text string | source string |
|---|---|
confusions), ensuring alignment with prior analysis (Appendix C, C.1). TME-DAG uses a graph-based memory structure to track subtasks and dependencies, ensuring precise updates, while TME-Flat and TME-RandomTRIM struggle with linear context and random intents, respectively. Notes •Scenario Overview: The cooking scenario... | https://arxiv.org/abs/2505.19436v1 |
prepared for dumplings.TME-DAG and TME- RandomTRIM create subtask nodes for each item; TME- Flat tracks in context. 3 Also use celery in dumplings.I’ve up- dated the dumplings recipe to include celery.I’ve updated the task to include using celery in the dumplings.I’ve up- dated the dumplings recipe to include celery.TM... | https://arxiv.org/abs/2505.19436v1 |
Round 1) Output: [ { "intent_type": "new", "subtask_title": "schedule team meeting", "parent_node": null, "dependency_nodes": [] } ] Classification Schema The TRIM module classifies user inputs into structured representations for DAG operations. The schema includes the following fields: •intent_type ∈{new,update ,check... | https://arxiv.org/abs/2505.19436v1 |
arXiv:2505.19439v1 [cs.CL] 26 May 2025 Surrogate Signals from Format and Length: Reinforcement Learning for Solving Mathematical Problems without Ground Truth Answers Rihui Xin†1, Han Liu†1,2, Zecheng Wang†1,3, Yupeng Zhang1, Dianbo Sui3,Xiaolin Hu*2,Bingning Wang*1 1Baichuan Inc.2Tsinghua University3Harbin Institute o... | https://arxiv.org/abs/2505.19439v1 |
limitation has spurred the exploration of alternative training strategies to enable effective RL without relying on explicit ground truth information. Motivated by these challenges, our research endeavors to explore the possibility of training LLMs for mathematical problem-solving using alternative signals instead of g... | https://arxiv.org/abs/2505.19439v1 |
evaluation metrics increase with RL training steps until an "Aha moment". Other studies explore length reward functions’ impacts. Yeo et al. Yeo et al. (2025) observed response lengths decline due to model size and KL divergence penalties. Chen et al. Chen et al. (2025) argued direct length extension training may harm ... | https://arxiv.org/abs/2505.19439v1 |
length may be, they can receive at most 0. Therefore, the final format-length reward can be expressed as: Rfl=Rf+Rl if the format is right. min(0,Rf+Rl)else.(4) 3 By combining the format reward and length reward, we provides an "surrogate signals" for the model’s reinforcement learning, helping to alleviate the issue ... | https://arxiv.org/abs/2505.19439v1 |
of the overall performance improvement in this early phase. However, as the training progresses beyond the 15-step mark, a significant divergence emerges. The performance of the format-only model plateaus, barely showing any im- 1https://github.com/MARIO-Math-Reasoning/MARIO_EVAL 2https://github.com/sympy/sympy 3https:... | https://arxiv.org/abs/2505.19439v1 |
(80.0*) 61.5 (62.7*) 60.0 (62.0*) Correctness (baseline) ✗ 26.7 / 26.7 74.6 / 73.0 57.8 / 56.6 53.0 / 52.1 Format-Only ✓ 26.7 / 26.7 72.6 / 72.8 55.4 / 53.0 51.6 / 50.8 Format-Length ✓ 33.3 /40.0 76.8 /73.0 60.2 / 54.2 56.8 /55.7 Table 1: Accuracy comparison of different models on benchmark datasets (cyan rows denote o... | https://arxiv.org/abs/2505.19439v1 |
the frequency of re- flective words in the generated responses (Figure 3c). Reflective words, including those related to verification ( wait/verify/check ), retrospection ( recall/recheck ), branch ex- ploration ( alternatively ), logical turn or contrast ( however/but/since ), and problem decomposition and step-by-ste... | https://arxiv.org/abs/2505.19439v1 |
latent cognitive abili- ties, explicit ground truth answers become an optional component rather than an essential requirement for high-performance RL in mathematical reasoning tasks. 5.2 Format Learning in RL and SFT Since both traditional RL with ground truth rewards and our format-based RL mainly learn answer formatt... | https://arxiv.org/abs/2505.19439v1 |
instances of ’\\boxed’. However, this problem was resolved after just 15 training steps. The repetition rate then dropped significantly and remained stable throughout the subsequent training process. These findings demonstrate that the format-length reward mechanism effectively balances response length, reasoning depth... | https://arxiv.org/abs/2505.19439v1 |
challenges the notion that ground truth answers are essential for LLM training. The format-length signals offer a practical, efficient alternative, reducing data collection costs. Applicable across mathematical and logical tasks, this approach opens new avenues for LLM training. Future work will optimize signal utiliza... | https://arxiv.org/abs/2505.19439v1 |
Jingyang Yuan, Junjie Qiu, Junlong Li, J. L. Cai, Jiaqi Ni, Jian Liang, Jin Chen, Kai Dong, Kai Hu, Kaige Gao, Kang Guan, Kexin Huang, Kuai Yu, Lean Wang, Lecong Zhang, Liang Zhao, Litong Wang, Liyue Zhang, Lei Xu, Leyi Xia, Mingchuan Zhang, Minghua Zhang, Minghui Tang, Meng Li, Miaojun Wang, Mingming Li, Ning Tian, Pa... | https://arxiv.org/abs/2505.19439v1 |
be aha moment in r1-zero-like training — a pilot study. https://oatllm.notion.site/ oat-zero , 2025a. Notion Blog. Zichen Liu, Changyu Chen, Wenjun Li, Penghui Qi, Tianyu Pang, Chao Du, Wee Sun Lee, and Min Lin. Understanding r1-zero-like training: A critical perspective, 2025b. URL https://arxiv.org/abs/2503.20783 . M... | https://arxiv.org/abs/2505.19439v1 |
Zhiwei He, Linfeng Song, Dian Yu, Juntao Li, Zhuosheng Zhang, Rui Wang, Zhaopeng Tu, Haitao Mi, and Dong Yu. Thoughts are all over the place: On the underthinking of o1-like llms, 2025b. URL https://arxiv.org/abs/2501.18585 . Aiyuan Yang, Bin Xiao, Bingning Wang, Borong Zhang, Ce Bian, Chao Yin, Chenxu Lv, Da Pan, Dian... | https://arxiv.org/abs/2505.19439v1 |
the policy from diverging significantly from the old policy in a single update step. To compute the advantage function At, PPO typically relies on value function estimation combined with Generalized Advantage Estimation (GAE). The value function V(s)parame- terized by ϕ, predicts the expected cumulative reward from sta... | https://arxiv.org/abs/2505.19439v1 |
Figure S2: Accuracy curves on (a) MATH500, (b) AIME2024, and (c) AMC2023 benchmarks training onMath-train dataset. F Case Study See Table S1 and S2. 3 PROMPT :system \nPlease reason step by step, and put your final answer within \\boxed{}. \n user \nYou have seven bags of gold coins. Each bag has the same number of gol... | https://arxiv.org/abs/2505.19439v1 |
smallest x is x = 13 + 8 = 21 . \n- If x = 21 , the total number of coins in the original seven bags is 7 \\times 21 = 147 . \n- Adding the 53 coins from the new bag, the total number of coins is 147 + 53 = 200 . \n\n 10.**Final Check :**\n-Since 200 is not more than 200, we need to find the next smallest x . \n- The n... | https://arxiv.org/abs/2505.19439v1 |
responsible for 12 dogs. Under the company’s new policy, they will be assigned or unassigned new dogs in groups of x dogs. The ratio of Denali’s pay to Nate’s pay would be the same if Denali started walking 4x more dogs and Nate stayed at 12 dogs or if x of Nate’s dogs were reassigned to Denali. Find x if x \\neq0. \na... | https://arxiv.org/abs/2505.19439v1 |
the possibility that the problem might have a different interpretation or constraint. Let’s recheck the problem statement and constraints. The only way the ratio can remain the same is if the number of dogs reassigned is such that the ratio of the new number of dogs is the same as the original ratio. Let’s solve it aga... | https://arxiv.org/abs/2505.19439v1 |
arXiv:2505.19440v1 [cs.CL] 26 May 2025The Birth of Knowledge: Emergent Features across Time, Space, and Scale in Large Language Models Shashata Sawmya1Micah Adler1Nir Shavit1,2 1Massachusetts Institute of Technology 2Red Hat, Inc. {shashata, micah, shanir}@mit.edu Abstract This paper studies the emergence of interpreta... | https://arxiv.org/abs/2505.19440v1 |
residual streams of LLMs across three different axes . Second, we conduct an extensive and fine-grained analysis of feature emergence and evolution in those dimensions across nine broad topical domains that span both the sciences and the arts. By probing into the feature dynamics associated with each topic, our analysi... | https://arxiv.org/abs/2505.19440v1 |
where sparsity emerges only through gradual weight adjustment. In the standard top- ksparse auto-encoder only the kstrongest latents update, so unused dictionary columns “die.” The OpenAI implementation adds two small tricks: multi- k, which reruns the same forward pass with extra seeds of multi- kunits receive gradien... | https://arxiv.org/abs/2505.19440v1 |
(via F1) against parameter count, we chose h= 512 . Why h≪d.The residual-stream dimensionality of the Pythia-12B model is d= 5120 . We intentionally select h≪dfor two reasons. First, the combined benchmark contains only 57 + 14 nominal subject labels, several of which overlap (e.g. high_school_physics vs.college_physic... | https://arxiv.org/abs/2505.19440v1 |
whose labels express the most semantically aligned concepts. The precise formulation is given below. Denote by φtheall-mpnet-base-v2 embedding function. For each neuron j∈Nhiwe compute vj=φ(ℓj). Given a subject q∈ Q with embedding uq=φ(q), the cosine similarity is sj,q=u⊤ qvj ∥uq∥2∥vj∥2. (3) Fixing a threshold τ= 0.3, ... | https://arxiv.org/abs/2505.19440v1 |
coherence yet significantly different from both input and output spaces; and the output block (Layers 36), realigning representations toward the prediction task. Figure 4: Cosine similarity (left) reveals three macro blocks (embedding, processing core, output), while SAE probes (right) show that feature directions are ... | https://arxiv.org/abs/2505.19440v1 |
than 5%of the labelled concepts. A single transition—between the 160 M and 410 M checkpoints—raises the activation rate by +92.9 pp to∼95%. Beyond this point activation saturates, peaking at 2.8 B parameters and remaining above 98% for all larger scales considered. Figure 5: Concept-activation saturation with model sca... | https://arxiv.org/abs/2505.19440v1 |
One limitation is that we did not conduct a finer-grained analysis to elucidate the underlying reasons for these emergent patterns, such as the specific contributions of training data distributions or internal network circuitry to the observed feature activations. Additionally, our choice of sparse autoencoders and spe... | https://arxiv.org/abs/2505.19440v1 |
arXiv preprint arXiv:2406.04093 , 2024. [15] Trenton Bricken, Adly Templeton, Joshua Batson, Brian Chen, Adam Jermyn, Tom Conerly, Nick Turner, Cem Anil, Carson Denison, Amanda Askell, et al. Towards monosemanticity: Decomposing language models with dictionary learning. Transformer Circuits Thread , 2, 2023. [16] Peter... | https://arxiv.org/abs/2505.19440v1 |
Influence 0.476 1.00 305 Decision-Making in Institutions 0.418 1.00 23 Environmental Ethics 0.405 1.00 269 Human-related Processes 0.371 0.91 457 Mill’s Utilitarian Philosophy 0.369 1.00 195 Mathematical Problem Solving 0.368 1.00 492 Interest Group Influence 0.357 1.00 294 Long-term Consequences 0.349 1.00 264 Cost an... | https://arxiv.org/abs/2505.19440v1 |
mid-stack, and then re-emerge when the Layer-36 SAE is probed. However, the precise silence window and reactivation point vary by domain—e.g., History and Business neurons vanish after Block 3 but resurface sharply at the final block, whereas Philosophy keeps a faint 3–4 % signal until about Block 11 before dropping of... | https://arxiv.org/abs/2505.19440v1 |
Vibe Coding vs. Agentic Coding: Fundamentals and Practical Implications of Agentic AI Ranjan Sapkota∗‡, Konstantinos I. Roumeliotis†, Manoj Karkee∗‡ ∗Cornell University, Department of Biological and Environmental Engineering, USA †University of the Peloponnese, Department of Informatics and Telecommunications, Tripoli,... | https://arxiv.org/abs/2505.19443v1 |
and aligned with human rea- soning, rather than merely a transcription of formal logic into text? Vibe coding attempts to answer in the affirmative proposing a new semiotic contract between the human mind and generative machines. The rise of vibe coding parallels the rapid advance- ment of foundation models and the gro... | https://arxiv.org/abs/2505.19443v1 |
intent. Moreover, the increasing prevalence of agentic systems raises critical questions around safety, reliability, and trustworthiness in software generated or managed with minimal human oversight. As organizations, developers, and researchers embrace these technologies, a dynamic understanding of the boundaries, aff... | https://arxiv.org/abs/2505.19443v1 |
Formulation): This involves a multi-layered approach to defining the problem for the LLM. It begins with Logical Thinking (the core what) [36], [37], progresses toAnalytical Thinking (how users interact, high- level components) [38]–[40], then to Computational Thinking [41] (structuring the problem into mod- ules, rule... | https://arxiv.org/abs/2505.19443v1 |
development environment, often managed by the developer. This sep- aration between generation and execution necessitates careful testing and integration, as the LLM does not inherently possess a runtime understanding of the code it produces in most Vibe Coding scenarios. This model thrives in creative and exploratory d... | https://arxiv.org/abs/2505.19443v1 |
step software development workflow typical of advanced agentic coding systems. •Maintain Long-Term Context: Codex maintains session state over complex multi-step tasks, man- aging API keys, dependencies, and environment variables Codex, OpenAI. Persistent memory and vector store integration enable agents to reference e... | https://arxiv.org/abs/2505.19443v1 |
[110], [111]. This includes capabili- ties such as modifying configuration files, running shell commands, or interacting with Git repositories [112]. A critical feature is the presence of memory and context management [113], [114], facilitating persis- tent state tracking across multi-step workflows. Agents leverage bo... | https://arxiv.org/abs/2505.19443v1 |
and may attempt self-correction, with human intervention required for complex or novel failures [93], [94]. Typical Task Scope Component-level tasks: generating functions, UI elements, simple scripts, unit tests for specific functions, drafting documentation, refactoring small code sections [95].Feature-level or even s... | https://arxiv.org/abs/2505.19443v1 |
strategic guide , reviewing real-time logs, intermediate artifacts, and agent-generated plans [129], [130]. This includes evaluating execution traces, test results, and change diffs [131], [132]. Intervention may be necessary when the agent encounters ambiguous requirements, edge cases beyond its training distribu- tio... | https://arxiv.org/abs/2505.19443v1 |
A. Execution Models: Comparative Analysis of Archi- tectural Design 1) Vibe Coding Interfaces and Developer-Driven Ex- ecution: Vibe coding architectures operate primar- ily through lightweight, stateless interfaces where LLMs serve as code-generation engines embedded in developer-centric environments such as IDEs, bro... | https://arxiv.org/abs/2505.19443v1 |
testing pipeline.Integrated QA loop with automatic test synthesis, execution, and patching. Agents evaluate outputs, revise plans, and regenerate on failure. Security and Guardrails Basic static analysis or organizational linters applied post-generation. Security remains human-enforced.Embedded static and dynamic scann... | https://arxiv.org/abs/2505.19443v1 |
of situational awareness and task- level memory often leads to hallucinated or ambiguous outputs. Prompt Engineering Insight Specific prompt: “Add role-based access control using JWTs and restrict admin endpoints.” Vague prompt: “Make this more secure.” Outcome: The former yields focused mid- dleware code with user rol... | https://arxiv.org/abs/2505.19443v1 |
the AI injects code with hardcoded credentials, insecure API calls [99], [171], or unsafe permission scopes problems often observed in rapid prototyping workflows [89]. Common Vibe Coding Risks •Hardcoded secrets: Generated code may embed plaintext API keys or passwords. •Insecure defaults: Lack of input sanitiza- tion... | https://arxiv.org/abs/2505.19443v1 |
mech- anisms [19], whereas vibe coding tools operate in a generation-first manner, relying on human oversight for execution, testing, and risk mitigation [182]. To illustrate these distinctions, consider the task of implementing a RESTful (RESTful API JWT-based au- thentication system. In a vibe coding workflow, the de... | https://arxiv.org/abs/2505.19443v1 |
A. Developer Roles and Mental Models 1) Vibe Coding: Dialogic Creation and Exploratory Interaction: Vibe Coding emphasizes an interactive, conversational dynamic between the developer and the LLM. Developers are engaged as co-creators, navigating design and implementation decisions through iterative prompt-response cyc... | https://arxiv.org/abs/2505.19443v1 |
Developer: Reviews logs and approves pull request. C. Comparative Analysis: Developer Engagement and Workflow Suitability The differing interaction paradigms of Vibe Coding and Agentic Coding are reflected not only in architec- tural and cognitive models, but also in their practical workflow characteristics. From the r... | https://arxiv.org/abs/2505.19443v1 |
This highlights how agentic systems can automate extensive codebase transformations with minimal oversight. Agentic in Practice: Python 2 to 3 Migration Prompt: “Upgrade all Python code to version 3.x.” Agent: •Parses source files for deprecated syntax. •Applies automatic and rule-based refactor- ing. •Executes test su... | https://arxiv.org/abs/2505.19443v1 |
conducive to exploratory development. In contrast, Agentic Coding shifts much of this responsibility to autonomous agents equipped with runtime monitoring, log analysis, and roll- back mechanisms. This section systematically analyzes the validation strategies, developer effort, and debugging affordances inherent in eac... | https://arxiv.org/abs/2505.19443v1 |
Assistants: Vibe coding tools prioritize acces- sibility, immediacy, and seamless integration into the developer’s creative flow. These platforms typically use prompt-response loops and are embedded within IDEs, browsers, or chat interfaces, enabling on-the-fly code generation, inline explanations, and lightweight cont... | https://arxiv.org/abs/2505.19443v1 |
for production environments, and lack of built-in quality assurance make it less ideal for scalable engineering tasks. In contrast, Agentic Coding emphasizes automation, structure, and reliability. Agents can autonomously per- form task planning, testing, and code integration, re- ducing developer burden and increasing... | https://arxiv.org/abs/2505.19443v1 |
lution (MTTR) by running test suites, applying patches, and updating changelogs without developer intervention ideal for mission-critical services. C. Comparative Analysis: Vibe vs. Agentic in Practice TABLE VI: Practical Use Case Comparison of Agentic vs. Vibe Coding Scenario Agentic Strengths Vibe Strengths Codebase ... | https://arxiv.org/abs/2505.19443v1 |
integration of data fluency and visual storytelling [228], [229], helping domain experts bridge the gap between backend analytics and user-facing visualizations us- ing conversational coding workflows. 3)Daily Email Report Automation Vibe coding excels in automating routine workflows such as scheduled email reports [15... | https://arxiv.org/abs/2505.19443v1 |
Tailwind CSS with clearly delineated sections. It generates layouts with responsive flexbox or grid-based posi- tioning, embedded SVG icons, call-to-action (CTA) buttons, and form validation logic. The landing page is typically annotated with placeholder text and sample imagery, which the developer can re- place for cu... | https://arxiv.org/abs/2505.19443v1 |
(TDD) or onboarding new engineers to legacy systems who need to create regression tests. By automating repet- itive scaffolding of test logic, vibe coding helps enforce consistent testing structures, reducing the manual effort typically required for frontend valida- tion and increasing reliability in component-driven a... | https://arxiv.org/abs/2505.19443v1 |
environment. The model produces event-driven scripts using onEdit(e) handlers that evaluate cell values and apply conditional for- matting via setBackground() methods. It also includes logic to optimize performance (e.g., range limiting) and optional enhancements like logging or undo triggers. This application is espec... | https://arxiv.org/abs/2505.19443v1 |
during post-deployment phases. 4)CI/CD Pipeline Automation Setting up and maintaining CI/CD pipelines is es- sential but repetitive ideal for delegation to agentic systems. When asked to “Set up and maintain a CI/CD pipeline that builds, tests, and deploys our microservices to AWS,” the agent scaffolds GitHub Actions o... | https://arxiv.org/abs/2505.19443v1 |
The final report includes be- fore/after metrics, call graphs, and optimization rationale. This use case demonstrates how agentic coding tools can enforce non-functional require- ments through measurable performance metrics. 9)End-to-End Feature Implementation Agentic coding systems are capable of implement- ing comple... | https://arxiv.org/abs/2505.19443v1 |
continuous loop of refinement via real-time feedback. However, challenges remain in ensuring explainability, safe prompt handling, and seamless cross-platform integration. B. Enterprise and Educational Adoption The convergence of AI coding paradigms is not only theoretical but increasingly visible in industry and ed- u... | https://arxiv.org/abs/2505.19443v1 |
studies and interdisciplinary insightsfrom software engineering, human-computer interaction, and AI ethics. A. Limitations of Agentic Coding Agentic coding systems, while promising high degrees of autonomy, introduce risks that arise from reduced human oversight, opaque execution logic, and uncon- trolled access to cri... | https://arxiv.org/abs/2505.19443v1 |
that ignore runtime dependencies or system architecture constraints. This makes them ideal for scaffolding or ideation, but subop- timal for system-level implementation unless paired with structured review protocols and toolchain integration. Finally, the rapid, iterative style of vibe coding can erode long-term code q... | https://arxiv.org/abs/2505.19443v1 |
This architectural pattern will mirror human software teams, enabling software construction to scale without linear increases in human supervision. C. Memory, Context, and Long-Term Adaptation Agentic AI will only succeed in production settings if it can reason across time, projects, and usage con- texts. Future system... | https://arxiv.org/abs/2505.19443v1 |
[248]–[250], which delivered domain-specific instruction using scripted pedagogical rules. Concur- rently, mobile agents such as General Magic’s Tele- script [251] facilitated lightweight task execution across distributed networks [252]. Behavioral models like the Belief-Desire-Intention (BDI) framework (e.g., PRS, JAM... | https://arxiv.org/abs/2505.19443v1 |
software lifecycle. Our investigation into their execution models, safety protocols, and developer interaction revealed how agen- tic systems enable delegation of complex development tasks while introducing new challenges in explainability, governance, and secure operation. We further compared their autonomy mechanisms... | https://arxiv.org/abs/2505.19443v1 |
ICFP, pp. 1–29, 2017. [7] M. Balliu, M. Dam, and R. Guanciale, “Automating information flow analysis of low level code,” in Proceedings of the 2014 ACM SIGSAC Conference on Computer and Communications Security , pp. 1080–1091, 2014. [8] A. Gadde, “Democratizing software engineering through gen- erative ai and vibe codi... | https://arxiv.org/abs/2505.19443v1 |
Intelligent User Interfaces , pp. 1529–1547, 2025. [24] M. Yin and R. Xiao, “Vibes: Exploring viewer spatial interac- tions as direct input for livestreamed content,” arXiv preprint arXiv:2504.09016 , 2025. [25] T. De Jong, “Learning and instruction with computer simula- tions,” Education and Computing , vol. 6, no. 3-... | https://arxiv.org/abs/2505.19443v1 |
W. Choo, T. J.-J. Li, S. T. Perrault, and T. W. Malone, “A taxonomy for human-llm inter- action modes: An initial exploration,” in Extended Abstracts of the CHI Conference on Human Factors in Computing Systems , pp. 1–11, 2024. [40] X. Ma, S. Mishra, A. Liu, S. Y . Su, J. Chen, C. Kulkarni, H.- T. Cheng, Q. Le, and E. ... | https://arxiv.org/abs/2505.19443v1 |
Tian, Y . Ye, Y . Qin, X. Cong, Y . Lin, Y . Pan, Y . Wu, H. Hui, W. Liu, Z. Liu, et al. , “Debugbench: Evaluating de- bugging capability of large language models,” arXiv preprint arXiv:2401.04621 , 2024.[53] C. Lee, C. S. Xia, L. Yang, J.-t. Huang, Z. Zhu, L. Zhang, and M. R. Lyu, “A unified debugging approach via llm... | https://arxiv.org/abs/2505.19443v1 |
A. Khan, E. McLean, C. Smith, W. Barfuss, J. Foerster, T. Gaven ˇciak, et al. , “Multi-agent risks from advanced ai,” arXiv preprint arXiv:2502.14143 , 2025. [68] A. Kasirzadeh and I. Gabriel, “Characterizing ai agents for alignment and governance,” arXiv preprint arXiv:2504.21848 , 2025. [69] M. Tufano, A. Agarwal, J.... | https://arxiv.org/abs/2505.19443v1 |
ing, requirements elicitation, and software design,” in Genera- tive ai for effective software development , pp. 71–108, Springer, 2024. [84] J. Wang, Y . Dai, Y . Zhang, Z. Ma, W. Li, and J. Chai, “Training turn-by-turn verifiers for dialogue tutoring agents: The curious case of llms as your coding tutors,” arXiv prep... | https://arxiv.org/abs/2505.19443v1 |
an open-source multi- agentic system for automated end-to-end development of ra- diomics and deep learning models for medical imaging,” arXiv preprint arXiv:2505.03785 , 2025. [99] A. Mohsin, H. Janicke, A. Wood, I. H. Sarker, L. Maglaras, and N. Janjua, “Can we trust large language models gener- ated code? a framework... | https://arxiv.org/abs/2505.19443v1 |
, 2025. [114] E. Miehling, K. N. Ramamurthy, K. R. Varshney, M. Riemer, D. Bouneffouf, J. T. Richards, A. Dhurandhar, E. M. Daly, M. Hind, P. Sattigeri, et al. , “Agentic ai needs a systems theory,” arXiv preprint arXiv:2503.00237 , 2025. [115] M. Fabian and M. Dold, “Agentic preferences: a foundation for nudging when ... | https://arxiv.org/abs/2505.19443v1 |
Wang, D. Manocha, F. Huang, A. Koppel, and S. Ganesh, “Collab: Controlled decoding using mixture of agents for llm alignment,” arXiv preprint arXiv:2503.21720 , 2025. [130] D. Moshkovich, H. Mulian, S. Zeltyn, N. Eder, I. Skarbovsky, and R. Abitbol, “Beyond black-box benchmarking: Observ- ability, analytics, and optimi... | https://arxiv.org/abs/2505.19443v1 |
programming support,” in Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems , pp. 1–21, 2025. [145] J. Zamfirescu-Pereira, E. Jun, M. Terry, Q. Yang, and B. Hart- mann, “Beyond code generation: Llm-supported exploration of the program design space,” in Proceedings of the 2025 CHI Conference on... | https://arxiv.org/abs/2505.19443v1 |
“Aios: Llm agent operating system,” arXiv preprint arXiv:2403.16971 , 2024. [159] Y . Li, H. Wen, W. Wang, X. Li, Y . Yuan, G. Liu, J. Liu, W. Xu, X. Wang, Y . Sun, et al. , “Personal llm agents: Insights and survey about the capability, efficiency and security,” arXiv preprint arXiv:2401.05459 , 2024. [160] J. Liu, C.... | https://arxiv.org/abs/2505.19443v1 |
(ICRA) , pp. 14435–14442, IEEE, 2024. [173] J. He, R. Ghosh, K. Walia, J. Chen, T. Dhadiwal, A. Hazel, and C. Inguva, “Frontiers of large language model-based agentic systems-construction, efficacy and safety,” in Proceedings of the 33rd ACM International Conference on Information and Knowledge Management , pp. 5526–55... | https://arxiv.org/abs/2505.19443v1 |
S. Thakur, A. Araujo, M. Blis- tein, A. Fr ¨ommgen, V . Hellendoorn, and S. Chandra, “Prompt- ing llms for code editing: Struggles and remedies,” arXiv preprint arXiv:2504.20196 , 2025. [191] Y . Cui, “Tests as prompt: A test-driven-development benchmark for llm code generation,” arXiv preprint arXiv:2505.09027 , 2025.... | https://arxiv.org/abs/2505.19443v1 |
, “The prompt report: A systematic survey of prompting techniques,” arXiv preprint arXiv:2406.06608 , vol. 5, 2024. [208] J. Robertson, C. Ferreira, E. Botha, and K. Oosthuizen, “Game changers: A generative ai prompt protocol to enhance human-ai knowledge co-construction,” Business Horizons , vol. 67, no. 5, pp. 499–51... | https://arxiv.org/abs/2505.19443v1 |
html code,” Automated Software Engineering , vol. 32, no. 2, pp. 1–24, 2025. [224] R. T ´oth, T. Bisztray, and L. Erd ˝odi, “Llms in web development: Evaluating llm-generated php code unveiling vulnerabilities and limitations,” in International Conference on Computer Safety, Reliability, and Security , pp. 425–437, Spr... | https://arxiv.org/abs/2505.19443v1 |
deterministic programs and their properties,” Information and control , vol. 68, no. 1-3, pp. 254–270, 1986. [243] F. Wagner, R. Schmuki, T. Wagner, and P. Wolstenholme, Mod- eling software with finite state machines: a practical approach . Auerbach Publications, 2006. [244] A. Drumea and C. Popescu, “Finite state mach... | https://arxiv.org/abs/2505.19443v1 |
BizFinBench: A Business-Driven Real-World Financial Benchmark for Evaluating LLMs Guilong Lu∗1, Xuntao Guo∗1,2, Rongjunchen Zhang♠1, Wenqiao Zhu1, and Ji Liu1 1HiThink Research 2Harbin Institute of Technology Abstract Large language models excel in general tasks, yet assessing their reliability in logic -heavy, precisi... | https://arxiv.org/abs/2505.19457v1 |
multi-step reasoning over real-world data, generally with adversarial or noisy context Du et al. [2024]. Despite the emergence of financial benchmarks such as FinEval Zhang et al. [2023b], existing approaches treat financial tasks as general document Query-Answering (QA) Wang et al. [2024], lacking structured inputs an... | https://arxiv.org/abs/2505.19457v1 |
evaluation accuracy and reliability through three core mechanisms: evaluation dimension disentanglement, sequential correction generation, and reference-aligned as- sessment. By integrating IteraJudge into BizFinBench, we establish a rigorous and interpretable evaluation pipeline for LLM performance in high-stakes fina... | https://arxiv.org/abs/2505.19457v1 |
for evaluating models in the rapidly evolving field of digital finance. In the context of Chinese financial benchmarks, several recent datasets have been released, including CFBenchmark Lei et al. [2023], which focuses on Chinese financial text analysis; DISC-FINSFT Chen et al. [2023], designed for financial sentiment ... | https://arxiv.org/abs/2505.19457v1 |
design of BizFinBench, a comprehensive benchmark specialized for evaluating LLMs in financial domains. Compared to previous datasets, BizFinBench places a strong emphasis on business practicality and real-world applicability, aiming to bridge the gap between academic evaluation and the complex challenges encountered in... | https://arxiv.org/abs/2505.19457v1 |
435 533 Next, we collect relevant contextual data from internal financial databases and external sources based on the content of each user query. This includes stock prices, historical trading data, financial news, company disclosures, and so on. These sources are directly related to the query topic. For example, as il... | https://arxiv.org/abs/2505.19457v1 |
4,556 tokens. This broad range reflects the complexity and heterogeneity of real-world financial scenarios and presents a meaningful challenge for models to demonstrate their ability to process both short and long financial texts effectively. 3.3 IteraJudge: An Incremental Multi-Dimensional Evaluation Framework As show... | https://arxiv.org/abs/2505.19457v1 |
97.34 35.00 54.94 54.41 67.53 Qwen3-1.7B 77.40 35.80 33.40 75.82 73.81 78.62 22.40 48.53 11.23 50.78 Qwen3-4B 83.60 47.40 50.00 78.19 82.24 80.16 42.20 50.51 25.19 59.94 Qwen3-14B 84.20 58.20 65.80 82.19 84.12 92.91 33.00 52.31 50.70 67.05 Qwen3-32B 83.80 59.60 64.60 85.12 85.43 95.37 39.00 52.26 49.19 68.26 Xuanyuan3-... | https://arxiv.org/abs/2505.19457v1 |
smallest 1.7B to the largest 32B, across nearly all tasks. Second, FTR emerges as a particularly challenging task, with a substantial score gap of 32.19 points between the top performer, GPT-4o (76.20), and lower-performing models like Llama-3.1-8B-Instruct (2.91), highlighting the need for targeted optimization in tem... | https://arxiv.org/abs/2505.19457v1 |
Yang, Julian McAuley, Linda Petzold, and William Yang Wang. A survey on large language models for critical societal domains: Finance, healthcare, and law. arXiv preprint arXiv:2405.01769 , 2024. Che Liu, Yingji Zhang, Dong Zhang, Weijie Zhang, Chenggong Gong, Haohan Li, Yu Lu, Shilin Zhou, Yue Lu, Ziliang Gan, et al. N... | https://arxiv.org/abs/2505.19457v1 |
, 2023. Zhiyu Chen, Wenhu Chen, Charese Smiley, and et al. Sameena Shah. Finqa: A dataset of numerical reasoning over financial data, 2022a. Zhiyu Chen, Shiyang Li, Charese Smiley, Zhiqiang Ma, Sameena Shah, and William Yang Wang. Con- vfinqa: Exploring the chain of numerical reasoning in conversational finance questio... | https://arxiv.org/abs/2505.19457v1 |
described in Section 4. All evaluation metrics are clearly defined to facilitate independent verification and replication of our experiments by the research community. B Limitations In this work, we propose a novel benchmark and conduct a comprehensive analysis of different LLMs’ capabilities in solving financial busin... | https://arxiv.org/abs/2505.19457v1 |
09/06/2024 Finance Llama-3.1-Instruct Dubey et al. [2024] 8B,70B ✓ Local 07/24/2024 General Llama 4 109B* ✓ Local 04/05/2025 General DeepSeek-V3 Liu et al. [2024] 671B* ✓ Local 12/26/2024 General DeepSeek-R1 Liu et al. [2024] 671B* ✓ Local 12/26/2024 General QwQ-32B Team [2025] 32B ✓ Local 03/06/2025 General DeepSeek-R... | https://arxiv.org/abs/2505.19457v1 |
938.7 1,064 ER 1,919 2,569 2,178.5 600 FNC 287 2,698 650.5 581 FDD 26 645 310.9 1,461 FTR 203 8,265 1,162.0 514 FQA 5 45 21.7 990 SP 1,254 5,532 4,498.1 497 F Dataset example G Other 16 你是一位经验丰富的金融数据分析师，请根据以下 < 数据参考 > ，回答出 < 用户问题 > ，并给出你的理由。你需要以指定的输出格式回答 ### 数据参考 编号: 1 取数问句 : 百度 2024 年 8 月每日的收盘价 取数结果 : 为您找到 1 条数据 编号: 2... | https://arxiv.org/abs/2505.19457v1 |
1.55% 。 6. **2025 年 1 月 27 日 ** ：股票大幅上涨 10.06% ，显示出市场情绪的强烈反弹。 7. **2025 年 2 月 5 日至 2 月 6 日 ** ：股票再次大幅下跌 10.00% ，随后在 2 月 6 日继续下跌 6.19% ，市场情绪再次转为悲观。 ### 数值分析 - **2025年 2 月 6 日的收盘价 ** ： 5.91 元 - **2025年 2 月 6 日的开盘价 ** ： 5.67 元 - **2025年 2 月 6 日的最高价 ** ： 6.04 元 - **2025年 2 月 6 日的最低价 ** ： 5.67 元 ### 总体趋势 - **波动性 ** ：在这段时间内，... | https://arxiv.org/abs/2505.19457v1 |
日 举办2023 年财报业绩沟通会，并且贾跃亭表示愿意将乐视作为个人 IP 商业化的重要合作伙伴，助力乐视超级电视重回巅峰。这一消 息提振了市场对法拉第未来业务协同和未来战略方向的信心。第二，新闻 4 指出拜登政府计划对中国进口电动汽车征收四倍关税，而 法拉第未来作为唯一在美国本地发展的中国背景电动车概念股，不会受到该政策的直接影响，反而凸显其独特的并购价值和竞争优 势，吸引了大量游资炒作，推动股价暴涨。 " }Anomalous Event Attribution 序号 时间 新闻 1 2023年 9 月 5 日公司半年度业绩交流会中提到，目前隆基绿能 BC 产能主要为 30GW HPBC （高效复合钝化背接触技术）电池项 目，并... | https://arxiv.org/abs/2505.19457v1 |
": " 纺织行业景气度持续回升 ", "摘要 ": " 海关总署数据显示 12 月纺织品出口同比大增 18% ，行业迎来量价齐升 周期" } ], "相关的最新研报 ": [], "用户问句 ": "嘉麟杰这个量价配合，是不是有可能走出一波主升浪？ " }Emotion RecognitionFigure 15: An example instance from the Emotion Recognition. 23 Figure 16: An example instance from the Stock Price Prediction. 24 Question: 你是一位经验丰富的金融数据分析师，请根据以下 < 数据参考 >... | https://arxiv.org/abs/2505.19457v1 |
arXiv:2505.19472v2 [cs.CL] 28 May 2025Balancing Computation Load and Representation Expressivity in Parallel Hybrid Neural Networks Mohammad Mahdi Moradi Department of Computer Science, Concordia University Ascend Team, Huawei Technologies mohammad.mahdi.moradi@h-partners.comWalid Ahmed Ascend Team Toronto Research Cen... | https://arxiv.org/abs/2505.19472v2 |
interactions. Hybrid networks have thus been proposed to combine the strengths of both, attention and SSM mechanisms to capture both short and long range dependencies. The two can be operative in sequence or in parallel. Our work presented in this paper is in the parallel category. Hymba (Dong et al. [2024]) is the onl... | https://arxiv.org/abs/2505.19472v2 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.