Bridging Offline and Online Reinforcement Learning for LLMs

We investigate the effectiveness of reinforcement learning methods for finetuning large language models when transitioning from offline to semi-online to fully online regimes for both verifiable and non-verifiable tasks. Our experiments cover training on verifiable math as well as non-verifiable instruction following with a set of benchmark evaluations for both. Across these settings, we extensively compare online and semi-online Direct Preference Optimization and Group Reward Policy Optimization objectives, and surprisingly find similar performance and convergence between these variants, which all strongly outperform offline methods. We provide a detailed analysis of the training dynamics and hyperparameter selection strategies to achieve optimal results. Finally, we show that multi-tasking with verifiable and non-verifiable rewards jointly yields improved performance across both task types.

PersonaAgent: When Large Language Model Agents Meet Personalization at Test Time

Large Language Model (LLM) empowered agents have recently emerged as advanced paradigms that exhibit impressive capabilities in a wide range of domains and tasks. Despite their potential, current LLM agents often adopt a one-size-fits-all approach, lacking the flexibility to respond to users' varying needs and preferences. This limitation motivates us to develop PersonaAgent, the first personalized LLM agent framework designed to address versatile personalization tasks. Specifically, PersonaAgent integrates two complementary components - a personalized memory module that includes episodic and semantic memory mechanisms; a personalized action module that enables the agent to perform tool actions tailored to the user. At the core, the persona (defined as unique system prompt for each user) functions as an intermediary: it leverages insights from personalized memory to control agent actions, while the outcomes of these actions in turn refine the memory. Based on the framework, we propose a test-time user-preference alignment strategy that simulate the latest n interactions to optimize the persona prompt, ensuring real-time user preference alignment through textual loss feedback between simulated and ground-truth responses. Experimental evaluations demonstrate that PersonaAgent significantly outperforms other baseline methods by not only personalizing the action space effectively but also scaling during test-time real-world applications. These results underscore the feasibility and potential of our approach in delivering tailored, dynamic user experiences.

Recycling the Web: A Method to Enhance Pre-training Data Quality and Quantity for Language Models

Scaling laws predict that the performance of large language models improves with increasing model size and data size. In practice, pre-training has been relying on massive web crawls, using almost all data sources publicly available on the internet so far. However, this pool of natural data does not grow at the same rate as the compute supply. Furthermore, the availability of high-quality texts is even more limited: data filtering pipelines often remove up to 99% of the initial web scrapes to achieve state-of-the-art. To address the "data wall" of pre-training scaling, our work explores ways to transform and recycle data discarded in existing filtering processes. We propose REWIRE, REcycling the Web with guIded REwrite, a method to enrich low-quality documents so that they could become useful for training. This in turn allows us to increase the representation of synthetic data in the final pre-training set. Experiments at 1B, 3B and 7B scales of the DCLM benchmark show that mixing high-quality raw texts and our rewritten texts lead to 1.0, 1.3 and 2.5 percentage points improvement respectively across 22 diverse tasks, compared to training on only filtered web data. Training on the raw-synthetic data mix is also more effective than having access to 2x web data. Through further analysis, we demonstrate that about 82% of the mixed in texts come from transforming lower-quality documents that would otherwise be discarded. REWIRE also outperforms related approaches of generating synthetic data, including Wikipedia-style paraphrasing, question-answer synthesizing and knowledge extraction. These results suggest that recycling web texts holds the potential for being a simple and effective approach for scaling pre-training data.

DetailFlow: 1D Coarse-to-Fine Autoregressive Image Generation via Next-Detail Prediction

This paper presents DetailFlow, a coarse-to-fine 1D autoregressive (AR) image generation method that models images through a novel next-detail prediction strategy. By learning a resolution-aware token sequence supervised with progressively degraded images, DetailFlow enables the generation process to start from the global structure and incrementally refine details. This coarse-to-fine 1D token sequence aligns well with the autoregressive inference mechanism, providing a more natural and efficient way for the AR model to generate complex visual content. Our compact 1D AR model achieves high-quality image synthesis with significantly fewer tokens than previous approaches, i.e. VAR/VQGAN. We further propose a parallel inference mechanism with self-correction that accelerates generation speed by approximately 8x while reducing accumulation sampling error inherent in teacher-forcing supervision. On the ImageNet 256x256 benchmark, our method achieves 2.96 gFID with 128 tokens, outperforming VAR (3.3 FID) and FlexVAR (3.05 FID), which both require 680 tokens in their AR models. Moreover, due to the significantly reduced token count and parallel inference mechanism, our method runs nearly 2x faster inference speed compared to VAR and FlexVAR. Extensive experimental results demonstrate DetailFlow's superior generation quality and efficiency compared to existing state-of-the-art methods.

J1: Incentivizing Thinking in LLM-as-a-Judge via Reinforcement Learning

The progress of AI is bottlenecked by the quality of evaluation, and powerful LLM-as-a-Judge models have proved to be a core solution. Improved judgment ability is enabled by stronger chain-of-thought reasoning, motivating the need to find the best recipes for training such models to think. In this work we introduce J1, a reinforcement learning approach to training such models. Our method converts both verifiable and non-verifiable prompts to judgment tasks with verifiable rewards that incentivize thinking and mitigate judgment bias. In particular, our approach outperforms all other existing 8B or 70B models when trained at those sizes, including models distilled from DeepSeek-R1. J1 also outperforms o1-mini, and even R1 on some benchmarks, despite training a smaller model. We provide analysis and ablations comparing Pairwise-J1 vs Pointwise-J1 models, offline vs online training recipes, reward strategies, seed prompts, and variations in thought length and content. We find that our models make better judgments by learning to outline evaluation criteria, comparing against self-generated reference answers, and re-evaluating the correctness of model responses.

Scalable Multi-task Edge Sensing via Task-oriented Joint Information Gathering and Broadcast

The recent advance of edge computing technology enables significant sensing performance improvement of Internet of Things (IoT) networks. In particular, an edge server (ES) is responsible for gathering sensing data from distributed sensing devices, and immediately executing different sensing tasks to accommodate the heterogeneous service demands of mobile users. However, as the number of users surges and the sensing tasks become increasingly compute-intensive, the huge amount of computation workloads and data transmissions may overwhelm the edge system of limited resources. Accordingly, we propose in this paper a scalable edge sensing framework for multi-task execution, in the sense that the computation workload and communication overhead of the ES do not increase with the number of downstream users or tasks. By exploiting the task-relevant correlations, the proposed scheme implements a unified encoder at the ES, which produces a common low-dimensional message from the sensing data and broadcasts it to all users to execute their individual tasks. To achieve high sensing accuracy, we extend the well-known information bottleneck theory to a multi-task scenario to jointly optimize the information gathering and broadcast processes. We also develop an efficient two-step training procedure to optimize the parameters of the neural network-based codecs deployed in the edge sensing system. Experiment results show that the proposed scheme significantly outperforms the considered representative benchmark methods in multi-task inference accuracy. Besides, the proposed scheme is scalable to the network size, which maintains almost constant computation delay with less than 1% degradation of inference performance when the user number increases by four times.

Transferable Deployment of Semantic Edge Inference Systems via Unsupervised Domain Adaption

This paper investigates deploying semantic edge inference systems for performing a common image clarification task. In particular, each system consists of multiple Internet of Things (IoT) devices that first locally encode the sensing data into semantic features and then transmit them to an edge server for subsequent data fusion and task inference. The inference accuracy is determined by efficient training of the feature encoder/decoder using labeled data samples. Due to the difference in sensing data and communication channel distributions, deploying the system in a new environment may induce high costs in annotating data labels and re-training the encoder/decoder models. To achieve cost-effective transferable system deployment, we propose an efficient Domain Adaptation method for Semantic Edge INference systems (DASEIN) that can maintain high inference accuracy in a new environment without the need for labeled samples. Specifically, DASEIN exploits the task-relevant data correlation between different deployment scenarios by leveraging the techniques of unsupervised domain adaptation and knowledge distillation. It devises an efficient two-step adaptation procedure that sequentially aligns the data distributions and adapts to the channel variations. Numerical results show that, under a substantial change in sensing data distributions, the proposed DASEIN outperforms the best-performing benchmark method by 7.09% and 21.33% in inference accuracy when the new environment has similar or 25 dB lower channel signal to noise power ratios (SNRs), respectively. This verifies the effectiveness of the proposed method in adapting both data and channel distributions in practical transfer deployment applications.

Pillar-Voxel Fusion Network for 3D Object Detection in Airborne Hyperspectral Point Clouds

Hyperspectral point clouds (HPCs) can simultaneously characterize 3D spatial and spectral information of ground objects, offering excellent 3D perception and target recognition capabilities. Current approaches for generating HPCs often involve fusion techniques with hyperspectral images and LiDAR point clouds, which inevitably lead to geometric-spectral distortions due to fusion errors and obstacle occlusions. These adverse effects limit their performance in downstream fine-grained tasks across multiple scenarios, particularly in airborne applications. To address these issues, we propose PiV-AHPC, a 3D object detection network for airborne HPCs. To the best of our knowledge, this is the first attempt at this HPCs task. Specifically, we first develop a pillar-voxel dual-branch encoder, where the former captures spectral and vertical structural features from HPCs to overcome spectral distortion, while the latter emphasizes extracting accurate 3D spatial features from point clouds. A multi-level feature fusion mechanism is devised to enhance information interaction between the two branches, achieving neighborhood feature alignment and channel-adaptive selection, thereby organically integrating heterogeneous features and mitigating geometric distortion. Extensive experiments on two airborne HPCs datasets demonstrate that PiV-AHPC possesses state-of-the-art detection performance and high generalization capability.

Forensics-Bench: A Comprehensive Forgery Detection Benchmark Suite for Large Vision Language Models

Recently, the rapid development of AIGC has significantly boosted the diversities of fake media spread in the Internet, posing unprecedented threats to social security, politics, law, and etc. To detect the ever-increasingly diverse malicious fake media in the new era of AIGC, recent studies have proposed to exploit Large Vision Language Models (LVLMs) to design robust forgery detectors due to their impressive performance on a wide range of multimodal tasks. However, it still lacks a comprehensive benchmark designed to comprehensively assess LVLMs' discerning capabilities on forgery media. To fill this gap, we present Forensics-Bench, a new forgery detection evaluation benchmark suite to assess LVLMs across massive forgery detection tasks, requiring comprehensive recognition, location and reasoning capabilities on diverse forgeries. Forensics-Bench comprises 63,292 meticulously curated multi-choice visual questions, covering 112 unique forgery detection types from 5 perspectives: forgery semantics, forgery modalities, forgery tasks, forgery types and forgery models. We conduct thorough evaluations on 22 open-sourced LVLMs and 3 proprietary models GPT-4o, Gemini 1.5 Pro, and Claude 3.5 Sonnet, highlighting the significant challenges of comprehensive forgery detection posed by Forensics-Bench. We anticipate that Forensics-Bench will motivate the community to advance the frontier of LVLMs, striving for all-around forgery detectors in the era of AIGC. The deliverables will be updated at https://Forensics-Bench.github.io/.

SWEET-RL: Training Multi-Turn LLM Agents on Collaborative Reasoning Tasks

Large language model (LLM) agents need to perform multi-turn interactions in real-world tasks. However, existing multi-turn RL algorithms for optimizing LLM agents fail to perform effective credit assignment over multiple turns while leveraging the generalization capabilities of LLMs and it remains unclear how to develop such algorithms. To study this, we first introduce a new benchmark, ColBench, where an LLM agent interacts with a human collaborator over multiple turns to solve realistic tasks in backend programming and frontend design. Building on this benchmark, we propose a novel RL algorithm, SWEET-RL (RL with Step-WisE Evaluation from Training-time information), that uses a carefully designed optimization objective to train a critic model with access to additional training-time information. The critic provides step-level rewards for improving the policy model. Our experiments demonstrate that SWEET-RL achieves a 6% absolute improvement in success and win rates on ColBench compared to other state-of-the-art multi-turn RL algorithms, enabling Llama-3.1-8B to match or exceed the performance of GPT4-o in realistic collaborative content creation.