Verifiers provide important reward signals for reinforcement learning of large language models (LLMs). However, it is challenging to develop or create reliable verifiers, especially for code generation tasks. A well-disguised wrong solution program may only be detected by carefully human-written edge cases that are difficult to synthesize automatically. To address this issue, we propose HARDTESTGEN, an approach to synthesize high-quality test cases for algorithmic coding problems. We curate a comprehensive algorithmic programming dataset HARDTESTS with 26.6k problems and high-quality synthetic tests. Compared with existing tests, HARDTESTGEN tests demonstrate significantly higher accuracy in verifying LLM-generated code (+11.22 percentage points in precision, the percentage of actually correct code within the predicted correct ones). We also show that downstream post-training -- including rejection sampling and reinforcement learning (RL) -- using HARDTESTS verifier results in improved performance of LLM code generation. We open-source our dataset and synthesis pipeline at https://leililab.github.io/HardTests/.
Recent studies have revealed the vulnerability of large language models to adversarial attacks, where adversaries craft specific inputs to induce incorrect or harmful outputs. Although various empirical defenses have been proposed, their worst-case robustness remains poorly understood, raising concerns about how they would behave against stronger future attacks. Here, we study the worst-case robustness of LLMs from both empirical and theoretical perspectives. First, we provide tighter empirical upper bounds for the tested deterministic defenses using stronger white-box attacks, under which most tested configurations achieve nearly 0\% robustness. Second, we formulate randomized-smoothing certification for bounded and binary base functions as fractional and 0-1 knapsack problems, respectively, and derive grouped solvers for two token kernels. For example, on AdvBench, we certify a GPT-4o safety detector smoothed by a uniform kernel to an average certified $\ell_0$ radius of 2.02. In our stated suffix-attack setting, the same kernel certifies a Llama-3-8B safety detector to an average certified radius of 6.41.
As neural language models approach human performance on NLP benchmark tasks, their advances are widely seen as evidence of an increasingly complex understanding of syntax. This view rests upon a hypothesis that has not yet been empirically tested: that word order encodes meaning essential to performing these tasks. We refute this hypothesis in many cases: in the GLUE suite and in various genres of English text, the words in a sentence or phrase can rarely be permuted to form a phrase carrying substantially different information. Our surprising result relies on inference by iterative shuffling (IBIS), a novel, efficient procedure that finds the ordering of a bag of words having the highest likelihood under a fixed language model. IBIS can use any black-box model without additional training and is superior to existing word ordering algorithms. Coalescing our findings, we discuss how shuffling inference procedures such as IBIS can benefit language modeling and constrained generation.
While large language models (LLMs) have rapidly improved their performance on a broad number of tasks, they still often fall short on reasoning tasks. As LLMs become more integrated in diverse real-world tasks, advancing their reasoning capabilities is crucial to their effectiveness in nuanced, complex problems. Wang et al.'s self-consistency framework reveals that sampling multiple rationales before taking a majority vote reliably improves model performance across various closed-answer reasoning tasks. Standard methods based on this framework aggregate the final decisions of these rationales but fail to utilize the semantic information detailed in the step-by-step reasoning paths. Our work introduces semantic self-consistency, enhancing this approach by incorporating and analyzing both the reasoning paths of these rationales in addition to their final decisions before taking a majority vote. These methods not only improve the reliability of reasoning paths but also cause more robust performance on complex reasoning tasks.
Effectively addressing the joint effects of non-IID data and long-tailed distribution remains a key challenge in federated learning. While fine-tuning vision-language models (VLMs) like CLIP has shown promise in tackling non-IID scenarios, it often degrades tail-class performance when both heterogeneity and long-tailed distributions are present, even leading to overall performance falling below the zero-shot baseline. To overcome this, we propose Class-Aware Prompt Tuning (CAPT), a novel framework leveraging pre-trained VLMs to effectively learn from non-IID and long-tailed data. Specifically, CAPT introduces a dual-prompt mechanism that synergizes class-aware and general prompts, enabling the model to preserve class-specific information while capturing global trends. Furthermore, to guide the prompts in learning targeted knowledge, we propose a heterogeneity-aware clustering strategy with two key components: similarity-based clusters aggregate class-aware prompts to reinforce tail-specific knowledge, while heterogeneity-based clusters aggregate general prompts to balance diverse representations and promote robust generalization. Extensive experiments demonstrate that CAPT significantly outperforms state-of-the-art methods, e.g., by improving overall accuracy by 6.1% on ImageNet-LT and 8.8% on Places-LT over baselines, while also substantially enhancing tail-class performance across other benchmarks.
Automated code generation is gaining significant importance in intelligent computer programming and system deployment. However, current approaches often face challenges in computational efficiency and lack robust mechanisms for code parsing and error correction. In this work, we propose a novel framework, PyCapsule, with a simple yet effective two-agent pipeline and efficient self-debugging modules for Python code generation. PyCapsule features sophisticated prompt inference, iterative error handling, and case testing, ensuring high generation stability, safety, and correctness. Empirically, PyCapsule achieves up to 5.7% improvement in success rate on HumanEval, 14.6% on HumanEval-ET, and up to 24.5% on BigCodeBench compared to state-of-the-art methods. Notably, PyCapsule enables the open-source Qwen2.5-Coder-Instruct 7B model to surpass the standalone performance of GPT-4 and GPT-3.5 on HumanEval and MBPP while requiring an average of fewer than 2 API calls per problem compared to up to 17 in competing frameworks. PyCapsule demonstrates broader impacts on advancing lightweight and efficient code generation, making competitive automated programming attainable with cost-effective, locally deployable open-source models.
Byte-Pair Encoding (BPE) is a widely used method for subword tokenization, with origins in grammar-based text compression. It is employed in a variety of language processing tasks such as machine translation or large language model (LLM) pretraining, to create a token dictionary of a prescribed size. Most evaluations of BPE to date are empirical, and the reasons for its good practical performance are not well understood. In this paper we focus on the optimization problem underlying BPE: finding a pair encoding that achieves optimal compression utility. We show that this problem is APX-complete, indicating that it is unlikely to admit a polynomial-time approximation scheme. This answers, in a stronger form, a question recently raised by Zouhar et al. On the positive side, we show that BPE approximates the compression utility of the optimal pair encoding to a worst-case factor between $1/3$ and $(1-e^{-2})/2 \approx 0.4324$. Our results aim to explain the ongoing success of BPE and are, to our knowledge, the first rigorous guarantees on its compression utility that hold for all inputs.
PaECTER is an open-source document-level encoder specific for patents. We fine-tune BERT for Patents with examiner-added citation information to generate numerical representations for patent documents. PaECTER performs better in similarity tasks than current state-of-the-art models used in the patent domain. More specifically, our model outperforms the patent specific pre-trained language model (BERT for Patents) and general-purpose text embedding models (e.g., E5, GTE, and BGE) on our patent citation prediction test dataset on different rank evaluation metrics. PaECTER predicts at least one most similar patent at a rank of 1.32 on average when compared against 25 irrelevant patents. Numerical representations generated by PaECTER from patent text can be used for downstream tasks such as classification, tracing knowledge flows, or semantic similarity search. Semantic similarity search is especially relevant in the context of prior art search for both inventors and patent examiners.
Image Quality Assessment (IQA) is a core task in computer vision. Multimodal methods based on vision-language models, such as CLIP, have demonstrated exceptional generalization capabilities in IQA tasks. To address the issues of excessive parameter burden and insufficient ability to identify local distorted features in CLIP for IQA, this study proposes a visual-language model knowledge distillation method aimed at guiding the training of models with architectural advantages using CLIP's IQA knowledge. First, quality-graded prompt templates were designed to guide CLIP to output quality scores. Then, CLIP is fine-tuned to enhance its capabilities in IQA tasks. Finally, a modality-adaptive knowledge distillation strategy is proposed to achieve guidance from the CLIP teacher model to the student model. Our experiments were conducted on multiple IQA datasets, and the results show that the proposed method significantly reduces model complexity while outperforming existing IQA methods, demonstrating strong potential for practical deployment.
The evaluation of large language models is a complex task, in which several approaches have been proposed. The most common is the use of automated benchmarks in which LLMs have to answer multiple-choice questions of different topics. However, this method has certain limitations, being the most concerning, the poor correlation with the humans. An alternative approach, is to have humans evaluate the LLMs. This poses scalability issues as there is a large and growing number of models to evaluate making it impractical (and costly) to run traditional studies based on recruiting a number of evaluators and having them rank the responses of the models. An alternative approach is the use of public arenas, such as the popular LM arena, on which any user can freely evaluate models on any question and rank the responses of two models. The results are then elaborated into a model ranking. An increasingly important aspect of LLMs is their energy consumption and, therefore, evaluating how energy awareness influences the decisions of humans in selecting a model is of interest. In this paper, we present GEA, the Generative Energy Arena, an arena that incorporates information on the energy consumption of the model in the evaluation process. Preliminary results obtained with GEA are also presented, showing that for most questions, when users are aware of the energy consumption, they favor smaller and more energy efficient models. This suggests that for most user interactions, the extra cost and energy incurred by the more complex and top-performing models do not provide an increase in the perceived quality of the responses that justifies their use.
In the era of Large Language Models (LLMs), alignment has emerged as a fundamental yet challenging problem in the pursuit of more reliable, controllable, and capable machine intelligence. The recent success of reasoning models and conversational AI systems has underscored the critical role of reinforcement learning (RL) in enhancing these systems, driving increased research interest at the intersection of RL and LLM alignment. This paper provides a comprehensive review of recent advances in LLM alignment through the lens of inverse reinforcement learning (IRL), emphasizing the distinctions between RL techniques employed in LLM alignment and those in conventional RL tasks. In particular, we highlight the necessity of constructing neural reward models from human data and discuss the formal and practical implications of this paradigm shift. We begin by introducing fundamental concepts in RL to provide a foundation for readers unfamiliar with the field. We then examine recent advances in this research agenda, discussing key challenges and opportunities in conducting IRL for LLM alignment. Beyond methodological considerations, we explore practical aspects, including datasets, benchmarks, evaluation metrics, infrastructure, and computationally efficient training and inference techniques. Finally, we draw insights from the literature on sparse-reward RL to identify open questions and potential research directions. By synthesizing findings from diverse studies, we aim to provide a structured and critical overview of the field, highlight unresolved challenges, and outline promising future directions for improving LLM alignment through RL and IRL techniques.
Language models (LMs) are used for a diverse range of tasks, from question answering to writing fantastical stories. In order to reliably accomplish these tasks, LMs must be able to discern the modal category of a sentence (i.e., whether it describes something that is possible, impossible, completely nonsensical, etc.). However, recent studies have called into question the ability of LMs to categorize sentences according to modality (Michaelov et al., 2025; Kauf et al., 2023). In this work, we identify linear representations that discriminate between modal categories within a variety of LMs, or modal difference vectors. Analysis of modal difference vectors reveals that LMs have access to more reliable modal categorization judgments than previously reported. Furthermore, we find that modal difference vectors emerge in a consistent order as models become more competent (i.e., through training steps, layers, and parameter count). Notably, we find that modal difference vectors identified within LM activations can be used to model fine-grained human categorization behavior. This potentially provides a novel view into how human participants distinguish between modal categories, which we explore by correlating projections along modal difference vectors with human participants' ratings of interpretable features. In summary, we derive new insights into LM modal categorization using techniques from mechanistic interpretability, with the potential to inform our understanding of modal categorization in humans.
Recent advancements in vision-language models (VLMs) have improved performance by increasing the number of visual tokens, which are often significantly longer than text tokens. However, we observe that most real-world scenarios do not require such an extensive number of visual tokens. While the performance drops significantly in a small subset of OCR-related tasks, models still perform accurately in most other general VQA tasks with only 1/4 resolution. Therefore, we propose to dynamically process distinct samples with different resolutions, and present a new paradigm for visual token compression, namely, VisionThink. It starts with a downsampled image and smartly decides whether it is sufficient for problem solving. Otherwise, the model could output a special token to request the higher-resolution image. Compared to existing Efficient VLM methods that compress tokens using fixed pruning ratios or thresholds, VisionThink autonomously decides whether to compress tokens case by case. As a result, it demonstrates strong fine-grained visual understanding capability on OCR-related tasks, and meanwhile saves substantial visual tokens on simpler tasks. We adopt reinforcement learning and propose the LLM-as-Judge strategy to successfully apply RL to general VQA tasks. Moreover, we carefully design a reward function and penalty mechanism to achieve a stable and reasonable image resize call ratio. Extensive experiments demonstrate the superiority, efficiency, and effectiveness of our method. Our code is available at https://github.com/dvlab-research/VisionThink.
Decoder-only language models, such as GPT and LLaMA, generally decode on the last layer. Motivated by human's hierarchical thinking capability, we propose that a hierarchical decoder architecture could be built with different layers decoding texts simultaneously. Due to limited time and computationally resources, we choose to adapt a pretrained language model into this form of hierarchical decoder. Language heads of the last layer are copied to different selected intermediate layers, and fine-tuned with different task inputs. By thorough experiments, we validate that these selective intermediate layers could be adapted to speak meaningful and reasonable contents, and this paradigm of hierarchical decoder can obtain state-of-the-art performances on multiple tasks such as hierarchical text classification, classification-guided generation, and hierarchical text generation. This study suggests the possibility of a generalized hierarchical reasoner, pretraining from scratch.
Large Language Models (LLMs) have advanced code generation and software automation, but are fundamentally constrained by limited inference-time context and lack of explicit code structure reasoning. We introduce Kodezi Chronos, a next-generation architecture for autonomous code understanding, debugging, and maintenance, designed to operate across ultra-long contexts comprising entire codebases, histories, and documentation, all without fixed window limits. Kodezi Chronos leverages a multi-level embedding memory engine, combining vector and graph-based indexing with continuous code-aware retrieval. This enables efficient and accurate reasoning over millions of lines of code, supporting repository-scale comprehension, multi-file refactoring, and real-time self-healing actions. Our evaluation introduces a novel Multi Random Retrieval benchmark, specifically tailored to the software engineering domain. Unlike classical retrieval benchmarks, this method requires the model to resolve arbitrarily distant and obfuscated associations across code artifacts, simulating realistic tasks such as variable tracing, dependency migration, and semantic bug localization. Chronos outperforms prior LLMs and code models, demonstrating a 23% improvement in real-world bug detection and reducing debugging cycles by up to 40% compared to traditional sequence-based approaches. By natively interfacing with IDEs and CI/CD workflows, Chronos enables seamless, autonomous software maintenance, elevating code reliability and productivity while reducing manual effort. These results mark a critical advance toward self-sustaining, continuously optimized software ecosystems.
Recent progress has been made in region-aware vision-language modeling, particularly with the emergence of the Describe Anything Model (DAM). DAM is capable of generating detailed descriptions of any specific image areas or objects without the need for additional localized image-text alignment supervision. We hypothesize that such region-level descriptive capability is beneficial for the task of Visual Question Answering (VQA), especially in challenging scenarios involving images with dense text. In such settings, the fine-grained extraction of textual information is crucial to producing correct answers. Motivated by this, we introduce DAM-QA, a framework with a tailored evaluation protocol, developed to investigate and harness the region-aware capabilities from DAM for the text-rich VQA problem that requires reasoning over text-based information within images. DAM-QA incorporates a mechanism that aggregates answers from multiple regional views of image content, enabling more effective identification of evidence that may be tied to text-related elements. Experiments on six VQA benchmarks show that our approach consistently outperforms the baseline DAM, with a notable 7+ point gain on DocVQA. DAM-QA also achieves the best overall performance among region-aware models with fewer parameters, significantly narrowing the gap with strong generalist VLMs. These results highlight the potential of DAM-like models for text-rich and broader VQA tasks when paired with efficient usage and integration strategies. Our code is publicly available at https://github.com/Linvyl/DAM-QA.git.
Face anti-spoofing (FAS) aims to construct a robust system that can withstand diverse attacks. While recent efforts have concentrated mainly on cross-domain generalization, two significant challenges persist: limited semantic understanding of attack types and training redundancy across domains. We address the first by integrating vision-language models (VLMs) to enhance the perception of visual input. For the second challenge, we employ a meta-domain strategy to learn a unified model that generalizes well across multiple domains. Our proposed InstructFLIP is a novel instruction-tuned framework that leverages VLMs to enhance generalization via textual guidance trained solely on a single domain. At its core, InstructFLIP explicitly decouples instructions into content and style components, where content-based instructions focus on the essential semantics of spoofing, and style-based instructions consider variations related to the environment and camera characteristics. Extensive experiments demonstrate the effectiveness of InstructFLIP by outperforming SOTA models in accuracy and substantially reducing training redundancy across diverse domains in FAS. Project website is available at https://kunkunlin1221.github.io/InstructFLIP.
Fine, regular, and star triangulations (FRSTs) of four-dimensional reflexive polytopes give rise to toric varieties, within which generic anticanonical hypersurfaces yield smooth Calabi-Yau threefolds. We employ transformers -- deep learning models originally developed for language modeling -- to generate FRSTs across a range of polytope sizes. Our models exhibit efficient and unbiased sampling, and can self-improve through retraining on their own output. These results lay the foundation for AICY: a community-driven platform that combines self-improving machine learning models with a continuously expanding FRST database to explore and catalog the Calabi-Yau landscape.
Chain-of-thought traces have been shown to improve performance of large language models in a plethora of reasoning tasks, yet there is no consensus on the mechanism through which this performance boost is achieved. To shed more light on this, we introduce Causal CoT Graphs (CCGs), which are directed acyclic graphs automatically extracted from reasoning traces that model fine-grained causal dependencies in the language model output. A collection of $1671$ mathematical reasoning problems from MATH500, GSM8K and AIME, and their associated CCGs are compiled into our dataset -- \textbf{KisMATH}. Our detailed empirical analysis with 15 open-weight LLMs shows that (i) reasoning nodes in the CCG are mediators for the final answer, a condition necessary for reasoning; and (ii) LLMs emphasise reasoning paths given by the CCG, indicating that models internally realise structures akin to our graphs. KisMATH enables controlled, graph-aligned interventions and opens up avenues for further investigation into the role of chain-of-thought in LLM reasoning.
Verifying the credibility of Cyber Threat Intelligence (CTI) is essential for reliable cybersecurity defense. However, traditional approaches typically treat this task as a static classification problem, relying on handcrafted features or isolated deep learning models. These methods often lack the robustness needed to handle incomplete, heterogeneous, or noisy intelligence, and they provide limited transparency in decision-making-factors that reduce their effectiveness in real-world threat environments. To address these limitations, we propose LRCTI, a Large Language Model (LLM)-based framework designed for multi-step CTI credibility verification. The framework first employs a text summarization module to distill complex intelligence reports into concise and actionable threat claims. It then uses an adaptive multi-step evidence retrieval mechanism that iteratively identifies and refines supporting information from a CTI-specific corpus, guided by LLM feedback. Finally, a prompt-based Natural Language Inference (NLI) module is applied to evaluate the credibility of each claim while generating interpretable justifications for the classification outcome. Experiments conducted on two benchmark datasets, CTI-200 and PolitiFact show that LRCTI improves F1-Macro and F1-Micro scores by over 5%, reaching 90.9% and 93.6%, respectively, compared to state-of-the-art baselines. These results demonstrate that LRCTI effectively addresses the core limitations of prior methods, offering a scalable, accurate, and explainable solution for automated CTI credibility verification