Chain-of-Thought prompting has significantly enhanced the reasoning capabilities of large language models, with numerous studies exploring factors influencing its performance. However, the underlying mechanisms remain poorly understood. To further demystify the operational principles, this work examines three key aspects: decoding, projection, and activation, aiming to elucidate the changes that occur within models when employing Chainof-Thought. Our findings reveal that LLMs effectively imitate exemplar formats while integrating them with their understanding of the question, exhibiting fluctuations in token logits during generation but ultimately producing a more concentrated logits distribution, and activating a broader set of neurons in the final layers, indicating more extensive knowledge retrieval compared to standard prompts. Our code and data will be publicly avialable when the paper is accepted.
Integration-by-parts reductions of Feynman integrals pose a frequent bottle-neck in state-of-the-art calculations in theoretical particle and gravitational-wave physics, and rely on heuristic approaches for selecting integration-by-parts identities, whose quality heavily influences the performance. In this paper, we investigate the use of machine-learning techniques to find improved heuristics. We use funsearch, a genetic programming variant based on code generation by a Large Language Model, in order to explore possible approaches, then use strongly typed genetic programming to zero in on useful solutions. Both approaches manage to re-discover the state-of-the-art heuristics recently incorporated into integration-by-parts solvers, and in one example find a small advance on this state of the art.
Several techniques have been proposed to detect vulnerable Spectre gadgets in widely deployed commercial software. Unfortunately, detection techniques proposed so far rely on hand-written rules which fall short in covering subtle variations of known Spectre gadgets as well as demand a huge amount of time to analyze each conditional branch in software. Moreover, detection tool evaluations are based only on a handful of these gadgets, as it requires arduous effort to craft new gadgets manually. In this work, we employ both fuzzing and deep learning techniques to automate the generation and detection of Spectre gadgets. We first create a diverse set of Spectre-V1 gadgets by introducing perturbations to the known gadgets. Using mutational fuzzing, we produce a data set with more than 1 million Spectre-V1 gadgets which is the largest Spectre gadget data set built to date. Next, we conduct the first empirical usability study of Generative Adversarial Networks (GANs) in the context of assembly code generation without any human interaction. We introduce SpectreGAN which leverages masking implementation of GANs for both learning the gadget structures and generating new gadgets. This provides the first scalable solution to extend the variety of Spectre gadgets. Finally, we propose FastSpec which builds a classifier with the generated Spectre gadgets based on a novel high dimensional Neural Embeddings technique (BERT). For the case studies, we demonstrate that FastSpec discovers potential gadgets with a high success rate in OpenSSL libraries and Phoronix benchmarks. Further, FastSpec offers much greater flexibility and time-related performance gain compared to the existing tools and therefore can be used for gadget detection in large-scale software.
In recent years, owing to the outstanding performance in graph representation learning, graph neural network (GNN) techniques have gained considerable interests in many real-world scenarios, such as recommender systems and social networks. In recommender systems, the main challenge is to learn the effective user/item representations from their interactions. However, many recent publications using GNNs for recommender systems cannot be directly compared, due to their difference on datasets and evaluation metrics. Furthermore, many of them only provide a demo to conduct experiments on small datasets, which is far away to be applied in real-world recommender systems. To address this problem, we introduce Graph4Rec, a universal toolkit that unifies the paradigm to train GNN models into the following parts: graphs input, random walk generation, ego graphs generation, pairs generation and GNNs selection. From this training pipeline, one can easily establish his own GNN model with a few configurations. Besides, we develop a large-scale graph engine and a parameter server to support distributed GNN training. We conduct a systematic and comprehensive experiment to compare the performance of different GNN models on several scenarios in different scale. Extensive experiments are demonstrated to identify the key components of GNNs. We also try to figure out how the sparse and dense parameters affect the performance of GNNs. Finally, we investigate methods including negative sampling, ego graph construction order, and warm start strategy to find a more effective and efficient GNNs practice on recommender systems. Our toolkit is based on PGL https://github.com/PaddlePaddle/PGL and the code is opened source in https://github.com/PaddlePaddle/PGL/tree/main/apps/Graph4Rec.
Pre-trained sequence-to-sequence language models have led to widespread success in many natural language generation tasks. However, there has been relatively less work on analyzing their ability to generate structured outputs such as graphs. Unlike natural language, graphs have distinct structural and semantic properties in the context of a downstream NLP task, e.g., generating a graph that is connected and acyclic can be attributed to its structural constraints, while the semantics of a graph can refer to how meaningfully an edge represents the relation between two node concepts. In this work, we study pre-trained language models that generate explanation graphs in an end-to-end manner and analyze their ability to learn the structural constraints and semantics of such graphs. We first show that with limited supervision, pre-trained language models often generate graphs that either violate these constraints or are semantically incoherent. Since curating large amount of human-annotated graphs is expensive and tedious, we propose simple yet effective ways of graph perturbations via node and edge edit operations that lead to structurally and semantically positive and negative graphs. Next, we leverage these graphs in different contrastive learning models with Max-Margin and InfoNCE losses. Our methods lead to significant improvements in both structural and semantic accuracy of explanation graphs and also generalize to other similar graph generation tasks. Lastly, we show that human errors are the best negatives for contrastive learning and also that automatically generating more such human-like negative graphs can lead to further improvements. Our code and models are publicly available at https://github.com/swarnaHub/ExplagraphGen
Current large language models (LLMs) often struggle to produce accurate responses on the first attempt for complex reasoning tasks like code generation. Prior research tackles this challenge by generating multiple candidate solutions and validating them with LLM-generated unit tests. The execution results of unit tests serve as reward signals to identify correct solutions. As LLMs always confidently make mistakes, these unit tests are not reliable, thereby diminishing the quality of reward signals. Motivated by the observation that scaling the number of solutions improves LLM performance, we explore the impact of scaling unit tests to enhance reward signal quality. Our pioneer experiment reveals a positive correlation between the number of unit tests and reward signal quality, with greater benefits observed in more challenging problems. Based on these insights, we propose CodeRM-8B, a lightweight yet effective unit test generator that enables efficient and high-quality unit test scaling. Additionally, we implement a dynamic scaling mechanism that adapts the number of unit tests based on problem difficulty, further improving efficiency. Experimental results show that our approach significantly improves performance across various models on three benchmarks (e.g., with gains of 18.43% for Llama3-8B and 3.42% for GPT-4o-mini on HumanEval Plus).
Reinforcement learning with verifiable rewards (RLVR) has substantially improved the reasoning ability of LLMs, but often depends on external supervision from human annotations or gold-standard solutions. Reinforcement learning from internal feedback (RLIF) has recently emerged as a scalable unsupervised alternative, using signals extracted from the model itself. However, existing RLIF methods typically rely on a single internal reward, which can lead to reward hacking, entropy collapse, and degraded reasoning structure. We propose a multi-reward RLIF framework that decomposes the training signal into two complementary components: an answer-level reward based on cluster voting and a completion-level reward based on token-wise self-certainty. To combine these signals robustly, we apply GDPO-based normalization to reduce reward-scale imbalance. We further introduce KL-Cov regularization, which targets low-entropy token distributions responsible for disproportionate entropy reduction, preserving exploration and preventing late-stage collapse. Across mathematical reasoning and code-generation benchmarks, our method improves stability and robustness over prior unsupervised RL approaches, while achieving performance close to supervised RLVR methods. These results show that complementary internal rewards, combined with targeted regularization, can support stable long-horizon reasoning without relying on external ground-truth supervision. Code will be released soon.
Layout is a fundamental component of any graphic design. Creating large varieties of plausible document layouts can be a tedious task, requiring numerous constraints to be satisfied, including local ones relating different semantic elements and global constraints on the general appearance and spacing. In this paper, we present a novel framework, coined READ, for REcursive Autoencoders for Document layout generation, to generate plausible 2D layouts of documents in large quantities and varieties. First, we devise an exploratory recursive method to extract a structural decomposition of a single document. Leveraging a dataset of documents annotated with labeled bounding boxes, our recursive neural network learns to map the structural representation, given in the form of a simple hierarchy, to a compact code, the space of which is approximated by a Gaussian distribution. Novel hierarchies can be sampled from this space, obtaining new document layouts. Moreover, we introduce a combinatorial metric to measure structural similarity among document layouts. We deploy it to show that our method is able to generate highly variable and realistic layouts. We further demonstrate the utility of our generated layouts in the context of standard detection tasks on documents, showing that detection performance improves when the training data is augmented with generated documents whose layouts are produced by READ.
The recent trend in multiple object tracking (MOT) is heading towards leveraging deep learning to boost the tracking performance. However, it is not trivial to solve the data-association problem in an end-to-end fashion. In this paper, we propose a novel proposal-based learnable framework, which models MOT as a proposal generation, proposal scoring and trajectory inference paradigm on an affinity graph. This framework is similar to the two-stage object detector Faster RCNN, and can solve the MOT problem in a data-driven way. For proposal generation, we propose an iterative graph clustering method to reduce the computational cost while maintaining the quality of the generated proposals. For proposal scoring, we deploy a trainable graph-convolutional-network (GCN) to learn the structural patterns of the generated proposals and rank them according to the estimated quality scores. For trajectory inference, a simple deoverlapping strategy is adopted to generate tracking output while complying with the constraints that no detection can be assigned to more than one track. We experimentally demonstrate that the proposed method achieves a clear performance improvement in both MOTA and IDF1 with respect to previous state-of-the-art on two public benchmarks. Our code is available at https://github.com/daip13/LPC_MOT.git.
Multirotor aerial robots are becoming widely used for the inspection of powerlines. To enable continuous, robust inspection without human intervention, the robots must be able to perch on the powerlines to recharge their batteries. Highly versatile perching capabilities are necessary to adapt to the variety of configurations and constraints that are present in real powerline systems. This paper presents a novel perching trajectory generation framework that computes perception-aware, collision-free, and dynamically-feasible maneuvers to guide the robot to the desired final state. Trajectory generation is achieved via solving a Nonlinear Programming problem using the Primal-Dual Interior Point method. The problem considers the full dynamic model of the robot down to its single rotor thrusts and minimizes the final pose and velocity errors while avoiding collisions and maximizing the visibility of the powerline during the maneuver. The generated maneuvers consider both the perching and the posterior recovery trajectories. The framework adopts costs and constraints defined by efficient mathematical representations of powerlines, enabling online onboard execution in resource-constrained hardware. The method is validated on-board an agile quadrotor conducting powerline inspection and various perching maneuvers with final pitch values of up to 180 degrees. The developed code is available online at: https://github.com/grvcPerception/pa_powerline_perching
Large language models (LLMs) can perform complex reasoning in few- and zero-shot settings by generating intermediate chain of thought (CoT) reasoning steps. Further, each reasoning step can rely on external tools to support computation beyond the core LLM capabilities (e.g. search/running code). Prior work on CoT prompting and tool use typically requires hand-crafting task-specific demonstrations and carefully scripted interleaving of model generations with tool use. We introduce Automatic Reasoning and Tool-use (ART), a framework that uses frozen LLMs to automatically generate intermediate reasoning steps as a program. Given a new task to solve, ART selects demonstrations of multi-step reasoning and tool use from a task library. At test time, ART seamlessly pauses generation whenever external tools are called, and integrates their output before resuming generation. ART achieves a substantial improvement over few-shot prompting and automatic CoT on unseen tasks in the BigBench and MMLU benchmarks, and matches performance of hand-crafted CoT prompts on a majority of these tasks. ART is also extensible, and makes it easy for humans to improve performance by correcting errors in task-specific programs or incorporating new tools, which we demonstrate by drastically improving performance on select tasks with minimal human intervention.
Multimodal reasoning is a challenging task that requires models to reason across multiple modalities to answer questions. Existing approaches have made progress by incorporating language and visual modalities into a two-stage reasoning framework, separating rationale generation from answer inference. However, these approaches often fall short due to the inadequate quality of the generated rationales. In this work, we delve into the importance of rationales in model reasoning. We observe that when rationales are completely accurate, the model's accuracy significantly improves, highlighting the need for high-quality rationale generation. Motivated by this, we propose MC-CoT, a self-consistency training strategy that generates multiple rationales and answers, subsequently selecting the most accurate through a voting process. This approach not only enhances the quality of generated rationales but also leads to more accurate and robust answers. Through extensive experiments, we demonstrate that our approach significantly improves model performance across various benchmarks. Remarkably, we show that even smaller base models, when equipped with our proposed approach, can achieve results comparable to those of larger models, illustrating the potential of our approach in harnessing the power of rationales for improved multimodal reasoning. The code is available at https://github.com/chengtan9907/mc-cot.
Norwegian, spoken by only 5 million population, is under-representative within the most impressive breakthroughs in NLP tasks. To the best of our knowledge, there has not yet been a comprehensive evaluation of the existing language models (LMs) on Norwegian generation tasks during the article writing process. To fill this gap, we 1) compiled the existing Norwegian dataset and pre-trained 4 Norwegian Open Language Models varied from parameter scales and architectures, collectively called NorGLM; 2) introduced a comprehensive benchmark, NLEBench, for evaluating natural language generation capabilities in Norwegian, encompassing translation and human annotation. Based on the investigation, we find that: 1) the mainstream, English-dominated LM GPT-3.5 has limited capability in understanding the Norwegian context; 2) the increase in model parameter scales demonstrates limited impact on the performance of downstream tasks when the pre-training dataset is constrained in size; 3) smaller models also demonstrate the reasoning capability through Chain-of-Thought; 4) a multi-task dataset that includes synergy tasks can be used to verify the generalizability of LLMs on natural language understanding and, meanwhile, test the interconnectedness of these NLP tasks. We share our resources and code for reproducibility under a CC BY-NC 4.0 license.
The deployment and scaling of large language models (LLMs) have become critical as they permeate various applications, demanding high-throughput and low-latency serving systems. Existing frameworks struggle to balance these requirements, especially for workloads with long prompts. This paper introduces DeepSpeed-FastGen, a system that employs Dynamic SplitFuse, a novel prompt and generation composition strategy, to deliver up to 2.3x higher effective throughput, 2x lower latency on average, and up to 3.7x lower (token-level) tail latency, compared to state-of-the-art systems like vLLM. We leverage a synergistic combination of DeepSpeed-MII and DeepSpeed-Inference to provide an efficient and easy-to-use serving system for LLMs. DeepSpeed-FastGen's advanced implementation supports a range of models and offers both non-persistent and persistent deployment options, catering to diverse user scenarios from interactive sessions to long-running applications. We present a detailed benchmarking methodology, analyze the performance through latency-throughput curves, and investigate scalability via load balancing. Our evaluations demonstrate substantial improvements in throughput and latency across various models and hardware configurations. We discuss our roadmap for future enhancements, including broader model support and new hardware backends. The DeepSpeed-FastGen code is readily available for community engagement and contribution.
Recent advancements in text-to-image generation using diffusion models have significantly improved the quality of generated images and expanded the ability to depict a wide range of objects. However, ensuring that these models adhere closely to the text prompts remains a considerable challenge. This issue is particularly pronounced when trying to generate photorealistic images of humans. Without significant prompt engineering efforts models often produce unrealistic images and typically fail to incorporate the full extent of the prompt information. This limitation can be largely attributed to the nature of captions accompanying the images used in training large scale diffusion models, which typically prioritize contextual information over details related to the person's appearance. In this paper we address this issue by introducing a training-free pipeline designed to generate accurate appearance descriptions from images of people. We apply this method to create approximately 250,000 captions for publicly available face datasets. We then use these synthetic captions to fine-tune a text-to-image diffusion model. Our results demonstrate that this approach significantly improves the model's ability to generate high-quality, realistic human faces and enhances adherence to the given prompts, compared to the baseline model. We share our synthetic captions, pretrained checkpoints and training code.
The scarcity of domain-specific dialogue datasets limits the development of dialogue systems across applications. Existing research is constrained by general or niche datasets that lack sufficient scale for training dialogue systems. To address this gap, we introduce DiaSynth - a synthetic dialogue generation framework capable of generating high-quality, contextually rich dialogues across a wide range of domains. Unlike existing frameworks, DiaSynth uses Large Language Models (LLMs) and Chain of Thought (CoT) reasoning to generate dynamic, domain-specific dialogues with simulated personas and diverse conversational features. We perform our experiments by generating synthetic data using different LLMs and few-shot examples from DialogSum and SAMSum. The pretrained language models fine-tuned on the synthetic data outperform the base models by 16.47% on dialogue summarization, while the comparison between models fine-tuned on in-domain data and synthetic data shows that the synthetic data is able to capture 90.48% of the performance distribution of the in-domain data on dialogue summarization. The quality of the data generated also increases as we increase the size of LLM from 3B to 8B. These results validate DiaSynth's potential as a robust alternative to traditional data collection methods. We open source the code and data generated for future research.
Recent advancements in Large Language Models (LLMs) have significantly enhanced their ability to process long contexts, yet a notable gap remains in generating long, aligned outputs. This limitation stems from a training gap where pre-training lacks effective instructions for long-text generation, and post-training data primarily consists of short query-response pairs. Current approaches, such as instruction backtranslation and behavior imitation, face challenges including data quality, copyright issues, and constraints on proprietary model usage. In this paper, we introduce an innovative iterative training framework called Self-Lengthen that leverages only the intrinsic knowledge and skills of LLMs without the need for auxiliary data or proprietary models. The framework consists of two roles: the Generator and the Extender. The Generator produces the initial response, which is then split and expanded by the Extender. This process results in a new, longer response, which is used to train both the Generator and the Extender iteratively. Through this process, the models are progressively trained to handle increasingly longer responses. Experiments on benchmarks and human evaluations show that Self-Lengthen outperforms existing methods in long-text generation, when applied to top open-source LLMs such as Qwen2 and LLaMA3. Our code is publicly available at https://github.com/QwenLM/Self-Lengthen.
The field of Artificial Intelligence (AI) continues to drive transformative innovations, with significant progress in conversational interfaces, autonomous vehicles, and intelligent content creation. Since the launch of ChatGPT in late 2022, the rise of Generative AI has marked a pivotal era, with the term Large Language Models (LLMs) becoming a ubiquitous part of daily life. LLMs have demonstrated exceptional capabilities in tasks such as text summarization, code generation, and creative writing. However, these models are inherently limited by their token-level processing, which restricts their ability to perform abstract reasoning, conceptual understanding, and efficient generation of long-form content. To address these limitations, Meta has introduced Large Concept Models (LCMs), representing a significant shift from traditional token-based frameworks. LCMs use concepts as foundational units of understanding, enabling more sophisticated semantic reasoning and context-aware decision-making. Given the limited academic research on this emerging technology, our study aims to bridge the knowledge gap by collecting, analyzing, and synthesizing existing grey literature to provide a comprehensive understanding of LCMs. Specifically, we (i) identify and describe the features that distinguish LCMs from LLMs, (ii) explore potential applications of LCMs across multiple domains, and (iii) propose future research directions and practical strategies to advance LCM development and adoption.
Large language models have achieved remarkable success in various tasks but suffer from high computational costs during inference, limiting their deployment in resource-constrained applications. To address this issue, we propose a novel Collaborative Inference with Token-lEvel Routing (CITER) framework that enables efficient collaboration between small and large language models (SLMs \& LLMs) through a token-level routing strategy. Specifically, CITER routes non-critical tokens to an SLM for efficiency and routes critical tokens to an LLM for generalization quality. We formulate router training as a policy optimization, where the router receives rewards based on both the quality of predictions and the inference costs of generation. This allows the router to learn to predict token-level routing scores and make routing decisions based on both the current token and the future impact of its decisions. To further accelerate the reward evaluation process, we introduce a shortcut which significantly reduces the costs of the reward estimation and improving the practicality of our approach. Extensive experiments on five benchmark datasets demonstrate that CITER reduces the inference costs while preserving high-quality generation, offering a promising solution for real-time and resource-constrained applications. Our data and code are available at https://github.com/aiming-lab/CITER.
In recent years, accurately and quickly deploying medical large language models (LLMs) has become a significant trend. Among these, retrieval-augmented generation (RAG) has garnered significant attention due to its features of rapid deployment and privacy protection. However, existing medical RAG frameworks still have shortcomings. Most existing medical RAG frameworks are designed for single-round question answering tasks and are not suitable for multi-round diagnostic dialogue. On the other hand, existing medical multi-round RAG frameworks do not consider the interconnections between potential diseases to inquire precisely like a doctor. To address these issues, we propose a Multi-Round Diagnostic RAG (MRD-RAG) framework that mimics the doctor's diagnostic process. This RAG framework can analyze diagnosis information of potential diseases and accurately conduct multi-round diagnosis like a doctor. To evaluate the effectiveness of our proposed frameworks, we conduct experiments on two modern medical datasets and two traditional Chinese medicine datasets, with evaluations by GPT and human doctors on different methods. The results indicate that our RAG framework can significantly enhance the diagnostic performance of LLMs, highlighting the potential of our approach in medical diagnosis. The code and data can be found in our project website https://github.com/YixiangCh/MRD-RAG/tree/master.