Source-linked AI summary
The Hidden Life of Tokens: Reducing Hallucination of Large Vision-Language Models via Visual Information Steering
Zhuowei Li, Haizhou Shi, Yunhe Gao, Di Liu, Zhenting Wang, Yuxiao Chen, Ting Liu, Long Zhao, Hao Wang, Dimitris N. Metaxas
TL;DR
LVLMs can generate visually ungrounded hallucinations, while the internal emergence and propagation of hallucination remain unclear. The paper analyzes token-logit rankings and proposes VISTA, a training-free inference intervention combining visual steering with early-layer logit augmentation. VISTA reduces hallucination by about 40% on evaluated open-ended generation and outperforms existing methods across the reported benchmarks, architectures, and decoding strategies.
Problem
LVLMs frequently produce semantically coherent but visually ungrounded content, and how hallucination emerges and propagates during generation remains unclear.
Method
VISTA is a training-free inference-time framework that reinforces visual cues in activation space and uses early-layer logits to promote semantically meaningful decoding.
Results
About 40% average hallucination reduction is reported for evaluated open-ended generation, with consistent outperformance across four benchmarks, four architectures, and three decoding strategies.
Takeaways & Limitations
Token-logit analysis identifies visual information loss, early excitation, and hidden genuine information as patterns that VISTA uses to reduce hallucination while promoting genuine information.
Takeaways & Limitations
Optimal hyperparameters may vary across architectures, VSV depends on visual-encoder quality, and the current implementation focuses on single-round tasks.
Abstract
from arXiv · showhide
Large Vision-Language Models (LVLMs) can reason effectively over both textual and visual inputs, but they tend to hallucinate syntactically coherent yet visually ungrounded contents. In this paper, we investigate the internal dynamics of hallucination by examining the tokens logits ranking throughout the generation process, revealing three key patterns in how LVLMs process information: (1) gradual visual information loss - visually grounded tokens gradually become less favored throughout generation, and (2) early excitation - semantically meaningful tokens achieve peak activation in the layers earlier than the final layer. (3) hidden genuine information - visually grounded tokens though not being eventually decoded still retain relatively high rankings at inference. Based on these insights, we propose VISTA (Visual Information Steering with Token-logit Augmentation), a training-free inference-time intervention framework that reduces hallucination while promoting genuine information. VISTA works by combining two complementary approaches: reinforcing visual information in activation space and leveraging early layer activations to promote semantically meaningful decoding. Compared to existing methods, VISTA requires no external supervision and is applicable to various decoding strategies. Extensive experiments show that VISTA on average reduces hallucination by about 40% on evaluated open-ended generation task, and it consistently outperforms existing methods on four benchmarks across four architectures under three decoding strategies. Code is available at https://github.com/LzVv123456/VISTA.
1. Introduction
The paper analyzes how visual information is lost during LVLM generation and introduces VISTA, a training-free framework that steers token activations and logits to reduce hallucination.
- LVLMs often produce semantically coherent but visually ungrounded content, limiting reliability in real-world applications.
- Inspecting Token Dynamics in LVLMs: The analysis tracks hidden genuine, decoded genuine, and hallucinated tokens through temporal and layer-wise logit-ranking patterns.Higher token rankings indicate higher generation probability.
- Gradual Visual Information Loss: As generation progresses, genuine-token rankings decline while hallucinated tokens surface, indicating gradual visual information loss.The analysis uses early, mid, and late generation stages across 500 randomly selected MSCOCO images.
- Early Excitation of Semantically Meaningful Tokens: Semantically meaningful tokens peak in the penultimate or preceding layers, while the final layer prioritizes syntactic tokens such as “this” and “a”.
- Hidden Genuine Information: Hidden genuine tokens remain relatively highly ranked despite not being decoded, suggesting that LVLMs retain visual information they do not express.Their rankings remain around 5K in a 32K vocabulary during generation.
- Reducing Hallucination of LVLMs: VISTA combines Visual Steering Vector and Self-Logits Augmentation to reinforce visual cues and promote semantically meaningful decoding without additional training or model modifications.Experiments span multiple architectures and protocols, with hallucination reductions of up to around 40% for open-ended generation.
2. Methodology
The methodology analyzes token-ranking dynamics across generation and layers, then uses those findings to design VISTA’s visual-steering and early-layer logit-augmentation modules. VSV reinforces image-specific visual cues, while SLA integrates earlier-layer logits to guide decoding.
- Token Ranking Analysis: The analysis tracks token rankings across time and layers, grouping decoded genuine, hidden genuine, and hallucinated tokens to characterize LVLM generation dynamics.Rankings are computed with a logit lens and aggregated into category-specific temporal and layer-wise patterns.
- Token Ranking Analysis: Temporal analysis averages rankings over early, mid, and late generation buckets across 500 MS COCO images to examine visual-information dynamics.The resulting temporal view shows visually grounded information sinking while hallucinated contents are surfaced.
- Visual Steering Vector (VSV): VSV constructs an image-specific steering vector by contrasting contexts with and without visual tokens, then injects it into residual streams during inference.Unlike amortized LLM steering vectors, VSV is computed per image to preserve visual details unique to each input.
- Visual Steering Vector (VSV): VSV normalizes modified hidden states and uses an intervention strength λ to balance visual fidelity against natural generation.The context-based vector can be computed in advance and applied across layers, with cached prompts improving efficiency.
- Self-Logits Augmentation (SLA): SLA averages logits from layers before the final layer and combines them with final-layer logits to promote semantically meaningful tokens during decoding.The window size w controls the averaged early-layer range, while γ controls early-layer influence; γ = 0 gives standard generation and γ = 1 relies entirely on early-layer logits.
3. Experiments
VISTA is evaluated across four LVLM architectures, three decoding strategies, and hallucination-specific and general-purpose benchmarks. Across these settings, it reduces hallucination, improves broader visual-language performance, and remains efficient, with moderate intervention strengths balancing quality and mitigation.
- Object Hallucination Benchmarks: Around 40% relative improvement is achieved over corresponding vanilla decoding methods on the CHAIR open-ended generation task.VISTA also outperforms other inference-time intervention and contrastive decoding methods across decoding strategies.
- Object Hallucination Benchmarks: VISTA achieves higher POPE accuracy across models and decoding strategies, including gains from 81.26% to 85.35% for LLAVA-1.5 under nucleus sampling.For MiniGPT-4 under nucleus sampling, accuracy increases from 60.56% to 66.96%.
- Comprehensive Benchmarks: VISTA consistently improves MMHal-Bench scores across architectures and question types, especially environmental inference, attribute perception, and counting.The strongest reported relative improvements are approximately 20% for LLAVA-1.5 and 30% for InstructBLIP.
- Comprehensive Benchmarks: VISTA improves MME performance across architectures and decoding methods, including MiniGPT-4 rising from 845.30 to 1069.37 under nucleus sampling.LLAVA-1.5 also improves from 1625.22 to 1738.56 under nucleus sampling.
- Analysis and Ablations: VISTA improves hidden genuine-token rankings and suppresses hallucinated-token prominence during the mid and late generation stages.Its greedy decoding latency is lower than PAI and VCD; moderate λ and γ values balance hallucination reduction with generation quality, while extreme settings degrade quality.
- Analysis and Ablations: Larger SLA windows generally reduce hallucination, with an optimal configuration achieving the highest reported F1 score of 78.4.The optimal mixing coefficient becomes more conservative as the window spans more layers.
4. Related Work
Prior work attributes LVLM hallucination to visual encoder limitations, excessive reliance on learned knowledge, and other mechanisms, motivating training, post-processing, and inference-time remedies. Contrastive approaches operate across visual inputs, time steps, or model layers to reduce hallucination.
- Hallucination Mitigation in LVLMs: LVLM hallucination has been linked to visual encoder limitations and excessive reliance on learned parametric knowledge.Related mitigation approaches include refined training datasets, revision, verification, and inference-time interventions.
- Contrastive Decoding in LVLMs: Contrastive decoding methods reduce hallucination by contrasting alternative logits or inputs across visual tokens, time steps, or model layers.VCD uses noisy visual tokens as negative samples, while DoLa contrasts logits distributions from different layers.
5. Conclusion and Limitations
VISTA mitigates LVLM hallucination by restoring visual information and using early-layer logits, while its effectiveness depends on architecture, visual encoding quality, and task setting.
- VISTA combines visual information steering with early-layer logits to reduce hallucination while preserving generation quality.
- VISTA’s optimal hyperparameter settings may vary across architectures.
- VSV may provide reduced benefits for models with weak visual encoding capabilities.
- The current implementation addresses hallucination in single-round tasks, while interactive visual dialogue may require additional considerations.
Impact Statement
The study presents hallucination mitigation as a step toward more trustworthy LVLMs, while its analysis relies on several methodological procedures for token classification and ranking.
- The research advances methods for making LVLMs more trustworthy and reliable by mitigating hallucination.
- GPT-4o serves as an oracle model for identifying hidden genuine, decoded genuine, and hallucinated tokens.
- Multi-token words are represented by their first token to maintain consistent and tractable token analysis.
- Cross-stage ranking aggregation uses only the final five layers to reduce embedding disparity with the decoding layer.
A.3. Advantages of Token Ranking Analysis
Token ranking analysis directly tracks visual-information preservation and reveals degradation patterns, but its interpretation is constrained by oracle reliability and earlier-layer embedding mismatch.
- Token ranking analysis quantitatively tracks visual information preservation throughout generation and reveals degradation patterns that attention analyses may obscure.
- The framework generates descriptions, classifies token categories, tokenizes them, and aggregates rankings across time stages and final layers.
- GPT-4o-based token classifications may diverge from human judgments, especially for nuanced genuine-versus-hallucinated cases.
- Applying the decoding head to earlier layers can produce less reliable rankings because their representations are misaligned with the final-layer embedding space.
- Despite these limitations, the analysis provides insights that motivate the hallucination mitigation approach.
A.4. Additional Token Ranking Analysis
Additional analyses on Shikra reproduce the token-ranking patterns found on LLaVA-1.5 and show that VISTA maintains genuine-token rankings while suppressing hallucinated-token promotion.
- Shikra shows gradual genuine-token rank degradation and increasing prioritization of hallucinated tokens across generation stages.
- Cross-stage Shikra analysis visualizes token-ranking behavior across generation stages.
- Layer-wise Shikra analysis provides a complementary view of token rankings across model layers.
- On Shikra, VISTA maintains genuine-token rankings while suppressing the promotion of hallucinated tokens, consistent with findings for LLaVA-1.5.
B.1. MMHal-Bench Results For Other Decoding Strategies
VISTA consistently improves MMHal-Bench performance under beam search and nucleus sampling across evaluated LVLMs, extending the trends observed with greedy decoding. These results support its robustness across inference strategies.
- VISTA consistently improves overall MMHal-Bench performance across evaluated LVLMs under beam search and nucleus sampling.The reported trends remain consistent with those observed under greedy decoding.
- The results indicate that VISTA’s visual-grounding and semantic-promotion mechanisms remain effective across decoding strategies.
- Figures 10 and 11 report MMHal-Bench comparisons using beam search and nucleus sampling, respectively.
- Figures 12–14 visualize ablations of VSV strength (λ) and SLA mixing ratio (γ) across MiniGPT-4, LLAVA-1.5, and InstructBLIP.Brighter colors indicate better performance, red boxes mark parameter combinations used in Table 1, and F1 measures overall generation quality.
B.2. Additional Ablation Results
Additional ablations show that moderate VSV strength and SLA mixing ratios generally balance hallucination reduction with generation quality, while optimal settings vary by architecture.
- MiniGPT-4 ablations examine VSV strength (λ) and SLA mixing ratio (γ) using CHAIRS, CHAIRI, and F1 scores.
- MiniGPT-4 has a slightly lower optimal VSV strength of λ = 0.1 than the setting observed for Shikra.The passage presents this difference as evidence that hyperparameter choices can be architecture-specific.
- Moderate values of VSV strength and SLA mixing ratio yield the best balance between hallucination reduction and generation quality across architectures.
- The consistent ablation pattern supports the generality of VISTA while retaining a need for architecture-specific hyperparameter fine-tuning.
C. Case Study
Qualitative case studies across four LVLM architectures compare vanilla decoding with VISTA to illustrate reduced hallucination and recovery of genuine visual information.
- Figures 15–18 compare vanilla decoding and VISTA for LLAVA-1.5, MiniGPT-4, Shikra, and InstructBLIP.
- Hallucinated content is highlighted in red in the qualitative examples.
- The case studies highlight genuine information that VISTA successfully recovers after being omitted or wrongly described by vanilla decoding.