text string | source string |
|---|---|
about communi- cating it to the teacher when gym class starts?”Responding to Emotions "I don’t feel very confident about talking to the teacher, honestly. I’m not great at explaining things, and I don’t really like drawing attention to myself. I’d rather just keep quiet and hope it works out on its own. ”"You’ve made a... | https://arxiv.org/abs/2505.20201v2 |
(SD = 0.75) com- pared to score of 2.45 (SD = 0.63) obtained by o1 (Mann–Whitney U test; p <0.001; U=597057 ). Similarly, R1 obtained significantly higher scores forPerceived Severity (2.75; SD=0.67) compared to o1 (2.56; SD=0.54) (Mann–Whitney U test; p<0.001; U=559539 ). Lastly, R1 also obtained significantly higher ... | https://arxiv.org/abs/2505.20201v2 |
metrics (ref. Sec- tion 4) based on patient’s dominant personality traits (Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism) and health literacy level (Basic/Advanced). Table 2 presents results across these strata for three personality types. Impact of Personality Traits : Among patient with advanc... | https://arxiv.org/abs/2505.20201v2 |
Accuracy metric. 5 5-10 10-152.32.42.52.6 Mean Score Perceived Susceptibility o1 R1 5 5-10 10-152.52.62.72.8 Perceived Severity o1 R1 5 5-10 10-152.72.82.93.0 Perceived Benefits o1 R1 5 5-10 10-153.23.43.6 Flow Correctness o1 R1 5 5-10 10-154060 Diagnostic Accuracy o1 (Hard) o1 (Soft) R1 (Hard) R1 (Soft) 5 5-10 10-1512... | https://arxiv.org/abs/2505.20201v2 |
generation framework based on dialogue ranking (Li et al., 2025). The Ask Patients with Patience (APP) framework (Zhu and Wu, 2025) allow LLMs to generate multi-turn conversations based on medical guidelines and entropy minimization. In contrast, (Liu et al., 2025b) proposed a patient simulator for multi-turn diagnosti... | https://arxiv.org/abs/2505.20201v2 |
work presents novel contribution on mul- tiple fronts, it is important to acknowledge the lim- itations of our work. While MedAgent framework provides a method for generating realistic mental health sensemaking conversations, it is important to recognize both patient and sensemaker actors are modeled by LLMs which may ... | https://arxiv.org/abs/2505.20201v2 |
clin- ical safety and hallucination rates of llms for medical text summarisation. npj Digital Medicine , 8(1):1–15. Catherine L Auriemma, Anne Song, Lake Walsh, Ja- son J Han, Sophia R Yapalater, Alexander Bain, Lindsay Haines, Stefania Scott, Casey Whitman, Stephanie P Taylor, and 1 others. 2024. Classification of doc... | https://arxiv.org/abs/2505.20201v2 |
Janz and Marshall H Becker. 1984. The health belief model: A decade later. Health education quar- terly, 11(1):1–47. Di Jin, Eileen Pan, Nassim Oufattole, Wei-Hung Weng, Hanyi Fang, and Peter Szolovits. 2021. What disease does this patient have? a large-scale open domain question answering dataset from medical exams. A... | https://arxiv.org/abs/2505.20201v2 |
Health promotion glossary 2021. Health promotion interna- tional , 36(6):1578–1598. OpenAI. 2024. Introducing Openai o1. https:// openai.com/o1/ . [Accessed 23-04-2025]. OpenAI. 2025. Chatgpt. OpenAI-GPT-4o. 2024. Hello gpt-4o. https:// openai.com/index/hello-gpt-4o/ . [Accessed 07-09-2024]. David Restrepo, Chenwei Wu,... | https://arxiv.org/abs/2505.20201v2 |
Sensemaker-patient Conversational Style Openness Imagination, curiosity, com- fort with novel, unusual, and abstract ideas.High : Ask why/how questions about symptoms and medicines. Explore alternatives. Receptive to innovative therapies. Low: Prefer familiar terms and concrete instruc- tions. May resist complex or les... | https://arxiv.org/abs/2505.20201v2 |
by the sensemaker). The intent adds a unique vignette to help guide the conversa- tion with the sensemaker at the early stages. To mimic a real-world setting, we refer to Aydin et al. (2024), which analyzed 201 works studying usage of AI in healthcare to characterize six themes of how patients use LLMs in these setting... | https://arxiv.org/abs/2505.20201v2 |
Table 12 with slight modifications in the gather- ing information and providing information stage as shown in Table 13 and Table 14 system prompts, respectively. Table 15 shows the goals associated with each state that are passed to the system prompt as CUR- RENT_STAGE_GOAL. H Sensemaker Message Redundancy Checking Sub... | https://arxiv.org/abs/2505.20201v2 |
therapist who can help create a safety plan and explore treatment options." showcase that while the sensemaker can provide a potential diagnosis and it fails to explicitly address the patient’s sus- ceptibility to their condition and the likelihood of experiencing a health problem due to the underly- ing condition. In ... | https://arxiv.org/abs/2505.20201v2 |
this diagnosis is really affecting you. It’s completely under- standable to feel this way. I’m here to support you, and we can discuss ways to help manage both your symptoms and the emotional impact. Table 7: The 6-function model of medical communication (De Haes and Bensing, 2009; King and Hoppe, 2013), which characte... | https://arxiv.org/abs/2505.20201v2 |
has been meaningfully completed, ensuring a natural transition. Warnings: - Do not keep the conversation stuck in the same stage for multiple iterations unless necessary. If progression is unclear, consider whether the user is engaging sufficiently before deciding. User PromptPOTENTIAL_NEXT_STAGE_REASONING: <next_stage... | https://arxiv.org/abs/2505.20201v2 |
or direction, stay in ’Decision Making’ to provide additional support. Responding to EmotionsThe next stage would be ’exit’ because the conversation has reached its end. Move forward from the ’Responding to Emotions’ when you validated their emotions with an empathetic response and the patient has replied with an affir... | https://arxiv.org/abs/2505.20201v2 |
should be independent of other NEW_FACTS. Your output should strictly follow the format: # <fact> : <’FactNotPresent’ or ’FactPresent’>. Where fact presents one fact from NEW_FACTS. Output the labels for each fact in NEW_FACTS and keep the fact text as it is (do not change the words). You should not change the words of... | https://arxiv.org/abs/2505.20201v2 |
gathering more information to either confirm or reject the diagnosis hypotheses. Your output should strictly be in the following format: OUTPUT_REASONING: <your step-by-step reasoning> OUTPUT_MESSAGE: <3 plausible messages to the patient each on a new line and starting with "# ">. Do not include other formatting. All t... | https://arxiv.org/abs/2505.20201v2 |
respond with empathy and compassion. Gathering InformationYour goal is to develop a comprehensive understanding of the patient’s needs, concerns, and medical history by exploring their condition from both biological and psychosocial perspectives. This understanding will allow you to support the patient in achieving the... | https://arxiv.org/abs/2505.20201v2 |
some additional instructions: 1. You should not add any new statement which was not present in the STATEMENT_MEMORY . 2. If a new statement in the CANDIDATE_STATEMENTS has a different phrasing but serves a similar context to any of the statements present in the STATEMENT_MEMORY , it should be considered ’Redun- dantSta... | https://arxiv.org/abs/2505.20201v2 |
then at the last output the VIGNETTE_NAME for the most logical vignette for the given case study. User PromptCASE_STUDY: <case_study> Strictly follow the format: <ASSIGNED_VIGNETTE_NAME: VIGNETTE_NAME>. Use the exact vignette name and nothing else. Table 19: (Patient Module) Prompt used to attach a high-level conversat... | https://arxiv.org/abs/2505.20201v2 |
Big 5 personality traits are defined as: <personality_trait_definitions> Instructions: 1. Using the list of medical facts given below, generate a conversational human-like response to an LLM in first person, using the emotions, tone, word choice and intensity of a patient who has the levels of the Big 5 Personality Tra... | https://arxiv.org/abs/2505.20201v2 |
explore some coping strategies related to your condition? REASONING: This is a question about next steps, hence the answer is in the affirmative. ANSWER: Yes, I would like to learn about some coping strategies. EXAMPLE 3: QUESTION: Looks like you are experiencing insomnia. When did you first start feeling these symptom... | https://arxiv.org/abs/2505.20201v2 |
for automated Perceived Susceptibility evaluation. 28 Type Prompt System PromptTask: You are an expert LLM-as-a-judge model and your task is to evaluate a sensemaker’s communication of "Perceived Severity" (definition given as SEVERITY_DEFINITION) within a patient-sensemaker conver- sation. SEVERITY_DEFINITION: "Percei... | https://arxiv.org/abs/2505.20201v2 |
The sensemaker failed to convey any benefits or efficacy of the advised action in reducing the risk or seriousness of the health impact. However, the tone was empathetic and engaging. 3. (Good: Score = 3): The sensemaker conveyed basic information about the benefits or efficacy of the advised action in reducing the ris... | https://arxiv.org/abs/2505.20201v2 |
be given a specific part from a patient-sensemaker conversation, the sensemaker’s diagnosis statement (given as PATIENT_DIAGNOSIS) and the ground truth diagnosis of the patient (given as GT_DIAGNOSIS). Your task is to evaluate the sensemaker’s "Diagnostic Accuracy" from this conversation snippet and assign a binary lab... | https://arxiv.org/abs/2505.20201v2 |
thorough understanding of the patient’s condition by exploring both biological and psychosocial factors, including their needs, concerns, and medical history. - Providing Information: Offer a potential diagnosis that directly addresses the patient’s concerns, ensuring clarity and sensitivity in communication. - Decisio... | https://arxiv.org/abs/2505.20201v2 |
How to Improve the Robustness of Closed-Source Models on NLI Joe Stacey1, Lisa Alazraki1, Aran Ubhi1, Beyza Ermis2, Aaron Mueller3, Marek Rei1 1Imperial College London,2Cohere Labs,3Northeastern University & Technion – IIT {j.stacey20, lisa.alazraki20, marek.rei}@imperial.ac.uk aran.ubhi@me.com, beyza@cohere.com, aa.mu... | https://arxiv.org/abs/2505.20209v1 |
A). As fine-tuning closed-source models with large- scale NLI datasets can be cost prohibitive, we in- stead consider a fixed training budget of 10,000 instances. We find that for NLI, closed-source autoregressive LLMs fine-tuned with this number of examples perform similarly in-distribution, but with substantially bet... | https://arxiv.org/abs/2505.20209v1 |
during training. The rise of LLMs has made data augmentation a popular approach for improving performance and robustness, using large-scale synthetic datasets as additional training data (Wu et al., 2022; Liu et al., 2022; Wang et al., 2023b; Chen et al., 2023; Hos- seini et al., 2024; Banerjee et al., 2024a). One chal... | https://arxiv.org/abs/2505.20209v1 |
class suggesting a lack of diffi- culty, while high confidence in the wrong class may instead suggest annotation errors (Swayamdipta et al., 2020). We therefore choose examples based on maximising the uncertainty of the model pre- dictions, choosing Dc upas the top Kexamples in Dc potentialwith the highest entropy over... | https://arxiv.org/abs/2505.20209v1 |
this method to using Dpotential 3It is common for large-scale NLI datasets to involve instances with repeated premises (Bowman et al., 2015; Williams et al., 2018; Nie et al., 2020a) 4See Appendix Q for the additional filtering we do on this data to improve the quality Premise Hypothesis Label: Neutral 🔄 Hypothesis Co... | https://arxiv.org/abs/2505.20209v1 |
2024). Inspired by the if in doubt, discard approach from task-oriented dialogue (Stacey et al., 2024a), we generate eight few-shot predictions from Mper instance, and re- tain only those for which all predictions agree. 4 Experiments We evaluate M,Mbaseand each of our proposed methods on a diverse set of out-of-distri... | https://arxiv.org/abs/2505.20209v1 |
PoE and Example Reweighting for a DeBERTa-base model. The best results are in bold .↑and↓show better or worse average performance compared to the DeBERTa-large baseline. models using different random seeds. Each of our proposed methods is designed with- out access to the out-of-distribution (OOD) data and uses no OOD-b... | https://arxiv.org/abs/2505.20209v1 |
59.62 86.93 87.24 71.21 72.55 80.21 79.63 Random Sampling 92.55 65.80 58.66 55.37 56.48 60.60 60.15 59.51 ↓86.94 87.00 71.17 69.56 81.15 79.16 ↓ Sampling: Misclassified Sampling 92.32 65.48 57.98 52.73 62.94 56.88 60.32 59.39 ↓86.87 87.06 71.85 67.52 79.80 78.62 ↓ Hypothesis Concat Sampling 92.56 65.72 58.14 55.88 62.8... | https://arxiv.org/abs/2505.20209v1 |
the best-performing meth- ods. For Difficulty Score Sampling , we test alter- native scoring functions, and for Uncertainty Sam- pling , we restrict upsampling to correctly predicted but uncertain examples. Neither adaptation yields further improvements (Appendix I). We also apply Uncertainty Sampling to maths datasets... | https://arxiv.org/abs/2505.20209v1 |
single domain datasets. 8As a result of the slower few-shot inference with Com- mand R, we fine-tune this model with the data generated by GPT-4o-mini5.5 Analysis of Dup We analyse the examples selected for Dupby man- ually inspecting 50 examples per method and as- signing a difficulty score between 1-10. We also revie... | https://arxiv.org/abs/2505.20209v1 |
75.76 58.82 62.61 65.90 88.30 88.54 71.56 73.42 85.09 81.38 Short & Simple Generation 92.71 75.90 65.38 62.47 83.46 67.08 64.73 69.84 ↑88.82 88.53 72.94 77.94 67.08 82.94↑ Long & Complex Generation 92.54 76.28 67.22 64.28 68.28 66.36 62.78 67.53 ↑88.61 89.01 73.01 75.30 87.67 82.72↑ Table 4: Testing our data generation... | https://arxiv.org/abs/2505.20209v1 |
best strategy is to train with more challenging training examples. However, for less challenging OOD data, replacing some of the training examples with LLM-generated data is the best strategy and can lead to substantial improvements. Finally, our results show that LLMs are consid- erably more robust than existing encod... | https://arxiv.org/abs/2505.20209v1 |
Choi. 2020a. Adversarial filters of dataset biases. In Proceedings of the 37th Inter- national Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event , volume 119 of Proceedings of Machine Learning Research , pages 1078–1088. PMLR. Ronan Le Bras, Swabha Swayamdipta, Chandra Bha- gavatula, Rowan Zelle... | https://arxiv.org/abs/2505.20209v1 |
the 2024 Conference on Empiri- cal Methods in Natural Language Processing , pages 2990–3001, Miami, Florida, USA. Association for Computational Linguistics. Mengnan Du, Varun Manjunatha, Rajiv Jain, Ruchi Deshpande, Franck Dernoncourt, Jiuxiang Gu, Tong Sun, and Xia Hu. 2021. Towards interpreting and mitigating shortcu... | https://arxiv.org/abs/2505.20209v1 |
Ghaddar, Khalil Bibi, Phillippe Langlais, and Pascal Poupart. 2022. CILDA: Contrastive data augmen- tation using intermediate layer knowledge distilla- tion. In Proceedings of the 29th International Con- ference on Computational Linguistics , pages 4707– 4713, Gyeongju, Republic of Korea. International Committee on Com... | https://arxiv.org/abs/2505.20209v1 |
ˇrej Sotolá ˇr, and Vlastimil Martinek. 2023. Calc-x and calcformers: Empowering arithmetical chain-of-thought through interaction with symbolic systems. In Proceedings of the The 2023 Conference on Empirical Methods in Natural Language Processing: Main track , Sin- gapore, Singapore. Association for Computational Ling... | https://arxiv.org/abs/2505.20209v1 |
one among all ? an empirical study towards the robustness of knowledge distillation in natural language understanding. In Findings of the Associ- ation for Computational Linguistics: EMNLP 2021 , pages 750–762, Punta Cana, Dominican Republic. Association for Computational Linguistics. Hunter Lightman, Vineet Kosaraju, ... | https://arxiv.org/abs/2505.20209v1 |
Linzen. 2019. Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference. In Proceed- ings of the 57th Annual Meeting of the Association for Computational Linguistics , pages 3428–3448, Flo- rence, Italy. Association for Computational Linguis- tics. Junghyun Min, Thomas McCoy, Dipanjan D... | https://arxiv.org/abs/2505.20209v1 |
Ahmad, Ilge Akkaya, Florencia Leoni Ale- man, Diogo Almeida, Janko Altenschmidt, Sam Alt- man, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haim- ing Bao, Mohammad Bavarian, Jeff Belgum, and 262 others. 2024. Gpt-4 technical report. Preprint , arXiv:2303.08774. Bhargavi Par... | https://arxiv.org/abs/2505.20209v1 |
Association for Computa- tional Linguistics: Human Language Technologies (Volume 4: Student Research Workshop) , pages 56– 74, Mexico City, Mexico. Association for Computa- tional Linguistics.Joe Stacey, Pasquale Minervini, Haim Dubossarsky, Oana-Maria Camburu, and Marek Rei. 2024b. Atomic inference for NLI with genera... | https://arxiv.org/abs/2505.20209v1 |
2022. Mitigating spurious cor- relation in natural language understanding with coun- terfactual inference. In Proceedings of the 2022 Con- ference on Empirical Methods in Natural Language Processing , pages 11308–11321, Abu Dhabi, United Arab Emirates. Association for Computational Lin- guistics. Prasetya Ajie Utama, N... | https://arxiv.org/abs/2505.20209v1 |
of neural models in monotonicity reasoning. InProceedings of the Eighth Joint Conference on Lexical and Computational Semantics (*SEM 2019) , pages 250–255, Minneapolis, Minnesota. Associa- tion for Computational Linguistics. An Yang, Beichen Zhang, Binyuan Hui, Bofei Gao, Bowen Yu, Chengpeng Li, Dayiheng Liu, Jian- ho... | https://arxiv.org/abs/2505.20209v1 |
al., 2020) or em- beddings (Zhou and Bansal, 2020). Despite the success of adversarial training with LSTM mod- els, we are not aware of any work showing that these methods can also be successfully applied to transformer-based NLI models. A.2 Data Augmentation An increasingly common approach for improv- ing model robust... | https://arxiv.org/abs/2505.20209v1 |
al., 2022a; Ross et al., 2022; Koulakos et al., 2024; Zang and Liu, 2024) with the e-SNLI human annotated explanations (Camburu et al., 2018). Models can learn from the human explanations by supervising the CLS token in the final layer of the model (Stacey et al., 2022a), or by training models to generate explanations ... | https://arxiv.org/abs/2505.20209v1 |
are adjusted, with this adjustment depending on the loss so far during training for the groups they belong to. Sagawa et al. (2020) find that DRO improves the robust- ness on instances belonging to the worst group during inference, although with lower overall per- formance. Rather than using known heuristics, sub- sequ... | https://arxiv.org/abs/2505.20209v1 |
more fine- tuning data improves in-distribution performance, but without improvements in model robustness. Our few-shot baseline using Mis substantially worse on SNLI-test, but also substantially more robust on Standard-OOD and Challenge-OOD. D Experiments Training on MNLI We additionally experiment with using MNLI as ... | https://arxiv.org/abs/2505.20209v1 |
on the Hard test set, whereas for Hypothesis Concat Sampling and Uncertainty Sampling the largest in- crease is on the Ambiguous test set. For SNLI, the differences between the baseline and our method are smaller, with our methods providing no clear advantage over the baseline model. For the data generation methods, we... | https://arxiv.org/abs/2505.20209v1 |
our best performing methods when training on MNLI, a dataset with five different training domains. SNLI MNLI-m Method Hard Amb. Easy Hard Amb. Easy Baselines: GPT-4o-mini 67.98 89.68 97.47 59.27 80.90 91.89 Our methods: Misclassified Sampling 68.20 89.44 97.27 60.49 80.65 91.69 Difficulty Score Sampling 67.85 89.98 97.... | https://arxiv.org/abs/2505.20209v1 |
our model to generate chain- of-thought answers. We experiment with K= |Dup| = 0.15 ×|Dinit|, similar to our NLI experiments. But unlike NLI, there is less variety in the complexity of the dif- ferent training examples in Calc-Ape210k, so we also experiment with K= |Dup| = 0.50 ×|Dinit|. We find that when we set K= |Du... | https://arxiv.org/abs/2505.20209v1 |
SNLI r1 r2 r3 COPA INLI-I WANLI Avg. MNLI-m MNLI-mm FEVER Scitail INLI-NLI Avg. GPT-4o-mini - Difficulty Score Sampling: C + D 92.55 67.42 58.34 54.98 57.58 59.54 61.08 59.82 87.93 87.79 71.59 70.62 80.61 79.71 D 92.66 67.48 59.10 56.42 61.40 60.46 61.51 60.68 87.64 87.66 71.64 71.88 80.81 79.92 C + D + P + F 92.41 67.... | https://arxiv.org/abs/2505.20209v1 |
expected, the variance is often large for out-of-distribution test sets (Mc- Coy et al., 2020). We therefore use 5 random seeds for every experiment (in both the main paper and the appendix), and we compare average per- formance across multiple different OOD datasets (using our average scores for the Challenge-OOD and ... | https://arxiv.org/abs/2505.20209v1 |
the hypothesis statement. When the task is to determine the cause of the premise, we use the following template for the hypothesis: “[choice_1] is a more likely cause of this than [choice_2]", with a label of entailment or non-entailment. Then, for instances when we determine which sentence is the most likely effect of... | https://arxiv.org/abs/2505.20209v1 |
arXiv:2505.20215v1 [cs.CL] 26 May 2025Dependency Parsing is More Parameter-Efficient with Normalization Paolo Gajo University of Bologna paolo.gajo2@unibo.itDomenic Rosati Dalhousie University Domenic.Rosati@Dal.Ca Hassan Sajjad Dalhousie University HSajjad@dal.caAlberto Barrón-Cedeño University of Bologna a.barron@uni... | https://arxiv.org/abs/2505.20215v1 |
which ensures that each entry in the score matrix has a standard deviation of 1, since Std “?dk. Consequently, lower input variance will result in more stable outputs. We observe that there is no consistency in the literature on the use of this scaling term for DP. Most works do not use this scaling term [ 7,8,11,3,8,1... | https://arxiv.org/abs/2505.20215v1 |
In this work, we focus on encoder-style models, since they currently achieve the best performance on DP tasks [ 30,12,3] and are much more parameter efficient than LLM-based solutions. These models approach entity (node) prediction analogously to NER, while edges and relations are handled via MLP projection of node-pai... | https://arxiv.org/abs/2505.20215v1 |
rank-rapproximation Ar“řr i“1σiuivJ i. This means that increasing the rank from r´1toradds a new PSD component ∆r: ArKxxArJ“Ar´1KxxAr´1J`∆r 3 Table 1: Samples per partition and entity/relation classes for the datasets used in this paper. Data (Train / Dev / Test) Entities Relations ADE 2,563 / 854 / 300 [10] disease, d... | https://arxiv.org/abs/2505.20215v1 |
SynDP on the Penn Treebank [ 24] it is closed access and prohibitively expensive. Instead, we use SciDTB [38], a discourse analysis dataset comprising 798 abstracts extracted from the ACL Anthology. It was processed for the syntax dependency parsing task using Stanza [ 26]. From these two datasets, we consider the xPOS... | https://arxiv.org/abs/2505.20215v1 |
is then used in the MST algorithm, producing trees with a single root and no cycles. Prior to energy calculation, edge scores and relation scores are scaled so that low values are squished and high values are increased, making the log softmax produce a hard adjacency matrix. The model is trained end-to-end jointly on t... | https://arxiv.org/abs/2505.20215v1 |
Following the best results obtained by [ 7], they use three BiLSTM layers in the parser with a hidden size of 400, while the four MLPs following the stacked BiLSTM have an output size of 500 for the edge representations and 100 for the relations. 4.4 Evaluation Following [ 3,8,12], we measure tagging and parsing perfor... | https://arxiv.org/abs/2505.20215v1 |
to 85%. As laid out in Section 3, score variance tends to decay with deeper BiLSTM stacks. This in turn produces a converging trend as Lψincreases. When looking at Figure 3, this is especially evident for ERFGC, enEWT, and SciDTB, for which the beneficial effect of score normalization shrinks smoothly with higher value... | https://arxiv.org/abs/2505.20215v1 |
a single BiLSTM layer can be sufficient to match or surpass the results of state-of-the-art architectures with a decrease in trained parameters of up to 85%. In the case of SciERC, a challenging dataset for semantic dependency parsing, we found that the performance boost was particularly great when only training the bi... | https://arxiv.org/abs/2505.20215v1 |
for Computational Linguistics. [7]Timothy Dozat and Christopher D. Manning. Deep Biaffine Attention for Neural Depen- dency Parsing. In International Conference on Learning Representations . arXiv, March 2017. arXiv:1611.01734 [cs]. [8]Timothy Dozat and Christopher D. Manning. Simpler but More Accurate Semantic Depende... | https://arxiv.org/abs/2505.20215v1 |
Joakim Nivre and Chiao-Ting Fang. Universal dependency evaluation. In Proceedings of the NoDaLiDa 2017 Workshop on Universal Dependencies (UDW 2017) , pages 86–95, 2017. [21] Giovanni Paolini, Ben Athiwaratkun, Jason Krone, Jie Ma, Alessandro Achille, RISHITA ANUBHAI, Cicero Nogueira dos Santos, Bing Xiang, and Stefano... | https://arxiv.org/abs/2505.20215v1 |
span representations. In Kentaro Inui, Jing Jiang, Vincent Ng, and Xiaojun Wan, editors, Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Lan- guage Processing (EMNLP-IJCNLP) , pages 5784–5789, Hong Kong, China, November 2019. A... | https://arxiv.org/abs/2505.20215v1 |
Korhonen, David Traum, and Lluís Màrquez, editors, Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , pages 1331–1339, Florence, Italy, July 2019. Association for Computational Linguistics. 13 Table 4: Micro-averaged test F 1performance on all tasks ( ϕ=✓,etag i“✓). Best in bold. ... | https://arxiv.org/abs/2505.20215v1 |
of possible relations only slightly affects labeled performance. As far as enEWT and SciDTB are concerned, the performance of edges (UAS) and relations (LAS) is also rather similar. Since in this case the predictions involve syntactic rather than semantic relations, the similar performance hints at relations being easi... | https://arxiv.org/abs/2505.20215v1 |
tasks. Surprisingly, we notice that for CoNLL04 and ERFGC the tagging performance is also seemingly higher with score normalization. However, the overlap of the standard deviations is too high to be able to make any claims on the matter. 15 0123456789100.000.250.500.75ADE (p= 0.91) 0123456789100.10.20.3SciERC (p<0.05) ... | https://arxiv.org/abs/2505.20215v1 |
˘0.021 0.580 ˘0.031 0.291 ˘0.021 0.693 ˘0.011 200 0.662 ˘0.027 0.582 ˘0.006 0.286 ˘0.032 0.703 ˘0.007 300 0.674 ˘0.029 0.585 ˘0.026 0.299 ˘0.023 0.703 ˘0.006 400 0.663 ˘0.032 0.562 ˘0.021 0.289 ˘0.038 0.705 ˘0.011 1? d100 0.685 ˘0.020 0.600 ˘0.018 0.302 ˘0.013 0.698 ˘0.008 200 0.686 ˘0.025 0.610 ˘0.018 0.314 ˘0.019 0.7... | https://arxiv.org/abs/2505.20215v1 |
˘0.011 ■ 3 0.675 ˘0.019 0.239 ˘0.327 0.111 ˘0.152 0.703 ˘0.006 N□ 10 0.545 ˘0.017 0.415 ˘0.014 0.155 ˘0.019 0.558 ˘0.009 0.520□ 1 0.667 ˘0.014 0.543 ˘0.017 0.275 ˘0.020 0.680 ˘0.015 □ 2 0.674 ˘0.025 0.576 ˘0.023 0.272 ˘0.014 0.699 ˘0.006 □ 3 0.672 ˘0.028 0.580 ˘0.022 0.297 ˘0.019 0.705 ˘0.006 □ 1? d0 0.570 ˘0.013 0.454... | https://arxiv.org/abs/2505.20215v1 |
for the base models ( η“1ˆ10´4) and the large models ( η“3ˆ10´5) and we apply gradient norm clipping with }∇}max“1.0. In the case of the large models, we also use a cosine schedule with warm-up over 6% of the steps. We adopt these measures because during early trials we experienced sudden mid-run gradient explosions. N... | https://arxiv.org/abs/2505.20215v1 |
raw. The p-values refer to the performance being greater with score normalization (one-tailed Wilcoxon signed-rank test). full fine-tuning ablation (Appendix B.6), training and evaluation took „1 hour for each of the 40 base models and „2-3 hours for each of the 40 large models, for an additional „140 GPU hours. 20 0 2... | https://arxiv.org/abs/2505.20215v1 |
articulated. •The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for on... | https://arxiv.org/abs/2505.20215v1 |
and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general. releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to re... | https://arxiv.org/abs/2505.20215v1 |
experiments are reproducible, they should state which ones are omitted from the script and why. •At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). •Providing as much information as possible in supplemental material (appended to the paper) is recommended, but incl... | https://arxiv.org/abs/2505.20215v1 |
the paper does not include experiments. •The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. •The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compu... | https://arxiv.org/abs/2505.20215v1 |
addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). 11.Safeguards Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high r... | https://arxiv.org/abs/2505.20215v1 |
with human subjects Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? Answer: [NA] Justification: Our work does not involve any human subject... | https://arxiv.org/abs/2505.20215v1 |
arXiv:2505.20225v1 [cs.CL] 26 May 2025FLAME-MoE: A Transparent End-to-End Research Platform for Mixture-of-Experts Language Models Hao Kang2∗Zichun Yu1∗Chenyan Xiong1,3 1Language Technologies Institute 2School of Computer Science 3Foundation and Language Model Center Carnegie Mellon University {haok, zichunyu, cx}@andr... | https://arxiv.org/abs/2505.20225v1 |
per layer, top-8 gating, and two shared experts, following DeepSeek- V2 [ 20] and OLMoE [ 25]. Using empirical scaling laws, we allocate compute-optimal training budgets to ensure fair, efficient pretraining. As summarized in Table 1, FLAME-MoE is the only MoE platform offering full openness—code, data, checkpoints, ro... | https://arxiv.org/abs/2505.20225v1 |
only architectural efficiency but also the stability of sparse computation. 2 In parallel, large-scale dense model platforms have played a critical role in advancing empirical research by enhancing reproducibility and transparency. Pythia [ 3], for example, comprises a suite of 16 decoder-only models trained on a commo... | https://arxiv.org/abs/2505.20225v1 |
as shared experts, meaning they are activated for every token, providing a baseline computation path. The remaining 6 experts are routed experts, selected dynamically by the router based on the input token. 3.2 Training Loss Training MoE models effectively often requires auxiliary loss functions to encourage a balanced... | https://arxiv.org/abs/2505.20225v1 |
of active experts per token for FLAME-MoE, we concentrate on the number of active parameters as the primary variable for our scaling law study. This simplification allows us to adapt established methodologies in dense models for finding compute-optimal MoE configurations. 4 Specifically, we investigate two primary appr... | https://arxiv.org/abs/2505.20225v1 |
Lval(Nactive, D)subject to the constraint C=κN activeD, we can derive the optimal N∗ active(C)andD∗(C). Similar to the IsoFLOP approach, we can then fit power laws to these derived optimal values as a function of C. 4.3 FLAME-MoE Model Family We present our fitted scaling law in Figure 1c and Table 4, where the results... | https://arxiv.org/abs/2505.20225v1 |
a decay ratio of 0.1 relative to the total number of training steps. We use 32 NVIDIA H100 GPUs for training and store 10 checkpoints across evenly splitted training steps to study its performance trends. We provide a summary of all training configurations in Table 5. For evaluation, we adapt lm-evaluation-harness [ 8]... | https://arxiv.org/abs/2505.20225v1 |
presented in Appendix A, the overall FLOPs throughput still lags behind dense models. This discrepancy primarily arises from the inherent sparsity and communication overheads introduced by MoE architectures, which pose unique infrastructure challenges. These limitations highlight an area for improvement in open-source ... | https://arxiv.org/abs/2505.20225v1 |
To understand expert interactions under top- krouting, we analyze expert co-activation—how often expert pairs are selected together for the same token. This reveals whether experts behave indepen- dently or tend to co-operate. Following OLMoE [ 25], we define the directional co-activation score from expert Eito expert ... | https://arxiv.org/abs/2505.20225v1 |
provided in Appendix D. 7 Conclusion We present FLAME-MoE, a transparent and reproducible research platform built to advance the study of Mixture-of-Experts language models. By releasing a family of seven compute-optimal models along with complete training artifacts—including logs, checkpoints, routing traces, and eval... | https://arxiv.org/abs/2505.20225v1 |
massive multitask language understanding. In Proc. of ICLR , 2021. [11] Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, Tom Hennigan, Eric Noland, Katherine Millican, George van den Driessche, Bogd... | https://arxiv.org/abs/2505.20225v1 |
mode collapse in the fine-tuning of large language models. In ICLR 2024 Workshop on Mathematical and Empirical Understanding of Foundation Models , 2024. [27] Samyam Rajbhandari, Conglong Li, Zhewei Yao, Minjia Zhang, Reza Yazdani Aminabadi, Ammar Ahmad Awan, Jeff Rasley, and Yuxiong He. Deepspeed-moe: Advancing mixtur... | https://arxiv.org/abs/2505.20225v1 |
(a) EP=2, PP=1 (b) EP=8, PP=1 Figure 7: GPU utilization under different parallelization strategies (EP = Expert Parallel, PP = Pipeline Parallel) using Megatron-LM, collected by Wandb. Figure 7 illustrates the variation in GPU utilization under different levels of expert parallelism while keeping the pipeline paralleli... | https://arxiv.org/abs/2505.20225v1 |
shallow layers (e.g., layer 2), larger models exhibit broader and more intense expert co-activation, as seen in the FLAME-MoE-1.7B-10.3B model, which shows darker and more widespread patterns. This suggests that larger models tend to engage more experts early in processing, likely to capture more diverse or complex inp... | https://arxiv.org/abs/2505.20225v1 |
saturation trend of a specific MoE layer, measured as the average overlap in expert selection with the final checkpoint. We observe that larger models exhibit slower saturation, particularly in early layers. For example, in FLAME-MoE-1.7B-10.3B under top-1 routing, the saturation in layer 2 (blue line) starts notably l... | https://arxiv.org/abs/2505.20225v1 |
arXiv:2505.20231v1 [cs.CL] 26 May 2025Bridging the Long-Term Gap: A Memory-Active Policy for Multi-Session Task-Oriented Dialogue Yiming Du1*, Bingbing Wang2*, Yang He3, Bin Liang1, Baojun Wang4, Zhongyang Li4,Lin Gui5,Jeff Z. Pan6,Ruifeng Xu2,Kam-Fai Wong1 1The Chinese University of Hong Kong2Harbin Institute of Techn... | https://arxiv.org/abs/2505.20231v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.