| Challenge: | Autoregressive models (ARMs) are prone to hallucinations due to their sequential text generation and high latency. |
| Approach: | They propose a training-free decoding strategy that augments the attention key space with a static, distance-aware matrix to reduce the attention sink effect on semantic anchors. |
| Outcome: | The proposed method reduces the attention sink effect on semantic anchors while enhancing their ability to aggregate global visual information. |
Similar Papers
TokenPenalty: Alleviating Attention Sinks and Positional Decay in LVLMs (2026.findings-acl)
Copied to clipboard
| Challenge: | Multimodal large language models (MLLMs) often hallucinate due to two relevant phenomena: massive activation phenomenon and positional information decay. |
| Approach: | They propose a token-level intervention strategy that dynamically suppresses irrelevant visual tokens while preserving key contextual signals. |
| Outcome: | Experiments show that TokenTruth significantly improves factual consistency across MLLMs on standard image understanding benchmarks. |
Revisit What You See: Revealing Visual Semantics in Vision Tokens to Guide LVLM Decoding (2026.acl-long)
Copied to clipboard
| Challenge: | Large Vision–Language Models (LVLMs) integrate visual perception with language understanding, but how vision information contributes to the model’s decoding process remains under-explored. |
| Approach: | They propose a simple training-free decoding method that guides text generation in Large Vision–Language Models by Referencing Vision Tokens. |
| Outcome: | The proposed method leverages the semantic information embedded within vision tokens by projecting it into the text token distribution. |
Mask Tokens as Prophet: Fine-Grained Cache Eviction for Efficient dLLM Inference (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing cache eviction strategies for autoregressive language models fail to account for the role of mask tokens and specific characteristics in dLLMs. |
| Approach: | They propose a training-free cache eviction framework tailored to dLLMs that denies a fully masked sequence and allows parallel decoding at the expense of memory and computation. |
| Outcome: | The proposed framework reduces the cost of memory and cache eviction and improves efficiency by reducing allocation in intermediate layers and concentrating resources on prompt-preferring heads. |
Diffusion-CAM: Faithful Visual Explanations for dMLLMs (2026.acl-long)
Copied to clipboard
| Challenge: | Existing Class Activation Mapping methods are ill-suited for interpreting non-autoregressive behaviors of diffusion-based architectures. |
| Approach: | They propose to use a method to generate parallel activation maps by probing intermediate representations in the transformer backbone to capture latent features and their class-specific gradients. |
| Outcome: | Experiments show that Diffusion-CAM significantly outperforms SoTA methods in localization accuracy and visual fidelity. |
Revealing and Enhancing Core Visual Regions: Harnessing Internal Attention Dynamics for Hallucination Mitigation in LVLMs (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing training-free methods are vulnerable to the attention sink phenomenon . Existing methods include contrastive decoding and auxiliary expert models . |
| Approach: | They propose a training-free attention intervention that constructs a PAD map to identify semantically core visual regions and applies per-head Median Absolute Deviation Scaling to adaptively control the intervention strength. |
| Outcome: | The proposed intervention improves visual grounding and reduces hallucinations on multiple LVLMs and benchmarks. |
Vocabulary Hijacking in LVLMs: Unveiling Critical Attention Heads by Excluding Inert Tokens to Mitigate Hallucination (2026.acl-long)
Copied to clipboard
| Challenge: | Large Vision-Language Models (LVLMs) are capable of processing visual inputs, but are susceptible to hallucinations. |
| Approach: | They propose a method to localize and localize specific visual tokens, which are defined as **Inert Tokens**, across layers, revealing a rigid semantic collapse. |
| Outcome: | The proposed approach reduces the likelihood of LVLMs being hijacked by visual inputs while maintaining general capabilities. |
Beyond Logit Lens: Contextual Embeddings for Robust Hallucination Detection & Grounding in VLMs (2025.naacl-long)
Copied to clipboard
Anirudh Phukan, Divyansh Divyansh, Harshit Kumar Morj, Vaishnavi Vaishnavi, Apoorv Saxena, Koustava Goswami
| Challenge: | Large Multimodal Models are plagued by hallucinations that limit their reliability and adoption. |
| Approach: | They propose a method that leverages contextual token embeddings from LMMs to detect hallucinations. |
| Outcome: | The proposed method improves hallucination detection and grounding across diverse categories while excelling in tasks requiring contextual understanding. |
SCOPE: Boosting LLM Efficiency with Scoped Position Encoding (2026.acl-long)
Copied to clipboard
| Challenge: | Positional encodings are fundamental to Transformers, but explicit methods like RoPE can degrade under length extrapolation and incur extra arithmetic and memory-access overhead. |
| Approach: | They propose a framework that reimagines structured sparsity as an intrinsic position encoding mechanism. |
| Outcome: | The proposed framework reduces the number of attention FLOPs by 8x compared to RoPE on LLaMA-3-8B architectures while reducing training and inference latency. |
Focus-dLLM: Accelerating Long-Context Diffusion LLM Inference via Confidence-Guided Context Focusing (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for estimating attention importance for tokens are ineffective . dLLMs require bidirectional attention, which limits inference efficiency . |
| Approach: | They propose a training-free attention sparsification framework for efficient long-context inference . they propose 'sink-aware pruning strategy' to accurately estimate and remove redundant computation . |
| Outcome: | The proposed approach offers 29 lossless speedup under 32K context length. |
Shallow Focus, Deep Fixes: Enhancing Shallow Layers Vision Attention Sinks to Alleviate Hallucination in LVLMs (2025.emnlp-main)
Copied to clipboard
Xiaofeng Zhang, Yihao Quan, Chen Shen, Chaochen Gu, Xiaosong Yuan, Shaotian Yan, Jiawei Cao, Hao Cheng, Kaijie Wu, Jieping Ye
| Challenge: | Multimodal large language models (MLLMs) demonstrate excellent abilities for understanding visual information, but the hallucination remains a challenging problem. |
| Approach: | They propose a training-free approach to enhance vision attention sinks to facilitate convergence of the image token attention sink within shallow layers. |
| Outcome: | The proposed approach improves the convergence of the image token attention sink within shallow layers and strengthens the layer’s focus on the image itself. |