Source-linked AI summary
Does Playing it Safe Count as Faithfulness? Reassessing LVLM Hallucination Mitigation Methods
Mehrdad Fazli, Sina Mansouri, Mohit Marvania, Ziwei Zhu
TL;DR
The paper asks whether lower hallucination scores reflect improved grounding or conservative generation. It evaluates six inference-time methods across three LVLMs and three benchmarks, finding that hallucination reductions often reduce informativeness and fail to transfer reliably to broader multimodal capabilities.
Problem
It remains unclear whether lower hallucination scores reflect improved multimodal grounding or more conservative generation.
Method
The study diagnostically evaluates six inference-time mitigation methods across three LVLMs using CHAIR, AMBER, and MMSTAR with paired informativeness and capability measures.
Results
Hallucination reductions on CHAIR and AMBER often coincide with reduced informativeness and do not reliably transfer to broader multimodal capabilities on MMSTAR.
Takeaways & Limitations
Hallucination mitigation should be assessed as a faithfulness–informativeness–capability trade-off rather than through hallucination scores alone.
Takeaways & Limitations
The conclusions apply most directly to inference-time distribution-shaping methods and do not cover training-based, preference-optimization, or retrieval-augmented strategies.
Abstract
from arXiv · showhide
Recent inference-time hallucination mitigation methods for large vision-language models (LVLMs) report strong gains on hallucination benchmarks. However, it remains unclear whether lower hallucination scores reflect improved multimodal grounding or more conservative generation. We evaluate six mitigation methods across three LVLMs and four benchmarks, including hallucination-focused evaluation and the diverse capability benchmark MMStar. Our analysis reveals two consistent patterns. First, hallucination reduction is often coupled with reduced informativeness: methods that lower hallucination rates also reduce object recall, visual coverage, or response detailedness. Second, improvements on hallucination benchmarks do not reliably transfer to broader multimodal capabilities, with methods showing inconsistent or degraded performance on fine-grained perception and reasoning tasks. Our findings suggest that current evaluation protocols may overestimate progress by rewarding conservative generation. We argue that hallucination mitigation should be evaluated as a faithfulness--informativeness--capability trade-off rather than through hallucination scores alone.
1 Introduction
The paper questions whether lower hallucination scores indicate better multimodal grounding, arguing that conservative responses can suppress risk without preserving informativeness or broader capability.
- Training-free decoding-time methods adjust LVLM output distributions without retraining and report gains on CHAIR, POPE, and AMBER.
- Hallucination benchmarks may reward shorter, non-committal responses that mention fewer visual entities rather than improve visual grounding.
- The evaluation combines hallucination metrics with object recall, visual coverage, and MMSTAR capability testing across six mitigation methods and three LVLMs.
- Lower hallucination rates often coincide with lower object recall and visual coverage, indicating risk suppression rather than greater informativeness.
- Hallucination gains do not reliably transfer to broader capabilities, with inconsistent or degraded performance on fine-grained perception, reasoning, science, and mathematics.
- The paper argues that mitigation should be evaluated as a faithfulness–informativeness–capability trade-off rather than by hallucination scores alone.
2 Diagnostic Evaluation Setup
The diagnostic setup compares six inference-time mitigation methods with vanilla decoding across three 7B LVLMs and benchmarks designed to separate hallucination reduction from information retention and general capability.
- The study evaluates 3 LVLMs × 6 methods × 3 benchmarks, totaling 54 experimental configurations.
- The models are LLaVA-1.5, LLaVA-NeXT, and InstructBLIP, each compared with a vanilla decoding baseline.
- The six methods span contrastive decoding, attention calibration, and recent state-of-the-art inference-time approaches.
- CHAIR and AMBER provide paired hallucination and informativeness metrics, including CHAIRs/CHAIRi with Object Recall and hallucination rates with Cover.
- MMSTAR probes coarse and fine-grained perception, instance and logical reasoning, science and technology, and mathematics after removing textual shortcuts.
3 Hallucination Reduction as Risk Suppression
The study finds that hallucination reduction frequently follows a risk-hedging path: methods lower errors by reducing visual informativeness rather than reaching the low-error, high-recall target.
- The target is the upper-left trade-off region, combining minimized hallucination errors with maximized informativeness or recall.
- Hallucination rates and informativeness show a positive correlation across both CHAIR and AMBER.
- On CHAIR, CHAIRs and object recall correlate at Pearson r = 0.73.
- CAAC and CEI preserve recall while mitigating hallucination, whereas AFTER and AGLA reduce recall alongside hallucination.
- On AMBER, Hal and Cover correlate at r = 0.70; AFTER achieves the lowest nominal hallucination rate but reduced coverage.
- Contrastive decoding can amplify vision-conditioned tokens, producing over-descriptive responses that inflate both hallucination and informativeness.
4 Hallucination Gains Do Not Reliably Transfer to General Capabilities
On MMSTAR, inference-time mitigation does not reliably preserve broader multimodal competence: fine-grained perception and reasoning often degrade, while category-level effects remain inconsistent.
- MMSTAR accuracy changes are measured against vanilla decoding across six multimodal task categories and 54 configurations.
- Fine-grained perception degrades across methods: only two of 18 configurations produced isolated gains exceeding 1%.
- Instance and logical reasoning also frequently degrade, particularly for InstructBLIP and LLaVA-1.5.
- No method improves MMSTAR macro averages on more than one of the three base architectures.
- A logical-reasoning example shows vanilla LLaVA-1.5 answering correctly while all six mitigation methods select incorrect options.
- AFTER raises InstructBLIP coarse perception by 7.2% but lowers mathematics by 4.4% on the same architecture.
- AGLA increases LLaVA-1.5 mathematics by 6.4% while reducing Science & Technology by 10.4 percentage points.
5 Related Work and Discussion
Inference-time hallucination mitigation methods are commonly evaluated through decoding and attention interventions, with standard benchmarks emphasizing object-level hallucination and hallucination rates.
- Training-free methods modify token selection or visual attention at inference time rather than retraining the LVLM.Contrastive decoding uses alternative distributions, while attention-based methods amplify, calibrate, or penalize attention patterns.
- Hallucination evaluation includes discriminative yes/no benchmarks and generative free-form benchmarks such as CHAIR and AMBER.Generative benchmarks remain largely object-centric, whereas discriminative benchmarks provide objective binary-style evaluation.
- Most hallucination-mitigation studies primarily report hallucination rates on these benchmarks.
6 Conclusion
The study finds that lower hallucination rates often coincide with reduced informativeness and do not reliably transfer to broader multimodal capabilities. It therefore recommends evaluating faithfulness alongside informativeness and capability.
- The study evaluates six inference-time mitigation methods across three LVLMs and three benchmarks.
- Hallucination reductions on CHAIR and AMBER often coincide with reduced informativeness.
- Gains on hallucination benchmarks do not reliably transfer to broader multimodal capabilities on MMSTAR.
- Hallucination mitigation should be evaluated as a faithfulness–informativeness–capability trade-off rather than by hallucination scores alone.
Limitations
The study’s conclusions are bounded by its benchmark coverage, inference-time method scope, model scale, hyperparameter protocol, and limits on causal interpretation. It also notes potential risks associated with how conservative generation findings might be used.
- Benchmark coverage: The benchmarks do not cover all relevant LVLM faithfulness axes, including attribute-level or relational hallucination, video understanding, and long-form instruction following.The reported patterns may not extend uniformly to these settings.
- Scope of mitigation methods: The study focuses on representative training-free inference-time interventions and excludes training-based methods, preference optimization, and retrieval-augmented grounding.Its conclusions therefore apply most directly to inference-time distribution-shaping methods.
- Model scale: The evaluation uses three 7B-scale LVLMs, so larger or newer models may respond differently and exhibit different trade-offs across scale and architecture.
- Hyperparameter scoping: When methods are applied outside their original studies, the analysis retains author-specified default hyperparameters instead of exhaustively optimizing each model.Custom tuning might alter individual performance profiles, while the protocol reflects standard out-of-the-box deployment conditions.
- Causal interpretation: The analysis does not directly measure grounding as a latent property, and its risk-suppression description is observational rather than mechanistic.
- Potential risks: The authors identify a potential risk that their findings could guide superficially accurate generation strategies that mask underlying unreliability.
A Full Experimental Details
The study evaluates six inference-time mitigation methods across three LVLMs using paired hallucination–informativeness metrics and MMSTAR capability accuracy. The appendix details method implementations, decoding settings, benchmark construction, and complete result reporting.
- Decoding and implementation: The vanilla baseline uses native generation without external modifications and is restricted to deterministic greedy decoding.Runs use seed=42; max_new_tokens is 512 for CHAIR and AMBER and 32 for MMSTAR.
- Benchmarks and metrics: CHAIR extracts mentioned object nouns from generated captions and compares them with image-level MSCOCO annotations.It reports sentence-level CHAIRs and instance-level CHAIRi error rates.
- Benchmarks and metrics: AMBER uses 1004 generative items with human-annotated object lists and reports CHAIR, Hal, and Cover.The paper pairs Hal with Cover as its primary hallucination–informativeness signal.
- Benchmarks and metrics: MMSTAR contains 1500 audited multiple-choice questions across six capability axes, with per-category accuracy and a macro average.Its items are filtered to remove cases solvable from text priors alone.
B.1 Per-Model Trade-off Scatters
Across LLaVA-1.5 and InstructBLIP, lower hallucination rates generally coincide with reduced informativeness. Complete CHAIR results likewise show that no method simultaneously minimizes hallucinations and maximizes object coverage.
- Cross-model scatter patterns: Across both additional architectures, localized hallucination reductions are generally entangled with losses in response detailedness.AFTER is an exception on LLaVA-1.5 with AMBER, reducing hallucination while increasing Cover.
- CHAIR results: No single configuration simultaneously minimizes CHAIRs and CHAIRi while maximizing object recall across the three LVLMs.The complete CHAIR table instead reveals distinct hallucination–coverage trade-offs among methods.
B.3 AMBER
AMBER results reproduce the hallucination–informativeness trade-off: methods with low hallucination often have lower Cover. MMSTAR results further show inconsistent or degraded broader capability performance.
- AMBER trade-offs: Methods that reduce AMBER CHAIR and Hal often lag significantly in Cover, the fraction of ground-truth objects mentioned.The paper pairs Hal with Cover as its main AMBER signal.
- MMSTAR capability results: No mitigation method consistently improves MMSTAR across models or categories, and vanilla decoding ranks second once and third twice in overall accuracy.Category gains are commonly offset by losses elsewhere.
C Qualitative Examples
Qualitative MMSTAR cases show mitigation methods turning correct vanilla predictions into systematic errors across perception, reasoning, and mathematics. The examples illustrate capability degradation through incorrect visual commitments and disrupted token selection.
- Perception: Mitigation methods over-count crowd targets when they overcompensate for background density.This example links a perception failure to altered visual interpretation.
- Reasoning: In a complex reasoning case, mitigation collapses all six options onto a false generic attribute by suppressing vital semantic pathways.The resulting answer is incorrect across the evaluated frameworks.
- Mathematics: Flattening the token distribution disrupts sequential commitments needed for exact arithmetic, causing every framework to select an incorrect ratio fraction.The described failure occurs in the mathematics example.
- Mathematics: VCD predicts 15 instead of the ground-truth 12 on a LLaVA-NeXT mathematics question, illustrating confident commitment to a fluent visual option.The passage attributes the error potentially to amplifying image-dependent tokens.
- Logical reasoning: M3ID selects salmon instead of “none of the above” on a LLaVA-1.5 food-web question, favoring an image-mentioned option over relational reasoning.The example is presented as a side effect of amplifying image-dependent tokens.
- Overview: Across MMSTAR task logs, vanilla decoding gives correct choices while mitigation strategies consistently produce systematic prediction failures.The failures span perception, instance tracking, and fractional calculation.