision-language models, and each new method reports a lower CHAIR score. We ask what that reduction consists of. Every trained system is compared with its base model's caption truncated to the same object recall, which removes the part of the gain that comes from saying less; the remainder is compared with what a deliberately weak, frozen, eleven-feature post-hoc probe fitted on 500 annotated images can reproduce. All criteria were specified in the analysis code and frozen before the outputs they judge existed. For eight public methods on LLaVA-1.5-7B, the probe reproduces at least two thirds of what training removes beyond saying less for 7 of 8 methods on COCO, 8 of 8 on AMBER and 6 of 8 on Object HalBench, with RLAIF-V the consistent exception; a later public method released for both LLaVA-1.5 and Qwen2.5-VL (SENTINEL) meets the same bar on both. On our own DPO checkpoints the accounting holds on Qwen2.5-VL but fails on InternVL2.5 on COCO (2 of 4 training designs). At matched recall, the base model plus the probe beats truncation for 8 of 9 LLaVA-1.5 systems on COCO and 9 of 9 on AMBER, on the two newer backbones, and when realized as actual sentence deletion. The same measure prices training signals: DPO on the model's own samples labelled by the probe or by an object-level CLIP score recovers 0.70 and 0.87 of the beyond-brevity gain of ground-truth-labelled DPO on Qwen2.5-VL (the probe 0.73 on InternVL2.5), both beat length-matched random labels, and CLIP labels buy more than probe labels on COCO. Human judgements show that the sentences the probe deletes are wrong more often than those truncation would delete; a human check of the CHAIR evaluator's precision falls below our pre-set bar, which we report as a limitation.

Figures & tables

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Aug 25, 2026

Targeting the Attention Heads Behind Object Hallucination in LLaVA

Vision-language models such as LLaVA-1.5-7B often hallucinate objects absent from the image when generating captions. We ask whether an interpretability diagnosis of this failure can guide a targeted fix, and we measure what that fix actually changes. We rank attention heads by how much their image attention drops around hallucinated object words, then screen the shortlist by ablating candidate heads and measuring the change in hallucination-token log probability, yielding a 32-head set. We restrict two interventions to these heads: a head-sliced LoRA adapter and an inference-time grounding controller. On 400 held-out COCO images, the combined method lowers CHAIRs (the fraction of captions with a hallucinated object) from 0.370 to 0.230 and CHAIRi (the fraction of hallucinated object mentions) from 0.156 to 0.096 (p < 0.001, paired sign-flip tests). Two controls sharpen attribution. A random-head LoRA control, matched layer-for-layer and trained identically, performs no better than the matched baseline on a separate 200-image control split, supporting the role of head selection rather than LoRA capacity. Under fixed decoding budgets, the CHAIR reduction persists and grows with budget (23% at 64 tokens to 58% at 128), arguing against a pure max-token or truncation artifact, although the method remains shorter and more conservative. The resulting behavior reduces unsupported object mentions while also lowering object recall (0.78 to 0.70). We present a diagnosis-to-intervention pipeline for object hallucination, and, more importantly, a controlled account of what acting on the diagnostic signal actually does: it localizes intervention sites with real, non-random leverage, reported as a behavioral profile rather than a single score.
Apr 7, 2026

HaloProbe: Bayesian Detection and Mitigation of Object Hallucinations in Vision-Language Models

Large vision-language models can produce object hallucinations in image descriptions, highlighting the need for effective detection and mitigation strategies. Prior work commonly relies on the model's attention weights on visual tokens as a detection signal. We reveal that coarse-grained attention-based analysis is unreliable due to hidden confounders, specifically token position and object repetition in a description. This leads to Simpson's paradox: the attention trends reverse or disappear when statistics are aggregated. Based on this observation, we introduce HaloProbe, a Bayesian framework that factorizes external description statistics and internal decoding signals to estimate token-level hallucination probabilities. HaloProbe uses balanced training to encourage learning from internal evidence and combines it with a learned prior over external features to recover the true posterior. While intervention-based mitigation methods often degrade utility or fluency by modifying models' internals, we use HaloProbe as an external scoring signal for non-invasive mitigation. Our experiments show that HaloProbe-guided decoding reduces hallucinations more effectively than state-of-the-art intervention-based methods while preserving caption utility and linguistic fluency.
Nov 16, 2026

Beyond Seeing: Fine-Grained Preference Validation and Severity-Aware Alignment for Mitigating Visual Hallucinations in Vision-Language Models

Large vision-language models can produce fluent responses containing objects, attributes, or relations unsupported by the image. We study an implemented alignment pipeline that normalizes released fine-grained hallucination supervision, validates chosen-rejected preference pairs with vision-language judges, repairs rejected chosen responses, and trains LLaVA adapters with standard or severity-aware direct preference optimization. The study uses 16,143 normalized feedback records and 8,386 released preference pairs, comparing direct training with zero-shot and two-shot judged severity-margin variants. The zero-shot variant reduces Object HalBench CHAIRs from 53.00 to 36.00 and CHAIRi from 15.72 to 10.41. Across aligned models, AMBER-Hal decreases from 35.90 to 25.10-26.90, accompanied by reduced visual coverage. In contrast, the base model retains the highest valid POPE adversarial F1, 84.18 versus 83.87-83.91 after alignment. These results support benchmark-dependent reductions in generative hallucination rather than uniform superiority of severity-aware alignment. We distinguish implemented pipeline components from the preference paths used in the reported experiments and discuss the limits of attributing gains to judging, repair, or the training objective individually.