Multi-Level Aware Preference Learning: Enhancing RLHF for Complex Multi-Instruction Tasks

RLHF has emerged as a predominant approach for aligning artificial intelligence systems with human preferences, demonstrating exceptional and measurable efficacy in instruction following tasks; however, it exhibits insufficient compliance capabilities when confronted with complex multi-instruction tasks. Conventional approaches rely heavily on human annotation or more sophisticated large language models, thereby introducing substantial resource expenditure or potential bias concerns. Meanwhile, alternative synthetic methods that augment standard preference datasets often compromise the model's semantic quality. Our research identifies a critical oversight in existing techniques, which predominantly focus on comparing responses while neglecting valuable latent signals embedded within prompt inputs, and which only focus on preference disparities at the intra-sample level, while neglecting to account for the inter-sample level preference differentials that exist among preference data. To leverage these previously neglected indicators, we propose a novel Multi-level Aware Preference Learning (MAPL) framework, capable of enhancing multi-instruction capabilities. Specifically, for any given response in original preference data pairs, we construct varied prompts with a preference relation under different conditions, in order to learn intra-sample level preference disparities. Furthermore, for any given original preference pair, we synthesize multi-instruction preference pairs to capture preference discrepancies at the inter-sample level. Building on the two datasets constructed above, we consequently devise two sophisticated training objective functions. Subsequently, our framework integrates seamlessly into both Reward Modeling and Direct Preference Optimization paradigms. Through rigorous evaluation across multiple benchmarks, we empirically validate the efficacy of our framework.

Mitigating Hallucination in VideoLLMs via Temporal-Aware Activation Engineering

Multimodal large language models (MLLMs) have achieved remarkable progress in video understanding.However, hallucination, where the model generates plausible yet incorrect outputs, persists as a significant and under-addressed challenge in the video domain. Among existing solutions, activation engineering has proven successful in mitigating hallucinations in LLMs and ImageLLMs, yet its applicability to VideoLLMs remains largely unexplored. In this work, we are the first to systematically investigate the effectiveness and underlying mechanisms of activation engineering for mitigating hallucinations in VideoLLMs. We initially conduct an investigation of the key factors affecting the performance of activation engineering and find that a model's sensitivity to hallucination depends on $\textbf{temporal variation}$ rather than task type. Moreover, selecting appropriate internal modules and dataset for activation engineering is critical for reducing hallucination. Guided by these findings, we propose a temporal-aware activation engineering framework for VideoLLMs, which adaptively identifies and manipulates hallucination-sensitive modules based on the temporal variation characteristic, substantially mitigating hallucinations without additional LLM fine-tuning. Experiments across multiple models and benchmarks demonstrate that our method markedly reduces hallucination in VideoLLMs, thereby validating the robustness of our findings.

Bias Fitting to Mitigate Length Bias of Reward Model in RLHF

Reinforcement Learning from Human Feedback relies on reward models to align large language models with human preferences. However, RLHF often suffers from reward hacking, wherein policy learning exploits flaws in the trained reward model to maximize reward scores without genuinely aligning with human preferences. A significant example of such reward hacking is length bias, where reward models usually favor longer responses irrespective of actual response quality. Previous works on length bias have notable limitations, these approaches either mitigate bias without characterizing the bias form, or simply assume a linear length-reward relation. To accurately model the intricate nature of length bias and facilitate more effective bias mitigation, we propose FiMi-RM (Bias Fitting to Mitigate Length Bias of Reward Model in RLHF), a framework that autonomously learns and corrects underlying bias patterns. Our approach consists of three stages: First, we train a standard reward model which inherently contains length bias. Next, we deploy a lightweight fitting model to explicitly capture the non-linear relation between length and reward. Finally, we incorporate this learned relation into the reward model to debias. Experimental results demonstrate that FiMi-RM achieves a more balanced length-reward distribution. Furthermore, when applied to alignment algorithms, our debiased reward model improves length-controlled win rate and reduces verbosity without compromising its performance.

VisuLogic: A Benchmark for Evaluating Visual Reasoning in Multi-modal Large Language Models

Visual reasoning is a core component of human intelligence and a critical capability for advanced multimodal models. Yet current reasoning evaluations of multimodal large language models (MLLMs) often rely on text descriptions and allow language-based reasoning shortcuts, failing to measure genuine vision-centric reasoning. To address this, we introduce VisuLogic: a benchmark of 1,000 human-verified problems across six categories (e.g., quantitative shifts, spatial relations, attribute comparisons). These various types of questions can be evaluated to assess the visual reasoning capabilities of MLLMs from multiple perspectives. We evaluate leading MLLMs on this benchmark and analyze their results to identify common failure modes. Most models score below 30% accuracy-only slightly above the 25% random baseline and far below the 51.4% achieved by humans-revealing significant gaps in visual reasoning. Furthermore, we provide a supplementary training dataset and a reinforcement-learning baseline to support further progress.

Multi-Scale Invertible Neural Network for Wide-Range Variable-Rate Learned Image Compression

Autoencoder-based structures have dominated recent learned image compression methods. However, the inherent information loss associated with autoencoders limits their rate-distortion performance at high bit rates and restricts their flexibility of rate adaptation. In this paper, we present a variable-rate image compression model based on invertible transform to overcome these limitations. Specifically, we design a lightweight multi-scale invertible neural network, which bijectively maps the input image into multi-scale latent representations. To improve the compression efficiency, a multi-scale spatial-channel context model with extended gain units is devised to estimate the entropy of the latent representation from high to low levels. Experimental results demonstrate that the proposed method achieves state-of-the-art performance compared to existing variable-rate methods, and remains competitive with recent multi-model approaches. Notably, our method is the first learned image compression solution that outperforms VVC across a very wide range of bit rates using a single model, especially at high bit rates.The source code is available at \href{https://github.com/hytu99/MSINN-VRLIC}{https://github.com/hytu99/MSINN-VRLIC}.

Cross-Modal Consistency Learning for Sign Language Recognition

Pre-training has been proven to be effective in boosting the performance of Isolated Sign Language Recognition (ISLR). Existing pre-training methods solely focus on the compact pose data, which eliminate background perturbation but inevitably suffer from insufficient semantic cues compared to raw RGB videos. Nevertheless, direct representation learning only from RGB videos remains challenging due to the presence of sign-independent visual features. To address this dilemma, we propose a Cross-modal Consistency Learning framework (CCL-SLR), which leverages the cross-modal consistency from both RGB and pose modalities based on self-supervised pre-training. First, CCL-SLR employs contrastive learning for instance discrimination within and across modalities. Through the single-modal and cross-modal contrastive learning, CCL-SLR gradually aligns the feature spaces of RGB and pose modalities, thereby extracting consistent sign representations. Second, we further introduce Motion-Preserving Masking (MPM) and Semantic Positive Mining (SPM) techniques to improve cross-modal consistency from the perspective of data augmentation and sample similarity, respectively. Extensive experiments on four ISLR benchmarks show that CCL-SLR achieves impressive performance, demonstrating its effectiveness. The code will be released to the public.

Multi-Cue Adaptive Visual Token Pruning for Large Vision-Language Models

As the computational needs of Large Vision-Language Models (LVLMs) increase, visual token pruning has proven effective in improving inference speed and memory efficiency. Traditional pruning methods in LVLMs predominantly focus on attention scores to determine token relevance, overlooking critical aspects such as spatial position and token similarity. To this end, we introduce AdaptPrune, a novel plug-and-play training-free pruning method that builds on conventional attention-based pruning by integrating spatial distance and token similarity with an adaptive NMS approach. Our method is based on several observed phenomena in large models: the positional bias in the model's image attention and the redundancy of token information ignored by previous approaches. By integrating attention, spatial, and similarity information, our approach ensures a comprehensive evaluation of token importance and substantially refines the pruning decisions. Our method has been extensively tested across various LVLMs and benchmarks, confirming its robustness and adaptability. The results demonstrate that AdaptPrune consistently outperforms existing methods across various pruning ratios. Code is available at https://github.com/bzluan/AdaptPrune.

DesignDiffusion: High-Quality Text-to-Design Image Generation with Diffusion Models

In this paper, we present DesignDiffusion, a simple yet effective framework for the novel task of synthesizing design images from textual descriptions. A primary challenge lies in generating accurate and style-consistent textual and visual content. Existing works in a related task of visual text generation often focus on generating text within given specific regions, which limits the creativity of generation models, resulting in style or color inconsistencies between textual and visual elements if applied to design image generation. To address this issue, we propose an end-to-end, one-stage diffusion-based framework that avoids intricate components like position and layout modeling. Specifically, the proposed framework directly synthesizes textual and visual design elements from user prompts. It utilizes a distinctive character embedding derived from the visual text to enhance the input prompt, along with a character localization loss for enhanced supervision during text generation. Furthermore, we employ a self-play Direct Preference Optimization fine-tuning strategy to improve the quality and accuracy of the synthesized visual text. Extensive experiments demonstrate that DesignDiffusion achieves state-of-the-art performance in design image generation.

Model Evolution Framework with Genetic Algorithm for Multi-Task Reinforcement Learning

Multi-task reinforcement learning employs a single policy to complete various tasks, aiming to develop an agent with generalizability across different scenarios. Given the shared characteristics of tasks, the agent's learning efficiency can be enhanced through parameter sharing. Existing approaches typically use a routing network to generate specific routes for each task and reconstruct a set of modules into diverse models to complete multiple tasks simultaneously. However, due to the inherent difference between tasks, it is crucial to allocate resources based on task difficulty, which is constrained by the model's structure. To this end, we propose a Model Evolution framework with Genetic Algorithm (MEGA), which enables the model to evolve during training according to the difficulty of the tasks. When the current model is insufficient for certain tasks, the framework will automatically incorporate additional modules, enhancing the model's capabilities. Moreover, to adapt to our model evolution framework, we introduce a genotype module-level model, using binary sequences as genotype policies for model reconstruction, while leveraging a non-gradient genetic algorithm to optimize these genotype policies. Unlike routing networks with fixed output dimensions, our approach allows for the dynamic adjustment of the genotype policy length, enabling it to accommodate models with a varying number of modules. We conducted experiments on various robotics manipulation tasks in the Meta-World benchmark. Our state-of-the-art performance demonstrated the effectiveness of the MEGA framework. We will release our source code to the public.

Disentangling Length Bias In Preference Learning Via Response-Conditioned Modeling

Reinforcement Learning from Human Feedback (RLHF) has achieved considerable success in aligning large language models (LLMs) by modeling human preferences with a learnable reward model and employing a reinforcement learning algorithm to maximize the reward model's scores. However, these reward models are susceptible to exploitation through various superficial confounding factors, with length bias emerging as a particularly significant concern. Moreover, while the pronounced impact of length bias on preference modeling suggests that LLMs possess an inherent sensitivity to length perception, our preliminary investigations reveal that fine-tuned LLMs consistently struggle to adhere to explicit length instructions. To address these two limitations, we propose a novel framework wherein the reward model explicitly differentiates between human semantic preferences and response length requirements. Specifically, we introduce a Response-conditioned Bradley-Terry (Rc-BT) model that enhances the reward model's capability in length bias mitigating and length instruction following, through training on our augmented dataset. Furthermore, we propose the Rc-DPO algorithm to leverage the Rc-BT model for direct policy optimization (DPO) of LLMs, simultaneously mitigating length bias and promoting adherence to length instructions. Extensive evaluations demonstrate that our approach substantially improves both preference modeling and length instruction compliance, with its effectiveness validated across various foundational models and preference datasets.