KVSmooth: Mitigating Hallucination in Multi-modal Large Language Models through Key-Value Smoothing¶

Conference: CVPR2026 arXiv: 2602.04268 Code: To be confirmed Area: Multimodal VLM Keywords: Multimodal hallucination mitigation, KV-Cache smoothing, attention entropy, exponential moving average, training-free inference

TL;DR¶

This paper proposes KVSmooth, a training-free plug-and-play inference-time method that applies adaptive exponential moving average (EMA) smoothing to KV-Cache guided by attention row entropy, effectively suppressing semantic drift and hallucination generation caused by sink tokens during decoding in multimodal large language models (MLLMs). On LLaVA-1.5, CHAIR_S is reduced from 41.8 to 18.2 (a 56% reduction), while F1 improves from 77.5 to 79.2.

Background & Motivation¶

Pervasive multimodal hallucination: MLLMs frequently generate objects, attributes, or relations inconsistent with visual content in tasks such as image captioning and VQA, severely hindering trustworthy deployment.
Long-range decay of visual dependency: As the decoding sequence grows, the influence of early visual tokens gradually diminishes in hidden representations, causing generated text to progressively deviate from image content.
Cumulative semantic drift: Minor inaccuracies at early generation steps accumulate and amplify over time, widening the gap between generated descriptions and visual facts.
Sink tokens exacerbate hallucination: Attention concentrates on a small number of "aggregation tokens" whose hidden states, produced via global averaging, deviate from the visual context and systematically inflate the logit scores of hallucinated objects.
Precision–recall trade-off in existing methods: Fine-tuning approaches require substantial data and computation; contrastive decoding incurs high computational overhead; attention redistribution methods often suppress correct object descriptions alongside hallucinations, sacrificing recall.
Insufficient understanding of the sink-token hallucination mechanism: Prior work focuses on reducing the occurrence of sink tokens or diminishing their attention weights, but does not explain why they trigger hallucinations.

Method¶

Overall Architecture¶

KVSmooth is a training-free, plug-and-play inference-time method. Its core mechanism applies adaptive EMA smoothing to the Key and Value vectors of the KV-Cache during decoding, stabilizing hidden-state evolution and suppressing hallucinations. The method consists of two core components.

Three Key Observations¶

Obs1 (Logit dynamic divergence): The mean logit of ground-truth (GT) objects decreases monotonically with stable variance, whereas the mean logit of hallucinated objects rises continuously with increasing variance—indicating that hallucination candidates accumulate instability in the hidden representations.
Obs2 (Row entropy and sink behavior): Attention row entropy is proposed as a real-time metric for measuring token sink intensity. Tokens with high row entropy have diffuse attention distributions; their hidden states approximate historical averages and exhibit small angular distances from most other states, causing them to attract disproportionately high attention in subsequent steps and form attention sinks. The cosine similarity between row entropy and the conventional column-sum metric concentrates around 0.79.
Obs3 (Entropy–rank coupling): The logit rank of hallucinated objects is positively correlated with attention row entropy—the more uniform a token's attention distribution (higher entropy), the higher the logits of hallucinated objects; the opposite holds for GT objects. This reveals a direct causal chain through which sink tokens systematically inflate hallucination scores via global context averaging.

Component 1: EMA Smoothing on KV-Cache¶

Based on the assumption that an ideal decoding trajectory should be smooth, hidden-state transitions are modeled as a Gaussian random walk $h_t = h_{t-1} + \epsilon_t$. Bayesian MAP estimation yields:

\[\hat{h_t} = (1-\lambda_t) o_t + \lambda_t h_{t-1}\]

which is exactly the EMA form. A key design choice is to apply EMA simultaneously to both Key and Value (rather than directly to hidden states), as this maximally regularizes both the mean and variance of logits, achieving the strongest hallucination suppression without degrading recall.

Component 2: Entropy-Guided Adaptive Coefficient¶

The attention row entropy $z_t^l$ of the current token at each layer is computed and maintained in a FIFO queue of length $M$.
The smoothing coefficient is determined by the percentile rank of the row entropy: $\hat{\lambda}_t^l = k/M$, assigning larger smoothing coefficients to higher-entropy tokens.
A clipping mechanism is introduced: centered at hyperparameter $\lambda_{\text{ref}}$, the coefficient is clipped to $[\lambda_{\text{ref}}-0.2, \lambda_{\text{ref}}+0.2]$, stabilizing generation while preserving representational diversity.

Final Update Rule¶

For token $x_t$ at designated layer $l$:

\[\hat{K_t^l} = (1-\tilde{\lambda}_t^l) K_t^l + \tilde{\lambda}_t^l K_{t-1}^l$$ $$\hat{V_t^l} = (1-\tilde{\lambda}_t^l) V_t^l + \tilde{\lambda}_t^l V_{t-1}^l\]

Applied to layers 3–31, with FIFO queue length 15 and $\lambda_{\text{ref}}$ set to 0.9 / 0.5 / 0.7 for LLaVA-1.5 / MiniGPT-4 / InstructBLIP, respectively.

Key Experimental Results¶

CHAIR Benchmark (Image Captioning Hallucination)¶

Method	LLaVA-1.5 CHAIR_S↓	LLaVA-1.5 F1↑	MiniGPT-4 CHAIR_S↓	MiniGPT-4 F1↑	InstructBLIP CHAIR_S↓	InstructBLIP F1↑
Baseline	41.8	77.5	31.8	69.9	61.4	71.6
PAI	22.6	75.5	24.6	71.0	63.4	71.1
OPERA	44.2	78.6	27.4	69.4	68.0	69.2
MiddleLayer	17.8	75.9	24.6	71.2	75.0	67.2
KVSmooth	18.2	79.2	17.0	71.7	42.2	75.1

On LLaVA-1.5, CHAIR_S is reduced by 56% while F1 improves from 77.5 to 79.2, making KVSmooth the only method to simultaneously improve both precision and recall.
On MiniGPT-4, CHAIR_S is reduced from 31.8 to 17.0 (a 47% reduction).

Object HalBench¶

On LLaVA-1.5, CHAIR_SR is reduced from 45.3% to 16.7% (a 63.1% reduction). Sentence-level hallucination rates are reduced by 63.1%, 40.3%, and 41.6% across the three models, respectively.

Ablation Study¶

Smoothing Location	LLaVA-1.5 Cs↓	F1↑
Attention output $o_t$ only	33.8	74.7
Key $K_t$ only	35.6	79.4
Key + Value	18.2	79.2

Jointly smoothing Key and Value yields the best performance; smoothing only hidden states leads to severe recall degradation.
The adaptive coefficient (Ada.) further reduces CHAIR_S on LLaVA-1.5 from 36.2 to 18.2 compared to a fixed coefficient, validating the precise identification capability of the entropy-guided mechanism.

Highlights & Insights¶

Training-free plug-and-play: Operates directly on the KV-Cache at inference time without fine-tuning or modifying model parameters, offering inherent generality.
Theory- and empirical-driven: The EMA form is derived from Bayesian MAP estimation, and the three observations provide a clear causal explanation chain (logit divergence → row-entropy sink → entropy–hallucination coupling).
Breaking the precision–recall dilemma: PR curve analysis demonstrates that KVSmooth is the only method that substantially reduces hallucinations while maintaining or improving F1.
Introducing the sink degree concept: Row entropy serves as a continuous, real-time sink metric that is more efficient than the conventional column-sum approach, requiring no multi-step look-back.
Broad validation: Evaluated across 3 models (LLaVA-1.5 / MiniGPT-4 / InstructBLIP) × 4 benchmarks (CHAIR / OPOPE / AMBER / Object HalBench) with consistent results.

Limitations & Future Work¶

Model-dependent hyperparameters: $\lambda_{\text{ref}}$ requires separate tuning for each model (0.9 / 0.5 / 0.7), with no automatic selection scheme.
Evaluation limited to 7B models: Performance on larger scales (13B / 70B) or newer architectures (e.g., Qwen2.5-VL) has not been verified.
Generation length constraint: Maximum generation is capped at 512 tokens; the effectiveness of EMA in ultra-long generation scenarios remains unexplored.
Fixed layer range: Application to layers 3–31 is empirically determined without a systematic layer selection criterion.
Focus on object hallucination only: Finer-grained hallucination types such as attribute and relation hallucinations are not evaluated.
High entropy is not always detrimental: At semantic transition points, high entropy may reflect legitimate context switching; uniform smoothing carries a theoretical risk of information loss.

Category	Representative Methods	Core Idea	KVSmooth Advantage
Contrastive decoding	VCD, OPERA	Noise-augmented views for contrast / rollback penalty	Lower computational overhead; no multiple forward passes required
Attention redistribution	PAI, SPARC, MiddleLayer	Enhance attention on visual tokens	Does not sacrifice recall; superior PR curve
Fine-tuning alignment	POVID, RLHF	Fine-tune with preference data	Training-free; requires no additional data
KV-Cache pruning	PruneHal	Remove redundant visual tokens	Preserves information more completely; regulates via smoothing rather than deletion

Rating¶

Novelty: ⭐⭐⭐⭐ — The row-entropy sink degree concept is novel; deriving EMA smoothing from a Bayesian perspective is theoretically elegant.
Experimental Thoroughness: ⭐⭐⭐⭐ — 3 models × 4 benchmarks with detailed ablations and PR analysis, though evaluation on larger models and additional hallucination types is absent.
Writing Quality: ⭐⭐⭐⭐ — The observation–method–experiment logical chain is clear; mathematical derivations are concise and accessible.
Value: ⭐⭐⭐⭐ — The lightweight, training-free approach offers strong practical utility and introduces a new direction for inference-time hallucination mitigation.

Smoothing Location	LLaVA-1.5 Cs↓	F1↑
Attention output \(o_t\) only	33.8	74.7
Key \(K_t\) only	35.6	79.4
Key + Value	18.2	79.2