text string | source string |
|---|---|
as evidence retrieval, extraction, synthesis, and summarization, as well as evidence adoption and evidence-based research, such as question-answering, clinical trial design and identification, and other cutting-edge studies across various clinical specialties. Furthermore, we outline key benchmarks to fa- cilitate the ... | https://arxiv.org/abs/2505.22280v1 |
metadata was extracted from each paper, including models, disease, tasks involved, results, and limitations. Two annotators cross-verified the study selection and metadata ex- traction processes and consulted a third in cases of disagreement. 2.4 Study Statistics From an initial pool of 601 papers retrieved from databa... | https://arxiv.org/abs/2505.22280v1 |
cycle to corresponding NLP tasks. evant information from large text corpora based on user queries. Early heuristic methods involved structured, keyword-based queries to retrieve arti- cles from repositories like MEDLINE or PubMed. These methods, while foundational, are limited by the high cost of expert annotation, mai... | https://arxiv.org/abs/2505.22280v1 |
within studies. Initially, rule-based and machine-learning meth- ods were used to extract meaningful relationshipsfrom medical literature (Alodadi and Janeja, 2019; Borchert et al., 2022). By 2021, transformative methodologies were developed, integrating deep learning frameworks like BERT and Augment Min- ing (AM). For... | https://arxiv.org/abs/2505.22280v1 |
introduced the platform A2A, which used Okapi Best Match 25 (BM25) that assigned scores to doc- uments based on term frequency and document length and Divergence from Randomness (DFR) that quantified informativeness as the divergence of a term’s distribution from randomness for doc- ument ranking. Additionally, machine... | https://arxiv.org/abs/2505.22280v1 |
Abstractive Summa- rization Sequence-to-sequence) to resolve the chal- lenge of obtaining useful information from a vast amount of clinical documents. With the increased demand for user-interacted summarization, Ramprasad et al. (2023b) presented TrialsSummarizer, a system that helps automate summarizing the most relev... | https://arxiv.org/abs/2505.22280v1 |
with reduced and preserved ejection fraction and atrial fibrillation, this model was deployed and achieved an impressive accuracy of 87.3%. 7.2 Clinical trial design and identification Not all medical specialties are fully addressed by current research, and even in those with significant focus, the integration of findi... | https://arxiv.org/abs/2505.22280v1 |
need for drugs for treatment. To quickly meet this requirement, the CovidX Net- work Algorithm Gates and Hamed (2020) was developed, which utilized NLP to analyze vast COVID-19 biomedical literature. It ranked poten- tial drug candidates for repurposing, highlighting NLP’s power in automating and accelerating evi- denc... | https://arxiv.org/abs/2505.22280v1 |
rank- ing the evidence with the highest confidence, sum- marizing the information, and answering questions. At the same time, as in any other evolving area, there remain challenges ahead. For example, gener- ative models in EBM tasks have demonstrated im- pressive fluency and scalability, yet their tendency to hallucin... | https://arxiv.org/abs/2505.22280v1 |
selection of 129 studies that focus on critical aspects of NLP within EBM. We first provide an overview of EBM, followed by a sur- vey of NLP methods and techniques that address each step of the EBM process. We also explore use cases that demonstrate the application of EBM in various scenarios. Additionally, we review ... | https://arxiv.org/abs/2505.22280v1 |
Philipp Sachs, Udo Hahn, and Matthieu-P. Schapranow. 2020. GG- PONC: A corpus of German medical text with rich metadata based on clinical practice guidelines. In Proceedings of the 11th International Workshop on Health Text Mining and Information Analysis , pages 38–48, Online. Association for Computational Lin- guisti... | https://arxiv.org/abs/2505.22280v1 |
Xiaoyan Wang. 2024. Autocriteria: a general- izable clinical trial eligibility criteria extraction sys- tem powered by large language models. Journal of the American Medical Informatics Association , 31(2):375–385. Published: 11 November 2023.Yang Deng, Yaliang Li, Ying Shen, Nan Du, Wei Fan, Min Yang, and Kai Lei. 201... | https://arxiv.org/abs/2505.22280v1 |
Brophy. 2024. Matching patients to accelerate clini- cal trials (MPACT): Enabling technology for oncol- ogy clinical trial workflow. Stud. Health Technol. Inform. , 310:1086–1090. Nicholas J. Dobbins, Bin Han, Weipeng Zhou, Kris- tine F. Lan, H. Nina Kim, Robert Harrington, Özlem Uzuner, and Meliha Yetisgen. 2023. Leaf... | https://arxiv.org/abs/2505.22280v1 |
tion of clinical guidelines: A case study of diabetic ketoacidosis guidelines. Cureus , 15(5):e38784. Hamed Hassanzadeh, Sarvnaz Karimi, and Anthony Nguyen. 2020. Matching patients to clinical trials using semantically enriched document representation. Journal of Biomedical Informatics , 105:103406. Hendrik Ter Horst, ... | https://arxiv.org/abs/2505.22280v1 |
literature. Journal of the American Medical Informatics Association , 28(4):812–823. Tian Kang, Yingcheng Sun, Jae Hyun Kim, Casey Ta, Adler Perotte, Kayla Schiffer, Mutong Wu, YangZhao, Nour Moustafa-Fahmy, Yifan Peng, and Chun- hua Weng. 2023. EvidenceMap: a three-level knowl- edge representation for medical evidence... | https://arxiv.org/abs/2505.22280v1 |
cat- egorization of self-acknowledged limitations in ran- domized controlled trial publications. J. Biomed. Inform. , 152:104628. Jinhyuk Lee, Wonjin Yoon, Sungdong Kim, Donghyeon Kim, Sunkyu Kim, Chan Ho So, and Jaewoo Kang. 2020. BioBERT: a pre-trained biomedical language representation model for biomedical text mini... | https://arxiv.org/abs/2505.22280v1 |
Inform. Assoc. , 27(12):1903–1912. Iain J Marshall, Thomas A Trikalinos, Frank Soboczen- ski, Hye Sun Yun, Gregory Kell, Rachel Marshall, and Byron C Wallace. 2023. In a pilot study, auto- mated real-time systematic review updates were fea- sible, accurate, and work-saving. J. Clin. Epidemiol. , 153:26–33. Tobias Mayer... | https://arxiv.org/abs/2505.22280v1 |
of pubmed queries: a study on quality, effi- ciency, satisfaction. Journal of biomedical informat- ics, 44(2):310–318. Abigail Newbury, Hao Liu, Betina Idnay, and Chun- hua Weng. 2023. The suitability of UMLS and SNOMED-CT for encoding outcome concepts. J. Am. Med. Inform. Assoc. , 30(12):1895–1903. Vincent Nguyen, Sar... | https://arxiv.org/abs/2505.22280v1 |
Evidence- based medicine: History, review, criticisms, and pit- falls. Cureus , 15(2):e35266. Omid Rohanian, Mohammadmahdi Nouriborji, Samaneh Kouchaki, Farhad Nooralahzadeh, Lei Clifton, and David A. Clifton. 2024. Exploring the effectiveness of instruction tuning in biomedical language processing. Artificial Intellig... | https://arxiv.org/abs/2505.22280v1 |
Alessandro Lenci. 2023. We understand elliptical sentences, and language models should too: A new dataset for study- ing ellipsis and its interaction with thematic fit. In Proceedings of the 61st Annual Meeting of the As- sociation for Computational Linguistics (Volume 1: Long Papers) , pages 3340–3353, Toronto, Canada... | https://arxiv.org/abs/2505.22280v1 |
An, Qi Zhang, and Li Liu. 2024. Evaluation of an artificial intelligence- based clinical trial matching system in chinese pa- tients with hepatocellular carcinoma: a retrospective study. BMC Cancer , 24(1):246. Yu Wang, Yuan Wang, Zhenwan Peng, Feifan Zhang, Luyao Zhou, and Fei Yang. 2023a. Medical text classification ... | https://arxiv.org/abs/2505.22280v1 |
classification for clinical trial recruitment: Algorithm development and validation. JMIR Medi- cal Informatics , 8(7):e17832. Gongbo Zhang, Qiao Jin, Yiliang Zhou, Song Wang, Betina Idnay, Yiming Luo, Elizabeth Park, Jor- dan G. Nestor, Matthew E. Spotnitz, Ali Soroush, Thomas R. Campion Jr, Zhiyong Lu, Chunhua Weng, ... | https://arxiv.org/abs/2505.22280v1 |
(2020) Graph, Statistical SARS NL Ghosh et al. (2024a) LLM – NE Ghosh et al. (2024b) Transformer, RNN – NE Górska and Tacconelli (2024)Transformer, LLM – NLY Gulden et al. (2019) Graph – NES Gwon et al. (2024) LLM Peyronie Disease NEIA Hamed et al. (2023) LLM Diabetic Ketoacidosis SEQ Hassanzadeh et al. (2020) Support ... | https://arxiv.org/abs/2505.22280v1 |
et al. (2024) LLM – TE Norman et al. (2019b) Statistical – L Nurmambetova et al. (2023) Random Forest, XGBoost Acquired Pressure Injuries NE Pan et al. (2021) Transformer, Graph COVID-19 NESQR Ramprasad et al. (2023a) Transformer, Longformer – SI Rony et al. (2023) LLM – EQ Rybinski et al. (2020a) Rules – I Rybinski et... | https://arxiv.org/abs/2505.22280v1 |
or therapeutic proceduresManual 1,200 texts about clinical trials with entities EBM-COMET (Ghosh et al., 2024b)Public Physiological or clinical, Death, Life impact, Resource use, Adverse eventsManual 300 RCT abstracts EBM-NLP (Nye et al., 2018)Public P, I, O Manual 4,993 medical abstracts from literatures on PubMed Eli... | https://arxiv.org/abs/2505.22280v1 |
arXiv:2505.22287v1 [cs.CY] 28 May 2025New Tools are Needed for Tracking Adherence to AI Model Behavioral Use Clauses Daniel McDuff∗Tim Korjakow Kevin Klyman Danish Contractor Abstract Foundation models have had a transformative impact on AI. A combination of large investments in research and development, growing source... | https://arxiv.org/abs/2505.22287v1 |
have highlighted the challenges of regulating AI, including the complexity of AI supply chains and the uncertainty surrounding copyright rules. Furthermore, [26] argue that the lack of standardization in licenses with behavioral use clauses can lead to confusion and inconsistency and suggest the use of tooling to help ... | https://arxiv.org/abs/2505.22287v1 |
increase. We argue that the availability, or unavailability, of such tools will have significant implications on the positive impact of Responsible AI Licenses. Furthermore, our findings can inform policymakers and regulators seeking to establish frameworks for the governance of AI, ensuring that these frameworks are g... | https://arxiv.org/abs/2505.22287v1 |
for each generated license. Furthermore, the inclusion of unique identifiers and immutable license records enables distribution not only in text form, but also through QR codes. The license generator has been publicly deployed, making it openly accessible to the general community. 2.2 Analysis Number of Licenses Create... | https://arxiv.org/abs/2505.22287v1 |
suggests that appli- cations, because they are designed for more specific purposes, rather than more generic models/code, are viewed as needing fewer behavioral-use restrictions. Behavioral-use clauses selected by users. The license generator had a minimum set of clauses indicated in Table 1 by a green dot. The most po... | https://arxiv.org/abs/2505.22287v1 |
informed consent of said natural person is obtained✓✓✓✓ ✓ ✓✓ ✓ 100% (7) To generate or disseminate information (including - but not limited to - images, code, posts, articles), and place the information in any public context without expressly and intelligibly disclaiming that the information and/or content is machine g... | https://arxiv.org/abs/2505.22287v1 |
mislead others, including failing to appropriately disclose to end users any known dangers of your system.✓ ✓ 25% (25) In connection with any academic dishonesty, including submitting any informational content or output of a Model as Your own work in any academic setting.✓ ✓ 17% Table 1: Summary of Behavioral-Use Claus... | https://arxiv.org/abs/2505.22287v1 |
behind clauses and their specific wording. The release of models such as Llama 2 [34], Gemma [32] and DeepSeek V2 and V3 [22, 23] accelerated the adoption of certain types of licenses. Our license generator is a demonstration of the demand for well-tailored licenses: the availability of our tool organically led to the ... | https://arxiv.org/abs/2505.22287v1 |
models or their outputs, coupled with little infrastructure for identifying violations of license terms even when a model and its outputs can be correctly tracked, may lead developers to seek out ways to “short-circuit” model performance in certain domains in an effort to promote adherence to behavioral use clauses [38... | https://arxiv.org/abs/2505.22287v1 |
with behavioral-use restrictions) and then switched to an ODC By 2.0 License,*which enabled the OLMo family of models (a model trained on this dataset) to be released under an Apache 2.0 license. 6 Conclusion Increasingly, AI software assets (models, source code, applications) are being released with licenses that incl... | https://arxiv.org/abs/2505.22287v1 |
Team. Granite 3.0 language models, 2024. [13] Alfonso Guarino, Nicola Lettieri, Delfina Malandrino, and Rocco Zaccagnino. A machine learning-based approach to identify unlawful practices in online terms of service: analysis, implementation and evaluation. Neural Computing and Applications , 33:17569–17587, 2021. [14] T... | https://arxiv.org/abs/2505.22287v1 |
Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res. , 21(1), January 2020. [28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High- resolution image synthesis with latent... | https://arxiv.org/abs/2505.22287v1 |
arXiv:2505.22288v1 [cs.AI] 28 May 2025Compression versus Accuracy: A Hierarchy of Lifted Models Jan Spellera,*, Malte Luttermannb,c, Marcel Gehrkecand Tanya Brauna aComputer Science Department, University of Münster, Germany bGerman Research Center for Artificial Intelligence (DFKI), Lübeck, Germany cInstitute for Huma... | https://arxiv.org/abs/2505.22288v1 |
εdoes not guarantee a model consistent with the previous one. E.g., with a larger ε, mak- ing more factors ε-equivalent, factors that were previously grouped together might no longer be part of the same group, because a differ- ent grouping appears more suitable. That is, the models do not form a hierarchy, where group... | https://arxiv.org/abs/2505.22288v1 |
V=R∪Φ, where R= {R1, . . . , R n}is a set of randvars and Φ={ϕ1, . . . , ϕ m}is a set of factors (functions), as well as a set of edges E⊆R×Φ. There is an edge between a randvar Ri∈Rand a factor ϕj∈ΦinE ifRiappears in the argument list of ϕj. A factor ϕj(Rj)defines a function ϕj:×R∈Rjrange (R)7→R>0that maps the ranges ... | https://arxiv.org/abs/2505.22288v1 |
positive real number. Two potentials φ1, φ2∈R>0areε-equivalent , denoted asφ1=εφ2, ifφ1∈[φ2·(1−ε), φ2·(1 + ε)]andφ2∈ [φ1·(1−ε), φ1·(1 + ε)]. Further, two factors ϕ1(R1, . . . , R n) andϕ2(R′ 1, . . . , R′ n)areε-equivalent , denoted as ϕ1=εϕ2, if there exists a permutation πof{1, . . . , n }such that for all assignment... | https://arxiv.org/abs/2505.22288v1 |
the potential associated with the k-th row in the potential table of ϕ. For example, factor ϕ1(A, B) in Fig. 1 is represented as the vector (φ1, φ2, φ3, φ4), with, e.g., ϕ1(true,false) =ϕ1(2) = φ2. After introducing 1DEED, we use it to set up a hierarchical ordering of ε-equivalent factors, which forms the backbone to ... | https://arxiv.org/abs/2505.22288v1 |
smallest admissible εfor each pair of factors, facilitating both storage and comparison. Corollary 4. For two vectors ϕ1, ϕ2∈Rn >0we have ϕ1=εϕ2if and only if ε≥ε0, where ε0=d∞(ϕ1, ϕ2)is the threshold value below which ε-equivalence no longer holds. Proof. It follows from Def. 4 that there exists a j∈ {1, . . . , n }su... | https://arxiv.org/abs/2505.22288v1 |
creates a matrix Λas shown in Table 1. Its cell entries are Λij:=εi,j:=d∞(ϕi, ϕj)for1≤i < j ≤m, where m=|Φ| is the number of factors in the FG. Due to the symmetric property ofd∞, additional information is not required, allowing for filling the matrix with zeros, thus forming an upper triangular matrix. Next, we examin... | https://arxiv.org/abs/2505.22288v1 |
to the maximum value. Regarding L, a new entry is formed, which is essentially a 3-tuple l= (ei′, ej′, h): one element ei′for the first (group of) factor(s), one element ej′for the second (group of) factor(s), and the last element hbeing the current hierarchy level shifted by m. Ifi′orj′identify a single factor, then e... | https://arxiv.org/abs/2505.22288v1 |
which groups get the same colour assigned, which is then the input to standard ACP, which runs independent of ε. Specifically, HACP proceeds in three phases, loading groups, running ACP, and updating potentials. Alg. 2 shows an overview, which is specified for a given hierarchy level ifor the sake of brevity but could ... | https://arxiv.org/abs/2505.22288v1 |
the deviation-wise worst-case sce- nario for an assignment is bounded. Notably, HACP relies on the mean values of the potentials of pairwise ε-equivalent factors. To quantify the difference in probabilistic queries between the original FG and the hierarchical processed FG after applying the HACP al- gorithm, we use the... | https://arxiv.org/abs/2505.22288v1 |
an initial model Mand a modified model M′obtained by running HACP (Alg. 2) or ε-ACP is bounded by pmax ∆≤√ ed1−1√ ed1+ 1withd1=DCD(PM, P′ M) ≤√ ed2−1√ ed2+ 1withd2= ln 1 +m−1 mε 1 +ε 1 +1 mε!m ≤√ ed3−1√ ed3+ 1withd3= ln 1 +ε2m ≤√ ed4−1√ ed4+ 1withd4= ln1 +ε 1−εm . This implies that for a given ε >0, we can det... | https://arxiv.org/abs/2505.22288v1 |
the indistin- guishability of objects. Over the past years, lifted variable elimina- tion has continuously been refined by many researchers to reach its current form [3, 5, 6, 11, 15, 19]. To construct a lifted (i.e., first-order) representation such as a parametric FG, the ACP algorithm [13], which generalises the Com... | https://arxiv.org/abs/2505.22288v1 |
This work provides a foundation for future advances in efficient and interpretable probabilistic inference. Acknowledgements This work was partially funded by the Ministry of Culture and Sci- ence of the German State of North Rhine-Westphalia. The research of Malte Luttermann was funded by the BMBF project AnoMed 16KIS... | https://arxiv.org/abs/2505.22288v1 |
Press, 2008. [16] M. Niepert and G. Van den Broeck. Tractability through Exchangeabil- ity: A New Perspective on Efficient Probabilistic Inference. In Proceed- ings of the Twenty-Eighth AAAI Conference on Artificial Intelligence (AAAI-2014) , pages 2467–2475. AAAI Press, 2014. [17] D. Poole. First-order Probabilistic I... | https://arxiv.org/abs/2505.22288v1 |
comparison to pfor different dvalues depending on mandε(Eq. (5) of the main paper). Circles/ crosses are the maximum distances from those bounds to the function p. Distances to function f(p) =pare later also referred as f=fupper forp∈[0,1 2]and f=flower forp∈(1 2,1]. running Hierarchical Advanced Colour Passing (HACP ,... | https://arxiv.org/abs/2505.22288v1 |
boundary values 0and1 are no possible points for a global maximum, because both functions fupperandflowertake on the value 0there. Therefore, the only possi- ble extreme point for the global maximum for fupperisp1=1√ ed+1 andflower isp2=√ ed√ ed+1. Note that p1andp2are symmetrically distanced to p= 1/2. Both reach exac... | https://arxiv.org/abs/2505.22288v1 |
(in particular, it holds thatϕ1(a,true) =ϕ2(c,true)for all assignments where a=c), we can exploit this property to simplify the computation and get P(B=true) =1 ZX a∈range(A)X c∈range(C)ϕ1(a,true)·ϕ2(c,true) =1 ZX a∈range(A)ϕ1(a,true)X c∈range(C)ϕ2(c,true) =1 Z X a∈range(A)ϕ1(a,true)!2 =1 Z X c∈range(C)ϕ2(c,true)!2 =1 ... | https://arxiv.org/abs/2505.22288v1 |
arXiv:2505.22290v1 [cs.AI] 28 May 2025Rethinking the Unsolvable: When In-Context Search Meets Test-Time Scaling Fanzeng Xia*1,Yidong Luo*1,Tinko Sebastian Bartels1,Yaqi Xu2, andTongxin Li†1 1The Chinese University of Hong Kong, Shenzhen 2Beijing University of Posts and Telecommunications Abstract Recent research has hi... | https://arxiv.org/abs/2505.22290v1 |
†Corresponding Author. 1 )BSE3FBTPOJOH1SPCMFN--.6OTPMWBCMF *O$POUFYU4FBSDI5FTU5JNF4DBMJOH"MHPSJUINJD4FBSDI "P5 (SFFEZ4FBSDI $P5 %JSFDU1SPNQUJOH4PMWBCMF _ 1BSBMMFM4DBMJOH4FRVFOUJBM4DBMJOH*OUFSOBM4DBMJOH&MFWBUFE3FBTPOJOH"CJMJUZ$PNNPOMZ6TFE&WBMVBUJPO$POGJHVSBUJPOT6OMPDLFE3FBTPOJOH1PUFOUJB... | https://arxiv.org/abs/2505.22290v1 |
more advanced in-context search techniques unexplored, concluding that these test-time scaling schemes yielded only small gains on hard problems; (iii) In the context of super-hard tasks, (Chen et al., 2024b) examined sequential scaling in conjunction with external pipelines for complex real-world planning. Other inves... | https://arxiv.org/abs/2505.22290v1 |
to find solutions. Specifically, we select Vertex Cover and 3-Dimensional Matching (3DM) and generate 100 controlled problem instances according to the methods in (Yang et al., 2025), enabling an understanding of their average behaviors and reliable performance estimates. We evaluate the LLMs’ reasoning capabilities on... | https://arxiv.org/abs/2505.22290v1 |
ensure a comprehensive understanding of their capabilities and responses to different experimental conditions. Prompts. We investigate three in-context search prompting strategies to instruct the LLMs for the tasks outlined above: •Direct Prompting : this approach provides the model with a few problem-solution pairs (W... | https://arxiv.org/abs/2505.22290v1 |
used evaluation configuration: direct prompting augmented with four test-time scaling variants: without scaling (Direct-WS), parallel scaling (Direct-PS), sequential scaling (Direct-SS), and internal scaling (Direct-IS). Figure 1 shows that: (i) Both Qwen 3 and Claude 3.7 achieve a 0%success rate under Direct-WS/PS/SS.... | https://arxiv.org/abs/2505.22290v1 |
CoT-IS and AoT-IS with internal scaling. We observe substantial improvements: Qwen 3’s success rate jumps to 24% (CoT-IS) and 30% (AoT-IS), while Claude 3.7 reaches 26% (CoT-IS) and 40% (AoT-IS). This represents up to a 30-fold improvement in success rate compared to the previous commonly used configurations. These res... | https://arxiv.org/abs/2505.22290v1 |
and all previously generated tokens: sj= Transformer θ x, s 1, . . . , s j−1 , j = 1, . . . , t (n). The class of languages recognizable by a Transformer that uses at most t(n)such intermediate decoding steps (i.e., generates up to t(n)intermediate tokens) is denoted CoT(t(n)). In essence, CoT allows the model to use... | https://arxiv.org/abs/2505.22290v1 |
internal scaling to modulate Tfrom polynomial to exponential scales is what allows such models to potentially address correspondingly more complex problems. Further theoretical definitions, including Turing machine, computational complexity classes, and decoder- only transformers are provided in Appendix E. 3.2 Theoret... | https://arxiv.org/abs/2505.22290v1 |
the architectural assumptions in Assumption 2), when augmented with a Chain of Thought (CoT) of length t(n) =exp(n)(i.e., t(n) =O(2p(n))for some polynomial p(n)in the input length n), is the complexity class EXP. Similarly, thisexponentialscalingofreasoningstepsalsoenhancesthepowerofAoT,enablingTransformers to tackle p... | https://arxiv.org/abs/2505.22290v1 |
Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. In International Conference on Learning Representations , 2021. Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray... | https://arxiv.org/abs/2505.22290v1 |
A survey on few-shot learning. ACM computing surveys (csur) , 53(3):1–34, 2020. Zhuosheng Zhang, Aston Zhang, Mu Li, and Alex Smola. Automatic chain of thought prompting in large language models. arXiv preprint arXiv:2210.03493 , 2022. Jinlan Fu, Shenzhen Huangfu, Hang Yan, See-Kiong Ng, and Xipeng Qiu. Hint-before-sol... | https://arxiv.org/abs/2505.22290v1 |
longer. arXiv preprint arXiv:2502.15631 , 2025. Jinyan Su, Jennifer Healey, Preslav Nakov, and Claire Cardie. Between underthinking and overthinking: An empirical study of reasoning length and correctness in llms. arXiv preprint arXiv:2505.00127 , 2025. Gabriel Maher. Llmpc: Large language model predictive control. arX... | https://arxiv.org/abs/2505.22290v1 |
and Test-Time Scaling, and positions our work within this broader landscape. Prompt Design. Appendix C presents the specific prompt templates employed for Direct Prompting, Chain of Thought (CoT) Prompting, and Algorithm of Thought (AoT) Prompting, using the Trip Planning task as an illustrative example. Experimental R... | https://arxiv.org/abs/2505.22290v1 |
LLMs through explicit algorithmic search pathways. This is achieved by employing detailed examples demonstrating algorithmic operations. External Search. In contrast with in-context search methods, external search involves an external algorithmic pipeline designed for halting, modifying, and then resuming the LLM’s gen... | https://arxiv.org/abs/2505.22290v1 |
to commute between them. ### Solution1 ### Here is the trip plan for visiting the 6 European cities for 16 days:\n\n**Day 1-4:** Arriving in Edinburgh and visit Edinburgh for 4 days.\n**Day 4:** Fly from Edinburgh to Milan.\n**Day 4-5:** Visit Milan for 2 days.\n**Day 5:** Fly from Milan to Copenhagen.\n**Day 5-8:** Vi... | https://arxiv.org/abs/2505.22290v1 |
equal **16** exactly. 3. **Flights requirements** A direct flight exists **only** when explicitly listed: Riga→Vilnius Riga →Edinburgh Riga →Milan Riga →Copenhagen Riga →Brussels Edinburgh →Copenhagen Edinburgh →Milan Edinburgh →Riga Edinburgh →Brussels Milan→Copenhagen Milan →Vilnius Milan →Brussels Milan →Riga Copenh... | https://arxiv.org/abs/2505.22290v1 |
them. Algorithm of Thought Prompting Problem Description: You are an expert at planning trips. You are given a few constraints regarding the cities to visit and the durations of staying at each city. You are also given the flight information between the cities. Depth-First Search Thinking Process: ### Objective ### Pla... | https://arxiv.org/abs/2505.22290v1 |
→**Riga** Calendar preview & test: Riga Day 5-7; relatives window lost Outcome: **Prune** Step: C3 Transition tried: ... →**Vilnius** Calendar preview & test: Vilnius Day 5-9; relatives & workshop windows broken Outcome: **Prune** Step: C4 Transition tried: ... →**Copenhagen** Calendar preview & test: Copenhagen Day 5–... | https://arxiv.org/abs/2505.22290v1 |
the complex real-world planning tasks (Zheng et al., 2024), supplementing those presented in the main body of the paper and featuring detailed tables and line graphs. 20 D.1 Tables Table 4: Performance on 3-Dimensional Matching (3DM) (Difficulty Level = 10) for controlled NP-hard task. Qwen3 showed almost no improvemen... | https://arxiv.org/abs/2505.22290v1 |
It then follows transitions until it reaches either qaccept(accept) or qreject(reject). •A Deterministic Turing Machine (DTM) Mdecidesa language L⊆Σ∗if on every input w, it halts in either qacceptorqreject, and accepts exactly those w∈L. •An Non-deterministic Turing Machine (NTM) Mdecides Lif on every w∈Lthere is at le... | https://arxiv.org/abs/2505.22290v1 |
>1, so that fthat recognizes L along some successful computational trace of a(n)decoding steps, using O ba(n) intermediate tokens to generate a tree of guesses (e.g. when using DFS). Remark 1 (On the Total Computational Effort for AoT) .It is important to distinguish between the length of the successful computational... | https://arxiv.org/abs/2505.22290v1 |
in decoder-only models. •Advanced Normalization Schemes: The architecture is assumed to employ specific layer normal- ization variants, such as projected pre-norm (where a learnable linear projection Pprojis applied to the input hsubof a sublayer before layer normalization LN) or multi-pre-norm. Outputsub= Sublayer( LN... | https://arxiv.org/abs/2505.22290v1 |
hτ iduring the i-th CoT step, the Transformer attends to the CoT history. The layer-norm hash ϕ(hτ i/i,1/i) =ϕ(hτ i,1)is used. An attention query, such as ⟨ϕ(hτ i,1), e1⟩, can be matched against keys derived from previous CoT steps, e.g., ⟨ϕ(hτ j,1),−ϕ(f(j),1)⟩ (where j < i), to identify and retrieve the most recent sy... | https://arxiv.org/abs/2505.22290v1 |
3 in Merrill and Sabharwal (2024): The relationship between CoT and Turing machine time complexity is given by: CoT(t(n))⊆^TIME (n2+t(n)2).(The tilde in ^TIMEhides polylogarithmic factors, specifically logk(n+t(n))for some constant k). Corresponding Proof. The proof proceeds by constructing a DTM that simulates the ope... | https://arxiv.org/abs/2505.22290v1 |
encapsulates polylogarithmic factors, which do not alter the overall polynomial nature of this time bound. Therefore, the DTM simulates MCoTin time that is polynomial with respect to the input length n. This signifies that the language L, decided by MCoT, belongs to the complexity class P. AsLwas an arbitrary language ... | https://arxiv.org/abs/2505.22290v1 |
u⟩. The verifier Vruns in p′ V(n)time on this input. According to Theorem 2 in Merrill and Sabharwal (2024) (as cited in the proof of Theorem 3.1), a 28 decoder-only Transformer satisfying Assumption 2 can simulate a DTM (like V) running in p′ V(n)steps by generating p′ V(n)intermediate tokens. These tokens effectively... | https://arxiv.org/abs/2505.22290v1 |
length of this certificate, measured in the number of tokens, is k, which is polynomial in n. The bit length of ucertisk·Pbits. With Pbits=O(log(n+k))(from Assumption 2 on Logarithmic Precision for Parameters and Activations, applied to input nand CoT/AoT length k), and since k≤a(n) =poly(n),n+kis also polynomial in n.... | https://arxiv.org/abs/2505.22290v1 |
check is part of the simulation of the k-th step (or an implicit (k+ 1)-th decision step based on the state after sk) and is therefore completed within the polynomial time bound. If the path SAoTleads to an acceptance state for MAoT, then VAoTaccepts ⟨w, S AoT⟩. Otherwise (e.g., if the path does not end in acceptance, ... | https://arxiv.org/abs/2505.22290v1 |
polylogarithmic factors of its argument, i.e., the actual DTM simulation time is O((n2+t(n)2)·(log(n+t(n)))j)for some constant j. Substitute tCoT(n) =O(2pM(n))into this time bound. The term (tCoT(n))2becomes: (tCoT(n))2= (O(2pM(n)))2=O(22pM(n)). Letp′ M(n) = 2 pM(n), which is also a polynomial. So, (tCoT(n))2=O(2p′ M(n... | https://arxiv.org/abs/2505.22290v1 |
denoted MAoT, satisfying Assumption 2, that decides L. On input w, the AoT process for MAoTproceeds as follows: 1. Certificate Generation (Simulating Nondeterministic Guess via AoT Exploration): The AoT prompting supplies MAoTwith examples demonstrating algorithmic search or construction. This capability is leveraged t... | https://arxiv.org/abs/2505.22290v1 |
A(n) =pA(n) +log2(pA(n)). Since pA(n)is polynomial, p′ A(n)is also polynomial (as log2of a polynomial grows slower than the polynomial). Thus, the bit length |ucert|=O(2p′ A(n)), which is exp(n). 2. Verifier DTM ( VAoT) Construction: The DTM VAoTtakes ⟨w, u cert⟩=⟨w, S AoT⟩as input. It deterministically simulates MAoTs... | https://arxiv.org/abs/2505.22290v1 |
defined as: Cpoly(AMT) =( PifAMTis deterministic , NPifAMTis nondeterministic , Cexp(AMT) =( EXPifAMTis deterministic , NEXPifAMTis nondeterministic . The following hold: 34 (1)Ifkcore(n) =O(poly(n)), then Llies in Cpoly AMT , regardless of any super-polynomial overhead inktotal(n). (2)To decide a language L∈Cexp AM... | https://arxiv.org/abs/2505.22290v1 |
Neural Restoration of Greening Defects in Historical Autochrome Photographs Based on Purely Synthetic Data Saptarshi Neil Sinhaa, P. Julius K ¨uhna, Johannes Koppeb, Arjan Kuijpera,b, Michael Weinmannc aFraunhofer IGD, Darmstadt, Germany bTU Darmstadt, Darmstadt, Germany cDelft University of Technology, Delft, Netherla... | https://arxiv.org/abs/2505.22291v1 |
as greening, occurs when the green potato starch grains bleed into adjacent areas due to their high sensi- tivity to water, resulting in unwanted green spots in the final image (see Fig. 2). Such artifacts distort the original appear- ance of these historical images, complicating their interpreta- tion and diminishing ... | https://arxiv.org/abs/2505.22291v1 |
ring, and deraining by introducing a dual-domain channel atten- tion mechanism that enhances interactions through lightweight convolutions in the spatial domain and integrates information from various frequency components. This paper explores the training, application, evaluation, and discussion of generative AI method... | https://arxiv.org/abs/2505.22291v1 |
seven selected autochromes for closer evaluation, with diameters ranging from 1% to 5% of the image width. Larger defects can appear alone, covering up to one-third of the image, and may merge with one another. We also analyzed greening defects and their e ffects on individual color channels (see Fig. 3) based on a per... | https://arxiv.org/abs/2505.22291v1 |
the Fast Fourier transform. The final loss is computed as l=ls+0.1lf. However, the main chal- lenge for autochrome restoration is the correct representation of the original colors of the defective area and to place a stronger focus on the proper color representation of the defective area, we therefore replace the spati... | https://arxiv.org/abs/2505.22291v1 |
models with the v1 dataset (see Fig. 6) clearly showed that the image-image models impacted regions beyond the defect areas. Quantitative results: The results in Table 1 present PSNR and SSIM scores between synthetic references and outputs, divided into two groups based on datasets v1 and v2 for comparability. Metrics ... | https://arxiv.org/abs/2505.22291v1 |
6. Conclusion Our method e ffectively identifies defects of all sizes and lo- cations by adjusting the colors of detected areas, making them less recognizable. However, it struggles to accurately repro- duce the original colors of autochrome images, often result- ing in bluish tones and missing very small defects. In c... | https://arxiv.org/abs/2505.22291v1 |
with conditional adversarial networks, in: CVPR, 2017, pp. 5967–5976. 2, 3 [14] M. Afifi, B. L. Price, S. Cohen, M. S. Brown, When color constancy goes wrong: Correcting improperly white-balanced images, in: CVPR, 2019, pp. 1535–1544. 2 [15] M. Afifi, M. S. Brown, Deep white-balance editing, in: CVPR, 2020, pp. 1394–14... | https://arxiv.org/abs/2505.22291v1 |
arXiv:2505.22303v1 [cs.HC] 28 May 2025VOICE CMS: UPDATING THE KNOWLEDGE BASE OF A DIGITAL ASSISTANT THROUGH CONVERSATION A P REPRINT Grzegorz Wolny Orange Research Warsaw, Poland grzegorz.wolny@orange.com Michał K. Szczerbak Orange Research Warsaw, Poland michal.szczerbak@orange.com May 29, 2025 ABSTRACT In this study,... | https://arxiv.org/abs/2505.22303v1 |
by the specialized service provider, through an API, for instance, assuring that the digital assistant’s knowledge is always valid falls into the responsibility of the hotel staff, who has the information in the first place. V oice CMS - G. Wolny and M. Szczerbak A P REPRINT Indeed, keeping a bot degenerating in time w... | https://arxiv.org/abs/2505.22303v1 |
Background Management of the knowledge base for a digital assistant is still more often than not restricted to its off-line preparation, where a assistant is designed either to have some fixed and infrequently changing information base or to be linked to an external data source, i.e. in RAG solutions for LLMs. Hence, a... | https://arxiv.org/abs/2505.22303v1 |
of interfaces on a very different maturity level at that time. A more recent study from the authors of Cha and Ji [2024] emphasizes the importance of activity context on the preference for choosing voice or touch. The reference GUI in that experiment consisted of buttons with Korean text, and researchers were observing... | https://arxiv.org/abs/2505.22303v1 |
define workflows aimed at responding to user queries, with most of the nodes using large language models, specifically Gemini 1.5 Flash2, for natural language interpretation and generation. To ensure the assistant remains a reliable source of information, a method for providing and maintaining its knowledge base is req... | https://arxiv.org/abs/2505.22303v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.