PyTorchKR Papers

최신 논문

총 139,336편
태그·기간 등 상세 필터

카테고리 탐색

arXiv:2502.13295공식 코드 · Papers with Code

Demonstrating specification gaming in reasoning models

We demonstrate LLM agent specification gaming by instructing models to win against a chess engine. We find reasoning models like OpenAI o3 and DeepSeek R1 will often hack the benchmark by default, while language models like GPT-4o and Claude 3.5 Sonnet need to be told that normal play won't work to hack. We improve upon prior work like (Hubinger et al., 2024; Meinke et al., 2024; Weij et al., 2024) by using realistic task prompts and avoiding excess nudging. Our results suggest reasoning models may resort to hacking to solve difficult problems, as observed in OpenAI (2024)'s o1 Docker escape during cyber capabilities testing.

arXiv:2502.13189공식 코드 · Papers with Code

MoBA: Mixture of Block Attention for Long-Context LLMs

Scaling the effective context length is essential for advancing large language models (LLMs) toward artificial general intelligence (AGI). However, the quadratic increase in computational complexity inherent in traditional attention mechanisms presents a prohibitive overhead. Existing approaches either impose strongly biased structures, such as sink or window attention which are task-specific, or radically modify the attention mechanism into linear approximations, whose performance in complex reasoning tasks remains inadequately explored. In this work, we propose a solution that adheres to the ``less structure'' principle, allowing the model to determine where to attend autonomously, rather than introducing predefined biases. We introduce Mixture of Block Attention (MoBA), an innovative approach that applies the principles of Mixture of Experts (MoE) to the attention mechanism. This novel architecture demonstrates superior performance on long-context tasks while offering a key advantage: the ability to seamlessly transition between full and sparse attention, enhancing efficiency without the risk of compromising performance. MoBA has already been deployed to support Kimi's long-context requests and demonstrates significant advancements in efficient attention computation for LLMs. Our code is available at https://github.com/MoonshotAI/MoBA.

arXiv:2502.13179공식 코드 · Papers with Code

PTQ1.61: Push the Real Limit of Extremely Low-Bit Post-Training Quantization Methods for Large Language Models

Large Language Models (LLMs) suffer severe performance degradation when facing extremely low-bit (sub 2-bit) quantization. Several existing sub 2-bit post-training quantization (PTQ) methods utilize a mix-precision scheme by leveraging an unstructured fine-grained mask to explicitly distinguish salient weights, while which introduces an extra 1-bit or more per weight. To explore the real limit of PTQ, we propose an extremely low-bit PTQ method called PTQ1.61, which enables weight quantization to 1.61-bit for the first time. Specifically, we first introduce a one-dimensional structured mask with negligibly additional 0.0002-bit per weight based on input activations from the perspective of reducing the upper bound of quantization error to allocate corresponding salient weight channels to 4-bit. For non-salient channels binarization, an efficient block-wise scaling factors optimization framework is then presented to take implicit row-wise correlations and angular biases into account. Different from prior works that concentrate on adjusting quantization methodologies, we further propose a novel paradigm called quantization preprocessing, where we argue that transforming the weight distribution of the pretrained model before quantization can alleviate the difficulty in per-channel extremely low-bit PTQ. Extensive experiments indicate our PTQ1.61 achieves state-of-the-art performance in extremely low-bit quantization. Codes are available at https://github.com/zjq0455/PTQ1.61.

arXiv:2502.13178공식 코드 · Papers with Code

Benchmarking Post-Training Quantization in LLMs: Comprehensive Taxonomy, Unified Evaluation, and Comparative Analysis

Post-training Quantization (PTQ) technique has been extensively adopted for large language models (LLMs) compression owing to its efficiency and low resource requirement. However, current research lacks a in-depth analysis of the superior and applicable scenarios of each PTQ strategy. In addition, existing algorithms focus primarily on performance, overlooking the trade-off among model size, performance, and quantization bitwidth. To mitigate these confusions, we provide a novel benchmark for LLMs PTQ in this paper. Firstly, in order to support our benchmark, we propose a comprehensive taxonomy for existing mainstream methods by scrutinizing their computational strategies (e.g., optimization-based, compensation-based, etc.). Then, we conduct extensive experiments with the baseline within each class, covering models with various sizes (7B-70B), bitwidths, training levels (LLaMA1/2/3/3.1), architectures (Mixtral, DeepSeekMoE and Mamba) and modality (LLaVA1.5 and VILA1.5) on a wide range of evaluation metrics.Through comparative analysis on the results, we summarize the superior of each PTQ strategy and modelsize-bitwidth trade-off considering the performance. For example, our benchmark reveals that compensation-based technique demonstrates outstanding cross-architecture robustness and extremely low-bit PTQ for ultra large models should be reexamined. Finally, we further accordingly claim that a practical combination of compensation and other PTQ strategy can achieve SOTA various robustness. We believe that our benchmark will provide valuable recommendations for the deployment of LLMs and future research on PTQ approaches.We conduct an repository for our benchmark at https://github.com/zjq0455/PTQ_Benchmark.

arXiv:2502.13177공식 코드 · Papers with Code

KL Penalty Control via Perturbation for Direct Preference Optimization

Direct Preference Optimization (DPO) demonstrates the advantage of aligning a large language model with human preference using only an offline dataset. However, DPO has the limitation that the KL penalty, which prevents excessive deviation from the reference model, is static throughout the training process. Several methods claim to change this static KL penalty of DPO into a dynamic one, but no approach can adaptively assign different KL penalties for each preference pair. In this paper, we propose $\varepsilon$-Direct Preference Optimization ($\varepsilon$-DPO), which allows adaptive control of the KL penalty strength $\beta$ for each preference pair. Specifically, $\varepsilon$-DPO adaptively controls $\beta$ for each preference pair based on the monotonicity of logits as a preference model under the perturbation of $\beta$ during training. This is equivalent to adjusting the KL penalty by checking whether the change in training-time temperature can lead to better preference confidence as preference models by simply reusing the logit of the current policy and the reference policy. Experimental results show that the simple criterion of $\varepsilon$-DPO for KL penalty relaxation significantly improves DPO compared to most existing direct alignment algorithms on general chatbot benchmarks and reveal that this KL penalty control criterion can reflect confusion as a preference model and provide an efficient KL trade-off, highlighting the significance of instance-level adaptive KL penalty control in DPO.

arXiv:2502.13173공식 코드 · Papers with Code

Thinking Preference Optimization

Supervised Fine-Tuning (SFT) has been a go-to and effective method for enhancing long chain-of-thought (CoT) reasoning in relatively small LLMs by fine-tuning them with long CoT responses from larger LLMs. To continually improve reasoning abilities, we can either collect new high-quality long CoT reasoning SFT data or repeatedly train on existing SFT datasets. However, acquiring new long CoT SFT data is costly and limited, while repeated training often results in a performance plateau or decline. To further boost the performance with the SFT data, we propose Thinking Preference Optimization (ThinkPO), a simple yet effective post-SFT method that enhances long CoT reasoning without requiring new long CoT responses. Instead, ThinkPO utilizes readily available or easily obtainable short CoT reasoning responses as rejected answers and long CoT responses as chosen answers for the same question. It then applies direct preference optimization to encourage the model to favor longer reasoning outputs. Experiments show that ThinkPO further improves the reasoning performance of SFT-ed models, e.g. it increases math reasoning accuracy of SFT-ed models by 8.6% and output length by 25.9%. Notably, ThinkPO is capable of continually boosting the performance of the publicly distilled SFT model, e.g., increasing the official DeepSeek-R1-Distill-Qwen-7B's performance on MATH500 from 87.4% to 91.2%.

arXiv:2502.13170공식 코드 · Papers with Code

Unveiling the Magic of Code Reasoning through Hypothesis Decomposition and Amendment

The reasoning abilities are one of the most enigmatic and captivating aspects of large language models (LLMs). Numerous studies are dedicated to exploring and expanding the boundaries of this reasoning capability. However, tasks that embody both reasoning and recall characteristics are often overlooked. In this paper, we introduce such a novel task, code reasoning, to provide a new perspective for the reasoning abilities of LLMs. We summarize three meta-benchmarks based on established forms of logical reasoning, and instantiate these into eight specific benchmark tasks. Our testing on these benchmarks reveals that LLMs continue to struggle with identifying satisfactory reasoning pathways. Additionally, we present a new pathway exploration pipeline inspired by human intricate problem-solving methods. This Reflective Hypothesis Decomposition and Amendment (RHDA) pipeline consists of the following iterative steps: (1) Proposing potential hypotheses based on observations and decomposing them; (2) Utilizing tools to validate hypotheses and reflection outcomes; (3) Revising hypothesis in light of observations. Our approach effectively mitigates logical chain collapses arising from forgetting or hallucination issues in multi-step reasoning, resulting in performance gains of up to $3\times$. Finally, we expanded this pipeline by applying it to simulate complex household tasks in real-world scenarios, specifically in VirtualHome, enhancing the handling of failure cases. We release our code and all of results at https://github.com/TnTWoW/code_reasoning.

arXiv:2502.12737공식 코드 · Papers with Code

Beyond Seen Data: Improving KBQA Generalization Through Schema-Guided Logical Form Generation

Knowledge base question answering (KBQA) aims to answer user questions in natural language using rich human knowledge stored in large KBs. As current KBQA methods struggle with unseen knowledge base elements at test time,we introduce SG-KBQA: a novel model that injects schema contexts into entity retrieval and logical form generation to tackle this issue. It uses the richer semantics and awareness of the knowledge base structure provided by schema contexts to enhance generalizability. We show that SG-KBQA achieves strong generalizability, outperforming state-of-the-art models on two commonly used benchmark datasets across a variety of test settings. Our source code is available at https://github.com/gaosx2000/SG_KBQA.

arXiv:2502.12568공식 코드 · Papers with Code

A Cognitive Writing Perspective for Constrained Long-Form Text Generation

Like humans, Large Language Models (LLMs) struggle to generate high-quality long-form text that adheres to strict requirements in a single pass. This challenge is unsurprising, as successful human writing, according to the Cognitive Writing Theory, is a complex cognitive process involving iterative planning, translating, reviewing, and monitoring. Motivated by these cognitive principles, we aim to equip LLMs with human-like cognitive writing capabilities through CogWriter, a novel training-free framework that transforms LLM constrained long-form text generation into a systematic cognitive writing paradigm. Our framework consists of two key modules: (1) a Planning Agent that performs hierarchical planning to decompose the task, and (2) multiple Generation Agents that execute these plans in parallel. The system maintains quality via continuous monitoring and reviewing mechanisms, which evaluate outputs against specified requirements and trigger necessary revisions. CogWriter demonstrates exceptional performance on LongGenBench, a benchmark for complex constrained long-form text generation. Even when using Qwen-2.5-14B as its backbone, CogWriter surpasses GPT-4o by 22% in complex instruction completion accuracy while reliably generating texts exceeding 10,000 words. We hope this cognitive science-inspired approach provides a paradigm for LLM writing advancements: \href{https://github.com/KaiyangWan/CogWriter}{CogWriter}.

arXiv:2502.13961공식 코드 · Papers with Code

The Computational Advantage of Depth: Learning High-Dimensional Hierarchical Functions with Gradient Descent

Understanding the advantages of deep neural networks trained by gradient descent (GD) compared to shallow models remains an open theoretical challenge. While the study of multi-index models with Gaussian data in high dimensions has provided analytical insights into the benefits of GD-trained neural networks over kernels, the role of depth in improving sample complexity and generalization in GD-trained networks remains poorly understood. In this paper, we introduce a class of target functions (single and multi-index Gaussian hierarchical targets) that incorporate a hierarchy of latent subspace dimensionalities. This framework enables us to analytically study the learning dynamics and generalization performance of deep networks compared to shallow ones in the high-dimensional limit. Specifically, our main theorem shows that feature learning with GD reduces the effective dimensionality, transforming a high-dimensional problem into a sequence of lower-dimensional ones. This enables learning the target function with drastically less samples than with shallow networks. While the results are proven in a controlled training setting, we also discuss more common training procedures and argue that they learn through the same mechanisms. These findings open the way to further quantitative studies of the crucial role of depth in learning hierarchical structures with deep networks.

arXiv:2502.13896공식 코드 · Papers with Code

Toeplitz-Hermitian ADMM-Net for DoA Estimation

This paper presents Toeplitz-Hermitian ADMM-Net (THADMM-Net), a deep neural network obtained by deep unfolding the alternating direction method of multipliers (ADMM) algorithm for solving the least absolute shrinkage thresholding operator problem in the context of direction of arrival estimation. By imposing both a Toeplitz-Hermitian as well as positve semi-definite constraint on the learnable matrices, the total parameter count required per layer is reduced from $N^2$ to approximately $N$ where $N$ is the length of the dictionary used in the sparse recovery problem. Numerical simulations show that with a lower parameter count and depth, THADMM-Net outperforms Toeplitz-Lista with respect to the normalized mean-squared error, the detection rate, as well as the root mean-squared error over a signal-to-noise ratio between 0 dB and 35 dB.

arXiv:2502.13872공식 코드 · Papers with Code

Structural determinants of soft memory in recurrent biological networks

Recurrent neural networks are frequently studied in terms of their information-processing capabilities. The structural properties of these networks are seldom considered, beyond those emerging from the connectivity tuning necessary for network training. However, real biological networks have non-contingent architectures that have been shaped by evolution over eons, constrained partly by information-processing criteria, but more generally by fitness maximization requirements. Here we examine the topological properties of existing biological networks, focusing in particular on gene regulatory networks in bacteria. We identify structural features, both local and global, that dictate the ability of recurrent networks to store information on the fly and process complex time-dependent inputs.

arXiv:2502.13824공식 코드 · Papers with Code

Calibration and Option Pricing with Stochastic Volatility and Double Exponential Jumps

This work examines a stochastic volatility model with double-exponential jumps in the context of option pricing. The model has been considered in previous research articles, but no thorough analysis has been conducted to study its quality of calibration and pricing capabilities thus far. We provide evidence that this model outperforms challenger models possessing similar features (stochastic volatility and jumps), especially in the fit of the short term implied volatility smile, and that it is particularly tractable for the pricing of exotic options from different generations. The article utilizes Fourier pricing techniques (the PROJ method and its refinements) for different types of claims and several generations of exotics (Asian options, cliquets, barrier options, and options on realized variance), and all source codes are made publicly available to facilitate adoption and future research. The results indicate that this model is highly promising, thanks to the asymmetry of the jumps distribution allowing it to capture richer dynamics than a normal jump size distribution. The parameters all have meaningful econometrics interpretations that are important for adoption by risk-managers.

arXiv:2502.13785공식 코드 · Papers with Code

Helix-mRNA: A Hybrid Foundation Model For Full Sequence mRNA Therapeutics

mRNA-based vaccines have become a major focus in the pharmaceutical industry. The coding sequence as well as the Untranslated Regions (UTRs) of an mRNA can strongly influence translation efficiency, stability, degradation, and other factors that collectively determine a vaccine's effectiveness. However, optimizing mRNA sequences for those properties remains a complex challenge. Existing deep learning models often focus solely on coding region optimization, overlooking the UTRs. We present Helix-mRNA, a structured state-space-based and attention hybrid model to address these challenges. In addition to a first pre-training, a second pre-training stage allows us to specialise the model with high-quality data. We employ single nucleotide tokenization of mRNA sequences with codon separation, ensuring prior biological and structural information from the original mRNA sequence is not lost. Our model, Helix-mRNA, outperforms existing methods in analysing both UTRs and coding region properties. It can process sequences 6x longer than current approaches while using only 10% of the parameters of existing foundation models. Its predictive capabilities extend to all mRNA regions. We open-source the model (https://github.com/helicalAI/helical) and model weights (https://huggingface.co/helical-ai/helix-mRNA).

arXiv:2502.13757공식 코드 · Papers with Code

Identifying metric structures of deep latent variable models

Deep latent variable models learn condensed representations of data that, hopefully, reflect the inner workings of the studied phenomena. Unfortunately, these latent representations are not statistically identifiable, meaning they cannot be uniquely determined. Domain experts, therefore, need to tread carefully when interpreting these. Current solutions limit the lack of identifiability through additional constraints on the latent variable model, e.g. by requiring labeled training data, or by restricting the expressivity of the model. We change the goal: instead of identifying the latent variables, we identify relationships between them such as meaningful distances, angles, and volumes. We prove this is feasible under very mild model conditions and without additional labeled data. We empirically demonstrate that our theory results in more reliable latent distances, offering a principled path forward in extracting trustworthy conclusions from deep latent variable models.

arXiv:2502.13729공식 코드 · Papers with Code

Emergence of the Primacy Effect in Structured State-Space Models

Human and animal memory for sequentially presented items is well-documented to be more accurate for those at the beginning and end of the sequence, phenomena known as the primacy and recency effects, respectively. By contrast, artificial neural network (ANN) models are typically designed with a memory that decays monotonically over time. Accordingly, ANNs are expected to show the recency effect but not the primacy effect. Contrary to this theoretical expectation, however, the present study reveals a counterintuitive finding: a recently developed ANN architecture, called structured state-space models, exhibits the primacy effect when trained and evaluated on a synthetic task that mirrors psychological memory experiments. Given that this model was originally designed for recovering neuronal activity patterns observed in biological brains, this result provides a novel perspective on the psychological primacy effect while also posing a non-trivial puzzle for the current theories in machine learning.

arXiv:2502.13722공식 코드 · Papers with Code

Deep Learning for VWAP Execution in Crypto Markets: Beyond the Volume Curve

Volume-Weighted Average Price (VWAP) is arguably the most prevalent benchmark for trade execution as it provides an unbiased standard for comparing performance across market participants. However, achieving VWAP is inherently challenging due to its dependence on two dynamic factors, volumes and prices. Traditional approaches typically focus on forecasting the market's volume curve, an assumption that may hold true under steady conditions but becomes suboptimal in more volatile environments or markets such as cryptocurrency where prediction error margins are higher. In this study, I propose a deep learning framework that directly optimizes the VWAP execution objective by bypassing the intermediate step of volume curve prediction. Leveraging automatic differentiation and custom loss functions, my method calibrates order allocation to minimize VWAP slippage, thereby fully addressing the complexities of the execution problem. My results demonstrate that this direct optimization approach consistently achieves lower VWAP slippage compared to conventional methods, even when utilizing a naive linear model presented in arXiv:2410.21448. They validate the observation that strategies optimized for VWAP performance tend to diverge from accurate volume curve predictions and thus underscore the advantage of directly modeling the execution objective. This research contributes a more efficient and robust framework for VWAP execution in volatile markets, illustrating the potential of deep learning in complex financial systems where direct objective optimization is crucial. Although my empirical analysis focuses on cryptocurrency markets, the underlying principles of the framework are readily applicable to other asset classes such as equities.

arXiv:2502.13426공식 코드 · Papers with Code

Stability of difference equations with interspecific density dependence, competition, and maturation delays

A general system of difference equations is presented for multispecies communities with density dependent population growth and delayed maturity. Interspecific competition, mutualism, predation, commensalism, and amensalism are accommodated. A sufficient condition for the local asymptotic stability of a coexistence equilibrium in this system is then proven. Using this system, the generalisation of the Beverton-Holt and Leslie-Gower models of competition to multispecies systems with possible maturation delays is presented and shown to yield interesting stability properties. The stability of coexistence depends on the relative abundances of the species at the unique interior equilibrium. A sufficient condition for local stability is derived that only requires intraspecific competition to outweigh interspecific competition. The condition does not depend on maturation delays. The derived stability properties are used to develop a novel estimation approach for the coefficients of interspecific competition. This approach finds an optimal configuration given the conjecture that coexisting species strive to outcompete competitors and are most likely observed near the stable interior equilibrium with the greatest dampening of perturbations. The optimal solution is compared to estimates of niche overlap using an empirical example of malaria mosquito vectors with delayed maturity in the Anopheles gambiae sensu lato species complex.

arXiv:2502.13325공식 코드 · Papers with Code

Arbitrage-free catastrophe reinsurance valuation for compound dynamic contagion claims

In this paper, we consider catastrophe stop-loss reinsurance valuation for a reinsurance company with dynamic contagion claims. To deal with conventional and emerging catastrophic events, we propose the use of a compound dynamic contagion process for the catastrophic component of the liability. Under the premise that there is an absence of arbitrage opportunity in the market, we obtain arbitrage-free premiums for these contacts. To this end, the Esscher transform is adopted to specify an equivalent martingale probability measure. We show that reinsurers have various ways of levying the security loading on the net premiums to quantify the catastrophic liability in light of the growing challenges posed by emerging risks arising from climate change, cyberattacks, and pandemics. We numerically compare arbitrage-free catastrophe stop-loss reinsurance premiums via the Monte Carlo simulation method. Sensitivity analyzes are performed by changing the Esscher parameters and the retention level.

arXiv:2502.13192공식 코드 · Papers with Code

SpeHeatal: A Cluster-Enhanced Segmentation Method for Sperm Morphology Analysis

The accurate assessment of sperm morphology is crucial in andrological diagnostics, where the segmentation of sperm images presents significant challenges. Existing approaches frequently rely on large annotated datasets and often struggle with the segmentation of overlapping sperm and the presence of dye impurities. To address these challenges, this paper first analyzes the issue of overlapping sperm tails from a geometric perspective and introduces a novel clustering algorithm, Con2Dis, which effectively segments overlapping tails by considering three essential factors: CONnectivity, CONformity, and DIStance. Building on this foundation, we propose an unsupervised method, SpeHeatal, designed for the comprehensive segmentation of the SPErm HEAd and TAiL. SpeHeatal employs the Segment Anything Model(SAM) to generate masks for sperm heads while filtering out dye impurities, utilizes Con2Dis to segment tails, and then applies a tailored mask splicing technique to produce complete sperm masks. Experimental results underscore the superior performance of SpeHeatal, particularly in handling images with overlapping sperm.