Shizhe Diao

Yong Lin, Hangyu Lin, Wei Xiong et al.

LLMs acquire a wide range of abilities during pre-training, but aligning LLMs under Reinforcement Learning with Human Feedback (RLHF) can lead to forgetting pretrained abilities, which is also known as the alignment tax. To investigate alignment tax, we conducted experiments with existing RLHF algorithms using OpenLLaMA-3B, which revealed a pronounced alignment tax in NLP tasks. Whereas, despite various techniques to mitigate forgetting, they are often at odds with the RLHF performance, leading to a trade-off between alignment performance and forgetting mitigation, leading to an alignment-forgetting trade-off. In this paper we show that model averaging, which simply interpolates between pre and post RLHF model weights, surprisingly achieves the most strongest alignment-forgetting Pareto front among a wide range of competing methods. To understand its effectiveness, we offer theoretical insights into model averaging, revealing that it enhances performance Pareto front by increasing feature diversity on the layers where tasks share overlapped feature spaces. Empirical evidence corroborates our analysis by showing the benefits of averaging low-level transformer layers. Building on the analysis and the observation that averaging different layers of the transformer leads to significantly different alignment-forgetting trade-offs, we propose Heterogeneous Model Averaging (HMA) to Heterogeneously find various combination ratios of model layers. HMA seeks to maximize the alignment performance while incurring minimal alignment tax. Moreover, we validate HMA's performance across a range of RLHF algorithms over OpenLLaMA-3B and further extend our findings to Mistral-7B which is evaluated by open-sourced preference model and GPT4. Code available here: https://github.com/avalonstrel/Mitigating-the-Alignment-Tax-of-RLHF.git.

Shizhe Diao, Sedrick Scott Keh, Liangming Pan et al. · cmu, pku

Social media classification tasks (e.g., tweet sentiment analysis, tweet stance detection) are challenging because social media posts are typically short, informal, and ambiguous. Thus, training on tweets is challenging and demands large-scale human-annotated labels, which are time-consuming and costly to obtain. In this paper, we find that providing hashtags to social media tweets can help alleviate this issue because hashtags can enrich short and ambiguous tweets in terms of various information, such as topic, sentiment, and stance. This motivates us to propose a novel Hashtag-guided Tweet Classification model (HashTation), which automatically generates meaningful hashtags for the input tweet to provide useful auxiliary signals for tweet classification. To generate high-quality and insightful hashtags, our hashtag generation model retrieves and encodes the post-level and entity-level information across the whole corpus. Experiments show that HashTation achieves significant improvements on seven low-resource tweet classification tasks, in which only a limited amount of training data is provided, showing that automatically enriching tweets with model-generated hashtags could significantly reduce the demand for large-scale human-labeled data. Further analysis demonstrates that HashTation is able to generate high-quality hashtags that are consistent with the tweets and their labels. The code is available at https://github.com/shizhediao/HashTation.

Shizhe Diao, Rui Pan, Hanze Dong et al.

Foundation models have demonstrated a great ability to achieve general human-level intelligence far beyond traditional approaches. As the technique keeps attracting attention from the AI community, an increasing number of foundation models are becoming publicly accessible. However, a significant shortcoming of most of these models lies in their performance in specialized-domain and task-specific applications, necessitating domain- and task-aware fine-tuning to develop effective scientific language models. As the number of available foundation models and specialized tasks keeps growing, the job of training scientific language models becomes highly nontrivial. In this paper, we initiate steps to tackle this issue. We introduce an extensible and lightweight toolkit, LMFlow, which aims to simplify the domain- and task-aware finetuning of general foundation models. LMFlow offers a complete finetuning workflow for a foundation model to support specialized training with limited computing resources. Furthermore, it supports continuous pretraining, instruction tuning, parameter-efficient finetuning, alignment tuning, inference acceleration, long context generalization, model customization, and even multimodal finetuning, along with carefully designed and extensible APIs. This toolkit has been thoroughly tested and is available at https://github.com/OptimalScale/LMFlow.

Shizhe Diao, Tianyang Xu, Ruijia Xu et al.

Pre-trained language models (PLMs) demonstrate excellent abilities to understand texts in the generic domain while struggling in a specific domain. Although continued pre-training on a large domain-specific corpus is effective, it is costly to tune all the parameters on the domain. In this paper, we investigate whether we can adapt PLMs both effectively and efficiently by only tuning a few parameters. Specifically, we decouple the feed-forward networks (FFNs) of the Transformer architecture into two parts: the original pre-trained FFNs to maintain the old-domain knowledge and our novel domain-specific adapters to inject domain-specific knowledge in parallel. Then we adopt a mixture-of-adapters gate to fuse the knowledge from different domain adapters dynamically. Our proposed Mixture-of-Domain-Adapters (MixDA) employs a two-stage adapter-tuning strategy that leverages both unlabeled data and labeled data to help the domain adaptation: i) domain-specific adapter on unlabeled data; followed by ii) the task-specific adapter on labeled data. MixDA can be seamlessly plugged into the pretraining-finetuning paradigm and our experiments demonstrate that MixDA achieves superior performance on in-domain tasks (GLUE), out-of-domain tasks (ChemProt, RCT, IMDB, Amazon), and knowledge-intensive tasks (KILT). Further analyses demonstrate the reliability, scalability, and efficiency of our method. The code is available at https://github.com/Amano-Aki/Mixture-of-Domain-Adapters.

Shizhe Diao, Pengcheng Wang, Yong Lin et al.

The increasing scale of large language models (LLMs) brings emergent abilities to various complex tasks requiring reasoning, such as arithmetic and commonsense reasoning. It is known that the effective design of task-specific prompts is critical for LLMs' ability to produce high-quality answers. In particular, an effective approach for complex question-and-answer tasks is example-based prompting with chain-of-thought (CoT) reasoning, which significantly improves the performance of LLMs. However, current CoT methods rely on a fixed set of human-annotated exemplars, which are not necessarily the most effective examples for different tasks. This paper proposes a new method, Active-Prompt, to adapt LLMs to different tasks with task-specific example prompts (annotated with human-designed CoT reasoning). For this purpose, we propose a solution to the key problem of determining which questions are the most important and helpful ones to annotate from a pool of task-specific queries. By borrowing ideas from the related problem of uncertainty-based active learning, we introduce several metrics to characterize the uncertainty so as to select the most uncertain questions for annotation. Experimental results demonstrate the superiority of our proposed method, achieving state-of-the-art on eight complex reasoning tasks. Further analyses of different uncertainty metrics, pool sizes, zero-shot learning, and accuracy-uncertainty relationship demonstrate the effectiveness of our method. Our code will be available at https://github.com/shizhediao/active-prompt.

KaShun Shum, Shizhe Diao, Tong Zhang

Chain-of-thought (CoT) advances the reasoning abilities of large language models (LLMs) and achieves superior performance in complex reasoning tasks. However, most CoT studies rely on carefully designed human-annotated rational chains to prompt LLMs, posing challenges for real-world applications where labeled data is available without rational chains. This paper proposes a new strategy, Automate-CoT (Automatic Prompt Augmentation and Selection with Chain-of-Thought), that can bypass human engineering of CoT by automatically augmenting rational chains from a small labeled dataset, and then pruning low-quality chains to construct a candidate pool of machine-generated rationale chains based on the labels. Finally, it selects the optimal combination of several rationale chains from the pool for CoT prompting by employing a variance-reduced policy gradient strategy to estimate the significance of each example. Automate-CoT enables a quick adaptation of the CoT technique to different tasks. Experimental results demonstrate the effectiveness of our method, where competitive results are achieved on arithmetic reasoning (+2.7%), commonsense reasoning (+3.4%), symbolic reasoning (+3.2%), and non-reasoning tasks (+2.5%). The code is available at https://github.com/SHUMKASHUN/Automate-CoT.

Rui Pan, Shizhe Diao, Jianlin Chen et al.

In this paper, we present ExtremeBERT, a toolkit for accelerating and customizing BERT pretraining. Our goal is to provide an easy-to-use BERT pretraining toolkit for the research community and industry. Thus, the pretraining of popular language models on customized datasets is affordable with limited resources. Experiments show that, to achieve the same or better GLUE scores, the time cost of our toolkit is over $6\times$ times less for BERT Base and $9\times$ times less for BERT Large when compared with the original BERT paper. The documentation and code are released at https://github.com/extreme-bert/extreme-bert under the Apache-2.0 license.

Hanning Zhang, Shizhe Diao, Yong Lin et al.

Large language models (LLMs) have revolutionized numerous domains with their impressive performance but still face their challenges. A predominant issue is the propensity for these models to generate non-existent facts, a concern termed hallucination. Our research is motivated by the observation that previous instruction tuning methods force the model to complete a sentence no matter whether the model knows the knowledge or not. When the question is out of the parametric knowledge, it will try to make up something and fail to indicate when it lacks knowledge. In this paper, we present a new approach called Refusal-Aware Instruction Tuning (R-Tuning). This approach is formalized by first identifying the disparity in knowledge encompassed by pre-trained parameters compared to that of instruction tuning data. Then, we construct the refusal-aware data based on the knowledge intersection, to tune LLMs to refrain from responding to questions beyond its parametric knowledge. Experimental results demonstrate R-Tuning effectively improves a model's ability to answer known questions and refrain from answering unknown questions. Furthermore, when tested on out-of-domain datasets, the refusal ability was found to be a meta-skill that could be generalized to other tasks. Further analysis surprisingly finds that learning the uncertainty results in better calibration and an improved ability to estimate the uncertainty than uncertainty-based testing. Our code is available at https://github.com/shizhediao/R-Tuning.

Hanze Dong, Wei Xiong, Deepanshu Goyal et al.

Generative foundation models are susceptible to implicit biases that can arise from extensive unsupervised training data. Such biases can produce suboptimal samples, skewed outcomes, and unfairness, with potentially serious consequences. Consequently, aligning these models with human ethics and preferences is an essential step toward ensuring their responsible and effective deployment in real-world applications. Prior research has primarily employed Reinforcement Learning from Human Feedback (RLHF) to address this problem, where generative models are fine-tuned with RL algorithms guided by a human-feedback-informed reward model. However, the inefficiencies and instabilities associated with RL algorithms frequently present substantial obstacles to the successful alignment, necessitating the development of a more robust and streamlined approach. To this end, we introduce a new framework, Reward rAnked FineTuning (RAFT), designed to align generative models effectively. Utilizing a reward model and a sufficient number of samples, our approach selects the high-quality samples, discarding those that exhibit undesired behavior, and subsequently enhancing the model by fine-tuning on these filtered samples. Our studies show that RAFT can effectively improve the model performance in both reward learning and other automated metrics in both large language models and diffusion models.

Hao Zhang, Mingjie Liu, Shaokun Zhang et al.

Multi-turn LLM agents are increasingly important for solving complex, interactive tasks, and reinforcement learning (RL) is a key ingredient for improving their long-horizon behavior. However, RL training requires generating large numbers of sandboxed rollout trajectories, and existing infrastructures often couple rollout orchestration with the training loop, making systems hard to migrate and maintain. Under the rollout-as-a-service philosophy, we present ProRL Agent , a scalable infrastructure that serves the full agentic rollout lifecycle through an API service. ProRL Agent also provides standardized and extensible sandbox environments that support diverse agentic tasks in rootless HPC settings. We validate ProRL Agent through RL training on software engineering, math, STEM, and coding tasks. ProRL Agent is open-sourced and integrated as part of NVIDIA NeMo Gym.

Shizhe Diao, Wangchunshu Zhou, Xinsong Zhang et al.

Recent advances in vision-language pre-training have pushed the state-of-the-art on various vision-language tasks, making machines more capable of multi-modal writing (image-to-text generation) and painting (text-to-image generation). However, few studies investigate if these two essential capabilities can be learned together and boost each other, making a versatile and powerful multi-modal foundation model. In this work, we disclose the potential of symmetric generative vision-language pre-training in learning to write and paint concurrently, and propose a new unified modal model, named DaVinci, trained with prefix language modeling and prefix image modeling, a simple generative self-supervised objective on image-text pairs. Thanks to the proposed prefix multi-modal modeling framework, DaVinci is simple to train, scalable to huge data, adaptable to both writing and painting tasks, and also strong on other vision, text, and multi-modal understanding tasks. DaVinci achieves competitive performance on a wide range of 27 generation/understanding tasks and demonstrates the superiority of combining vision/language generative pre-training. Furthermore, we carefully benchmark the performance of different vision-language pre-training objectives on different scales of pre-training datasets on a heterogeneous and broad distribution coverage. Our results demonstrate the potential of exploiting self-supervision in both language and vision inputs, and establish new, stronger baselines for future comparisons at different data scales. The code and pre-trained models are available at https://github.com/shizhediao/DaVinci.

Rui Pan, Shuo Xing, Shizhe Diao et al.

Since the emergence of large language models, prompt learning has become a popular method for optimizing and customizing these models. Special prompts, such as Chain-of-Thought, have even revealed previously unknown reasoning capabilities within these models. However, the progress of discovering effective prompts has been slow, driving a desire for general prompt optimization methods. Unfortunately, few existing prompt learning methods satisfy the criteria of being truly "general", i.e., automatic, discrete, black-box, gradient-free, and interpretable all at once. In this paper, we introduce metaheuristics, a branch of discrete non-convex optimization methods with over 100 options, as a promising approach to prompt learning. Within our paradigm, we test six typical methods: hill climbing, simulated annealing, genetic algorithms with/without crossover, tabu search, and harmony search, demonstrating their effectiveness in white-box and black-box prompt learning. Furthermore, we show that these methods can be used to discover more human-understandable prompts that were previously unknown in both reasoning and image generation tasks, opening the door to a cornucopia of possibilities in prompt optimization. We release all the codes in \url{https://github.com/research4pan/Plum}.

Xu Liu, Junfeng Hu, Yuan Li et al.

Multivariate time series forecasting plays a pivotal role in contemporary web technologies. In contrast to conventional methods that involve creating dedicated models for specific time series application domains, this research advocates for a unified model paradigm that transcends domain boundaries. However, learning an effective cross-domain model presents the following challenges. First, various domains exhibit disparities in data characteristics, e.g., the number of variables, posing hurdles for existing models that impose inflexible constraints on these factors. Second, the model may encounter difficulties in distinguishing data from various domains, leading to suboptimal performance in our assessments. Third, the diverse convergence rates of time series domains can also result in compromised empirical performance. To address these issues, we propose UniTime for effective cross-domain time series learning. Concretely, UniTime can flexibly adapt to data with varying characteristics. It also uses domain instructions and a Language-TS Transformer to offer identification information and align two modalities. In addition, UniTime employs masking to alleviate domain convergence speed imbalance issues. Our extensive experiments demonstrate the effectiveness of UniTime in advancing state-of-the-art forecasting performance and zero-shot transferability.

Zhihong Chen, Shizhe Diao, Benyou Wang et al.

Medical vision-and-language pre-training (Med-VLP) has shown promising improvements on many downstream medical tasks owing to its applicability to extracting generic representations from medical images and texts. Practically, there exist two typical types, \textit{i.e.}, the fusion-encoder type and the dual-encoder type, depending on whether a heavy fusion module is used. The former is superior at multi-modal tasks owing to the sufficient interaction between modalities; the latter is good at uni-modal and cross-modal tasks due to the single-modality encoding ability. To take advantage of these two types, we propose an effective yet straightforward scheme named PTUnifier to unify the two types. We first unify the input format by introducing visual and textual prompts, which serve as a feature bank that stores the most representative images/texts. By doing so, a single model could serve as a \textit{foundation model} that processes various tasks adopting different input formats (\textit{i.e.}, image-only, text-only, and image-text-pair). Furthermore, we construct a prompt pool (instead of static ones) to improve diversity and scalability. Experimental results show that our approach achieves state-of-the-art results on a broad range of tasks, spanning uni-modal tasks (\textit{i.e.}, image/text classification and text summarization), cross-modal tasks (\textit{i.e.}, image-to-text generation and image-text/text-image retrieval), and multi-modal tasks (\textit{i.e.}, visual question answering), demonstrating the effectiveness of our approach. Note that the adoption of prompts is orthogonal to most existing Med-VLP approaches and could be a beneficial and complementary extension to these approaches.

Ruida Wang, Jipeng Zhang, Yizhen Jia et al.

Proving mathematical theorems using computer-verifiable formal languages like Lean significantly impacts mathematical reasoning. One approach to formal theorem proving involves generating complete proofs using Large Language Models (LLMs) based on Natural Language (NL) proofs. However, due to the scarcity of aligned NL and Formal Language (FL) theorem-proving data most modern LLMs exhibit suboptimal performance.This scarcity results in a paucity of methodologies for training LLMs and techniques to fully utilize their capabilities in composing formal proofs. To address these challenges, this paper proposes TheoremLlama, an end-to-end framework that trains a general-purpose LLM to be a Lean4 expert. TheoremLlama includes NL-FL dataset generation and bootstrapping method to obtain aligned dataset, curriculum learning and block training techniques to train the model, and iterative proof writing method to write Lean4 proofs that work together synergistically. Using the dataset generation method in TheoremLlama, we provide Open Bootstrapped Theorems (OBT), an NL-FL aligned and bootstrapped dataset. Our novel NL-FL bootstrapping method, where NL proofs are integrated into Lean4 code for training datasets, leverages the NL reasoning ability of LLMs for formal reasoning. The TheoremLlama framework achieves cumulative accuracies of 36.48% and 33.61% on MiniF2F-Valid and Test datasets respectively, surpassing the GPT-4 baseline of 22.95% and 25.41%. Our code, model checkpoints, and the generated dataset is published in GitHub

Zhihong Chen, Guiming Hardy Chen, Shizhe Diao et al.

Masked language modeling (MLM) has been one of the most popular pretraining recipes in natural language processing, e.g., BERT, one of the representative models. Recently, contrastive language-image pretraining (CLIP) has also attracted attention, especially its vision models that achieve excellent performance on a broad range of vision tasks. However, few studies are dedicated to studying the text encoders learned by CLIP. In this paper, we analyze the difference between BERT-style and CLIP-style text encoders from three experiments: (i) general text understanding, (ii) vision-centric text understanding, and (iii) text-to-image generation. Experimental analyses show that although CLIP-style text encoders underperform BERT-style ones for general text understanding tasks, they are equipped with a unique ability, i.e., synesthesia, for the cross-modal association, which is more similar to the senses of humans.

Ziqiang Zheng, Jipeng Zhang, Tuan-Anh Vu et al.

Large language models (LLMs), such as ChatGPT/GPT-4, have proven to be powerful tools in promoting the user experience as an AI assistant. The continuous works are proposing multi-modal large language models (MLLM), empowering LLMs with the ability to sense multiple modality inputs through constructing a joint semantic space (e.g. visual-text space). Though significant success was achieved in LLMs and MLLMs, exploring LLMs and MLLMs in domain-specific applications that required domain-specific knowledge and expertise has been less conducted, especially for \textbf{marine domain}. Different from general-purpose MLLMs, the marine-specific MLLM is required to yield much more \textbf{sensitive}, \textbf{informative}, and \textbf{scientific} responses. In this work, we demonstrate that the existing MLLMs optimized on huge amounts of readily available general-purpose training data show a minimal ability to understand domain-specific intents and then generate informative and satisfactory responses. To address these issues, we propose \textbf{MarineGPT}, the first vision-language model specially designed for the marine domain, unlocking the secrets of the ocean to the public. We present our \textbf{Marine-5M} dataset with more than 5 million marine image-text pairs to inject domain-specific marine knowledge into our model and achieve better marine vision and language alignment. Our MarineGPT not only pushes the boundaries of marine understanding to the general public but also offers a standard protocol for adapting a general-purpose assistant to downstream domain-specific experts. We pave the way for a wide range of marine applications while setting valuable data and pre-trained models for future research in both academic and industrial communities.

Shih-Yang Liu, Xin Dong, Ximing Lu et al.

As language models become increasingly capable, users expect them to provide not only accurate responses but also behaviors aligned with diverse human preferences across a variety of scenarios. To achieve this, Reinforcement learning (RL) pipelines have begun incorporating multiple rewards, each capturing a distinct preference, to guide models toward these desired behaviors. However, recent work has defaulted to apply Group Relative Policy Optimization (GRPO) under multi-reward setting without examining its suitability. In this paper, we demonstrate that directly applying GRPO to normalize distinct rollout reward combinations causes them to collapse into identical advantage values, reducing the resolution of the training signal and resulting in suboptimal convergence and, in some cases, early training failure. We then introduce Group reward-Decoupled Normalization Policy Optimization (GDPO), a new policy optimization method to resolve these issues by decoupling the normalization of individual rewards, more faithfully preserving their relative differences and enabling more accurate multi-reward optimization, along with substantially improved training stability. We compare GDPO with GRPO across three tasks: tool calling, math reasoning, and coding reasoning, evaluating both correctness metrics (accuracy, bug ratio) and constraint adherence metrics (format, length). Across all settings, GDPO consistently outperforms GRPO, demonstrating its effectiveness and generalizability for multi-reward reinforcement learning optimization.

Hanze Dong, Shizhe Diao, Weizhong Zhang et al.

Normalizing flow (NF) has gained popularity over traditional maximum likelihood based methods due to its strong capability to model complex data distributions. However, the standard approach, which maps the observed data to a normal distribution, has difficulty in handling data distributions with multiple relatively isolated modes. To overcome this issue, we propose a new framework based on variational latent representation to improve the practical performance of NF. The idea is to replace the standard normal latent variable with a more general latent representation, jointly learned via Variational Bayes. For example, by taking the latent representation as a discrete sequence, our framework can learn a Transformer model that generates the latent sequence and an NF model that generates continuous data distribution conditioned on the sequence. The resulting method is significantly more powerful than the standard normalization flow approach for generating data distributions with multiple modes. Extensive experiments have shown the advantages of NF with variational latent representation.

Wangchunshu Zhou, Yan Zeng, Shizhe Diao et al.

Recent advances in vision-language pre-training (VLP) have demonstrated impressive performance in a range of vision-language (VL) tasks. However, there exist several challenges for measuring the community's progress in building general multi-modal intelligence. First, most of the downstream VL datasets are annotated using raw images that are already seen during pre-training, which may result in an overestimation of current VLP models' generalization ability. Second, recent VLP work mainly focuses on absolute performance but overlooks the efficiency-performance trade-off, which is also an important indicator for measuring progress. To this end, we introduce the Vision-Language Understanding Evaluation (VLUE) benchmark, a multi-task multi-dimension benchmark for evaluating the generalization capabilities and the efficiency-performance trade-off (``Pareto SOTA'') of VLP models. We demonstrate that there is a sizable generalization gap for all VLP models when testing on out-of-distribution test sets annotated on images from a more diverse distribution that spreads across cultures. Moreover, we find that measuring the efficiency-performance trade-off of VLP models leads to complementary insights for several design choices of VLP. We release the VLUE benchmark to promote research on building vision-language models that generalize well to more diverse images and concepts unseen during pre-training, and are practical in terms of efficiency-performance trade-off.

KaShun Shum, Minrui Xu, Jianshu Zhang et al.

Large language models (LLMs) have become increasingly prevalent in our daily lives, leading to an expectation for LLMs to be trustworthy -- - both accurate and well-calibrated (the prediction confidence should align with its ground truth correctness likelihood). Nowadays, fine-tuning has become the most popular method for adapting a model to practical usage by significantly increasing accuracy on downstream tasks. Despite the great accuracy it achieves, we found fine-tuning is still far away from satisfactory trustworthiness due to "tuning-induced mis-calibration". In this paper, we delve deeply into why and how mis-calibration exists in fine-tuned models, and how distillation can alleviate the issue. Then we further propose a brand new method named Efficient Trustworthy Distillation (FIRST), which utilizes a small portion of teacher's knowledge to obtain a reliable language model in a cost-efficient way. Specifically, we identify the "concentrated knowledge" phenomenon during distillation, which can significantly reduce the computational burden. Then we apply a "trustworthy maximization" process to optimize the utilization of this small portion of concentrated knowledge before transferring it to the student. Experimental results demonstrate the effectiveness of our method, where better accuracy (+2.3%) and less mis-calibration (-10%) are achieved on average across both in-domain and out-of-domain scenarios, indicating better trustworthiness.

Hongjin Su, Shizhe Diao, Ximing Lu et al.

Large language models are powerful generalists, yet solving deep and complex problems such as those of the Humanity's Last Exam (HLE) remains both conceptually challenging and computationally expensive. We show that small orchestrators managing other models and a variety of tools can both push the upper bound of intelligence and improve efficiency in solving difficult agentic tasks. We introduce ToolOrchestra, a method for training small orchestrators that coordinate intelligent tools. ToolOrchestra explicitly uses reinforcement learning with outcome-, efficiency-, and user-preference-aware rewards. Using ToolOrchestra, we produce Orchestrator, an 8B model that achieves higher accuracy at lower cost than previous tool-use agents while aligning with user preferences on which tools are to be used for a given query. On HLE, Orchestrator achieves a score of 37.1%, outperforming GPT-5 (35.1%) while being 2.5x more efficient. On tau2-Bench and FRAMES, Orchestrator surpasses GPT-5 by a wide margin while using only about 30% of the cost. Extensive analysis shows that Orchestrator achieves the best trade-off between performance and cost under multiple metrics, and generalizes robustly to unseen tools. These results demonstrate that composing diverse tools with a lightweight orchestration model is both more efficient and more effective than existing methods, paving the way for practical and scalable tool-augmented reasoning systems.

Mingjie Liu, Shizhe Diao, Ximing Lu et al. · uw

Recent advances in reasoning-centric language models have highlighted reinforcement learning (RL) as a promising method for aligning models with verifiable rewards. However, it remains contentious whether RL truly expands a model's reasoning capabilities or merely amplifies high-reward outputs already latent in the base model's distribution, and whether continually scaling up RL compute reliably leads to improved reasoning performance. In this work, we challenge prevailing assumptions by demonstrating that prolonged RL (ProRL) training can uncover novel reasoning strategies that are inaccessible to base models, even under extensive sampling. We introduce ProRL, a novel training methodology that incorporates KL divergence control, reference policy resetting, and a diverse suite of tasks. Our empirical analysis reveals that RL-trained models consistently outperform base models across a wide range of pass@k evaluations, including scenarios where base models fail entirely regardless of the number of attempts. We further show that reasoning boundary improvements correlates strongly with task competence of base model and training duration, suggesting that RL can explore and populate new regions of solution space over time. These findings offer new insights into the conditions under which RL meaningfully expands reasoning boundaries in language models and establish a foundation for future work on long-horizon RL for reasoning. We release model weights to support further research: https://huggingface.co/nvidia/Nemotron-Research-Reasoning-Qwen-1.5B

Qiaolong Cai, Zhaowei Wang, Shizhe Diao et al.

With the increasing popularity of large language models (LLMs), reasoning on basic graph algorithm problems is an essential intermediate step in assessing their abilities to process and infer complex graph reasoning tasks. Existing methods usually convert graph-structured data to textual descriptions and then use LLMs for reasoning and computation. However, LLMs often produce computation errors on arithmetic parts in basic graph algorithm problems, such as counting number of edges. In addition, they struggle to control or understand the output of the reasoning process, raising concerns about whether LLMs are simply guessing. In this paper, we introduce CodeGraph, a method that encodes graph problem solutions as code. The methods solve new graph problems by learning from exemplars, generating programs, and executing them via a program interpreter. Using the few-shot setting, we evaluate CodeGraph with the base LLM being GPT-3.5 Turbo, Llama3-70B Instruct, Mixtral-8x22B Instruct, and Mixtral-8x7B Instruct. Experimental results on six tasks with six graph encoding methods in the GraphQA dataset demonstrate that CodeGraph can boost performance on graph reasoning tasks inside LLMs by 1.3% to 58.6%, depending on the task. Compared to the existing methods, CodeGraph demonstrates strong performance on arithmetic problems in graph tasks and offers a more controllable and interpretable approach to the reasoning process.

Zhifan Ye, Kejing Xia, Yonggan Fu et al.

State space models (SSMs) have emerged as an efficient alternative to Transformer models for language modeling, offering linear computational complexity and constant memory usage as context length increases. However, despite their efficiency in handling long contexts, recent studies have shown that SSMs, such as Mamba models, generally underperform compared to Transformers in long-context understanding tasks. To address this significant shortfall and achieve both efficient and accurate long-context understanding, we propose LongMamba, a training-free technique that significantly enhances the long-context capabilities of Mamba models. LongMamba builds on our discovery that the hidden channels in Mamba can be categorized into local and global channels based on their receptive field lengths, with global channels primarily responsible for long-context capability. These global channels can become the key bottleneck as the input context lengthens. Specifically, when input lengths largely exceed the training sequence length, global channels exhibit limitations in adaptively extend their receptive fields, leading to Mamba's poor long-context performance. The key idea of LongMamba is to mitigate the hidden state memory decay in these global channels by preventing the accumulation of unimportant tokens in their memory. This is achieved by first identifying critical tokens in the global channels and then applying token filtering to accumulate only those critical tokens. Through extensive benchmarking across synthetic and real-world long-context scenarios, LongMamba sets a new standard for Mamba's long-context performance, significantly extending its operational range without requiring additional training. Our code is available at https://github.com/GATECH-EIC/LongMamba.

Xin Xu, Qiyun Xu, Tong Xiao et al.

Large language models (LLMs) have demonstrated remarkable capabilities in solving complex reasoning tasks, particularly in mathematics. However, the domain of physics reasoning presents unique challenges that have received significantly less attention. Existing benchmarks often fall short in evaluating LLMs' abilities on the breadth and depth of undergraduate-level physics, underscoring the need for a comprehensive evaluation. To fill this gap, we introduce UGPhysics, a large-scale and comprehensive benchmark specifically designed to evaluate UnderGraduate-level Physics (UGPhysics) reasoning with LLMs. UGPhysics includes 5,520 undergraduate-level physics problems in both English and Chinese, covering 13 subjects with seven different answer types and four distinct physics reasoning skills, all rigorously screened for data leakage. Additionally, we develop a Model-Assistant Rule-based Judgment (MARJ) pipeline specifically tailored for assessing answer correctness of physics problems, ensuring accurate evaluation. Our evaluation of 31 leading LLMs shows that the highest overall accuracy, 49.8% (achieved by OpenAI-o1-mini), emphasizes the necessity for models with stronger physics reasoning skills, beyond math abilities. We hope UGPhysics, along with MARJ, will drive future advancements in AI for physics reasoning. Codes and data are available at https://github.com/YangLabHKUST/UGPhysics .

Rui Pan, Shivanshu Shekhar, Boyao Wang et al.

Small language models (SLMs) have attracted considerable attention from both academia and industry due to their broad range of applications in edge devices. To obtain SLMs with strong performance, conventional approaches either pre-train the models from scratch, which incurs substantial computational costs, or compress/prune existing large language models (LLMs), which results in performance drops and falls short in comparison to pre-training. In this paper, we investigate the family of acceleration methods that involve both structured pruning and model training. We found 1) layer-wise adaptive pruning (Adapt-Pruner) is extremely effective in LLMs and yields significant improvements over existing pruning techniques, 2) adaptive pruning equipped with further training leads to models comparable to those pre-training from scratch, 3) incremental pruning brings non-trivial performance gain by interleaving pruning with training and only removing a small portion of neurons ($\sim$5%) at a time. Experimental results on LLaMA-3.1-8B demonstrate that Adapt-Pruner outperforms conventional pruning methods, such as LLM-Pruner, FLAP, and SliceGPT, by an average of 1%-7% in accuracy on commonsense benchmarks. Additionally, Adapt-Pruner restores the performance of MobileLLM-125M to 600M on the MMLU benchmark with 200$\times$ fewer tokens via pruning from its larger counterparts, and discovers a new 1B model that surpasses LLaMA-3.2-1B in multiple benchmarks. The official code is released at https://github.com/research4pan/AdaptPruner.

Tianyang Han, Qing Lian, Rui Pan et al.

Large language models (LLMs) have recently experienced remarkable progress, where the advent of multi-modal large language models (MLLMs) has endowed LLMs with visual capabilities, leading to impressive performances in various multi-modal tasks. However, those powerful MLLMs such as GPT-4V still fail spectacularly when presented with certain image and text inputs. In this paper, we identify a typical class of inputs that baffles MLLMs, which consist of images that are highly relevant but inconsistent with answers, causing MLLMs to suffer from visual illusion. To quantify the effect, we propose CorrelationQA, the first benchmark that assesses the visual illusion level given spurious images. This benchmark contains 7,308 text-image pairs across 13 categories. Based on the proposed CorrelationQA, we conduct a thorough analysis on 9 mainstream MLLMs, illustrating that they universally suffer from this instinctive bias to varying degrees. We hope that our curated benchmark and evaluation results aid in better assessments of the MLLMs' robustness in the presence of misleading images. The code and datasets are available at https://github.com/MasaiahHan/CorrelationQA.

Tianhao Chen, Xin Xu, Zijing Liu et al.

Modern Large Language Models, such as the LLaMA, Qwen and DeepSeek series, predominantly adopt the Pre-LayerNorm (Pre-LN) Transformer architecture. While being stable during pretraining and scalable to large model sizes, Pre-LN suffers from an exponential growth in activation variance across layers, causing the shortcut to dominate over sub-layer outputs in the residual connection and limiting the learning capacity of deeper layers. To mitigate this issue, we propose Gradient-Preserving Activation Scaling (GPAS), a simple technique that can be used in combination with existing approaches. GPAS works by scaling down the intermediate activations while keeping their gradients unchanged. This leaves information in the activations intact, and avoids the gradient vanishing problem associated with gradient downscaling. Extensive experiments across various model sizes from 71M to 1B show that GPAS achieves consistent performance gains. Beyond enhancing Pre-LN Transformers, GPAS also shows promise in improving alternative architectures such as Sandwich-LN and DeepNorm, demonstrating its versatility and potential for improving training dynamics in a wide range of settings. Our code is available at https://github.com/dandingsky/GPAS.

Haoxiang Wang, Yong Lin, Wei Xiong et al.

Fine-grained control over large language models (LLMs) remains a significant challenge, hindering their adaptability to diverse user needs. While Reinforcement Learning from Human Feedback (RLHF) shows promise in aligning LLMs, its reliance on scalar rewards often limits its ability to capture diverse user preferences in real-world applications. To address this limitation, we introduce the Directional Preference Alignment (DPA) framework. Unlike the scalar-reward RLHF, DPA incorporates multi-objective reward modeling to represent diverse preference profiles. Additionally, DPA models user preferences as directions (i.e., unit vectors) in the reward space to achieve user-dependent preference control. Our method involves training a multi-objective reward model and then fine-tuning the LLM with a preference-conditioned variant of Rejection Sampling Finetuning (RSF), an RLHF method adopted by Llama 2. This method enjoys a better performance trade-off across various reward objectives. In comparison with the scalar-reward RLHF, DPA offers users intuitive control over LLM generation: they can arithmetically specify their desired trade-offs (e.g., more helpfulness with less verbosity). We also validate the effectiveness of DPA with real-world alignment experiments on Mistral-7B. Our method provides straightforward arithmetic control over the trade-off between helpfulness and verbosity while maintaining competitive performance with strong baselines such as Direct Preference Optimization (DPO).

Zhilin Wang, Jaehun Jung, Ximing Lu et al. · uw

Evaluating progress in large language models (LLMs) is often constrained by the challenge of verifying responses, limiting assessments to tasks like mathematics, programming, and short-form question-answering. However, many real-world applications require evaluating LLMs in processing professional documents, synthesizing information, and generating comprehensive reports in response to user queries. We introduce ProfBench: a set of over 7000 response-criterion pairs as evaluated by human-experts with professional knowledge across Physics PhD, Chemistry PhD, Finance MBA and Consulting MBA. We build robust and affordable LLM-Judges to evaluate ProfBench rubrics, by mitigating self-enhancement bias and reducing the cost of evaluation by 2-3 orders of magnitude, to make it fair and accessible to the broader community. Our findings reveal that ProfBench poses significant challenges even for state-of-the-art LLMs, with top-performing models like GPT-5-high achieving only 65.9\% overall performance. Furthermore, we identify notable performance disparities between proprietary and open-weight models and provide insights into the role that extended thinking plays in addressing complex, professional-domain tasks. Data: https://huggingface.co/datasets/nvidia/ProfBench and Code: https://github.com/NVlabs/ProfBench

Quyet V. Do, Tianqing Fang, Shizhe Diao et al.

Reasoning over Commonsense Knowledge Bases (CSKB), i.e. CSKB reasoning, has been explored as a way to acquire new commonsense knowledge based on reference knowledge in the original CSKBs and external prior knowledge. Despite the advancement of Large Language Models (LLM) and prompt engineering techniques in various reasoning tasks, they still struggle to deal with CSKB reasoning. One of the problems is that it is hard for them to acquire explicit relational constraints in CSKBs from only in-context exemplars, due to a lack of symbolic reasoning capabilities (Bengio et al., 2021). To this end, we proposed **ConstraintChecker**, a plugin over prompting techniques to provide and check explicit constraints. When considering a new knowledge instance, ConstraintChecker employs a rule-based module to produce a list of constraints, then it uses a zero-shot learning module to check whether this knowledge instance satisfies all constraints. The acquired constraint-checking result is then aggregated with the output of the main prompting technique to produce the final output. Experimental results on CSKB Reasoning benchmarks demonstrate the effectiveness of our method by bringing consistent improvements over all prompting methods. Codes and data are available at \url{https://github.com/HKUST-KnowComp/ConstraintChecker}.

Shizhe Diao, Zhichao Huang, Ruijia Xu et al.

The increasing scale of general-purpose Pre-trained Language Models (PLMs) necessitates the study of more efficient adaptation across different downstream tasks. In this paper, we establish a Black-box Discrete Prompt Learning (BDPL) to resonate with pragmatic interactions between the cloud infrastructure and edge devices. Particularly, instead of fine-tuning the model in the cloud, we adapt PLMs by prompt learning, which efficiently optimizes only a few parameters of the discrete prompts. Moreover, we consider the scenario that we do not have access to the parameters and gradients of the pre-trained models, except for its outputs given inputs. This black-box setting secures the cloud infrastructure from potential attack and misuse to cause a single-point failure, which is preferable to the white-box counterpart by current infrastructures. Under this black-box constraint, we apply a variance-reduced policy gradient algorithm to estimate the gradients of parameters in the categorical distribution of each discrete prompt. In light of our method, the user devices can efficiently tune their tasks by querying the PLMs bounded by a range of API calls. Our experiments on RoBERTa and GPT-3 demonstrate that the proposed algorithm achieves significant improvement on eight benchmarks in a cloud-device collaboration manner. Finally, we conduct in-depth case studies to comprehensively analyze our method in terms of various data sizes, prompt lengths, training budgets, optimization objectives, prompt transferability, and explanations of the learned prompts. Our code will be available at https://github.com/shizhediao/Black-Box-Prompt-Learning.

Shizhe Diao, Yan Song, Tong Zhang

Keyphrase generation aims to produce a set of phrases summarizing the essentials of a given document. Conventional methods normally apply an encoder-decoder architecture to generate the output keyphrases for an input document, where they are designed to focus on each current document so they inevitably omit crucial corpus-level information carried by other similar documents, i.e., the cross-document dependency and latent topics. In this paper, we propose CDKGen, a Transformer-based keyphrase generator, which expands the Transformer to global attention with cross-document attention networks to incorporate available documents as references so as to generate better keyphrases with the guidance of topic information. On top of the proposed Transformer + cross-document attention architecture, we also adopt a copy mechanism to enhance our model via selecting appropriate words from documents to deal with out-of-vocabulary words in keyphrases. Experiment results on five benchmark datasets illustrate the validity and effectiveness of our model, which achieves the state-of-the-art performance on all datasets. Further analyses confirm that the proposed model is able to generate keyphrases consistent with references while keeping sufficient diversity. The code of CDKGen is available at https://github.com/SVAIGBA/CDKGen.

Shizhe Diao, Jiaxin Bai, Yan Song et al.

The pre-training of text encoders normally processes text as a sequence of tokens corresponding to small text units, such as word pieces in English and characters in Chinese. It omits information carried by larger text granularity, and thus the encoders cannot easily adapt to certain combinations of characters. This leads to a loss of important semantic information, which is especially problematic for Chinese because the language does not have explicit word boundaries. In this paper, we propose ZEN, a BERT-based Chinese (Z) text encoder Enhanced by N-gram representations, where different combinations of characters are considered during training. As a result, potential word or phase boundaries are explicitly pre-trained and fine-tuned with the character encoder (BERT). Therefore ZEN incorporates the comprehensive information of both the character sequence and words or phrases it contains. Experimental results illustrated the effectiveness of ZEN on a series of Chinese NLP tasks. We show that ZEN, using less resource than other published encoders, can achieve state-of-the-art performance on most tasks. Moreover, it is shown that reasonable performance can be obtained when ZEN is trained on a small corpus, which is important for applying pre-training techniques to scenarios with limited data. The code and pre-trained models of ZEN are available at https://github.com/sinovation/zen.

Peter Belcak, Greg Heinrich, Shizhe Diao et al.

Large language models (LLMs) are often praised for exhibiting near-human performance on a wide range of tasks and valued for their ability to hold a general conversation. The rise of agentic AI systems is, however, ushering in a mass of applications in which language models perform a small number of specialized tasks repetitively and with little variation. Here we lay out the position that small language models (SLMs) are sufficiently powerful, inherently more suitable, and necessarily more economical for many invocations in agentic systems, and are therefore the future of agentic AI. Our argumentation is grounded in the current level of capabilities exhibited by SLMs, the common architectures of agentic systems, and the economy of LM deployment. We further argue that in situations where general-purpose conversational abilities are essential, heterogeneous agentic systems (i.e., agents invoking multiple different models) are the natural choice. We discuss the potential barriers for the adoption of SLMs in agentic systems and outline a general LLM-to-SLM agent conversion algorithm. Our position, formulated as a value statement, highlights the significance of the operational and economic impact even a partial shift from LLMs to SLMs is to have on the AI agent industry. We aim to stimulate the discussion on the effective use of AI resources and hope to advance the efforts to lower the costs of AI of the present day. Calling for both contributions to and critique of our position, we commit to publishing all such correspondence at https://research.nvidia.com/labs/lpr/slm-agents.

Rui Pan, Xiang Liu, Shizhe Diao et al.

The machine learning community has witnessed impressive advancements since large language models (LLMs) first appeared. Yet, their massive memory consumption has become a significant roadblock to large-scale training. For instance, a 7B model typically requires at least 60 GB of GPU memory with full parameter training, which presents challenges for researchers without access to high-resource environments. Parameter Efficient Fine-Tuning techniques such as Low-Rank Adaptation (LoRA) have been proposed to alleviate this problem. However, in most large-scale fine-tuning settings, their performance does not reach the level of full parameter training because they confine the parameter search to a low-rank subspace. Attempting to complement this deficiency, we investigate the layerwise properties of LoRA on fine-tuning tasks and observe an unexpected but consistent skewness of weight norms across different layers. Utilizing this key observation, a surprisingly simple training strategy is discovered, which outperforms both LoRA and full parameter training in a wide range of settings with memory costs as low as LoRA. We name it Layerwise Importance Sampled AdamW (LISA), a promising alternative for LoRA, which applies the idea of importance sampling to different layers in LLMs and randomly freezes most middle layers during optimization. Experimental results show that with similar or less GPU memory consumption, LISA surpasses LoRA or even full parameter tuning in downstream fine-tuning tasks, where LISA consistently outperforms LoRA by over 10%-35% in terms of MT-Bench score while achieving on-par or better performance in MMLU, AGIEval and WinoGrande. On large models, specifically LLaMA-2-70B, LISA surpasses LoRA on MT-Bench, GSM8K, and PubMedQA, demonstrating its effectiveness across different domains.

Xin Xu, Shizhe Diao, Can Yang et al.

Chain-of-Thought (CoT) prompting has marked a significant advancement in enhancing the reasoning capabilities of large language models (LLMs). Previous studies have developed various extensions of CoT, which focus primarily on enhancing end-task performance. In addition, there has been research on assessing the quality of reasoning chains in CoT. This raises an intriguing question: Is it possible to predict the accuracy of LLM outputs by scrutinizing the reasoning chains they generate? To answer this research question, we introduce a benchmark, R2PE, designed specifically to explore the relationship between reasoning chains and performance in various reasoning tasks spanning five different domains. This benchmark aims to measure the falsehood of the final output of LLMs based on the reasoning steps. To make full use of information in multiple reasoning chains, we propose the process discernibility score (PDS) framework that beats the answer-checking baseline by a large margin. Concretely, this resulted in an average of $5.1\%$ increase in the F1 score and $2.97\%$ improvement in AUC-PR across all 45 subsets within R2PE. We further demonstrate our PDS's efficacy in advancing open-domain QA accuracy.

Shizhe Diao, Yu Yang, Yonggan Fu et al.

Pre-training datasets are typically collected from web content and lack inherent domain divisions. For instance, widely used datasets like Common Crawl do not include explicit domain labels, while manually curating labeled datasets such as The Pile is labor-intensive. Consequently, identifying an optimal pre-training data mixture remains a challenging problem, despite its significant benefits for pre-training performance. To address these challenges, we propose CLustering-based Iterative Data Mixture Bootstrapping (CLIMB), an automated framework that discovers, evaluates, and refines data mixtures in a pre-training setting. Specifically, CLIMB embeds and clusters large-scale datasets in a semantic space and then iteratively searches for optimal mixtures using a smaller proxy model and a predictor. When continuously trained on 400B tokens with this mixture, our 1B model exceeds the state-of-the-art Llama-3.2-1B by 2.0%. Moreover, we observe that optimizing for a specific domain (e.g., Social Sciences) yields a 5% improvement over random sampling. Finally, we introduce ClimbLab, a filtered 1.2-trillion-token corpus with 20 clusters as a research playground, and ClimbMix, a compact yet powerful 400-billion-token dataset designed for efficient pre-training that delivers superior performance under an equal token budget. We analyze the final data mixture, elucidating the characteristics of an optimal data mixture. Our data is available at: https://research.nvidia.com/labs/lpr/climb/

Ruida Wang, Rui Pan, Yuxin Li et al.

Solving mathematical problems using computer-verifiable languages like Lean has significantly impacted the mathematical and computer science communities. State-of-the-art methods utilize a single Large Language Model (LLM) to generate complete proof or perform tree search, but they fail to balance these tasks. We propose **MA-LoT**: *Model-CollAboration Lean-based Long Chain-of-Thought*, a comprehensive framework for Lean4 theorem proving to solve this issue. It separates the cognition tasks of general NL for whole-proof generation and error analysis for proof correction using the model-collaboration method. We achieve this by structured interaction of the LLM and Lean4 verifier in Long CoT. To implement the framework, we propose the novel *LoT-Transfer Learning* training-inference pipeline, which enables the Long CoT thinking capability to LLMs without special data annotation. Extensive experiment shows that our framework achieves a **61.07%** accuracy rate on the Lean4 version of the MiniF2F-Test dataset, largely outperforming DeepSeek-V3 (33.61%), single-model tree search (InternLM-Step-Prover, 50.70%), and whole-proof generation (Godel-Prover, 55.33%) baselines. Furthermore, our findings highlight the potential of combining Long CoT with formal verification for a more insightful generation in a broader perspective.

Shih-Yang Liu, Xin Dong, Ximing Lu et al. · uw

Reasoning language models such as OpenAI-o1, DeepSeek-R1, and Qwen achieve strong performance via extended chains of thought but often generate unnecessarily long outputs. Maximizing intelligence per token--accuracy relative to response length--remains an open problem. We revisit reinforcement learning (RL) with the simplest length penalty--truncation--and show that accuracy degradation arises not from the lack of sophisticated penalties but from inadequate RL optimization. We identify three key challenges: (i) large bias in advantage estimation, (ii) entropy collapse, and (iii) sparse reward signal. We address them with Doing Length pEnalty Right (DLER), a training recipe combining batch-wise reward normalization, higher clipping, dynamic sampling, and a simple truncation length penalty. DLER achieves state-of-the-art accuracy--efficiency trade-offs, cutting output length by over 70 percent while surpassing all previous baseline accuracy. It also improves test-time scaling: compared to DeepSeek-R1-7B, DLER-7B generates multiple concise responses in parallel with 28 percent higher accuracy and lower latency. We further introduce Difficulty-Aware DLER, which adaptively tightens truncation on easier questions for additional efficiency gains. We also propose an update-selective merging method that preserves baseline accuracy while retaining the concise reasoning ability of the DLER model, which is useful for scenarios where RL training data is scarce.

Yue Xin, Wenyuan Wang, Rui Pan et al.

Multimodal large language models have various practical applications that demand strong reasoning abilities. Despite recent advancements, these models still struggle to solve complex geometric problems. A key challenge stems from the lack of high-quality image-text pair datasets for understanding geometric images. Furthermore, most template-based data synthesis pipelines typically fail to generalize to questions beyond their predefined templates. In this paper, we bridge this gap by introducing a complementary process of Reinforcement Learning with Verifiable Rewards (RLVR) into the data generation pipeline. By adopting RLVR to refine captions for geometric images synthesized from 50 basic geometric relations and using reward signals derived from mathematical problem-solving tasks, our pipeline successfully captures the key features of geometry problem-solving. This enables better task generalization and yields non-trivial improvements. Furthermore, even in out-of-distribution scenarios, the generated dataset enhances the general reasoning capabilities of multimodal large language models, yielding accuracy improvements of $2.8\%\text{-}4.8\%$ in statistics, arithmetic, algebraic, and numerical tasks with non-geometric input images of MathVista and MathVerse, along with $2.4\%\text{-}3.9\%$ improvements in Art, Design, Tech, and Engineering tasks in MMMU.

Ruida Wang, Jiarui Yao, Rui Pan et al.

Solving math problems through verifiable languages such as Lean has significantly impacted both the mathematics and computer science communities. Current state-of-the-art models are often trained with expensive online Reinforcement Learning (RL) or expert iteration. However, these approaches rely on fixed problem sets, which causes inefficient training and limits the model to tackle complex problems. To overcome these limitations, we propose GAR: Generative Adversarial Reinforcement learning, a comprehensive RL training framework that jointly trains the problem composer and solver in an adversarial loop. GAR introduces an implicit curriculum learning mechanism, which aligns task difficulty with the prover's evolving capability. It thereby improves the training efficiency and enables stronger performance of proving advanced theorems. Experiments show that with GAR training, Goedel-Prover-V2-8B and DeepSeek-Prover-V2-7B achieve an average relative improvement in pass@32 of 4.20% on MiniF2F-Test benchmark, while DeepSeek-Prover-V2's pass@32 on ProofNet-Test increases from 22.58% to 25.81%. Beyond formal proving, GAR establishes a general RL paradigm for co-evolution of problem generation and solving under verifiable environments.

Jian Hu, Mingjie Liu, Ximing Lu et al. · uw

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a key ingredient for unlocking complex reasoning capabilities in large language models. Recent work ProRL has shown promise in scaling RL by increasing the number of training steps. However, performance plateaus after thousands of steps, with clear diminishing returns from allocating more computation to additional training. In this work, we investigate a complementary paradigm for scaling RL, BroR-Lincreasing the number of rollouts per example to hundreds to exhaustively Broaden exploration, which yields continuous performance gains beyond the saturation point observed in ProRL when scaling the number of training steps. Our approach is motivated by a mass balance equation analysis allowing us to characterize the rate of change in probability mass for correct and incorrect tokens during the reinforcement process. We show that under a one-step RL assumption, sampled rollout tokens always contribute to correct-mass expansion, while unsampled tokens outside rollouts may lead to gains or losses depending on their distribution and the net reward balance. Importantly, as the number of rollouts per example N increases, the effect of unsampled terms diminishes, ensuring overall correct-mass expansion. To validate our theoretical analysis, we conduct simulations under more relaxed conditions and find that a sufficiently large rollout size N-corresponding to ample exploration-guarantees an increase in the probability mass of all correct tokens. Empirically, BroRL revives models saturated after 3K ProRL training steps and demonstrates robust, continuous improvement, achieving state-of-the-art results for the 1.5B model across diverse benchmarks.

Mingjie Liu, Shizhe Diao, Jian Hu et al.

Recent advancements in reasoning-focused language models such as OpenAI's O1 and DeepSeek-R1 have shown that scaling test-time computation-through chain-of-thought reasoning and iterative exploration-can yield substantial improvements on complex tasks like mathematics and code generation. These breakthroughs have been driven by large-scale reinforcement learning (RL), particularly when combined with verifiable reward signals that provide objective and grounded supervision. In this report, we investigate the effects of prolonged reinforcement learning on a small language model across a diverse set of reasoning domains. Our work identifies several key ingredients for effective training, including the use of verifiable reward tasks, enhancements to Group Relative Policy Optimization (GRPO), and practical techniques to improve training stability and generalization. We introduce controlled KL regularization, clipping ratio, and periodic reference policy resets as critical components for unlocking long-term performance gains. Our model achieves significant improvements over strong baselines, including +14.7% on math, +13.9% on coding, and +54.8% on logic puzzle tasks. To facilitate continued research, we release our model publicly.

Wei Du, Branislav Kisacanin, George Armstrong et al.

Reasoning-capable language models achieve state-of-the-art performance in diverse complex tasks by generating long, explicit Chain-of-Thought (CoT) traces. While recent works show that base models can acquire such reasoning traces via reinforcement learning or distillation from stronger models like DeepSeek-R1, previous works demonstrate that even short CoT prompting without fine-tuning is able to improve reasoning. We ask whether long CoT can be induced in a base model using only prompting or minimal tuning. Using just 20 long CoT examples from the reasoning model \texttt{QwQ-32B-Preview}, we lightly fine-tune the base model \texttt{Qwen2.5-32B}. The resulting model outperforms the much larger \texttt{Qwen2.5-Math-72B-Instruct}, showing that a handful of high-quality examples can unlock strong reasoning capabilities. We further explore using CoT data from non-reasoning models and human annotators, enhanced with prompt engineering, multi-pass editing, and structural guidance. However, neither matches the performance of reasoning model traces, suggesting that certain latent qualities of expert CoT are difficult to replicate. We analyze key properties of reasoning data, such as problem difficulty, diversity, and answer length, that influence reasoning distillation. While challenges remain, we are optimistic that carefully curated human-written CoT, even in small quantities, can activate reasoning behaviors in base models. We release our human-authored dataset across refinement stages and invite further investigation into what makes small-scale reasoning supervision so effective.

Cheng Niu, Yang Guan, Yuanhao Wu et al.

The proliferation of fake news poses a significant threat not only by disseminating misleading information but also by undermining the very foundations of democracy. The recent advance of generative artificial intelligence has further exacerbated the challenge of distinguishing genuine news from fabricated stories. In response to this challenge, we introduce VeraCT Scan, a novel retrieval-augmented system for fake news detection. This system operates by extracting the core facts from a given piece of news and subsequently conducting an internet-wide search to identify corroborating or conflicting reports. Then sources' credibility is leveraged for information verification. Besides determining the veracity of news, we also provide transparent evidence and reasoning to support its conclusions, resulting in the interpretability and trust in the results. In addition to GPT-4 Turbo, Llama-2 13B is also fine-tuned for news content understanding, information verification, and reasoning. Both implementations have demonstrated state-of-the-art accuracy in the realm of fake news detection.

Dylan Zhang, Shizhe Diao, Xueyan Zou et al.

Preference learning provides a promising solution to address the limitations of supervised fine-tuning (SFT) for code language models, where the model is not explicitly trained to differentiate between correct and incorrect code. Recent findings demonstrate that on-policy data is the key to successful preference learning, where the preference data is collected using the same policy LM being trained. Inspired by this, we propose PLUM, an on-policy $\textbf{P}$reference $\textbf{L}$earning framework A$\textbf{u}$gmented with test cases for code L$\textbf{M}$ s. The framework operates in three key stages: (1) automatic generation of test cases from natural language instructions, (2) creation of a preference data by evaluating candidate code solutions sampled from the policy, which can then be used to (3) train the policy LM. PLUM levitates the need to train reward models, allowing for large scale on-policy and online preference data collation. PLUM is evaluated on both standard benchmarks (HumanEval, MBPP) and more challenging ones (LiveCodeBench), delivering substantial improvements over original SFT'ed models and other execution-feedback-driven approaches. We show PLUM's benefits are consistent across various widely-used code LMs even they have been well-trained with SFT. For example, PLUM increases pass rates by up to 4.8% on average on standard benchmarks and 11.8% on LiveCodeBench, demonstrating its effectiveness and generalizability. We also demonstrate the benefits of on-policy and online preference learning by comprehensive experimentation.

Renjie Pi, Jiahui Gao, Shizhe Diao et al.

In recent years, the field of computer vision has seen significant advancements thanks to the development of large language models (LLMs). These models have enabled more effective and sophisticated interactions between humans and machines, paving the way for novel techniques that blur the lines between human and machine intelligence. In this paper, we introduce a new paradigm for object detection that we call reasoning-based object detection. Unlike conventional object detection methods that rely on specific object names, our approach enables users to interact with the system using natural language instructions, allowing for a higher level of interactivity. Our proposed method, called DetGPT, leverages state-of-the-art multi-modal models and open-vocabulary object detectors to perform reasoning within the context of the user's instructions and the visual scene. This enables DetGPT to automatically locate the object of interest based on the user's expressed desires, even if the object is not explicitly mentioned. For instance, if a user expresses a desire for a cold beverage, DetGPT can analyze the image, identify a fridge, and use its knowledge of typical fridge contents to locate the beverage. This flexibility makes our system applicable across a wide range of fields, from robotics and automation to autonomous driving. Overall, our proposed paradigm and DetGPT demonstrate the potential for more sophisticated and intuitive interactions between humans and machines. We hope that our proposed paradigm and approach will provide inspiration to the community and open the door to more interative and versatile object detection systems. Our project page is launched at detgpt.github.io.

Xiao Zhou, Weizhong Zhang, Zonghao Chen et al.

Sparse training is a natural idea to accelerate the training speed of deep neural networks and save the memory usage, especially since large modern neural networks are significantly over-parameterized. However, most of the existing methods cannot achieve this goal in practice because the chain rule based gradient (w.r.t. structure parameters) estimators adopted by previous methods require dense computation at least in the backward propagation step. This paper solves this problem by proposing an efficient sparse training method with completely sparse forward and backward passes. We first formulate the training process as a continuous minimization problem under global sparsity constraint. We then separate the optimization process into two steps, corresponding to weight update and structure parameter update. For the former step, we use the conventional chain rule, which can be sparse via exploiting the sparse structure. For the latter step, instead of using the chain rule based gradient estimators as in existing methods, we propose a variance reduced policy gradient estimator, which only requires two forward passes without backward propagation, thus achieving completely sparse training. We prove that the variance of our gradient estimator is bounded. Extensive experimental results on real-world datasets demonstrate that compared to previous methods, our algorithm is much more effective in accelerating the training process, up to an order of magnitude faster.