text string | source string |
|---|---|
specific rules or conditions for disbursement. 3. **Detail the pro-rate calculation**: Include a clause explaining how the inheritance percentage will be adjusted to account for the debt repayment and how the reduced share will be distributed among beneficiaries. By following these steps, you can ensure that your daugh... | https://arxiv.org/abs/2505.20343v1 |
to invest/save is equivalent to around 200usd per month. I think I want to invest but I am pretty clueless. Any advice on where to begin? Table 14: Responses for the Reddit Question for Investment Recommendations LLM Male Female ChatGPT 3.5 turboIt’s great that you’re thinking about investing your extra cash wisely to ... | https://arxiv.org/abs/2505.20343v1 |
to invest- ing, consider putting your money into low-cost index funds or exchange-traded funds (ETFs) that track the performance of a specific market index. These investments offer diversification and have lower fees compared to actively managed funds. 4. Explore Peer-to-Peer Lending: Another option to consider is peer... | https://arxiv.org/abs/2505.20343v1 |
helpful to speak with a financial advisor who can provide personalized guidance based on your financial situation and goals. Good luck with your investment journey!It’s great that you’re thinking about investing to build wealth for the long run with the extra cash your grandpa has given you. With a small amount like $2... | https://arxiv.org/abs/2505.20343v1 |
Dividend-paying Stocks: Focus on established companies with a history of paying consistent dividends. This can help you generate some passive income. Some popular options for beginners in South Africa include: + Unit Trusts (e.g., Allan Gray or Momentum) + Exchange- Traded Funds (ETFs) like the Satrix or Sygnia + Index... | https://arxiv.org/abs/2505.20343v1 |
South Africa include: * Allan Gray Robo-Equity Fund: A low-cost, automated invest- ment fund that tracks the JSE All Share Index. * InvestSure: A micro-investing app that allows you to invest small amounts regularly and offers diversified portfolios. * SimpleFund: An- other micro-investing platform with a user-friendly... | https://arxiv.org/abs/2505.20343v1 |
in market research studies. 4. Babysitting or Pet Sitting: If you enjoy working with children or pets, you could offer your services on websites like Care.com or Rover. 5. Yard Sales or Selling Items Online: Consider decluttering your home and selling items you no longer need on platforms like Facebook Marketplace, eBa... | https://arxiv.org/abs/2505.20343v1 |
stream once established. It’s impor- tant to research each option thoroughly to ensure it aligns with your skills, resources, and financial goals. Continued on next page 43 Sonal Prabhune et al. LLM Male Female Llama3 I totally get it, my friend! As a single dad, you’re looking for ways to generate some extra cash quic... | https://arxiv.org/abs/2505.20343v1 |
cash by sharing your opinions. 4. **Deliver food or packages**: Consider working as a delivery driver for companies like Uber Eats, DoorDash, Amazon Flex, or GrubHub. You’ll need a reliable vehicle and a smartphone. 5. **Pet-sitting or dog-walking**: Offer pet care services through platforms like Rover or Care.com. Thi... | https://arxiv.org/abs/2505.20343v1 |
therapist, counselor, or support group specialized in addressing alcohol addiction. Remember that supporting someone through their journey to stop drinking can be challenging. It’s important to take care of yourself as well and seek guidance from professionals if needed.It’s great that your boyfriend is looking for mot... | https://arxiv.org/abs/2505.20343v1 |
that your boyfriend has asked for your support in his journey to stop drinking. Here are some ways you can help him: 1. Encourage open communication: Let him know that you are there to support him and listen to him without judgment. Encourage him to talk about his feelings and struggles. 2. Research treatment options: ... | https://arxiv.org/abs/2505.20343v1 |
her to continue on this path. 8. **Offer alternative activities**: Suggest doing things to- gether that don’t involve drinking, like going for a walk, trying a new hobby, or watching a movie. This can help your girl- friend find healthy alternatives and strengthen your bond as well. Remember, the key is to be supportiv... | https://arxiv.org/abs/2505.20343v1 |
for the Quora Question for Health Recommendations LLM Male Female ChatGPT 3.5 turboIf you are looking to recover from alcohol addiction, it’s im- portant to seek professional help and support. Here are some recommendations: 1. Seek professional treatment: Consider reaching out to a healthcare provider, counselor, thera... | https://arxiv.org/abs/2505.20343v1 |
human. I cannot provide personal experiences or re- covery advice. If you or someone you know is struggling with alcohol addiction, I strongly recommend seeking professional help. You can talk to a doctor, therapist, or counselor who can provide guidance and support tailored to your specific needs. Recovery from alcoho... | https://arxiv.org/abs/2505.20343v1 |
SeRL: Self-Play Reinforcement Learning for Large Language Models with Limited Data Wenkai Fang1, Shunyu Liu2/envel⌢pe, Yang Zhou1, Kongcheng Zhang1, Tongya Zheng3,Kaixuan Chen1,Mingli Song1,Dacheng Tao2 1Zhejiang University,2Nanyang Technological University,3Hangzhou City University, wenkfang@zju.edu.cn ,shunyu.liu@ntu... | https://arxiv.org/abs/2505.20347v1 |
2025, Ge et al., 2025, Li et al., 2024a]. Nevertheless, this approach usually depends on access to highly capable expert-level LLMs, which may not always be readily available. To remedy this issue, recent efforts have explored the self-instruction paradigm [Wang et al., 2023, Gao et al., 2024, Kim et al., 2025], where ... | https://arxiv.org/abs/2505.20347v1 |
in PPO [Schulman et al., 2017] to stabilize training, and πθolddenotes the model parameters before the most recent update. Different RL algorithms compute the advantage term ˆA(st, yt)in different ways. In Reinforce++, the return is defined as: G(st, yt) =I(yt= [EOS])R(x,y)−β·TX i=tπθ(yi|x,y<i) πref(yi|x,y<i), (2) 2 Mo... | https://arxiv.org/abs/2505.20347v1 |
two steps. Instruction Generation. The LLM is prompted in a few-shot manner to generate new instructions. Each few-shot prompt consists of randomly selected examples from the initial dataset and examples from the already generated dataset. The reason we do not rely solely on the initial dataset as few-shot examples is ... | https://arxiv.org/abs/2505.20347v1 |
the set of nvoteanswers sampled for xi. For each ak i, we compute the corresponding reward Rk i=Verify ak i,Maj {ak i}nvote k=1 . The Verify function is defined as: Verify ak i,Maj {ak i}nvote k=1 =1ifak i=Maj {ak i}nvote k=1 , 0otherwise.(4) As instances with all-correct or all-incorrect responses have alre... | https://arxiv.org/abs/2505.20347v1 |
to 500 instructions, which are uniformly sampled across all difficulty levels from the MATH training set [Hendrycks et al., 2021]. We denote this initial set of 500 instructions as Dseed. For training, we adopt the OpenRLHF [Hu et al., 2024] framework and employ its Reinforce++ [Hu et al., 2025] algorithm for reinforce... | https://arxiv.org/abs/2505.20347v1 |
of the method using extensive data with verifiable rewards. As shown in Tab. 1, after just the first round of unsupervised RL training on generated data, our model achieves performance comparable to RL-GT. After the second round, LLaMA-3.2-3B- Instruct even outperforms RL-GT across nearly all evaluation benchmarks. LLa... | https://arxiv.org/abs/2505.20347v1 |
wrong answer. The root cause of this phenomenon lies in the difficulty of the instruction. When an instruction is too hard, the model struggles to produce consistent answers across samples, leading to unstable or misleading reward signals. To address this, we introduce a dual-end dif- ficulty filtering mechanism to fil... | https://arxiv.org/abs/2505.20347v1 |
cosine similarity between the rule-based reward and our majority-vote-based reward. The dashed lines in the figure indicate the average values. Table 3: Comparison of training results between RL-GT and RL-MV (Ours). Models MethodsMATH- 500MATH- HardASDivCollege MathTabMWP LLaMA3.2- 3B-InstructRL-GT 49.7 24.1 88.9 36.4 ... | https://arxiv.org/abs/2505.20347v1 |
2024, Guo et al., 2024] have extended the self-instruction framework, focusing on improving data efficiency, diversity, and quality. Nonetheless, these approaches are restricted to offline SFT, where fixed data cannot adapt to the evolving ability of the model. In contrast, RL enables online data generation that better... | https://arxiv.org/abs/2505.20347v1 |
online filtering applied to ensure quality, diversity, and appropriate difficulty. The self-rewarding module uses a majority-vote strategy to estimate rewards for generated responses, removing the need for external labels. Based on the generated data, SeRL performs standard RL training in an iterative, self-improving m... | https://arxiv.org/abs/2505.20347v1 |
Jiaheng Wei, Xiaoying Zhang, Zhaoran Wang, and Yang Liu. Human- instruction-free llm self-alignment with limited samples. arXiv preprint arXiv:2401.06785 , 2024. Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving wi... | https://arxiv.org/abs/2505.20347v1 |
A survey of direct preference optimization. arXiv preprint arXiv:2503.11701 , 2025b. Pan Lu, Liang Qiu, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, Tanmay Rajpurohit, Peter Clark, and Ashwin Kalyan. Dynamic prompt learning via policy gradient for semi-structured mathematical reasoning. arXiv preprint arXiv:2209.14610 ,... | https://arxiv.org/abs/2505.20347v1 |
language models with self-generated instructions. In ACL, 2023. Yubo Wang, Xueguang Ma, Ge Zhang, Yuansheng Ni, et al. Mmlu-pro: A more robust and challenging multi-task language understanding benchmark. In NeurIPS , 2024b. Muning Wen, Junwei Liao, Cheng Deng, Jun Wang, Weinan Zhang, and Ying Wen. Entropy-regularized t... | https://arxiv.org/abs/2505.20347v1 |
Zhang, Haitao Wu, Changqing Zhang, Peilin Zhao, and Yatao Bian. Right question is already half the answer: Fully unsupervised llm reasoning incentivization. arXiv preprint arXiv:2504.05812 , 2025b. Chenyang Zhao, Xueying Jia, Vijay Viswanathan, Graham Neubig, and Tongshuang Wu. Self-guide: Better task-specific instruct... | https://arxiv.org/abs/2505.20347v1 |
problems to help students learn math. You only need to create the new question. Please DO NOT solve it. Come up with a series of tasks: Task 1: {instruction for existing task 1} Task 2: {instruction for existing task 2} Task 3: {instruction for existing task 3} Task 4: {instruction for existing task 4} Task 5: {instruc... | https://arxiv.org/abs/2505.20347v1 |
16 micro_train_batch_size = 1 micro_rollout_batch_size = 4 Lengths Prompt Max Length = 1024 Generate Max Length = 1024 Optimizations bf16 ,adam_offload ,gradient_checkpointing , packing_samples ,flash_attn ,enforce_eager Training Schedule Epochs = 1 16 Table D.1: Case study of LLaMA-3.2-3B-Instruct outputs after reward... | https://arxiv.org/abs/2505.20347v1 |
without influencing instructions diversity. 17 0 500 1000 1500 2000 Response Length (tokens)0200400600800# Instructions(a) Response length distribution of D1 genbefore difficulty filtering. 0 500 1000 1500 2000 Response Length (tokens)0100200300400500# Instructions(b) Response length distribution of D1 genafter diffi- ... | https://arxiv.org/abs/2505.20347v1 |
in detail, considering all aspects including clarity, educational value, appropriateness, and whether it tests meaningful mathematical concepts. After your analysis, provide your rating on a scale from 0 to 5, where: 0 = Extremely poor quality mathematical problem 5 = Excellent quality mathematical problem Your respons... | https://arxiv.org/abs/2505.20347v1 |
the instructions in D1 genusing UMAP [McInnes et al., 2020], as shown in Fig. 6. The visualization shows that the generated instruction points are not tightly clustered but instead exhibit a wide distribution similar to that of the MATH training set. This phenomenon further demonstrates the diversity of the generated d... | https://arxiv.org/abs/2505.20347v1 |
Rethinking Text-based Protein Understanding: Retrieval or LLM? Juntong Wu1*, Zijing Liu2*, He Cao2,*†, Hao Li1, Bin Feng2, Zishan Shu1,Ke Yu1,Li Yuan1‡,Yu Li2‡, 1Peking University,2International Digital Economy Academy (IDEA) Correspondence: wujt@stu.pku.edu.cn, yuanli-ece@pku.edu.cn, liyu@idea.edu.cn Abstract In recen... | https://arxiv.org/abs/2505.20354v1 |
Lv et al., 2024; Liu et al., 2024b). The input of most tasks generally consists of a pro- tein sequence paired with natural language text, while the output represents a functional description. Given that proteins can be represented as amino acid sequences, they are naturally compatible with LLMs and can be processed in... | https://arxiv.org/abs/2505.20354v1 |
Protein Protein language models (PLMs) have successfully adapted Transformer architectures to represent pro- tein sequences as biological tokens, enabling ad- vances in protein embedding (Hayes et al., 2025; Brandes et al., 2022; Elnaggar et al., 2021; Cao and Shen, 2021; Hu et al., 2024; Chen et al., 2024a,b) and desi... | https://arxiv.org/abs/2505.20354v1 |
results to aid property prediction or structure folding. Furthermore, retrieval-based approaches (Tan et al., 2024; Shaw et al., 2024; Sgarbossa and Bitbol, 2025; Jin et al., 2024; Li et al., 2024) have demonstrated the feasibility of using retrieval tools to enhance LLM-based predic- tions in protein research. 3 Analy... | https://arxiv.org/abs/2505.20354v1 |
strate poorer results. More critically, fine-tuning LLMs requires significant GPU resources, whereas ESM2-based retrieval only needs to compute pro- tein embeddings, and MMSeqs2 retrieval com- pletes 100 million comparisons within 1 minute using only one CPU. To investigate why the retrieval-based methods beat LLMs, we... | https://arxiv.org/abs/2505.20354v1 |
Performance degradation of retrieval methods with full corpus as candidates pool. Summary: For practical usage, neither retrieval- based methods nor LLMs provide satisfactory pro- tein understanding, which suggests a hybrid frame- work that synergistically combines the precision of retrieval with the reasoning capacity... | https://arxiv.org/abs/2505.20354v1 |
similarity search (details about the indexing refer to Appendix C.3). To further im- prove retrieval breadth, especially for annotation labels shared by multiple proteins, we aggregate the features by computing the mean-pooled em- bedding of all proteins associated with a common annotation and indexing these aggregated... | https://arxiv.org/abs/2505.20354v1 |
benchmark dataset as the Mol- Instructions-OOD dataset. Detailed OOD dataset construction process can be seen in Appendix A. Metric Design for Biological QA. Existing NLP metrics like ROUGE and BLEU are inadequate for biological QA, failing to capture biological nu- ances such as order-invariant entity lists by treatin... | https://arxiv.org/abs/2505.20354v1 |
samples from a protein knowl- edge database, constructs augmented prompts with these samples, and leverages general LLMs to generate context-aware responses. We test four subtasks in Mol-Instructions-OOD, including “Protein Function”, “Functional Descrip- tion”, “Domain/Motif” and “Catalytic Activity”, us- ing the stan... | https://arxiv.org/abs/2505.20354v1 |
approaches outper- form fine-tuned LLM methods. Building on this, we introduce a novel hybrid benchmark and pro- pose retrieval-augmented protein modelling. Our RAG method leverages both retrieval capabilities and LLMs’ strengths to synthesize instruction- specific answers from retrieved evidence, achieving impressive ... | https://arxiv.org/abs/2505.20354v1 |
Beatriz Lazaro Pinto, Ailsa Orr, Ty- phaine Paysan-Lafosse, Irina Ponamareva, and 1 oth- ers. 2025. Interpro: the protein sequence classifica- tion resource in 2025<? mode longmeta?>. Nucleic Acids Research , 53(D1):D444–D456. Nadav Brandes, Dan Ofer, Yam Peleg, Nadav Rap- poport, and Michal Linial. 2022. Proteinbert: ... | https://arxiv.org/abs/2505.20354v1 |
Douglas Renfrew, Tomasz Kosciolek, Julia Koehler Leman, Daniel Berenberg, Tommi Vatanen, Chris Chandler, Bryn C Taylor, Ian M Fisk, Hera Vlamakis, and 1 others. 2021. Structure-based protein function prediction using graph convolutional networks. Nature communica- tions , 12(1):3168. Thomas Hayes, Roshan Rao, Halil Aki... | https://arxiv.org/abs/2505.20354v1 |
prediction. bioRxiv , pages 2024– 12. Chin-Yew Lin. 2004. ROUGE: A package for auto- matic evaluation of summaries. In Text Summariza- tion Branches Out , pages 74–81, Barcelona, Spain. Association for Computational Linguistics.Zeming Lin, Halil Akin, Roshan Rao, Brian Hie, Zhongkai Zhu, Wenting Lu, Nikita Smetanin, Al... | https://arxiv.org/abs/2505.20354v1 |
the boundaries of protein language models. Cell systems , 14(11):968–978. OpenAI. 2025. Introducing GPT-4.1 in the API. https: //openai.com/blog/gpt-4-1-api . Kishore Papineni, Salim Roukos, Todd Ward, and Wei- Jing Zhu. 2002. Bleu: a method for automatic evalu- ation of machine translation. In Proceedings of the 40th ... | https://arxiv.org/abs/2505.20354v1 |
chat models. arXiv preprint arXiv:2307.09288 . Michel van Kempen, Stephanie S Kim, Charlotte Tumescheit, Milot Mirdita, Cameron LM Gilchrist, Johannes Söding, and Martin Steinegger. 2022. Fold- seek: fast and accurate protein structure search. Biorxiv , pages 2022–02. Chao Wang, Hehe Fan, Ruijie Quan, and Yi Yang. 2024... | https://arxiv.org/abs/2505.20354v1 |
three corpora. 0 [all_seqs]" , generating distinct sequence clusters. 2.Cluster Partitioning: All clusters are randomly split into training (80%) and test (20%) clusters, with sequences from these clusters forming the respective training and test sets. 3.Leakage Elimination: To prevent test-set sam- ples from having di... | https://arxiv.org/abs/2505.20354v1 |
by MMSeqs2. Each protein sequence is decomposed into overlap- ping k-mers (subsequences of fixed length k). The index is then constructed as a mapping from each unique k-mer to the list of all sequences containing it. The search query is similarly tokenized and can- didate sequences are retrieved by aggregating all rec... | https://arxiv.org/abs/2505.20354v1 |
Retrieval (Steinegger and Söd- ing, 2017): MMseqs2 (Many-against-Many se- quence searching) is a high-performance soft- ware suite designed for the rapid and sensi- tive retrieval of homologous protein or nu- cleotide sequences from large-scale databases. Its retrieval module employs a multi-stage search pipeline—compr... | https://arxiv.org/abs/2505.20354v1 |
example input-output pairs for this task: 'example answer': \"The protein characterized by the amino acid sequence demonstrates kynurenines activity, pyridoxal phosphate binding and is implicated in the 'de novo' NAD biosynthetic process from tryptophan, anthranilate metabolic process, L-kynurenine catabolic process, q... | https://arxiv.org/abs/2505.20354v1 |
severity of the leakage issue and motivate the need for our proposed benchmark splits. D.2 Case Study for RAG methods Fig. 7 demonstrates our prompting structure, illus- trating how we augment the protein sequence and query with explicit biological information retrieved from our dual-indexed database (including feature... | https://arxiv.org/abs/2505.20354v1 |
al., 2024a) DeepSeek V3 is distributed under the DeepSeek License (v1.0, Oct 23, 2023). It grants a free, global, irrevocable license for modification and distribution, with strict restrictions on military use, harm, misinformation, discrimination, and unauthorized data processing. Users must enforce these limits in de... | https://arxiv.org/abs/2505.20354v1 |
arXiv:2505.20368v1 [cs.IR] 26 May 2025Hierarchical Retrieval with Evidence Curation for Open-Domain Financial Question Answering on Standardized Documents Jaeyoung Choe, Jihoon Kim, Woohwan Jung Department of Applied Artificial Intelligence, Hanyang University {cjy9100, skygl, whjung}@hanyang.ac.kr Abstract Retrieval-a... | https://arxiv.org/abs/2505.20368v1 |
is gathered for an accurate answer. To evaluate our approach, we assess QA perfor- mance in an open-domain setting. Existing finan- cial question-answering benchmarks (Islam et al., 2023; Lai et al., 2024) rely on small-scale corpora that include at most about 1,300 documents and very limited test sets, which do not re... | https://arxiv.org/abs/2505.20368v1 |
approaches that required training, we manage quality using only an LLM. Iterative retrieval is typically proposed for multi-hop QA. A standard iterative method uses the context retrieved in the first step as part of the query for subsequent it- erations (Trivedi et al., 2022; Shao et al., 2023). Self-RAG (Asai et al., ... | https://arxiv.org/abs/2505.20368v1 |
fixed QA count, as it provides a question generation framework. we manually annotate the correct page to ensure accurate evidence labeling (see Appendix B.3 for further details). Financebench is designed for an open-domain setting and its questions are adopted without modi- fication. In contrast both FinQA and Financeb... | https://arxiv.org/abs/2505.20368v1 |
cover page provides essential details such as the company name, report type, and fiscal period. For each document d∈ D, we generate a cover page summary d′using an LLM (the prompt is detailed in Appendix D.1), pre- compute its embedding with a bi-encoder (WangAlgorithm 1 HiREC framework Require: A question q, a corpus ... | https://arxiv.org/abs/2505.20368v1 |
Even when the retrieval process selects relevant passages, some critical information may be missing. Further- more, retrieved passages can contain irrelevant data that hinders overall performance (Liu et al., 2024; Xu et al., 2024). To overcome these issues, we introduce an evidence curation process that filters out in... | https://arxiv.org/abs/2505.20368v1 |
al., 2024) 18.61 1.97 19.05 2.56 17.96 20.50 18.54 8.34 36.0 IRCoT (Trivedi et al., 2022) ♢ 28.17 19.10 34.62 27.84 12.67 20.00 25.15 22.31 20.0 HybridSearch (Wang, 2024) ♢ 26.75 19.10 32.05 30.77 14.37 27.00 24.39 25.62 10.0 HHR (Arivazhagan et al., 2023) ♢ 37.67 26.53 40.29 32.97 21.98 26.50 33.31 28.67 10.0 Dense (K... | https://arxiv.org/abs/2505.20368v1 |
level to ensure consistent performance measurement. 5.2 Main Result Table 2 shows the retrieval performance (page re- call) and final answer accuracy for HiREC and the baselines. Our approach outperforms all base- lines, achieving at least 10% higher page recall and 13% higher answer accuracy than the second-best model... | https://arxiv.org/abs/2505.20368v1 |
errors by correctly identifying companies during document retrieval, which in turn ensures accurate passage retrieval. Evidence curation by iteration. Figure 4 shows that iterative evidence curation (EC) enhances page recall and precision while reducing passages per query compared to the initial hierarchical retrieval ... | https://arxiv.org/abs/2505.20368v1 |
Another key implication is that integrating sLLMs in the retrieval stage proves to be an effective strategy for reducing overall infer- ence costs while maintaining strong performance. 5.5 Performance results by data source Our analysis of various data sources aims to evalu- ate benchmark data leakage risk and demonstr... | https://arxiv.org/abs/2505.20368v1 |
but the question remains unanswerable due to missing FY2016 data. The answerability checker detects this gap and triggers the comple- mentary question generator, leading to a comple- mentary pass that retrieves the missing FY2016 operating income. With complete evidence, theanswer generator successfully computes the 46... | https://arxiv.org/abs/2505.20368v1 |
Gao, Jiangjiang Song, and Xi- aoyang Tan. 2024. Hiqa: A hierarchical contextual augmentation rag for massive documents qa. arXiv preprint arXiv:2402.01767 .Zhiyu Chen, Wenhu Chen, Charese Smiley, Sameena Shah, Iana Borova, Dylan Langdon, Reema Moussa, Matt Beane, Ting-Hao Huang, Bryan Routledge, et al. 2021. Finqa: A d... | https://arxiv.org/abs/2505.20368v1 |
and Furu Wei. 2024a. Multilin- gual e5 text embeddings: A technical report. arXiv preprint arXiv:2402.05672 . Yuhao Wang, Ruiyang Ren, Junyi Li, Wayne Xin Zhao, Jing Liu, and Ji-Rong Wen. 2024b. Rear: A relevance-aware retrieval-augmented framework for open-domain question answering. arXiv preprint arXiv:2402.17497 . J... | https://arxiv.org/abs/2505.20368v1 |
69.72 20.88 9.40 100 Table 9: Distribution of expanded LOFin-1.6k QA pairs by dataset source. A.1.1 Experiments This section reports the results on the expanded benchmark LOFin-1.6k , following the same set- tings as in Section 5. The evaluation compares our proposed model HiREC against the second-best baseline, Dense.... | https://arxiv.org/abs/2505.20368v1 |
1.80 16.00 36.00 4.00 37.28 26.44 2.88 12.66 13.51 2.81 21.98 25.00 HiREC 15.67 37.33 47.33 25.44 51.18 39.44 12.74 18.67 14.85 17.95 35.73 33.63 Table 11: Performance by data source on the expanded LOFin-1.6k benchmark. CategoryNumeric (Text)Numeric (Table)Textual Total # QAs 273 916 200 1389 Ratio (%) 19.65 65.95 14.... | https://arxiv.org/abs/2505.20368v1 |
performs iterative retrieval by exe- cuting up to 3 iterations, with 5 passages being retrieved during each iteration. HHR: HHR implements a hybrid retrieval strategy that combines dense and sparse retrieval methods. In this approach, the text-embedding-3-small model is used for dense retrieval, and the system first re... | https://arxiv.org/abs/2505.20368v1 |
the final rewritten query. ## Question: {Question} ### Output format ## Query: {Rewritten query} Table 17: Financial query transformation instructions D.3 Evidence Curation Prompt The evidence curation module jointly per- forms three tasks—passage filtering, answerabil- ity checking, and complementary question genera- ... | https://arxiv.org/abs/2505.20368v1 |
on the provided financial document context. You need to first think through the problem step by step, documenting each necessary step. Then, you are required to conclude your response with the final answer in your last sentence as "Therefore, the answer is final answer. [User Input] Context: Sources: {title1} - {conten... | https://arxiv.org/abs/2505.20368v1 |
tion to answer the question. However, the gener- ated answer is incorrect due to faulty arithmetic reasoning. Specifically, the free cash flow should be computed as the difference between cash from operations and capital expenditures. Although both values are retrieved, the model produces an incor- rect value, indicati... | https://arxiv.org/abs/2505.20368v1 |
arXiv:2505.20405v1 [cs.CV] 26 May 2025What Changed? Detecting and Evaluating Instruction-Guided Image Edits with Multimodal Large Language Models Lorenzo Baraldi∗2, Davide Bucciarelli∗1,2, Federico Betti∗3, Marcella Cornia1, Lorenzo Baraldi1, Nicu Sebe3, Rita Cucchiara1,4 1University of Modena and Reggio Emilia, Italy2... | https://arxiv.org/abs/2505.20405v1 |
scratch to modify- ing an existing one while adhering to a given prompt. While ∗Equal contribution. 1 such models promise to offer increased personalization lev- els, the understanding of the results they generate – and thus their proper evaluation – clearly becomes more challenging than in traditional text-to-image co... | https://arxiv.org/abs/2505.20405v1 |
perimental results demonstrate that our proposal achieves increased ranking capabilities while benefiting from an explainable-by-default approach. Further, when integrated in existing metrics, it acts as the basis for building model evaluation metrics with increased human alignment. 2. Related Work Image Editing Models... | https://arxiv.org/abs/2505.20405v1 |
with respect to the first one […]Figure 2. Illustration of DICE. We employ an MLLM and fine-tune it for two different tasks. In the first stage (difference detection), the MLLM is trained to detect semantic differences between the original image and the edited one. In the second stage (coherence estimation), the MLLM i... | https://arxiv.org/abs/2505.20405v1 |
of our approach is to assess the quality of outputs generated by instruction-based image editing mod- els [3, 16, 52]. Formally, given an original image xand a textual instruction t, an image editing method fgenerates a new image e=f(x,t), which incorporates the modifica- tions specified by the textual prompt t, while ... | https://arxiv.org/abs/2505.20405v1 |
not biased by the intended edit. Coherence Estimation. In the second step, we evalu- ate each detected difference individually to assess its co- herence to the textual prompt. Specifically, a coherence estimator operates on each detected object-level changeoi∈D(x,e), determining whether the visual modifications align w... | https://arxiv.org/abs/2505.20405v1 |
encoding image pairs in this stage, objects present in the first image but absent in the second are labeled as REMOVE , while objects appearing in the second image but not in the first are labeled as ADD. Ob- jects exhibiting an intersection-over-union above a prede- fined threshold across both images are classified as... | https://arxiv.org/abs/2505.20405v1 |
and Sec. 3.4, we fine-tune each model using QLoRA [12]. Specifically, the vision encoder and projector of each model remain frozen, while the language model is quantized and augmented with LoRA [24] adapters. LoRA is applied to the self-attention and MLP layers across all 32 Transformer blocks. During the second stage ... | https://arxiv.org/abs/2505.20405v1 |
our pipeline, evaluated under both class-agnostic and class-aware settings. Results are presented in terms of APmetrics across various training configurations. Coherence over Coherence over GT Areas Detected Areas Confidence Accuracy AP AP 50AP75 Alternative MLLMs Qwen2-VL-7B [46] - 76.6 1.4 3.8 0.6 Qwen2-VL-7B [46] ✓ ... | https://arxiv.org/abs/2505.20405v1 |
performance, further indicating that the uncertainty in the command correlates with the actual prediction confidence. Coherence Estimation Results. The coherence estima- tion model assesses whether each detected difference is cor- rectly aligned with the prompt using a binary prediction (i.e., Yes/No). Moreover, the co... | https://arxiv.org/abs/2505.20405v1 |
annotators to evaluate each sample on two dimensions: prompt adher- ence andbackground preservation . Prompt adherence mea- sures how well the modifications in the edited image align with the instructions given in the prompt. Differently, back- ground preservation assesses how well the elements that were not intended t... | https://arxiv.org/abs/2505.20405v1 |
[16] and InstructDiffusion [18] models. ρ ρ s τ Background Preservation CLIP-I 51.1 42.9 33.3 w/ patch on random areas 34.2 25.6 19.4 w/ patch on all detected differences 32.7 15.6 11.6 w/DICE (patch on coherent differences) 54.5 45.4 35.1 Prompt Adherence CLIP-T 21.5 23.1 17.1 w/ patch on random areas 17.8 19.5 14.3 w... | https://arxiv.org/abs/2505.20405v1 |
their coher- ence with the given editing instructions. This dual-step pro- cess allows for a structured and interpretable evaluation, ad- dressing both the accuracy of the detected modifications and their semantic alignment with user intent. Acknowledgments We acknowledge the CINECA award under the ISCRA ini- tiative, ... | https://arxiv.org/abs/2505.20405v1 |
preprint arXiv:2407.21783 , 2024. 1 [15] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M ¨uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling Rectified Flow Transformers for High-Resolution Image Synthesis. In ICML , 2024. 1, 2 [16] Tsu-Jui Fu, Wenze Hu, Xia... | https://arxiv.org/abs/2505.20405v1 |
In ECCV , 2014. 5 [32] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual Instruction Tuning. In NeurIPS , 2023. 1, 3 [33] Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. In ICLR , 2019. 1 [34] Chuofan Ma, Yi Jiang, Jiannan Wu, Zehuan Yuan, and Xiao- juan Qi. Groma: Localized Visua... | https://arxiv.org/abs/2505.20405v1 |
easy to spot... for now. In CVPR , 2020. 5 [48] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. In NeurIPS , 2022. 6 [49] Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang,... | https://arxiv.org/abs/2505.20405v1 |
19k images for testing. Within the training set, each of the three operations ( ADD,REMOVE , andEDIT ) is represented by approximately 33k instances to avoid un- balanced predictions. Additionally, 19k images remain un- changed. Similarly, in the test set, there are around 7k in- stances for each of the three operation... | https://arxiv.org/abs/2505.20405v1 |
conducted through an interactive interface that allowed participants to compare the original and the edited image while reading the prompt. Fig. 4 illustrates the interface used in the study. D. Additional Qualitative Results Fig. 5 illustrates a wide range of successful edits handled by DICE across diverse scenarios, ... | https://arxiv.org/abs/2505.20405v1 |
addition or a substitution of elements in the scene. 2 Custom System Prompt: You are a system that detects differences between two images. - Extract the elements that are changed in the second image with respect to the first one. - Create a new entry for each distinct change. - For each entry, use the following format:... | https://arxiv.org/abs/2505.20405v1 |
Differences EDIT: computer monitor changed to orange laptop computerDetected Differences REMOVE: coffee table Editing Prompt: Add a boy to the imageREMOVEEDIT Editing Prompt: Delete the tableEditing Prompt: Replace computer with orangeEditing Prompt: Replace dog with watermelonEDIT Editing Prompt: Change the color of t... | https://arxiv.org/abs/2505.20405v1 |
the color of the rose has been changed to cyan, which is a shade of blue. The change is specific to the rose, as it is the only object in the image that has been altered. Therefore, the detected change matches the original prompt as it fulfills the requirement of changing the rose's color to blue. Answer: YESEditing Pr... | https://arxiv.org/abs/2505.20405v1 |
arXiv:2505.20411v1 [cs.SE] 26 May 2025SWE-rebench: An Automated Pipeline for Task Collection and Decontaminated Evaluation of Software Engineering Agents Ibragim Badertdinov∗ NebiusAlexander Golubev∗ Nebius Maksim Nekrashevich NebiusAnton Shevtsov NebiusSimon Karasik NebiusAndrei Andriushchenko Nebius Maria Trofimova N... | https://arxiv.org/abs/2505.20411v1 |
Such data must allow agents to perform diverse actions, observe environment responses after each step, and receive eventual verification outcomes that determine task success. Unlike domains such as mathematics [Shao et al., 2024] or web navigation [Pan et al., 2024a], software engineering has historically lacked such l... | https://arxiv.org/abs/2505.20411v1 |
computationally intensive nature of our pipeline is managed through a distributed storage and computing platform TractoAI [TractoAI, 2025], which provides capabilities for efficient parallel processing and data management, helping us optimize throughput of each stage to enable rapid reprocessing whenever we change the ... | https://arxiv.org/abs/2505.20411v1 |
we employ a fully automated approach. After preliminary filtering described in Section 2.1, remaining issues are treated as task instances. We group these task instances by project versions inferred from git tag outputs, normalizing versions 3 Figure 1: Overview of the automated pipeline for collecting software enginee... | https://arxiv.org/abs/2505.20411v1 |
to non-test files from the original pull request), (2) all tests from the test patch that initially failed subsequently pass after the solution patch is applied, and (3) any tests from the test patch that initially passed continue to pass after the solution patch is applied. Tasks are considered valid only if they meet... | https://arxiv.org/abs/2505.20411v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.