Alignment of large language models with constrained learning

We study the problem of computing an optimal large language model (LLM) policy for a constrained alignment problem, where the goal is to maximize a primary reward objective while satisfying constraints on secondary utilities. Despite the popularity of Lagrangian-based LLM policy search in constrained alignment, iterative primal-dual methods often fail to converge, and non-iterative dual-based methods do not achieve optimality in the LLM parameter space. To address these challenges, we employ Lagrangian duality to develop an iterative dual-based alignment method that alternates between updating the LLM policy via Lagrangian maximization and updating the dual variable via dual descent. In theory, we characterize the primal-dual gap between the primal value in the distribution space and the dual value in the LLM parameter space. We further quantify the optimality gap of the learned LLM policies at near-optimal dual variables with respect to both the objective and the constraint functions. These results prove that dual-based alignment methods can find an optimal constrained LLM policy, up to an LLM parametrization gap. We demonstrate the effectiveness and merits of our approach through extensive experiments conducted on the PKU-SafeRLHF dataset.

Effective Reinforcement Learning for Reasoning in Language Models

Reinforcement learning (RL) has emerged as a promising strategy for improving the reasoning capabilities of language models (LMs) in domains such as mathematics and coding. However, most modern RL algorithms were designed to target robotics applications, which differ significantly from LM reasoning. We analyze RL algorithm design decisions for LM reasoning, for both accuracy and computational efficiency, focusing on relatively small models due to computational constraints. Our findings are: (i) on-policy RL significantly outperforms supervised fine-tuning (SFT), (ii) PPO-based off-policy updates increase accuracy instead of reduce variance, and (iii) removing KL divergence can lead to more concise generations and higher accuracy. Furthermore, we find that a key bottleneck to computational efficiency is that the optimal batch sizes for inference and backpropagation are different. We propose a novel algorithm, DASH, that performs preemptive sampling (i.e., sample a large batch and accumulate gradient updates in small increments), and gradient filtering (i.e., drop samples with small advantage estimates). We show that DASH reduces training time by 83% compared to a standard implementation of GRPO without sacrificing accuracy. Our findings provide valuable insights on designing effective RL algorithms for LM reasoning.

Flash-VL 2B: Optimizing Vision-Language Model Performance for Ultra-Low Latency and High Throughput

In this paper, we introduce Flash-VL 2B, a novel approach to optimizing Vision-Language Models (VLMs) for real-time applications, targeting ultra-low latency and high throughput without sacrificing accuracy. Leveraging advanced architectural enhancements and efficient computational strategies, Flash-VL 2B is designed to maximize throughput by reducing processing time while maintaining competitive performance across multiple vision-language benchmarks. Our approach includes tailored architectural choices, token compression mechanisms, data curation, training schemes, and a novel image processing technique called implicit semantic stitching that effectively balances computational load and model performance. Through extensive evaluations on 11 standard VLM benchmarks, we demonstrate that Flash-VL 2B achieves state-of-the-art results in both speed and accuracy, making it a promising solution for deployment in resource-constrained environments and large-scale real-time applications.

Effective Length Extrapolation via Dimension-Wise Positional Embeddings Manipulation

Large Language Models (LLMs) often struggle to process and generate coherent context when the number of input tokens exceeds the pre-trained length. Recent advancements in long-context extension have significantly expanded the context window of LLMs but require expensive overhead to train the large-scale models with longer context. In this work, we propose Dimension-Wise Positional Embeddings Manipulation (DPE), a training-free framework to extrapolate the context window of LLMs by diving into RoPE's different hidden dimensions. Instead of manipulating all dimensions equally, DPE detects the effective length for every dimension and finds the key dimensions for context extension. We reuse the original position indices with their embeddings from the pre-trained model and manipulate the key dimensions' position indices to their most effective lengths. In this way, DPE adjusts the pre-trained models with minimal modifications while ensuring that each dimension reaches its optimal state for extrapolation. DPE significantly surpasses well-known baselines such as YaRN and Self-Extend. DPE enables Llama3-8k 8B to support context windows of 128k tokens without continual training and integrates seamlessly with Flash Attention 2. In addition to its impressive extrapolation capability, DPE also dramatically improves the models' performance within training length, such as Llama3.1 70B, by over 18 points on popular long-context benchmarks RULER. When compared with commercial models, Llama 3.1 70B with DPE even achieves better performance than GPT-4-128K.

MR. Guard: Multilingual Reasoning Guardrail using Curriculum Learning

Large Language Models (LLMs) are susceptible to adversarial attacks such as jailbreaking, which can elicit harmful or unsafe behaviors. This vulnerability is exacerbated in multilingual setting, where multilingual safety-aligned data are often limited. Thus, developing a guardrail capable of detecting and filtering unsafe content across diverse languages is critical for deploying LLMs in real-world applications. In this work, we propose an approach to build a multilingual guardrail with reasoning. Our method consists of: (1) synthetic multilingual data generation incorporating culturally and linguistically nuanced variants, (2) supervised fine-tuning, and (3) a curriculum-guided Group Relative Policy Optimization (GRPO) framework that further improves performance. Experimental results demonstrate that our multilingual guardrail consistently outperforms recent baselines across both in-domain and out-of-domain languages. The multilingual reasoning capability of our guardrail enables it to generate multilingual explanations, which are particularly useful for understanding language-specific risks and ambiguities in multilingual content moderation.

Logits DeConfusion with CLIP for Few-Shot Learning

With its powerful visual-language alignment capability, CLIP performs well in zero-shot and few-shot learning tasks. However, we found in experiments that CLIP's logits suffer from serious inter-class confusion problems in downstream tasks, and the ambiguity between categories seriously affects the accuracy. To address this challenge, we propose a novel method called Logits DeConfusion, which effectively learns and eliminates inter-class confusion in logits by combining our Multi-level Adapter Fusion (MAF) module with our Inter-Class Deconfusion (ICD) module. Our MAF extracts features from different levels and fuses them uniformly to enhance feature representation. Our ICD learnably eliminates inter-class confusion in logits with a residual structure. Experimental results show that our method can significantly improve the classification performance and alleviate the inter-class confusion problem. The code is available at https://github.com/LiShuo1001/LDC.

Evaluating the Diversity and Quality of LLM Generated Content

Recent work suggests that preference-tuning techniques--including Reinforcement Learning from Human Preferences (RLHF) methods like PPO and GRPO, as well as alternatives like DPO--reduce diversity, creating a dilemma given that such models are widely deployed in applications requiring diverse outputs. To address this, we introduce a framework for measuring effective semantic diversity--diversity among outputs that meet quality thresholds--which better reflects the practical utility of large language models (LLMs). Using open-ended tasks that require no human intervention, we find counterintuitive results: although preference-tuned models--especially those trained via RL--exhibit reduced lexical and syntactic diversity, they produce greater effective semantic diversity than SFT or base models, not from increasing diversity among high-quality outputs, but from generating more high-quality outputs overall. We discover that preference tuning reduces syntactic diversity while preserving semantic diversity--revealing a distinction between diversity in form and diversity in content that traditional metrics often overlook. Our analysis further shows that smaller models are consistently more parameter-efficient at generating unique content within a fixed sampling budget, offering insights into the relationship between model scaling and diversity. These findings have important implications for applications that require diverse yet high-quality outputs, from creative assistance to synthetic data generation.

Debate-Feedback: A Multi-Agent Framework for Efficient Legal Judgment Prediction

The use of AI in legal analysis and prediction (LegalAI) has gained widespread attention, with past research focusing on retrieval-based methods and fine-tuning large models. However, these approaches often require large datasets and underutilize the capabilities of modern large language models (LLMs). In this paper, inspired by the debate phase of real courtroom trials, we propose a novel legal judgment prediction model based on the Debate-Feedback architecture, which integrates LLM multi-agent debate and reliability evaluation models. Unlike traditional methods, our model achieves significant improvements in efficiency by minimizing the need for large historical datasets, thus offering a lightweight yet robust solution. Comparative experiments show that it outperforms several general-purpose and domain-specific legal models, offering a dynamic reasoning process and a promising direction for future LegalAI research.

ATM-Net: Anatomy-Aware Text-Guided Multi-Modal Fusion for Fine-Grained Lumbar Spine Segmentation

Accurate lumbar spine segmentation is crucial for diagnosing spinal disorders. Existing methods typically use coarse-grained segmentation strategies that lack the fine detail needed for precise diagnosis. Additionally, their reliance on visual-only models hinders the capture of anatomical semantics, leading to misclassified categories and poor segmentation details. To address these limitations, we present ATM-Net, an innovative framework that employs an anatomy-aware, text-guided, multi-modal fusion mechanism for fine-grained segmentation of lumbar substructures, i.e., vertebrae (VBs), intervertebral discs (IDs), and spinal canal (SC). ATM-Net adopts the Anatomy-aware Text Prompt Generator (ATPG) to adaptively convert image annotations into anatomy-aware prompts in different views. These insights are further integrated with image features via the Holistic Anatomy-aware Semantic Fusion (HASF) module, building a comprehensive anatomical context. The Channel-wise Contrastive Anatomy-Aware Enhancement (CCAE) module further enhances class discrimination and refines segmentation through class-wise channel-level multi-modal contrastive learning. Extensive experiments on the MRSpineSeg and SPIDER datasets demonstrate that ATM-Net significantly outperforms state-of-the-art methods, with consistent improvements regarding class discrimination and segmentation details. For example, ATM-Net achieves Dice of 79.39% and HD95 of 9.91 pixels on SPIDER, outperforming the competitive SpineParseNet by 8.31% and 4.14 pixels, respectively.

Dynamic Allocation Hypernetwork with Adaptive Model Recalibration for Federated Continual Learning

Federated continual learning (FCL) offers an emerging pattern to facilitate the applicability of federated learning (FL) in real-world scenarios, where tasks evolve dynamically and asynchronously across clients, especially in medical scenario. Existing server-side FCL methods in nature domain construct a continually learnable server model by client aggregation on all-involved tasks. However, they are challenged by: (1) Catastrophic forgetting for previously learned tasks, leading to error accumulation in server model, making it difficult to sustain comprehensive knowledge across all tasks. (2) Biased optimization due to asynchronous tasks handled across different clients, leading to the collision of optimization targets of different clients at the same time steps. In this work, we take the first step to propose a novel server-side FCL pattern in medical domain, Dynamic Allocation Hypernetwork with adaptive model recalibration (FedDAH). It is to facilitate collaborative learning under the distinct and dynamic task streams across clients. To alleviate the catastrophic forgetting, we propose a dynamic allocation hypernetwork (DAHyper) where a continually updated hypernetwork is designed to manage the mapping between task identities and their associated model parameters, enabling the dynamic allocation of the model across clients. For the biased optimization, we introduce a novel adaptive model recalibration (AMR) to incorporate the candidate changes of historical models into current server updates, and assign weights to identical tasks across different time steps based on the similarity for continual optimization. Extensive experiments on the AMOS dataset demonstrate the superiority of our FedDAH to other FCL methods on sites with different task streams. The code is available:https://github.com/jinlab-imvr/FedDAH.