Evaluating and Mitigating LLM-as-a-judge Bias in Communication Systems

Large Language Models (LLMs) are increasingly being used to autonomously evaluate the quality of content in communication systems, e.g., to assess responses in telecom customer support chatbots. However, the impartiality of these AI "judges" is not guaranteed, and any biases in their evaluation criteria could skew outcomes and undermine user trust. In this paper, we systematically investigate judgment biases in two LLM-as-a-judge models (i.e., GPT-Judge and JudgeLM) under the point-wise scoring setting, encompassing 11 types of biases that cover both implicit and explicit forms. We observed that state-of-the-art LLM judges demonstrate robustness to biased inputs, generally assigning them lower scores than the corresponding clean samples. Providing a detailed scoring rubric further enhances this robustness. We further found that fine-tuning an LLM on high-scoring yet biased responses can significantly degrade its performance, highlighting the risk of training on biased data. We also discovered that the judged scores correlate with task difficulty: a challenging dataset like GPQA yields lower average scores, whereas an open-ended reasoning dataset (e.g., JudgeLM-val) sees higher average scores. Finally, we proposed four potential mitigation strategies to ensure fair and reliable AI judging in practical communication scenarios.

A.I.R.: Enabling Adaptive, Iterative, and Reasoning-based Frame Selection For Video Question Answering

Effectively applying Vision-Language Models (VLMs) to Video Question Answering (VideoQA) hinges on selecting a concise yet comprehensive set of frames, as processing entire videos is computationally infeasible. However, current frame selection methods face a critical trade-off: approaches relying on lightweight similarity models, such as CLIP, often fail to capture the nuances of complex queries, resulting in inaccurate similarity scores that cannot reflect the authentic query-frame relevance, which further undermines frame selection. Meanwhile, methods that leverage a VLM for deeper analysis achieve higher accuracy but incur prohibitive computational costs. To address these limitations, we propose A.I.R., a training-free approach for Adaptive, Iterative, and Reasoning-based frame selection. We leverage a powerful VLM to perform deep, semantic analysis on complex queries, and this analysis is deployed within a cost-effective iterative loop that processes only a small batch of the most high-potential frames at a time. Extensive experiments on various VideoQA benchmarks demonstrate that our approach outperforms existing frame selection methods, significantly boosts the performance of the foundation VLM, and achieves substantial gains in computational efficiency over other VLM-based techniques.

From Frames to Clips: Efficient Key Clip Selection for Long-Form Video Understanding

Video Large Language Models (VLMs) have achieved remarkable results on a variety of vision language tasks, yet their practical use is limited by the "needle in a haystack" problem: the massive number of visual tokens produced from raw video frames exhausts the model's context window. Existing solutions alleviate this issue by selecting a sparse set of frames, thereby reducing token count, but such frame-wise selection discards essential temporal dynamics, leading to suboptimal reasoning about motion and event continuity. In this work we systematically explore the impact of temporal information and demonstrate that extending selection from isolated key frames to key clips, which are short, temporally coherent segments, improves video understanding. To maintain a fixed computational budget while accommodating the larger token footprint of clips, we propose an adaptive resolution strategy that dynamically balances spatial resolution and clip length, ensuring a constant token count per video. Experiments on three long-form video benchmarks demonstrate that our training-free approach, F2C, outperforms uniform sampling up to 8.1%, 5.6%, and 10.3% on Video-MME, LongVideoBench and MLVU benchmarks, respectively. These results highlight the importance of preserving temporal coherence in frame selection and provide a practical pathway for scaling Video LLMs to real world video understanding applications. Project webpage is available at https://guangyusun.com/f2c .

AToken: A Unified Tokenizer for Vision

We present AToken, the first unified visual tokenizer that achieves both high-fidelity reconstruction and semantic understanding across images, videos, and 3D assets. Unlike existing tokenizers that specialize in either reconstruction or understanding for single modalities, AToken encodes these diverse visual inputs into a shared 4D latent space, unifying both tasks and modalities in a single framework. Specifically, we introduce a pure transformer architecture with 4D rotary position embeddings to process visual inputs of arbitrary resolutions and temporal durations. To ensure stable training, we introduce an adversarial-free training objective that combines perceptual and Gram matrix losses, achieving state-of-the-art reconstruction quality. By employing a progressive training curriculum, AToken gradually expands from single images, videos, and 3D, and supports both continuous and discrete latent tokens. AToken achieves 0.21 rFID with 82.2% ImageNet accuracy for images, 3.01 rFVD with 40.2% MSRVTT retrieval for videos, and 28.28 PSNR with 90.9% classification accuracy for 3D.. In downstream applications, AToken enables both visual generation tasks (e.g., image generation with continuous and discrete tokens, text-to-video generation, image-to-3D synthesis) and understanding tasks (e.g., multimodal LLMs), achieving competitive performance across all benchmarks. These results shed light on the next-generation multimodal AI systems built upon unified visual tokenization.

MANZANO: A Simple and Scalable Unified Multimodal Model with a Hybrid Vision Tokenizer

Unified multimodal Large Language Models (LLMs) that can both understand and generate visual content hold immense potential. However, existing open-source models often suffer from a performance trade-off between these capabilities. We present Manzano, a simple and scalable unified framework that substantially reduces this tension by coupling a hybrid image tokenizer with a well-curated training recipe. A single shared vision encoder feeds two lightweight adapters that produce continuous embeddings for image-to-text understanding and discrete tokens for text-to-image generation within a common semantic space. A unified autoregressive LLM predicts high-level semantics in the form of text and image tokens, with an auxiliary diffusion decoder subsequently translating the image tokens into pixels. The architecture, together with a unified training recipe over understanding and generation data, enables scalable joint learning of both capabilities. Manzano achieves state-of-the-art results among unified models, and is competitive with specialist models, particularly on text-rich evaluation. Our studies show minimal task conflicts and consistent gains from scaling model size, validating our design choice of a hybrid tokenizer.

EvoEmpirBench: Dynamic Spatial Reasoning with Agent-ExpVer

Most existing spatial reasoning benchmarks focus on static or globally observable environments, failing to capture the challenges of long-horizon reasoning and memory utilization under partial observability and dynamic changes. We introduce two dynamic spatial benchmarks, locally observable maze navigation and match-2 elimination that systematically evaluate models' abilities in spatial understanding and adaptive planning when local perception, environment feedback, and global objectives are tightly coupled. Each action triggers structural changes in the environment, requiring continuous update of cognition and strategy. We further propose a subjective experience-based memory mechanism for cross-task experience transfer and validation. Experiments show that our benchmarks reveal key limitations of mainstream models in dynamic spatial reasoning and long-term memory, providing a comprehensive platform for future methodological advances. Our code and data are available at https://anonymous.4open.science/r/EvoEmpirBench-143C/.

Lethe: Purifying Backdoored Large Language Models with Knowledge Dilution

Large language models (LLMs) have seen significant advancements, achieving superior performance in various Natural Language Processing (NLP) tasks. However, they remain vulnerable to backdoor attacks, where models behave normally for standard queries but generate harmful responses or unintended output when specific triggers are activated. Existing backdoor defenses either lack comprehensiveness, focusing on narrow trigger settings, detection-only mechanisms, and limited domains, or fail to withstand advanced scenarios like model-editing-based, multi-trigger, and triggerless attacks. In this paper, we present LETHE, a novel method to eliminate backdoor behaviors from LLMs through knowledge dilution using both internal and external mechanisms. Internally, LETHE leverages a lightweight dataset to train a clean model, which is then merged with the backdoored model to neutralize malicious behaviors by diluting the backdoor impact within the model's parametric memory. Externally, LETHE incorporates benign and semantically relevant evidence into the prompt to distract LLM's attention from backdoor features. Experimental results on classification and generation domains across 5 widely used LLMs demonstrate that LETHE outperforms 8 state-of-the-art defense baselines against 8 backdoor attacks. LETHE reduces the attack success rate of advanced backdoor attacks by up to 98% while maintaining model utility. Furthermore, LETHE has proven to be cost-efficient and robust against adaptive backdoor attacks.

UNIFORM: Unifying Knowledge from Large-scale and Diverse Pre-trained Models

In the era of deep learning, the increasing number of pre-trained models available online presents a wealth of knowledge. These models, developed with diverse architectures and trained on varied datasets for different tasks, provide unique interpretations of the real world. Their collective consensus is likely universal and generalizable to unseen data. However, effectively harnessing this collective knowledge poses a fundamental challenge due to the heterogeneity of pre-trained models. Existing knowledge integration solutions typically rely on strong assumptions about training data distributions and network architectures, limiting them to learning only from specific types of models and resulting in data and/or inductive biases. In this work, we introduce a novel framework, namely UNIFORM, for knowledge transfer from a diverse set of off-the-shelf models into one student model without such constraints. Specifically, we propose a dedicated voting mechanism to capture the consensus of knowledge both at the logit level -- incorporating teacher models that are capable of predicting target classes of interest -- and at the feature level, utilizing visual representations learned on arbitrary label spaces. Extensive experiments demonstrate that UNIFORM effectively enhances unsupervised object recognition performance compared to strong knowledge transfer baselines. Notably, it exhibits remarkable scalability by benefiting from over one hundred teachers, while existing methods saturate at a much smaller scale.

Seeing Further on the Shoulders of Giants: Knowledge Inheritance for Vision Foundation Models

Vision foundation models (VFMs) are predominantly developed using data-centric methods. These methods require training on vast amounts of data usually with high-quality labels, which poses a bottleneck for most institutions that lack both large-scale data and high-end GPUs. On the other hand, many open-source vision models have been pretrained on domain-specific data, enabling them to distill and represent core knowledge in a form that is transferable across diverse applications. Even though these models are highly valuable assets, they remain largely under-explored in empowering the development of a general-purpose VFM. In this paper, we presents a new model-driven approach for training VFMs through joint knowledge transfer and preservation. Our method unifies multiple pre-trained teacher models in a shared latent space to mitigate the ``imbalanced transfer'' issue caused by their distributional gaps. Besides, we introduce a knowledge preservation strategy to take a general-purpose teacher as a knowledge base for integrating knowledge from the remaining purpose-specific teachers using an adapter module. By unifying and aggregating existing models, we build a powerful VFM to inherit teachers' expertise without needing to train on a large amount of labeled data. Our model not only provides generalizable visual features, but also inherently supports multiple downstream tasks. Extensive experiments demonstrate that our VFM outperforms existing data-centric models across four fundamental vision tasks, including image classification, object detection, semantic and instance segmentation.

FakeHunter: Multimodal Step-by-Step Reasoning for Explainable Video Forensics

FakeHunter is a multimodal deepfake detection framework that combines memory-guided retrieval, chain-of-thought (Observation-Thought-Action) reasoning, and tool-augmented verification to provide accurate and interpretable video forensics. FakeHunter encodes visual content using CLIP and audio using CLAP, generating joint audio-visual embeddings that retrieve semantically similar real exemplars from a FAISS-indexed memory bank for contextual grounding. Guided by the retrieved context, the system iteratively reasons over evidence to localize manipulations and explain them. When confidence is low, it automatically invokes specialized tools-such as zoom-in image forensics or mel-spectrogram inspection-for fine-grained verification. Built on Qwen2.5-Omni-7B, FakeHunter produces structured JSON verdicts that specify what was modified, where it occurs, and why it is judged fake. We also introduce X-AVFake, a benchmark comprising 5.7k+ manipulated and real videos (950+ min) annotated with manipulation type, region/entity, violated reasoning category, and free-form justification. On X-AVFake, FakeHunter achieves an accuracy of 34.75%, outperforming the vanilla Qwen2.5-Omni-7B by 16.87 percentage points and MiniCPM-2.6 by 25.56 percentage points. Ablation studies reveal that memory retrieval contributes a 7.75 percentage point gain, and tool-based inspection improves low-confidence cases to 46.50%. Despite its multi-stage design, the pipeline processes a 10-minute clip in 8 minutes on a single NVIDIA A800 (0.8x real-time) or 2 minutes on four GPUs (0.2x), demonstrating practical deployability.