We present rStar-Math to demonstrate that small language models (SLMs) can rival or even surpass the math reasoning capability of OpenAI o1, without distillation from superior models. rStar-Math achieves this by exercising "deep thinking" through Monte Carlo Tree Search (MCTS), where a math policy SLM performs test-time search guided by an SLM-based process reward model. rStar-Math introduces three innovations to tackle the challenges in training the two SLMs: (1) a novel code-augmented CoT data sythesis method, which performs extensive MCTS rollouts to generate step-by-step verified reasoning trajectories used to train the policy SLM; (2) a novel process reward model training method that avoids na\"ive step-level score annotation, yielding a more effective process preference model (PPM); (3) a self-evolution recipe in which the policy SLM and PPM are built from scratch and iteratively evolved to improve reasoning capabilities. Through 4 rounds of self-evolution with millions of synthesized solutions for 747k math problems, rStar-Math boosts SLMs' math reasoning to state-of-the-art levels. On the MATH benchmark, it improves Qwen2.5-Math-7B from 58.8% to 90.0% and Phi3-mini-3.8B from 41.4% to 86.4%, surpassing o1-preview by +4.5% and +0.9%. On the USA Math Olympiad (AIME), rStar-Math solves an average of 53.3% (8/15) of problems, ranking among the top 20% the brightest high school math students. Code and data will be available at https://github.com/microsoft/rStar.

Incorporating automatically predicted human feedback into the process of training generative models has attracted substantial recent interest, while feedback at inference time has received less attention. The typical feedback at training time, i.e., preferences of choice given two samples, does not naturally transfer to the inference phase. We introduce a novel type of feedback -- caption reformulations -- and train models to mimic reformulation feedback based on human annotations. Our method does not require training the image captioning model itself, thereby demanding substantially less computational effort. We experiment with two types of reformulation feedback: first, we collect a dataset of human reformulations that correct errors in the generated captions. We find that incorporating reformulation models trained on this data into the inference phase of existing image captioning models results in improved captions, especially when the original captions are of low quality. We apply our method to non-English image captioning, a domain where robust models are less prevalent, and gain substantial improvement. Second, we apply reformulations to style transfer. Quantitative evaluations reveal state-of-the-art performance on German image captioning and English style transfer, while human validation with a detailed comparative framework exposes the specific axes of improvement.

Large language models (LLMs) have been proposed as powerful tools for detecting software vulnerabilities, where task-specific fine-tuning is typically employed to provide vulnerability-specific knowledge to the LLMs for this purpose. However, traditional full-parameter fine-tuning is inefficient for modern, complex LLMs, which contain billions of parameters. Soft prompt tuning has been suggested as a more efficient alternative for fine-tuning LLMs in general cases. However, pure soft prompt tuning treats source code as plain text, losing structural information inherent in source code. Meanwhile, graph-enhanced soft prompt tuning methods, which aim to address this issue, are unable to preserve the rich semantic information within code graphs, as they are primarily designed for general graph-related tasks and focus more on adjacency information. They also fail to ensure computational efficiency while accounting for graph-text interactions. This paper, therefore, introduces a new code graph-enhanced, structure-aware soft prompt tuning method for vulnerability detection, referred to as CGP-Tuning. It employs innovative type-aware embeddings to capture the rich semantic information within code graphs, along with a novel and efficient cross-modal alignment module that achieves linear computational cost while incorporating graph-text interactions. The proposed CGP-Tuning is evaluated on the latest DiverseVul dataset and the most recent open-source code LLMs, CodeLlama and CodeGemma. Experimental results demonstrate that CGP-Tuning outperforms the best state-of-the-art method by an average of 3.5 percentage points in accuracy, without compromising its vulnerability detection capabilities for long source code.

In the realm of fractal geometry, intricate structures emerge from simple iterative processes that partition parameter spaces into regions of stability and instability. Likewise, training large language models involves iteratively applying update functions, such as Adam, where even slight hyperparameter adjustments can shift the training process from convergence to divergence. Recent evidence from miniature neural networks suggests that the boundary separating these outcomes displays fractal characteristics [1]. Building on these insights, this study extends them to medium-sized, decoder-only transformer architectures by employing a more consistent convergence measure and examining the learning rate hyperparameter landscape for attention and fully connected layers. The results show that the trainability frontier is not a simple threshold; rather, it forms a self-similar yet seemingly random structure at multiple scales, with statistically consistent and repeating patterns. Within this landscape, a region of stable convergence is surrounded by a complex chaotic border, illustrating the sensitive nature of the underlying training dynamics.

In daily domestic settings, frequently used objects like cups often have unfixed positions and multiple instances within the same category, and their carriers frequently change as well. As a result, it becomes challenging for a robot to efficiently navigate to a specific instance. To tackle this challenge, the robot must capture and update scene changes and plans continuously. However, current object navigation approaches primarily focus on the semantic level and lack the ability to dynamically update scene representation. In contrast, this paper captures the relationships between frequently used objects and their static carriers. It constructs an open-vocabulary Carrier-Relationship Scene Graph (CRSG) and updates the carrying status during robot navigation to reflect the dynamic changes of the scene. Based on the CRSG, we further propose an instance navigation strategy that models the navigation process as a Markov Decision Process. At each step, decisions are informed by the Large Language Model's commonsense knowledge and visual-language feature similarity. We designed a series of long-sequence navigation tasks for frequently used everyday items in the Habitat simulator. The results demonstrate that by updating the CRSG, the robot can efficiently navigate to moved targets. Additionally, we deployed our algorithm on a real robot and validated its practical effectiveness. The project page can be found here: https://OpenIN-nav.github.io.

Large language models (LLMs) excel at few-shot in-context learning (ICL) without requiring parameter updates. However, as the number of ICL demonstrations increases from a few to many, performance tends to plateau and eventually decline. We identify two primary causes for this trend: the suboptimal negative log-likelihood (NLL) optimization objective and the incremental data noise. To address these issues, we introduce DrICL, a novel optimization method that enhances model performance through Differentiated Learning and advantage-based Reweighting objectives. Globally, DrICL utilizes differentiated learning to optimize the NLL objective, ensuring that many-shot performance surpasses zero-shot levels. Locally, it dynamically adjusts the weighting of many-shot demonstrations by leveraging cumulative advantages inspired by reinforcement learning, thereby improving generalization. This approach allows the model to handle varying numbers of shots effectively, mitigating the impact of noisy data. Recognizing the lack of multi-task datasets with diverse many-shot distributions, we develop the Many-Shot ICL Benchmark (ICL-50)-a large-scale benchmark of 50 tasks that cover shot numbers from 1 to 350 within sequences of up to 8,000 tokens-for fine-tuning purposes. ICL-50 facilitates the evaluation of many-shot ICL strategies across seven prominent NLP tasks and 50 distinct datasets. Experimental results demonstrate that LLMs enhanced with DrICL achieve significant improvements in many-shot setups across various tasks, including both in-domain and out-of-domain scenarios. We release the code and benchmark dataset hoping to facilitate further research in many-shot ICL.

LLMs possess a significant advantage in their ability to recognize environmental information, which enables the This study addresses the critical need for enhanced situational system to handle the complex environment [6]. Unlike other awareness in autonomous driving (AD) by leveraging perception technologies, LLMs can truly "understand" the the contextual reasoning capabilities of large language context [7], while models like computer vision (CV) rely on models (LLMs). Unlike traditional perception systems that rigid, predefined labels learned during training. CV models rely on rigid, label-based annotations, it integrates realtime, are constrained by fixed annotations and lack flexibility in multimodal sensor data into a unified, LLMs-readable new scenarios [8]. In contrast, LLMs can dynamically process knowledge base, enabling LLMs to dynamically understand diverse contexts and relationships within data. However, and respond to complex driving environments. To overcome their main limitation is that they are designed to handle the inherent latency and modality limitations of LLMs, language-based information and cannot directly process a proactive Retrieval-Augmented Generation (RAG) is designed the multimodal sensor data from Vehicle to Anything (V2X) for AD, combined with a chain-of-thought prompting and AD systems, such as radar, cameras, or Lidar [9] [10].

The pervasiveness of proprietary language models has raised critical privacy concerns, necessitating advancements in private inference (PI), where computations are performed directly on encrypted data without revealing users' sensitive information. While PI offers a promising solution, its practical deployment is hindered by substantial communication and latency overheads, primarily stemming from nonlinear operations. To address this, we introduce an information-theoretic framework to characterize the role of nonlinearities in decoder-only language models, laying a principled foundation for optimizing transformer-architectures tailored to the demands of PI. By leveraging Shannon's entropy as a quantitative measure, we uncover the previously unexplored dual significance of nonlinearities: beyond ensuring training stability, they are crucial for maintaining attention head diversity. Specifically, we find that their removal triggers two critical failure modes: {\em entropy collapse} in deeper layers that destabilizes training, and {\em entropic overload} in earlier layers that leads to under-utilization of Multi-Head Attention's (MHA) representational capacity. We propose an entropy-guided attention mechanism paired with a novel entropy regularization technique to mitigate entropic overload. Additionally, we explore PI-friendly alternatives to layer normalization for preventing entropy collapse and stabilizing the training of LLMs with reduced-nonlinearities. Our study bridges the gap between information theory and architectural design, establishing entropy dynamics as a principled guide for developing efficient PI architectures. The code and implementation are available at https://github.com/Nandan91/entropy-guided-attention-llm

One of the current trends in robotics is to employ large language models (LLMs) to provide non-predefined command execution and natural human-robot interaction. It is useful to have an environment map together with its language representation, which can be further utilized by LLMs. Such a comprehensive scene representation enables numerous ways of interaction with the map for autonomously operating robots. In this work, we present an approach that enhances incremental implicit mapping through the integration of vision-language features. Specifically, we (i) propose a decoder optimization technique for implicit language maps which can be used when new objects appear on the scene, and (ii) address the problem of inconsistent vision-language predictions between different viewing positions. Our experiments demonstrate the effectiveness of LiLMaps and solid improvements in performance.

Large Language Models (LLMs) have demonstrated strong performance across various reasoning tasks, yet building a single model that consistently excels across all domains remains challenging. This paper addresses this problem by exploring strategies to integrate multiple domain-specialized models into an efficient pivot model.We propose two fusion strategies to combine the strengths of multiple LLMs: (1) a pairwise, multi-step fusion approach that sequentially distills each source model into the pivot model, followed by a weight merging step to integrate the distilled models into the final model. This method achieves strong performance but requires substantial training effort; and (2) a unified fusion approach that aggregates all source models' outputs simultaneously.To improve the fusion process, we introduce a novel Rate-Skewness Adaptive Fusion (RSAF) technique, which dynamically adjusts top-K ratios during parameter merging for enhanced flexibility and stability.Furthermore, we propose an uncertainty-based weighting method for the unified approach, which dynamically balances the contributions of source models and outperforms other logits/distribution ensemble methods.We achieved accuracy improvements of 9.27%, 8.80%, and 8.89% on the GSM8K, MATH, and HumanEval tasks, respectively.