text string | source string |
|---|---|
and action steps, teaching it a behavioral paradigm of alternating reasoning with action, while preserving its original reasoning capabilities as much as possible. Following the empirical findings of Chen et al. [28,6], Zhang et al. [22], to avoid interference from external feedback during learning, we mask out loss co... | https://arxiv.org/abs/2505.22648v1 |
interaction spans from <think> to </tool_response> . The rollout concludes with the generation of <answer> and </answer> , following the final thought . Reward Design The reward design plays a critical role during the RL training process [ 25]. Our reward system mainly consists of two types of rewards, score format and... | https://arxiv.org/abs/2505.22648v1 |
13.6 3.8 2.5 3.3 3.1 RAG 12.8 11.8 8.3 11.8 23.1 14.3 11.3 15.3 Qwen-2.5-72B Base 20.5 13.5 0.0 14.6 9.4 7.1 3.3 6.3 GPT-4o Base 23.1 15.4 8.3 17.5 6.7 6.0 4.2 5.5 QwQ-32BBase 30.8 15.4 25.0 22.3 7.5 2.1 4.6 4.3 RAG 33.3 36.5 8.3 32.0 36.9 26.1 33.5 31.2 DeepSeek-R1-671B Base 43.6 26.9 8.3 31.1 5.0 11.8 11.3 10.0 Close... | https://arxiv.org/abs/2505.22648v1 |
long trajectories [ 38,39]. Nevertheless, we observe a notable improvement in consistency, with a 30% increase in the number of problems answered correctly across all three attempts. This suggests that continued on-policy optimization yields limited benefits for LRMs in agentic tasks. Our best-performing model achieves... | https://arxiv.org/abs/2505.22648v1 |
ferent temperatures. Figure 5: Analysis on RL algorithm, emergent agency, and agent environments using GAIA bench- mark. RL enables longer reasoning processes and supports more complex agentic action. As demon- strated by the results on Qwen-32B in Figure 5b, we observe that SFT leads to more frequent action generation... | https://arxiv.org/abs/2505.22648v1 |
generalization performance when confronted with adaptive operational contexts [ 10,50–52]. Building upon these limitations, RL-based methods [ 7,10,10,50] have demonstrated remarkable potential in developing sophisticated search strategies through learned 9 exploration policies. Despite their theoretical advantages, pr... | https://arxiv.org/abs/2505.22648v1 |
Fulford, Hyung Won Chung, Alex Tachard Passos, William Fedus, and Ame lia Glaese. Browsecomp: A simple yet challenging benchmark for browsing agents. 10 [14] Runnan Fang, Xiaobin Wang, Yuan Liang, Shuofei Qiao, Jialong Wu, Zekun Xi, Ningyu Zhang, Yong Jiang, Pengjun Xie, Fei Huang, et al. Synworld: Virtual scenario syn... | https://arxiv.org/abs/2505.22648v1 |
Calvin Zhang, Mohamed Shaaban, John Ling, Sean Shi, et al. Humanity’s last exam. arXiv preprint arXiv:2501.14249 , 2025. [28] Baian Chen, Chang Shu, Ehsan Shareghi, Nigel Collier, Karthik Narasimhan, and Shunyu Yao. Fireact: Toward language agent fine-tuning. arXiv preprint arXiv:2310.05915 , 2023. [29] Chong Zhang, Yu... | https://arxiv.org/abs/2505.22648v1 |
for reasoning models, 2025. [43] Zekun Xi, Wenbiao Yin, Jizhan Fang, Jialong Wu, Runnan Fang, Ningyu Zhang, Jiang Yong, Pengjun Xie, Fei Huang, and Huajun Chen. Omnithink: Expanding knowledge boundaries in machine writing through thinking. arXiv preprint arXiv:2501.09751 , 2025. [44] Bang Liu, Xinfeng Li, Jiayi Zhang, ... | https://arxiv.org/abs/2505.22648v1 |
Theodore A Kanell, Peter Xu, Omar Khattab, and Monica S Lam. Assisting in writing wikipedia-like articles from scratch with large language models. In NAACL-HLT , 2024. [57] An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, ... | https://arxiv.org/abs/2505.22648v1 |
arXiv preprint arXiv: 2409.19256 , 2024. [71] Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Syst... | https://arxiv.org/abs/2505.22648v1 |
in automated content extraction or surveillance. We emphasize the importance of transparency, source attribution, and responsible deployment practices to mitigate potential harms. C Discussions C.1 Concurrent Work Comparison with the Training-based Methods The reported results of these two methods are shown in Table 4.... | https://arxiv.org/abs/2505.22648v1 |
GPT-4o, followed by a multi-stage filtering process to ensure quality, as described in Sec. 2.2. Table 5 separately reports the statistics for SFT data generated using Long-CoT and Short-CoT reasoning. We plan to scale this high-quality dataset further to investigate whether increasing the data volume leads to signif- ... | https://arxiv.org/abs/2505.22648v1 |
best you can. Use the following format: Question: the input question you must answer Thought: you should always think about what to do 6https://github.com/QwenLM/Qwen-Agent/ 17 Action: the action to take, should be one of [{tool_names}] Action Input: the input to the action, use JSON Schema with explicit parameters Obs... | https://arxiv.org/abs/2505.22648v1 |
it successfully retrieves the ZIP code. This demonstrates the Agent’s flexibility and adaptability when facing uncertainty. Case Trajectory in GAIA Question: I’m researching species that became invasive after people who kept them as pets released them. There’s a certain species of fish that was popularized as a pet by ... | https://arxiv.org/abs/2505.22648v1 |
The key point here is that the only entry before 2020 is in Florida, specifically Pinellas County, collected in July 2018 at Fred Howard Park.Now, the next step is converting that location into a zip code ... </think> <tool_call> {"name": "visit", "arguments": {"url": "https://nas.er.usgs.gov/queries/collectioninfo.asp... | https://arxiv.org/abs/2505.22648v1 |
arXiv:2505.22653v1 [cs.CL] 28 May 2025The Climb Carves Wisdom Deeper Than the Summit: On the Noisy Rewards in Learning to Reason Ang Lv1Ruobing Xie2∗Xingwu Sun2,3Zhanhui Kang2Rui Yan1,4∗ 1GSAI, Renmin University of China 2Large Language Model Department, Tencent 3University of Macau 4School of Computer Science, Wuhan U... | https://arxiv.org/abs/2505.22653v1 |
ence of substantial noise in the rewards, the model can still be effectively trained and achieve fast conver- gence. For example, when training on math problems and introducing noise by randomly flipping 40% of the reward function’s outputs (i.e., assigning positive rewards for incorrect answers), a Qwen-2.5-7B [ 32] m... | https://arxiv.org/abs/2505.22653v1 |
outperforms Llama [ 19] not only in downstream performance after RL but also in robustness to noisy rewards—a difference that aligns with Llama’s widely recognized weakness in reasoning [23, 33, 10]. 2 2. We present direct evidence showing that after RL, models’ enhanced performance on challenging tasks primarily stems... | https://arxiv.org/abs/2505.22653v1 |
main text is the Qwen-2.5-7B [ 32], which has demonstrated strong reasoning potential [10, 33], while experimental results on other models are provided in the appendix. Evaluation. We use three datasets—MATH-500 [ 12], GPQA [ 25], and AIME 2024 [ 2]—to assess the model’s reasoning ability on challenging tasks. We repor... | https://arxiv.org/abs/2505.22653v1 |
study. Takeaway 1. For models with strong reasoning potential, even with significant opposite noise in theverification rewards, the model can still be effectively trained during RL. Experiment 2. Given the surprising result from Experiment 1, the key question is why assigning a reward of 1 to outputs with genuinely inc... | https://arxiv.org/abs/2505.22653v1 |
this reasoning-then-revisiting cycle continues for too many iterations, resulting in an excessively long chain of reasoning that exceeds the context limit and is truncated before the final answer can be generated. The model produces these long reasoning steps from multiple viewpoints, effectively “escaping” the repetit... | https://arxiv.org/abs/2505.22653v1 |
helpfulness scale into a binary classification task: the more helpful response in each pair is labeled as 1, and the less helpful one as 0. For each response, we concatenate it with the chat history as the input to the reward model. Step05000100001500020000250005055606570758085Accuracy (%) Figure 5: Reward model’s accu... | https://arxiv.org/abs/2505.22653v1 |
provided in Appendix B. There, we show that human evaluation aligns with GPT assessments, with evaluators demonstrating moderate to substantial agreement. Loss42%Win 46%Tie12%RM (85% Acc) vsRM (75% Acc)Win55%Tie15%Loss30%RM (85% Acc) vsRM (65% Acc)Net Win: 4%Net Win: 25% Figure 7: Qwen-2.5-7B trained with an 85%- accur... | https://arxiv.org/abs/2505.22653v1 |
Recognizing the importance of reasoning patterns, we propose a method for calibrating RMs with RPR (Section 2). This approach overcomes the performance ceiling imposed by the limitations of the reward models at hand. 7 3.3 Calibrating noisy RMs with reasoning pattern reward Method. Considering that (1) it is impossible... | https://arxiv.org/abs/2505.22653v1 |
trained with a 65%-accurate RM calibrated by RPR to a model trained with an 85%-accurate RM. We set the threshold τto 0.5 and αto 0.1. The choice of αis discussed in Appendix E. The results in Figure 8 demonstrate the effectiveness of RPR calibration: 1. The calibrated 65%-accurate reward model lags only 8% behind the ... | https://arxiv.org/abs/2505.22653v1 |
The role of RL in post-training LLMs. This paper aligns with recent studies suggesting that pre-trained models already possess the fundamental reasoning abilities needed for complex tasks. Yeo et al. [ 33] found that pre-training data often includes long chain-of-thought patterns, establishing a foundation for reasonin... | https://arxiv.org/abs/2505.22653v1 |
Li, and Rui Yan. Fortify the shortest stave in attention: Enhancing context awareness of large language models for effective tool use. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar, editors, Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages 11160–... | https://arxiv.org/abs/2505.22653v1 |
Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Candès, and Tatsunori Hashimoto. s1: Simple test-time scaling, 2025. [21] Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang, Christina Kim, Christo- pher Hesse, Shantanu Jain, Vineet Kosara... | https://arxiv.org/abs/2505.22653v1 |
learning really incentivize reasoning capacity in llms beyond the base model?, 2025. [35] Dan Zhang, Sining Zhoubian, Ziniu Hu, Yisong Yue, Yuxiao Dong, and Jie Tang. ReST- MCTS*: LLM self-training via process reward guided tree search. In The Thirty-eighth Annual Conference on Neural Information Processing Systems , 2... | https://arxiv.org/abs/2505.22653v1 |
max(0, score) 59 return score Figure 9: An example code of reasoning pattern reward without checking answer correctness. 13 A conversation between User and Assistant. The user asks a question, and the Assistant solves it. The assistant first thinks about the reasoning process in the mind and then provides the user with... | https://arxiv.org/abs/2505.22653v1 |
two adjacent double bonds, let us try to identify those in this option. Upon closer inspection, we notice that there is indeed a diene present in this option, specifically 2,2-diiodoethen-1-one. This compound contains two adjacent double bonds, making it a diene. Therefore, we can conclude that option A contains a dien... | https://arxiv.org/abs/2505.22653v1 |
be the most reactive diene among the ones we have considered. Next, we need to determine the reactivity of the diene in option C.Finally, let us consider the reactivity of the diene in option C. As previously mentioned, the diene in option C lacks any substituents on the double bond. Consequently, we can infer that thi... | https://arxiv.org/abs/2505.22653v1 |
sole reward signal effectively enables the model to reach peak performance comparable to the noiseless baseline. Notably, on the most challenging AIME tasks, RPR yields the highest peak performance across all setups. 20406080100120140204060801001201402040608010012014001020304050602015105018100MATH-500GPQAAIME 𝑝=0𝑝=10... | https://arxiv.org/abs/2505.22653v1 |
For fractions, use the a/b format as it is easier for me to understand. As for the images, first translate the QUERY into English, and then use the following website with an English query to retrieve the images:https://source.unsplash.com/960x640/?QUERY Use Markdown syntax to display the images.” Llama models are widel... | https://arxiv.org/abs/2505.22653v1 |
easy to understand, and does it present information logically? –Thoroughness: Does the response cover all aspects of the user’s request? Is anything missing or incomplete? •Avoid Quick Judgment: We will randomize the response order from two models. You cannot infer which one is always better based on the order. Also, d... | https://arxiv.org/abs/2505.22653v1 |
could yield similar outcomes with RL. Additionally, we highlight the importance of continuing efforts to enhance fundamental reasoning abilities during the pre-training stage. This paper presents findings and insights into post-training LLMs using RL with noisy rewards, which do not raise safety concerns. Average Respo... | https://arxiv.org/abs/2505.22653v1 |
arXiv:2505.22661v1 [cs.CL] 28 May 2025 …GUESS ARENA : Guess Who I Am? A Self-Adaptive Framework for Evaluating LLMs in Domain-Specific Knowledge and Reasoning Qingchen Yu1*Zifan Zheng2*Ding Chen3* Simin Niu4Bo Tang1Feiyu Xiong1Zhiyu Li1† 1MemTensor (Shanghai) Technology Co., Ltd.2University of Sydney 3Research Institut... | https://arxiv.org/abs/2505.22661v1 |
frameworks (e.g., Chatbot Arena (Chi- ang et al., 2024)) improve evaluation flexibility through human interaction; however, their results remain inherently influenced by subjective judg- ments, posing challenges to standardization. Re- cently, GameArena (Hu et al., 2024) proposed a gamified evaluation mechanism, offeri... | https://arxiv.org/abs/2505.22661v1 |
reliance on human feedback introduces subjectivity and limits scalability, reducing overall evaluation efficiency. In contrast, GUESS ARENA provides a more auto- mated, reproducible, and flexible evaluation frame- work. By leveraging adaptively generated domain Example: Financial Industry ScenarioVarious Types of Finan... | https://arxiv.org/abs/2505.22661v1 |
Qd=Template (dmeta,T) (1) Here, dmetadenotes the document metadata, and Trefers to the predefined domain-specific termi- nology dictionary. For each document, we em- ploy GPT-4o as the retrieval-augmented generator MRAGto produce a keyword set by leveraging the document content: Kd=MRAG(Qd|dcontent ) (2) However, the i... | https://arxiv.org/abs/2505.22661v1 |
target card p. 3.3 Evaluation Metrics To comprehensively evaluate the knowledge ca- pability and reasoning ability of the tested model within a specific domain, we design a composite score that integrates the model’s domain reason- ing accuracy ( E), reasoning efficiency ( F), and knowledge applicability ( K). The comp... | https://arxiv.org/abs/2505.22661v1 |
Instruct (MetaAI, 2024), and QwQ-32B (Qwen- Team, 2025). Detailed information about each model is shown in Table 1. Model #Para. Type Date GPT-4o NaN Chat 2024.05 OpenAI-o1 NaN Chat 2024.09 Qwen2.5-32B 32B Instruct 2024.09 Qwen2.5-72B 72B Instruct 2024.09 Claude-3.5-Sonnet NaN Chat 2024.10 Llama3.3-70B 70B Instruct 202... | https://arxiv.org/abs/2505.22661v1 |
column is boldfaced , and the second-highest is underlined . tively worse overall, with particularly low scores in the finance and healthcare industries. To further verify the effectiveness of the GUES- SARENA method in distinguishing the reasoning and domain knowledge capabilities of different LLMs in specific fields,... | https://arxiv.org/abs/2505.22661v1 |
0.8 0.9 1.0Info TechFinanceEducationHealthcareManufacturing 0.90520.85330.89330.91060.9020 0.91240.87360.89430.90470.9033 0.88560.85180.87920.91330.8925Qwen2.5-72B-Instruct 0.6 0.7 0.8 0.9 1.0Info TechFinanceEducationHealthcareManufacturing 0.85430.85970.85960.89020.8991 0.86160.83860.84080.88710.8918 0.86120.82560.847... | https://arxiv.org/abs/2505.22661v1 |
we randomly sampled 1,200 instances from the full evaluation set and invited human annotators to provide gold-standard labels. We then designated several mainstream LLMs, including GPT-4o, as judge models and measured their agreement with the human annotations. The results, summarized in Table 5, show that GPT-4o attai... | https://arxiv.org/abs/2505.22661v1 |
benchmarks, our framework en- ables more efficient and cost-effective evaluation of domain-specific reasoning capabilities while al- leviating credibility concerns arising from question leakage in static benchmarks. In experiments conducted across five predefined vertical domains, GUESS ARENA effectively re- vealed per... | https://arxiv.org/abs/2505.22661v1 |
Wei-Lin Chiang, Lianmin Zheng, Ying Sheng, Anasta- sios Nikolas Angelopoulos, Tianle Li, Dacheng Li, Hao Zhang, Banghua Zhu, Michael Jordan, Joseph E Gonzalez, et al. 2024. Chatbot arena: An open plat- form for evaluating llms by human preference. arXiv preprint arXiv:2403.04132 . Peter Clark, Isaac Cowhey, Oren Etzion... | https://arxiv.org/abs/2505.22661v1 |
Qwen-Team. 2025. Qwq-32b: Embracing the power of reinforcement learning. N Reimers. 2019. Sentence-bert: Sentence embed- dings using siamese bert-networks. arXiv preprint arXiv:1908.10084 . Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. InProceedings of the 2019 C... | https://arxiv.org/abs/2505.22661v1 |
core of "Guess Who I Am?" lies in its strict turn-based questioning protocol. Players alternate turns posing a single question about a feature of the opponent’s secret character. Crucially, these questions must be structured to elicit a definitive "Yes" or "No" response (e.g., "Does your character have red hair?"). Bas... | https://arxiv.org/abs/2505.22661v1 |
any parentheses, such as “product lifecycle management” or “learning management system”. 3.Broad Coverage : The keywords should broadly cover various knowledge aspects of the field, including common terms, basic concepts, and specialized vocabulary from subfields. If appropriate, you may also include well-known entitie... | https://arxiv.org/abs/2505.22661v1 |
iot platform, nvdia vgpu, ai server integration, data center operations, data security management, communication latency, it infrastructure, it project management, cybersecurity framework, ai servers, pc industry trends, vsmp, offline data collection, data analytics, distributed storage system, task management, data ce... | https://arxiv.org/abs/2505.22661v1 |
Do not request additional hints. 2.Scoring Mechanism: The game score is inversely proportional to the number of questions asked. The fewer the questions, the higher the final score, assuming you correctly guess the chosen card. Successfully and quickly identifying the target card is key to achieving a high score. 3.Gue... | https://arxiv.org/abs/2505.22661v1 |
be answered with "[Yes]" or "[No]". ## Game Rules 1.Question Rules: You may ask only one clear and concise question at a time, which must be answerable with "[Yes]" or "[No]". The question must not contain line breaks, nor can it directly ask about the card’s specific identity. Do not request additional hints. 2.Scorin... | https://arxiv.org/abs/2505.22661v1 |
arXiv:2505.20694v1 [cs.CV] 27 May 2025Temporal Saliency-Guided Distillation: A Scalable Framework for Distilling Video Datasets Xulin Gu1∗Xinhao Zhong1∗Zhixing Wei1Yimin Zhou3Shuoyang Sun1 Bin Chen1,2†Hongpeng Wang1Yuan Luo4 1Harbin Institute of Technology, Shenzhen2Peng Cheng Laboratory 3Tsinghua Shenzhen Internationa... | https://arxiv.org/abs/2505.20694v1 |
a synthetic dataset introduces considerable time overhead and memory consumption. In contrast, training-free methods [ 19,18,10, 5] commonly utilize the priors from pre-trained classifiers or generative models to directly synthesize datasets, avoiding pixel-level optimization of individual images and thereby significan... | https://arxiv.org/abs/2505.20694v1 |
train a standard video classification model to encode video information into a pre-trained classifier. Subsequently, we align the distribution of the synthetic dataset with the model’s internal statistics under the guidance of classification loss. To further reinforce temporal information, we apply TSGF, 2 computed via... | https://arxiv.org/abs/2505.20694v1 |
and task-specific loss with those of the model. However, a key limitation of existing work [ 6] that directly applies decoupled methods to video datasets lies in the treatment of frames from the same video as independent and unrelated samples, resulting in redundant optimization and neglect of temporal structure. 2.2 V... | https://arxiv.org/abs/2505.20694v1 |
distillation process, the synthetic dataset Sis first initialized using random noise or randomly sampled instances. Then, the distillation loss is computed, and Sis optimized based on the loss. However, for video datasets, each video consists of multiple frames and can be regarded as image data with a large batch size ... | https://arxiv.org/abs/2505.20694v1 |
this end, we compute the temporal saliency of each frame and assign adaptive optimization magnitudes accordingly. Specifically, we calculate the inter-frame difference for each frame using the following formulation: di=|fi+1−fi|+|fi−fi−1| 2(6) where fidenotes the i-th frame, and direpresents the inter-frame difference ... | https://arxiv.org/abs/2505.20694v1 |
existing methods on small-scale datasets. Our method achieve significant performance improvements across various settings. Dataset MiniUCF HMDB51 IPC 1 5 1 5 Random 9.9±0.8 22 .9±1.1 4 .6±0.5 6 .6±0.7 Coreset Selection Herding 12.7±1.6 25 .8±0.3 3 .8±0.2 8 .5±0.4 K-Center 11.5±0.7 23 .0±1.3 3 .1±0.1 5 .2±0.3 DM 15.3±1.... | https://arxiv.org/abs/2505.20694v1 |
only 28.6%. The accuracy 7 Table 2: Comparison of top-5 accuracy with existing methods on large-scale datasets.†denotes the top-5 accuracy of our teacher models trained on the full dataset is lower than the baselines due to different hyper-parameters settings. Dataset Kinetics-400 SSv2 IPC 1 5 1 5 Random 3.0±0.1 5 .6±0... | https://arxiv.org/abs/2505.20694v1 |
the effectiveness of each component in the dis- tillation framework, we conduct ablation studies on the MiniUCF dataset under the IPC=5 setting. The detailed results are shown in Table 3, where TSGF Odenotes the temporal saliency-guided fil- ter, and TSGF Arefers to the temporal saliency- aware video augmentation strat... | https://arxiv.org/abs/2505.20694v1 |
and Pattern Recognition , pages 6299–6308, 2017. [3]George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A Efros, and Jun-Yan Zhu. Dataset distillation by matching training trajectories. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition , pages 4750–4759, 2022. [4]George Caze... | https://arxiv.org/abs/2505.20694v1 |
Yu, and Tao Lin. On the diversity and realism of distilled dataset: An efficient dataset distillation paradigm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 9390–9399, 2024. [20] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spa- ti... | https://arxiv.org/abs/2505.20694v1 |
Methods Acc random init 36.8±0.3 real init 54.8±0.5Table 6: Ablation study on data augmentation. Methods Acc image-based 43.1±0.4 TSGF A 54.8±0.5Table 7: Experimental results on static and dynamic group. IPC Acc S_Acc D_Acc 1 39.2±0.7 35 .5±1.0 34 .6±0.8 5 54.8±0.5 53 .6±1.2 58 .5±0.7 10 60.5±0.6 57 .1±0.9 64 .6±0.7 To... | https://arxiv.org/abs/2505.20694v1 |
transitions. This further validates the effectiveness of our temporal saliency-guided framework in maintaining critical dynamic information necessary for accurate video understanding. F Implementation Details F.1 Hyperparameters Dataset IPC lr r_bn MiniUCF1 0.25 0.001 5 0.25 0.005 HMDB511 0.25 0.001 5 0.25 0.005 Kineti... | https://arxiv.org/abs/2505.20694v1 |
arXiv:2505.20697v1 [cs.LG] 27 May 2025Generating Hypotheses of Dynamic Causal Graphs in Neuroscience: Leveraging Generative Factor Models of Observed Time Series Zachary C. Brown1David Carlson1 Abstract The field of hypothesis generation promises to re- duce costs in neuroscience by narrowing the range of interventiona... | https://arxiv.org/abs/2505.20697v1 |
has demonstrated that using automated hypothe- ses generation for these mechanisms can still yield useful insights which aid scientific discovery (Mague et al., 2022). Indeed, we accurately estimate causal graphs in observed systems without making restrictive assumptions as to the state-dependence or linearity/nonlinea... | https://arxiv.org/abs/2505.20697v1 |
including tasks from finding predictive graphical models (L ¨owe et al., 2022) to estimating a subject’s emotional state via recorded electroencephalogram (EEG) signals (Song et al., 2018). These frameworks tend to focus on estimating static causal graphs (Tank et al., 2021; Song et al., 2018; Bussmann et al., 2021; Pa... | https://arxiv.org/abs/2505.20697v1 |
al., 2017; Keshtkaran et al., 2022). In contrast, we largely ignore our models’ latent features beyond applying con- straints because our focus is on identifying candidates for causal relationships in recorded data. 2 Generating Hypotheses of Dynamic Causal Graphs in Neuroscience 3. Methods 3.1. Problem Statement Suppo... | https://arxiv.org/abs/2505.20697v1 |
of adjacency matrices ˆA∈Rnc×nc×τinwhere each ai,j,t∈ˆA is the estimated weight of the Granger causal relationship where the state of node νj∈ V from ttime steps in the past ‘causally’ effects current node νi∈ V. To make quan- titative comparisons, we standardized the representation(s) of baselines’ and our models’ est... | https://arxiv.org/abs/2505.20697v1 |
needed in each time series forecast. Thus gθis fundamentally dif- ferent from the generative factors of fϕ(which perform regression on the time series itself); indeed, we found that gθtypically required more historical information than the generative factors to sufficiently learn its task (see Section 3.6 and Appendix ... | https://arxiv.org/abs/2505.20697v1 |
ative information contained therein, and may be used to limit the effect gradients from the forecasting loss have on gθ. When the behavioral label ynis a known quantitative variable, we can generate ˆynvia supervised regression. Al- ternatively, if ynis categorical, we can predict a behavior as ‘present’ in the signal ... | https://arxiv.org/abs/2505.20697v1 |
We now report empirical comparisons of REDCLIFF-S against other methods3, with additional results given in Appendix C. All algorithms were trained on a local cluster featuring a variety of GPU devices, including A6000 GPUs. We focus much of our analysis on f1-score performance, partly because our neuroscientific applic... | https://arxiv.org/abs/2505.20697v1 |
system, we sampled recordings and their 6 Generating Hypotheses of Dynamic Causal Graphs in Neuroscience labels for the dataset. First, a random state vector was drawn from a uniform distribution for each V AR model in the sys- tem. This vector was recurrently fed through its correspond- ing V AR model for a number of ... | https://arxiv.org/abs/2505.20697v1 |
with four other system states/factors adding noise in the dataset marks a sea change in modeling capability over baselines. While we could have high- lighted examples with higher optimal f1-scores, we highlight this one due to its comparable complexity to our TST case study. As another test, we re-trained REDCLIFF-S mo... | https://arxiv.org/abs/2505.20697v1 |
but one DREAM4 folds to a recording from the dominant fold at three different signal-to-noise ratios: low (LSNR), moderate (MSNR), and high (HSNR). We refer to the resulting dataset as the “DREAM4 Insilico-Combined” (D4IC) dataset. Hyperparameters for each algorithm were tuned to a single repeat of the D4IC MSNR datase... | https://arxiv.org/abs/2505.20697v1 |
SYN.SYS. 12-11-2 ↓ ↓ ↓ ↓ SYN.SYS. 12-11-5 ↑ ↓ ↓ ↓ D4IC HSNR ↓ ↓ ↓ ↓ generates hypotheses about the presence of dynamic causal relationships that prove accurate at state-of-the-art rates, and this pattern holds true for systems with differing numbers of variables, inter-variable relationships, and factors. Results from ... | https://arxiv.org/abs/2505.20697v1 |
the Infralimbic Cortex and then to the Nucleus Accumbens Shell. 6. Conclusion We present the novel REDCLIFF-S hypothesis generation algorithm for temporal systems featuring dynamic causal interactions. By learning to combine nonlinear factors using weights conditioned on signal history, REDCLIFF-S attains state-of-the-... | https://arxiv.org/abs/2505.20697v1 |
Kiritoshi, K., Okawachi, T., Izumitani, T., and Shimizu, S. Causal discovery for non- stationary non-linear time series data using just-in-time modeling. In Conference on Causal Learning and Rea- soning , pp. 880–894. PMLR, 2023. Gallagher, N., Dzirasa, K., and Carlson, D. Directed spec- trum measures improve latent ne... | https://arxiv.org/abs/2505.20697v1 |
H. Causal discovery from conditionally stationary time-series. arXiv preprint arXiv:2110.06257 , 2021. Runge, J. Discovering contemporaneous and lagged causal relations in autocorrelated nonlinear time series datasets. InConference on Uncertainty in Artificial Intelligence , pp. 1388–1397. Pmlr, 2020. Sadeghi, A., Gopa... | https://arxiv.org/abs/2505.20697v1 |
this section we provide theoretical arguments as to why the systems at the center of our work - dynamical systems featuring nonlinear, dynamic causal graphs - are generally not identifiable. At the same time, we point out that prior work has demonstrated that utilizing causal discovery methods can still benefit scienti... | https://arxiv.org/abs/2505.20697v1 |
models of dynamical systems (Kugiumtzis, 1996). Here, we argue that the task of predicting the state of a signal requires more historical time steps than is required to generate the same time series in the presence of nontrivial noise, at least for some dynamical systems. We now present a summary of our theoretical arg... | https://arxiv.org/abs/2505.20697v1 |
enough capacity to accurately distinguish between states, Lemma 1 indicates that gθmay also have access (by necessity) to at least as much information as that required to forecast the evolution of Φ(assuming the proper noise profile) in one or more system states , at least for restricted (yet simple and arguably common... | https://arxiv.org/abs/2505.20697v1 |
(2021) and REDCLIFF-based models) learned multiple causal graph factors. The standard we chose to adopt was that each estimated (and true) Granger causal graph would need to be presented as a simple adjacency matrix; i.e. if there were ncvariables in the recorded system, then the true/estimated graph(s) from each syste... | https://arxiv.org/abs/2505.20697v1 |
✓ -0.654 - ✓ ✓ -0.201 ✓ ✓ ✓ -0.674 One open question is how to balance the significant differences in variation between the forecasting and state prediction penalties on one hand and the significantly less-variable cosine similarity term on the other. In selecting the final model for our D4IC experiment, we found that ... | https://arxiv.org/abs/2505.20697v1 |
experiments. We also report ROC-AUC statistics corresponding to D4IC-trained REDCLIFF-S, cMLP-v2, and dCSFA-NMF models in Table 5. C.3. Additional Synthetic Systems Results We now report additional results from our Synthetic Systems experiments for transparency. 17 Generating Hypotheses of Dynamic Causal Graphs in Neur... | https://arxiv.org/abs/2505.20697v1 |
LOWER UPPER LOWER AVG. PLACE . LASAR (S UP.) 0.457 ±0.043 4.9 ±1.466 2.9 ±0.434 4.7 ±1.331 2.9 ±0.434 3.8 QRBS (S UP.) 0.424 ±0.029 2.1 ±0.385 1.9 ±0.537 2.5 ±0.548 1.9 ±0.537 3.0 SLARAC (S UP.) 0.347 ±0.023 0.8 ±0.716 0.4 ±0.219 0.7 ±0.626 0.4 ±0.219 2.8 REGIME -PCMCI 0.775 ±0.018 3.8 ±0.540 3.6 ±0.518 4.1 ±0.477 3.8 ... | https://arxiv.org/abs/2505.20697v1 |
of Dynamic Causal Graphs in Neuroscience Table 8. Mean optimal f1-score ±SEM obtained by REDCLIFF-S ablations across various datasets (summarized in Table 3 of the main paper). Values should be compared against those depicted in Figures 2A and 4 of the main paper. ABLATION DATASET ρ= 0(EQ. 5, 10) nk= 1(≡EQ. 2) α= 1(EQ.... | https://arxiv.org/abs/2505.20697v1 |
Synthetic Systems experiments. Figure 12. Average improvement in optimal f1-scores ±SEM by the REDCLIFF-S algorithm over baselines from our Low- complexity Synthetic Systems experiments. Figure 13. Average optimal f1-scores ±SEM from our Low- complexity Synthetic Systems experiments. Figure 14. Average improvement in o... | https://arxiv.org/abs/2505.20697v1 |
of Dynamic Causal Graphs in Neuroscience D. Datasets Used: Additional Details In this section we report various hyperparameters used in curating/preparing the datasets used in our experiments. Sup- plemental Table 9 presents hyperparameters for the D4IC dataset (see Section 4.3). Supplemental Table 10 presents hyperpar... | https://arxiv.org/abs/2505.20697v1 |
of Dynamic Causal Graphs in Neuroscience Table 11. Region-Averaged TST 100 Hz Dataset (Section 4.4) - hyperparameters across experiments. PARAMETER NAME VALUE NUM PROCESSED SAMPLES 10000 SAMPLE TEMP WINDOW SIZE 1500 SAMPLE FREQ 1000 POST PROCESSING SAMPLE FREQ 100 CUTOFF 35 MAD THRESHOLD 15 Q 2 ORDER 3 FILTER TYPE LOWP... | https://arxiv.org/abs/2505.20697v1 |
search(es) on Repeat 3 of the dataset for a (nonlinear, noisy) system with nc= 12 ,ne= 33 , and nk= 3; code for generating this dataset is included in the project repository for the main paper. Models represented by columns with “D4IC” in the column name had their hyperparameters selected by grid search(es) performed o... | https://arxiv.org/abs/2505.20697v1 |
MAXLAGS 20 NUM HIDDEN 256 HIDDEN LAYERS 2 LEARNING RATE 0.0001 WEIGHT DECAY 0 LAMBDA 1 0.0 32 Generating Hypotheses of Dynamic Causal Graphs in Neuroscience Table 18. dCSFA - hyperparameters across experiments. PARAMETER NAME SYNTH . SYSTEMS D4IC BATCH SIZE 128 128 NEPOCHS 250 1000 NPRE EPOCHS 50 50 NMF MAX ITER 10 20 ... | https://arxiv.org/abs/2505.20697v1 |
Published in Transactions on Machine Learning Research (03/2025) Sparsified State-Space Models are Efficient Highway Networks Woomin Song woomin.song@kaist.ac.kr Korea Advanced Institute of Science & Technology (KAIST) Jihoon Tack jihoontack@kaist.ac.kr Korea Advanced Institute of Science & Technology (KAIST) Sangwoo M... | https://arxiv.org/abs/2505.20698v1 |
Simba-2.8b with Mamba and Pythia, all with the same number of FLOPS. We report the mean accuracy over 6 NLP benchmarks and perplexity on the PG-19 dataset with 2k context, following the setups in Section 4. Simba outperforms both models in accuracy and perplexity. 2023; Jelassi et al., 2024). However, this compromises ... | https://arxiv.org/abs/2505.20698v1 |
in Simba. Somewhat unexpectedly, we found that Simba performed better than its original unpruned Mamba in some of our experiments, potentially benefiting from the highways created at the upper layers. To further investigate the positive effect of highways, we examine the information flow across layers by assessing the ... | https://arxiv.org/abs/2505.20698v1 |
Transformers. To this end, we propose a novel token pruning criterion based on the global importance of tokens to the final output, derived from reformulating SSM equations to accumulate local recurrences. Second, by targeting SSMs, our hierarchical pruning scheme offers a novel interpretation of highway networks, conn... | https://arxiv.org/abs/2505.20698v1 |
model must compress all previ- ous information into the state of each token. Thus, the states of SSM tokens are likely highly redundant, as states at similar positions would compress a sim- ilar set of information. To verify this, we visualize the token redundancy of Mamba in Figure 2 by mea- suring the cosine similari... | https://arxiv.org/abs/2505.20698v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.