A training-free decoding method for large vision-language models reported gains on several benchmark measures used to track hallucinations, while showing a lower reported inference-time multiplier than VCD, PAI and SID. Called FLASH, it was tested on LLaVA-1.5 and Shikra, but its gains varied across generative metrics.
A different diagnosis
The study asks whether these failures are more complicated than a simple imbalance in where a model's attention falls, whether frequency-domain analysis can help diagnose and mitigate them, and whether that can be done without the computational overhead associated with contrastive decoding. FLASH is training-free, so it changes the decoding process without retraining the model. It uses a measure called the Spectral Vortex Score to identify vision heads and applies adaptive spectral modulation during decoding.
The diagnostic analysis challenged a simple fix based on spatial attention alone. The paper reported that linear amplification of visual attention was insufficient and could be counterproductive, and linked observed gains from existing methods mainly to contrastive decoding rather than to attention manipulation. It also described two hallucination patterns: Perceptual-Semantic Dissociation, a break between spatial localization and semantic interpretation, and Localized Fixation, in which attention is captured by irrelevant visual locations.
What the benchmark scores showed
On LLaVA-1.5 7B, FLASH reported POPE scores of 90.03/89.77 for the Random split, 86.53/86.62 for Popular and 80.53/81.75 for Adversarial, with each pair representing accuracy and F1. Its MME score was 636.67. The reported CHAIR sentence and instance scores were 47.50 and 12.55, while AMBER Cover, Hal and Cog were 51.00, 25.00 and 2.15. Against the unmodified baseline, the paper reported higher POPE and MME values, better results on several generative metrics and no change in AMBER Cover.
On Shikra, the corresponding POPE pairs were 86.03/84.99, 83.30/82.56 and 79.80/79.65 across the three splits, with an MME score of 468.33. Generative results were less consistent: CHAIR sentence and instance scores were 49.60 and 14.05, while AMBER Cover, Hal and Cog were 52.65, 35.75 and 3.10. The paper characterized the comparison as favorable on discriminative measures but mixed on generative ones.
With LLaVA-1.5 13B, FLASH reported POPE pairs of 90.60/90.71, 86.70/87.35 and 80.80/82.70, followed by an MME score of 601.67. Its CHAIR sentence and instance scores were 42.70/11.75, and its AMBER Cover, Hal and Cog scores were 50.45/23.00/2.25. In the paper's comparison, FLASH exceeded the vanilla model on all reported POPE entries and improved the reported CHAIR and AMBER hallucination measures, although it was not the best method on every comparison.
The efficiency comparison
On the POPE benchmark, the reported inference-time multiplier was 1.69 times the vanilla baseline for FLASH, compared with 1.98 for VCD, 1.87 for PAI and 2.58 for SID. FLASH therefore had the lowest reported relative multiplier, although the study gave relative figures rather than absolute latency.
What the evidence leaves open
The evaluation covered 9,000 POPE-MS-COCO queries, 500 randomly sampled CHAIR images and 200 randomly selected AMBER images from a 1,004-image generative split. It included POPE and MME discriminative tasks and CHAIR and AMBER generative tasks, using LLaVA-1.5 at 7B and 13B and the Shikra family.
These comparisons do not establish causal superiority, and the reported tables provide point estimates only, with no interval or replicate variability reported for the comparisons. The authors also state that FLASH still requires multiple spectral transformations and filtering operations, applies value and attention modulation simultaneously, and has unclear applicability to Q-Former-based models.
The authors interpret FLASH as a favorable performance-efficiency balance while bypassing standard contrastive decoding. The supplied document is labeled a preprint, although its front matter also carries a conference-proceedings statement.
Disclosure
The paper reports support from the National Natural Science Foundation of China under Grant 62273353 and declares no conflicts of interest.
Paper data and sources
Original title: Beyond Attention Imbalance: Mitigating Hallucinations via Spectral Surgery
Authors: Siqi Lu, Suo Wei, Yongbin Zheng et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-09-29
DOI: Not available
Original paper · Full text