text string | source string |
|---|---|
in-degree are often considered influential because they have been cited by many other works. Mathematically, for a directed graph with adjacency matrix A, where aij= 1 if paper i cites paper j(and 0 otherwise), the in-degree and out-degree are given by the following sums: kin(i) =X jaji, k out(i) =X jaij. (1) 2.Between... | https://arxiv.org/abs/2505.21162v1 |
incorporating both the quantity and quality (prestige) of citations. Papers receiving citations from highly influential sources will achieve higher PageRank scores, reflecting their increased visibility and authoritative status within the scientific community. B. Hyperparameter Configuration for GAN Training 1.SciCite ... | https://arxiv.org/abs/2505.21162v1 |
Powley, D. Radev, and Y. F. Tan. The ACL Anthology reference corpus: A reference dataset for bibliographic research in computational linguistics. In N. Calzolari, K. Choukri, B. Maegaard, J. Mariani, J. Odijk, S. Piperidis, and D. Tapias, editors, Proceedings of the Sixth International Conference on Language Resources ... | https://arxiv.org/abs/2505.21162v1 |
Thailand, Nov. 2011. Asian Federation of Natural Language Processing. [18] L. C. Freeman. Centrality in social networks conceptual clarification. Social Networks , 1(3):215–239, 1978. [19] E. Garfield et al. Can citation indexing be automated. In Statistical association methods for mechanized documentation, symposium p... | https://arxiv.org/abs/2505.21162v1 |
page 213–222, New York, NY, USA, 2008. Association for Computing Machinery. [40] T. Saier and M. F¨ arber. unarXive: A Large Scholarly Data Set with Publications’ Full-Text, Annotated In-Text Citations, and Links to Metadata. Scientometrics , 125(3):3085–3108, Dec. 2020. [41] A. Scharnhorst, K. Borner, and P. Van den B... | https://arxiv.org/abs/2505.21162v1 |
arXiv:2505.21172v1 [cs.CL] 27 May 2025TAT-R1: Terminology-Aware Translation with Reinforcement Learning and Word Alignment Zheng Li, Mao Zheng, Mingyang Song, Wenjie Yang Tencent Hunyuan jasonzli@tencent.com Abstract Recently, deep reasoning large language mod- els(LLMs) like DeepSeek-R1 have made sig- nificant progres... | https://arxiv.org/abs/2505.21172v1 |
best of our knowledge, no existing research has explored the integration of reinforce- ment learning and deep reasoning for terminology translation tasks. In this paper, we propose TAT-R1, a terminology- aware translation model trained with reinforcement learning and word alignment. First, using word alignment techniqu... | https://arxiv.org/abs/2505.21172v1 |
rewards has been vali- dated in papers in (He et al., 2025) and (Feng et al., 2025). In this work, we incorporate COMET-22 as one component of our reward functions. To main- tain training stability, we adopt a similar approach to that used in (He et al., 2025), specifically: Rcomet =round (comet, 2) (2) Word Alignment ... | https://arxiv.org/abs/2505.21172v1 |
aligned, which can be expressed in formulas re- spectively as Aref ij= (si/i, r j/j)andApre ik= (si/i, p k/k). Next, we perform Named Entity Recognition (NER) on the source tokens, retaining nouns as key elements requiring alignment. This is because noun translations typically exhibit less variability compared to sente... | https://arxiv.org/abs/2505.21172v1 |
. , r G})(12) GRPO then optimizes the policy parameters θ by maximizing the following objective: JGRPO (θ) =Eq∼P(Q),{oi}G i=1∼πθold(O|q) " 1 GGX i=1minπθ(oi|q) πθold(oi|q)Ai, clipπθ(oi|q) πθold(oi|q),1−ε,1 +ε Ai −β D KL πθ πref# , (13) where: •εcontrols the clipping range for policy up- dates, ensuring stable tra... | https://arxiv.org/abs/2505.21172v1 |
experiments are trained for three epochs. 3.2 Results and Analysis This section presents the main experimental results, demonstrating that our proposed word-alignment reward is highly effective. We then provide a de- tailed analysis of the experimental outcomes and supplement the findings with relevant ablation stud- i... | https://arxiv.org/abs/2505.21172v1 |
between SFT and RL. To demonstrate the effectiveness of RL, we fine-tune the model using the same training data with SFT. As shown in Figure 2, although the fine-tuned model showed a slight improvement over the baseline in Chinese-to- English (Zh->En) translation on WMT, there was a noticeable decline in English-to-Chi... | https://arxiv.org/abs/2505.21172v1 |
apparent degradation in translation fluency. In contrast, the word alignment rewards focus solely on the correctness of keyword transla- tions, demonstrating positive effects on lexical and semantic translation quality. 4 Related Work 4.1 Reason-based LLMs In recent years, reason-based large language mod- els, such as ... | https://arxiv.org/abs/2505.21172v1 |
first terminology-aware translation model trained with RL and word alignment. Empowered by word alignment in machine translation, we design three types of new rule-based rewards. Combining the word alignment rewards with format reward and comet reward, we train our model with GRPO. Ex- perimental results demonstrate th... | https://arxiv.org/abs/2505.21172v1 |
Feng, Shaosheng Cao, Jiahan Ren, Jiayuan Su, Ruizhe Chen, Yan Zhang, Zhe Xu, Yao Hu, Jian Wu, and Zuozhu Liu. 2025. Mt-r1-zero: Advanc- ing llm-based machine translation via r1-zero-like reinforcement learning. Preprint , arXiv:2504.10160. Nuno Miguel Guerreiro, Ricardo Rei, Daan van Stigt, Luísa Coheur, Pierre Colombo... | https://arxiv.org/abs/2505.21172v1 |
and Joohyung Han. 2024. Efficient technical term translation: A knowledge distillation approach for parenthetical terminology translation. In Proceed- ings of the Ninth Conference on Machine Translation , pages 1410–1427, Miami, Florida, USA. Association for Computational Linguistics. Kishore Papineni, Salim Roukos, To... | https://arxiv.org/abs/2505.21172v1 |
Bo Zheng, Bowen Yu, Chengyuan Li, Dayi- heng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Ji- axi Yang, Jingren Zhou, Junyang Lin, Kai Dang, and 22 others. 2024. Qwen2.5 technical report. CoRR , abs/2412.15115. Weihao Zeng, Yuzhen Huang, Wei Liu, Keqing He, Qian Liu, Zejun ... | https://arxiv.org/abs/2505.21172v1 |
arXiv:2505.21178v1 [cs.CL] 27 May 2025Walk Before You Run! Concise LLM Reasoning via Reinforcement Learning Mingyang Song, Mao Zheng Tencent Hunyuan nickmysong@tencent.com Abstract As test-time scaling becomes a pivotal research frontier in Large Language Models (LLMs) development, contemporary and advanced post-traini... | https://arxiv.org/abs/2505.21178v1 |
Such information and behaviors not only inefficiently consume computational resources but also compromise reasoning accuracy by causing models to deviate from valid logical pathways to incorrect conclusions. To address these issues, recent studies [ 8,11,12,13,14] are researching efficient reasoning method- ologies bas... | https://arxiv.org/abs/2505.21178v1 |
Policy Optimization (GRPO) [ 6] introduces a policy gradient framework that elimi- nates the reliance on explicit value function by utilizing comparative advantage estimation within a group of responses. This method samples multiple candidate outputs for each input question and computing advantages based on the relativ... | https://arxiv.org/abs/2505.21178v1 |
introduce an entropy bonus to encourage greater exploration capability in the model, named GRPO++. Similar to the original approach, GRPO++ estimates the advantage in a group-relative manner and optimizes the policy model using the following objective: JGRPO++ (θ) =Eq∼D,{oi}G i=1∼πθold(·|q) 1 GGX i=1n minh τi(θ)ˆAi,cli... | https://arxiv.org/abs/2505.21178v1 |
reshape the reward function in GRPO. When the model’s rollout results for a question are all correct, we further optimize the model’s reasoning length for that question by using the remaining maximum response length as a reward (under the specified context length, the more remaining context length, the higher the rewar... | https://arxiv.org/abs/2505.21178v1 |
17K problem-answer pairs, each paired with an integer as the answer. DAPO-Math-17K was compiled from the Art of Problem Solving (AoPS4) website and official competition websites using a combination of web scraping and manual annotation. MATH5(Level 3-5), which contains approximately 8K problem-answer pairs. Each proble... | https://arxiv.org/abs/2505.21178v1 |
= \frac{2y}{5}.\]Cross-multiplying gives:\[5 \log(y) = 2y (2\log(5) - \log(y)) \implies 5 \log(y) = 4y \log(5) - 2y \log(y).\]Rearranging terms:\[5 \log(y) + 2y \log(y) = 4y \log(5) \implies \log(y)(5 + 2y) = 4y \log(5).\]This equation must hold true for some \(y > 1\). Similarly, we check the second equation \(\log_y(... | https://arxiv.org/abs/2505.21178v1 |
widely recognized reasoning benchmarks. Specifically, ConciseR achieves an average accuracy improvement of 55.2% compared to the base model, Qwen2.5-Math-7B. Meanwhile, our method, GRPO++, also consistently surpasses all baselines, showing superior overall performance averaged across the five benchmarks. Figure 3 illus... | https://arxiv.org/abs/2505.21178v1 |
as illustrated by the following two equations, ˆri=ri+λˆLi,ˆLi=(Max({L1,L2,...,LG})−Li Max({L1,L2,...,LG})−Min({L1,L2,...,LG}),ifPG i=1ri=G 0, ifPG i=1ri̸=G, (10) 8 102030405060708090100110 Steps0200400600800 check 102030405060708090100110 Steps0123 rethink 102030405060708090100110 Steps024 reassess 1020304050607080901... | https://arxiv.org/abs/2505.21178v1 |
RL 9 post-training via GRPO, without the need for supervised warm-up, can directly induce robust reason- ing abilities. Remarkably, this kind of method not only achieves performance competitive with o1 but also exhibits emergent behaviors such as self-verification and multi-step planning. This paradigm shift significan... | https://arxiv.org/abs/2505.21178v1 |
llms. arXiv preprint arXiv:2501.12599 , 2025. [10] Alejandro Cuadron, Dacheng Li, Wenjie Ma, Xingyao Wang, Yichuan Wang, Siyuan Zhuang, Shu Liu, Luis Gaspar Schroeder, Tian Xia, Huanzhi Mao, Nicholas Thumiger, Aditya Desai, Ion Stoica, Ana Klimovic, Graham Neubig, and Joseph E. Gonzalez. The danger of overthinking: Exa... | https://arxiv.org/abs/2505.21178v1 |
Wan, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zihan Qiu. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115 , 2024. [22] An Yang, Beichen Zhang, Binyuan Hui, Bofei Gao, Bowen Yu, Chengpeng Li, Dayiheng Liu, Jianhong Tu, Jingren Zhou, Junyang Lin, Keming Lu, Mingfeng Xue, Runji Lin, Tianyu Liu, Xingzhang Ren, and... | https://arxiv.org/abs/2505.21178v1 |
arXiv:2505.21224v1 [cs.CL] 27 May 2025A Representation Level Analysis of NMT Model Robustness to Grammatical Errors Abderrahmane Issam Yusuf Can Semerci Jan Scholtes Gerasimos Spanakis Department of Advanced Computing Sciences Maastricht University {abderrahmane.issam, y.semerci, j.scholtes, jerry.spanakis}@maastrichtu... | https://arxiv.org/abs/2505.21224v1 |
that the cor- rection is happening. Our work addresses the following research ques- tions and makes the following contributions: RQ1: How do models represent and handle gram- matical errors to achieve robustness? NMT en- coders inherently implement a Grammatical Error Correction (GEC) setup in which they first detect t... | https://arxiv.org/abs/2505.21224v1 |
et al., 2021; Passban et al., 2021; Yang et al., 2022; Wang et al., 2023) and our experiments show that NMT encoders do this inherently when trained on synthetic noise. While analyzing the robustness of NMT transformer mod- els has been an area of exploration (Napoles et al., 2016; Khayrallah and Koehn, 2018; Belinkov ... | https://arxiv.org/abs/2505.21224v1 |
models, namely: OPUS-MT, M2M100, MBART and NLLB. We fine-tune these models on the ad- versarial dataset of one of the error types, and since this leads to improving their robustness to the error type, we also analyze the representations of fine- tuned models. To separate the effect of robustness from domain adaptation,... | https://arxiv.org/abs/2505.21224v1 |
ungrammatical word’s representation, we are led to introduce what we term Robustness Heads , which we define as heads that influence the ungrammat- ical word’s representation toward its grammatical form. This requires a simple redefinition of In- fluential Heads , where instead of computing the distance to the original... | https://arxiv.org/abs/2505.21224v1 |
on each error type. 4.2 Synthetic Errors Nounnum : We find nouns in the sentence then sample one of them and change its number from plural to singular or the opposite depending on its actual number. For English, we use Berkeley parser (Petrov et al., 2006) to identify nouns and their number, and for French we use Spacy... | https://arxiv.org/abs/2505.21224v1 |
noisy. As metrics, we use COMET (Rei et al., 2020), BLEU (Papineni et al., 2002) and ChrF (Popovi ´c, 2015). COMET is a neural based metric that was shown to be more aligned with human judgments (Freitag et al., 2022). We use the reference-based model wmt22-comet-da7, and we report BLEU and ChrF results in our reposito... | https://arxiv.org/abs/2505.21224v1 |
1.0 m2m100-noise 77.57 77.54 0.03 77.58 77.5 0.08 77.51 77.23 0.28 mbart-base 78.04 77.23 0.81 78.04 77.25 0.79 78.04 77.24 0.79 mbart-clean 78.51 77.77 0.74 78.64 77.74 0.9 78.49 77.73 0.76 mbart-noise 78.61 78.58 0.04 78.57 78.51 0.06 78.58 78.38 0.2 nllb-base 78.34 77.6 0.74 78.34 77.69 0.66 78.34 77.52 0.83 nllb-cl... | https://arxiv.org/abs/2505.21224v1 |
to explain this behavior in Noise-Finetuned Models. Figure 2: CKA distance of clean and noise word repre- sentations across models and errors on En-Es. Noise- Finetuned models drive the representation of the noisy word to be more similar to the clean word. 5.4 Robustness Heads 5.4.1 Attention to POS Tags Figure 3 (and ... | https://arxiv.org/abs/2505.21224v1 |
put more attention on nouns (from 0.051 and 0.065 to 0.053 and 0.069). On En-Es Article errors, NLLB and OPUS-MT have learned to put more attention on nouns and proper nouns respectively (going from 0.052 and 0.057 to 0.055 and 0.061), while on Prep errors, they have learned to put more attention on verbs (from 0.062 a... | https://arxiv.org/abs/2505.21224v1 |
the error then corrects it, and this behavior is more distinguishable in Noise-Finetuned models. Com- pared to Base or Clean-Finetuned models, Noise- Finetuned models maintain their error detection in lower layers, while they drive the representation of the ungrammatical word to be as closely similar to the grammatical... | https://arxiv.org/abs/2505.21224v1 |
understand this behavior, we propose a method for finding Robustness Heads - attention heads that attend to POS tags and help detect and correct the grammatical error. Addition- ally, we find that fine-tuning on grammatical errors leads the model to use more Robustness Heads especially in deeper layers. These findings ... | https://arxiv.org/abs/2505.21224v1 |
translation models. Computational Linguistics , 46(1):1–52. Yonatan Belinkov, Lluís Màrquez, Hassan Sajjad, Nadir Durrani, Fahim Dalvi, and James Glass. 2017. Evalu- ating layers of representation in neural machine trans- lation on part-of-speech and semantic tagging tasks. InProceedings of the Eighth International Joi... | https://arxiv.org/abs/2505.21224v1 |
automatic detection of learners’ er- rors. In Proceedings of the Fourth International Conference on Language Resources and Evaluation (LREC’04) , Lisbon, Portugal. European Language Resources Association (ELRA). Sai Muralidhar Jayanthi and Adithya Pratapa. 2021. A study of morphological robustness of neural machine tra... | https://arxiv.org/abs/2505.21224v1 |
in pre-trained transformers. In Proceedings of the Third BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP , pages 68–82, Online. Association for Computational Linguistics. Courtney Napoles, Aoife Cahill, and Nitin Madnani. 2016. The effect of multiple grammatical errors on processing non-nativ... | https://arxiv.org/abs/2505.21224v1 |
Stewart, Ana C Farinha, and Alon Lavie. 2020. COMET: A neural framework for MT evaluation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Process- ing (EMNLP) , pages 2685–2702, Online. Association for Computational Linguistics.Elizabeth Salesky, Matthias Sperber, and Alexander Waibel. 2... | https://arxiv.org/abs/2505.21224v1 |
Nadir Dur- rani, Fahim Dalvi, and James Glass. 2020. Similar- ity analysis of contextual word representation mod- els. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , pages 4638–4655, Online. Association for Computational Linguistics. Weiwen Xu, Ai Ti Aw, Yang Ding, Kui Wu, ... | https://arxiv.org/abs/2505.21224v1 |
parmi, pendant, pour, sans, sauf, selon, sous, suiv- ant, sur, vers}. B.2 Data Splits Table 2 shows the amount of data for each language direction in our experiments. Although the train set for Fr-Es is small in comparison to the other language directions, the COMET improvements af- ter fine-tuning shown in Table C.1 s... | https://arxiv.org/abs/2505.21224v1 |
75.7 73.11 2.58 opus-mt-noise 75.62 75.54 0.07 75.69 75.62 0.07 75.65 75.18 0.46 m2m100-base 73.04 72.21 0.83 73.04 72.45 0.59 73.04 70.06 2.98 m2m100-clean 74.01 73.45 0.55 73.98 73.65 0.33 74.27 72.21 2.07 m2m100-noise 74.24 74.23 0.01 74.22 74.16 0.06 74.2 73.84 0.36 mbart-base 69.13 68.27 0.87 69.13 68.82 0.31 69.1... | https://arxiv.org/abs/2505.21224v1 |
performance on the clean and noisy test sets and their difference ( ∆). (a) Fr-Es (b) En-De (c) En-It (d) En-Nl Figure 6: GED probing performance of Noise-Finetuned, Clean-Finetuned and Base models on Fr-Es, En-De, En-It and En-Nl. GED probing performance of Noise-Finetuned models witnesses a degradation in deeper laye... | https://arxiv.org/abs/2505.21224v1 |
NOUN PRON PROPN PUNCT VERB ADJ ADP ADV AUX DET NOUN PRON PROPN PUNCT VERB T ag ADJ ADP ADV AUX DET NOUN PRON PROPN PUNCT VERBAverage Attention Scores for POS T ags on En-It Article Nounnum PrepOPUS-MT M2M100 MBART NLLBFigure 10: Robustness Heads attention to the 10 most common POS tags in the test set on En-It. The sca... | https://arxiv.org/abs/2505.21224v1 |
74.4 2.85 mbart-noisy 77.38 77.28 0.1 nllb-base 72.61 69.73 2.87 nllb-clean 73.64 71.13 2.51 nllb-noisy 73.53 73.39 0.15 Table 4: COMET scores for Morpheus on En-Es, Fr-Es, En-De, En-It and En-NL. The Base model is the original model, Clean is fine-tuned on the clean version of the data, and Noise is fine-tuned on the ... | https://arxiv.org/abs/2505.21224v1 |
arXiv:2505.21239v1 [cs.CL] 27 May 2025LMCD: Language Models are Zeroshot Cognitive Diagnosis Learners Yu He1,2*, Zihan Yao2* Chentao Song2,Tianyu Qi2,Jun Liu1†,Ming Li2,Qing Huang2 1Xi’an Jiaotong University, Xi’an, Shaanxi, China 2TAL Education Group, Beijing, China heyucs@stu.xjtu.edu.cn, liukeen@mail.xjtu.edu.cn {ya... | https://arxiv.org/abs/2505.21239v1 |
emplified by works like ICDM (Liu et al., 2024a) and TechCD (Gao et al., 2023), construct relation- ships between exercises, KCs, and students to es- tablish connections between unseen and seen enti- ties through graph structures. Although innovative, these methods are limited by the accuracy of KC annotations and the ... | https://arxiv.org/abs/2505.21239v1 |
0nodes at depth 1, each of them is considered a distinct domain Di= K1 i,R1 i ,∀i∈[1, M1], where K1 i={kl ij|j∈ [1,2, ..., Ml i], l∈[1,2, ..., max _depth ]}, and R1 i={(kip, kiq)|kip, kiq∈K1 i},(kip, kiq)is a directed edge of the tree. Note that knowledge across different domains is entirely isolated, that is,K1 i∩K1 j... | https://arxiv.org/abs/2505.21239v1 |
en- hanced representations are subsequently processed by standard CDM heads to predict response proba- bilities and knowledge proficiency, as illustrated in Figure 2. 3 3.2 Knowledge Diffusion NLP-based CDMs are limited by brief exercise texts and vague KC labels, creating imprecise se- mantic relationships. Similar-lo... | https://arxiv.org/abs/2505.21239v1 |
feed Efusion into the LLM backbone for forward propagation to ob- tain the final representation Ofusion. h0=Efusion hl= FFN l(Attn l(hl−1)), l= 1,2, . . . , N Ofusion =hl (5) where Attn andFFN represent the attention layer and feed-forward network structure in the LLM based on the Transformer architecture (Vaswani et a... | https://arxiv.org/abs/2505.21239v1 |
•Random: Correspondingly, the lower bound of prediction skill is measured by randomly sampling from Uniform (0,1)as the student’s correct response probability. •TechCD (Gao et al., 2023): TechCD utilizes graphical relationships between KCs to estab- lish connections between students’ practiced and unseen exercises, gen... | https://arxiv.org/abs/2505.21239v1 |
problem representations. By contrast, TechCD failed completely in cross- domain scenarios mainly due to the fact that the cold-start data is isolated from the training set bothat the KC and exercise level, and thus no transfer- able knowledge among domains can be obtained through the graph structure. In summary, exper-... | https://arxiv.org/abs/2505.21239v1 |
0.4880 0.6505 0.7091 0.4857 0.6271 0.6726 0.4888 Our Method 0.6336 0.6837 0.4726 0.6518 0.7049 0.4658 0.6363 0.6843 0.4758 MIRTOracle 0.7223 0.7942 0.4272 0.7116 0.7694 0.4377 0.7180 0.7860 0.4324 Random 0.4354 0.4938 0.5328 0.5064 0.4951 0.5154 0.4770 0.5010 0.5224 Bert 0.6339 0.6712 0.4814 0.6318 0.6818 0.4761 0.6212... | https://arxiv.org/abs/2505.21239v1 |
lim- its such as oversimplifying exercise difficulty and misreading unclear texts, leading to inaccuracies in cross-domain settings. LLMs in Cognitive Diagnosis . LLMs have rev- olutionized natural language processing with ad- vanced comprehension (Brown et al., 2020) and reasoning capabilities (Kojima et al., 2022; Ma... | https://arxiv.org/abs/2505.21239v1 |
Chen, and Fei Wu. 2025. Knowledge is power: Harnessing large language models for enhanced cognitive diagnosis. In Pro- ceedings of the AAAI Conference on Artificial Intelli- gence .Weibo Gao, Qi Liu, Hao Wang, Linan Yue, Haoyang Bi, Yin Gu, Fangzhou Yao, Zheng Zhang, Xin Li, and Yuanjing He. 2024. Zero-1-to-3: Domain-l... | https://arxiv.org/abs/2505.21239v1 |
anomaly detection for enhancing cognitive diagnosis. In Pro- ceedings of the AAAI Conference on Artificial Intelli- gence , volume 39, pages 12337–12345. Jie Ma, Zhitao Gao, Qi Chai, Wangchun Sun, Pinghui Wang, Hongbin Pei, Jing Tao, Lingyun Song, Jun Liu, Chen Zhang, and 1 others. 2025b. Debate on graph: a flexible an... | https://arxiv.org/abs/2505.21239v1 |
: A sub-dataset of the NeurIPS Edu- cation Challenge, containing Tasks 3 and 4, is a powerful resource tailored to the advancement of educational data analytics and machine learn- ing applications within the education field. The dataset comprises crowdsourced diagnostic math- ematics exercises collected from the Eedi e... | https://arxiv.org/abs/2505.21239v1 |
We employ the default settings with- out any modifications to ensure the consistency of the experiment results. TechCD : The model architecture remains con- sistent with the original setup, with the addition of optional IRT and MIRT predict heads. Furthermore, undirected edges characterizing the similarity be- tween fi... | https://arxiv.org/abs/2505.21239v1 |
an ex- cessive focus on the content of example exercises leading to redundant information. In summary,the inclusion of distractors not only enriches the prompt but also enhances the robustness of the generated results by mitigating sensitivity to the selection of specific example exercises. KC Descriptions: Case 2 Targ... | https://arxiv.org/abs/2505.21239v1 |
arXiv:2505.21242v1 [cs.CL] 27 May 2025Evaluation of LLMs in Medical Text Summarization: The Role of Vocabulary Adaptation in High OOV Settings Gunjan Balde*, Soumyadeep Roy, Mainack Mondal and Niloy Ganguly Indian Institute of Technology Kharagpur balde.gunjan0812@kgpian.iitkgp.ac.in soumyadeep.roy9@iitkgp.ac.in {maina... | https://arxiv.org/abs/2505.21242v1 |
(Brown et al., 2020; Lampinen et al., 2022) where a fixed num- ber of exemplars is added in the prompt at infer- ence time only; and (ii) parameter-efficient fine- tuning using QLoRA (Dettmers et al., 2024) be- cause LLMs have billions of parameters which make complete finetuning computationally infea- sible. However, ... | https://arxiv.org/abs/2505.21242v1 |
1.28 1.46 Split>3 17% 28% Table 1: OOV concentration and fragment score ob- served for general domain dataset (CNN-DailyMail) and medical domain dataset (PAC) obtained using to- kenizers of models: Llama-2 and Mistral (V ocabulary size: 32K), and Llama-3.1 and Qwen-2 (V ocabulary size: 128K and 151K). Fragment score (R... | https://arxiv.org/abs/2505.21242v1 |
the target downstream task dataset for which the vo- cabulary is to be constructed. We now describe the different vocabulary adaptation methods used for our benchmarking study. MEDVOC. MEDVOC (Balde et al., 2024b) is a SoTA vocabulary adaptation strategy for adapt- ing PLMs like BART and PEGASUS, on medical summarizati... | https://arxiv.org/abs/2505.21242v1 |
fragment score to obtain the optimal vocabulary to be added. To offset the absence of such derivative tokens, we use AdaptBPE tokeniza- tion scheme (Balde et al., 2024a) instead of the standard Llama tokenizers. Instead of directly uti- lizing merge rules, AdaptBPE first checks whether a part of the input token (using ... | https://arxiv.org/abs/2505.21242v1 |
computed using PubMedBERT2with the given test data point. The template for prompting is shown in Appendix A (Table 9). We also describe the fine-grained evaluation setup used for this bench- marking study in Table 4. Next, we describe the benchmark medical text summarization datasets used, followed by details 2https://... | https://arxiv.org/abs/2505.21242v1 |
2022; Mangrulkar et al., 2022) to carry out the pretraining in this resource- constrained setting. The LoRA adapters are applied to all the linear modules in the model. These include {k_proj, q _proj, v _proj, ando_proj} modules from self attention layers along with {gate_proj, up _proj, anddown _proj} mod- ules from M... | https://arxiv.org/abs/2505.21242v1 |
leads to per- formance improvement in terms of Rouge-L for Llama-2 and Llama-3.1 in seven out of eight set- tings (except for Llama-2 on EBM dataset). In terms of Concept-Score which is a proxy measure for faithfulness (Zhang et al., 2023), Figure 2 shows that at least one vocabulary adaptation performs the best in fiv... | https://arxiv.org/abs/2505.21242v1 |
24.70 17.14 29.41 26.82 MEDVOC 32.26 28.17 18.18 27.40 26.50 37.01 29.79 18.18 32.43 29.35 MEDVOC-LLM 32.40 29.41 20.29 30.33 28.11 37.15 32.89 20.29 37.84 32.04 ScafFix 32.88 29.29 22.22 36.44 30.21 36.70 32.26 20.90 34.15 31.00 BioASQ- M BASE 28.50 27.27 21.53 28.00 26.33 29.28 26.67 21.51 28.57 26.51 CPT-only 27.22 ... | https://arxiv.org/abs/2505.21242v1 |
drop in the number of under-trained tokens in the LLM vocabulary. This makes the training phase less noisy and yields superior perfor- 7 Average Concept Score0510152025 Llama-2-7B Llama-3.1-8BBASE CPT-only MEDVOC MEDVOC-LLM ScafFix(a) Average Concept Score051015 Llama-2-7B Llama-3.1-8BBASE CPT-only MEDVOC MEDVOC-LLM Sc... | https://arxiv.org/abs/2505.21242v1 |
was compensated at a rate of 8UK pounds per hour (see Appendix C for more details), and each summary par was evaluated by three annotators. Figure 3 shows the human eval- uation results where the ScafFix method generates more faithful summaries ( 93.34% versus 83.34% of summaries are faithful), and more relevant sum- m... | https://arxiv.org/abs/2505.21242v1 |
certain challenging generation scenar- ios where reference summaries have high OOV con- centration and high novelty. We then benchmark the performance of three vocabulary adaptation strategies on four models: Llama-2 7B, Mistral 7B, Qwen-2 7B, and Llama-3.1 8B model; over three biomedical summarization datasets and two... | https://arxiv.org/abs/2505.21242v1 |
Florida, USA. Association for Computational Linguistics. Marco Cognetta, Tatsuya Hiraoka, et al. 2024. An anal- ysis of BPE vocabulary trimming in neural machine translation. In Proceedings of the Fifth Workshop on Insights from Negative Results in NLP , pages 48–50, Mexico City, Mexico. Association for Computational L... | https://arxiv.org/abs/2505.21242v1 |
Proceed- ings of the 15th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2022) - HEALTHINF , pages 180–188. IN- STICC. Haoran Lian, Yizhe Xiong, et al. 2025. Scaffold-bpe: Enhancing byte pair encoding for large language models with simple and effective scaffold token re- mova... | https://arxiv.org/abs/2505.21242v1 |
. George Tsatsaronis, Georgios Balikas, et al. 2015. An overview of the bioasq large-scale biomedical se- mantic indexing and question answering competition. BMC bioinformatics , 16:1–28. Dave Van Veen, Cara Van Uden, et al. 2024. Adapted large language models can outperform medical ex- perts in clinical text summariza... | https://arxiv.org/abs/2505.21242v1 |
examples concatenated using ‘##’– Query {i}: {Train-Query} Document {i}: {Train-Source Document} Summary {i} : {Train-Summary} ## –Test Example– Query: {Test-Query} Document: {Test-Source Document} Summary: Table 9: Prompt Template used to prompt the language models for the task of query focused summarization. Prompts ... | https://arxiv.org/abs/2505.21242v1 |
26.16 30.00 29.79 21.17 22.84 28.07 26.00 28.57 28.57 19.53 24.10 26.49 BioASQ-S BASE 32.12 35.72 26.32 20.59 33.33 27.92 32.12 35.72 26.12 20.53 33.33 27.92 CPT-only 33.09 29.35 25.53 19.64 23.65 24.12 33.30 30.33 29.22 20.00 30.30 21.92 MEDVOC 32.34 30.26 29.41 20.29 29.41 22.47 32.26 27.40 28.17 18.18 26.09 23.53 ME... | https://arxiv.org/abs/2505.21242v1 |
The task was conducted using Google Forms, with participants being shown a consent notice beforehand. The results are shown in Figure 3. Participation Criteria. The filtering criteria for participants were kept same as that of MED- 3https://www.prolific.com/ 13 End-To-End Two-Stage Test Full Difficult SD Difficult RS N... | https://arxiv.org/abs/2505.21242v1 |
clear, we provide few ex- amples of positive and negative examples to makepeople aware of what is a relevant, coherent, and factually consistent document vs what is not. The pdf is present in the github codebase. 14 End-To-End Two-Stage Test Full Difficult SD Difficult RS Novel RS AllSD AllRS Test Full Difficult SD Dif... | https://arxiv.org/abs/2505.21242v1 |
19.67 26.09 27.73 CPT-only 23.88 28.57 26.10 18.52 25.25 26.90 24.56 28.57 25.00 17.86 25.35 26.90 MEDVOC 25.00 30.20 26.67 20.00 26.67 26.85 24.49 28.57 25.00 19.05 24.70 26.32 MEDVOC-LLM 24.30 28.57 25.71 19.23 25.35 27.27 24.34 28.57 26.67 20.00 25.00 26.80 ScafFix 24.49 28.24 26.23 19.67 25.00 27.20 24.69 28.57 26.... | https://arxiv.org/abs/2505.21242v1 |
thousand five hundred and thirty-one measurements of ei- ther serum amylase and or serum pancreatic lipase were made ... One thousand eight hundred and twenty-five patients had either elevated serum amylase and or serum pancreatic lipase. The medical records coded for pancreatitis in a further 55 whose enzymes were not... | https://arxiv.org/abs/2505.21242v1 |
arXiv:2505.21250v1 [cs.CL] 27 May 2025ReSCORE: Label-free Iterative Retriever Training for Multi-hop Question Answering with Relevance-Consistency Supervision Dosung Lee1∗Wonjun Oh1∗Boyoung Kim1Minyoung Kim1 Joonsuk Park2,3,4†Paul Hongsuck Seo1† 1Dept. of CSE, Korea University, 2NA VER AI Lab,3NA VER Cloud, 4University... | https://arxiv.org/abs/2505.21250v1 |
matching, dense retrievers rely on query and document embeddings that need to be trained on the target domain (Karpukhin et al., 2020). For MHQA, however, it is cost- and labor-intensive to prepare documents labeled with their relevance to respective queries across iterations, because the queries—reformulated questions... | https://arxiv.org/abs/2505.21250v1 |
tic solution for MHQA. Iterative RAG Iterative RAG extends single- hop RAG to tasks requiring multiple reason- ing steps across documents (Xiong et al., 2021). FLARE (Jiang et al., 2023) focuses on adaptively retrieving documents when low-probability tokens are generated. To dynamically determine the need for external ... | https://arxiv.org/abs/2505.21250v1 |
. . . , d(1) k} ⊆ D .D(1)is then incorporated into a predefined prompt for the LLM. The prompt instructs the LLM to either defer answer generation to retrieve additional informa- tion, or predict an answer a(1)based on the suffi- ciency of the information in D(1), thereby terminat- ing the question-answering process, a... | https://arxiv.org/abs/2505.21250v1 |
Thought Generation⋯LLM & Normalization⋯𝑑1(2)𝑑2(2)𝑑3(2)𝑑𝑀(2)⋯Retrieval Distribution 𝑃𝑅2(𝒟2|𝑞(2)) Pseudo -GT Distribution 𝑄LM2𝒟2𝑞2 ∝𝑃LM2𝑎,𝑞𝒟2KL Divergence 𝐷KL(𝑄LM2||𝑃𝑅2) 𝑑1(2)𝑑2(2)𝑑3(2)𝑑𝑀(2)⋯ReSCORE training for iteration 2 ⋯ Figure 2: Overview of ReSCORE. At each iteration iwithin a iterative RA... | https://arxiv.org/abs/2505.21250v1 |
Story—was acquired by Disney in January 2006, and its first post-acquisition film, Cars, was released in May 2006. where Nis the number of QA pairs in the train- ing set, ηnis the number of iterations determined by the LLM for each question qn, and P(i) Ris the document distribution for retrieval at iteration i. The di... | https://arxiv.org/abs/2505.21250v1 |
total number of the GT supporting documents. Implementation Details We train the question embedder while keeping the document embedder frozen throughout the process. To compute the document distribution, we format the question, an- swer, and document into a predefined prompt, as described in Section 5. For loss calcula... | https://arxiv.org/abs/2505.21250v1 |
method, ReSCORE, with other existing iterative MHQA methods, including Self-RAG (Asai et al., 2023), FLARE (Jiang et al., 2023), and Adaptive- Note (Wang et al., 2024). These frameworks are re- implemented using Llama and Contriever to avoid costs for API calls. Tab. 2 presents the MHQA performance in terms of EM and F... | https://arxiv.org/abs/2505.21250v1 |
50.0 59.7 51.2 81.2 88.0 Table 2: Effects of ReSCORE with various iterative RAG systems on three MHQA benchmarks. All meth- ods are re-implemented using Llama 3.1 and Contriever, except for Self-RAG, which uses Llama-2-7B model from the original study. All hyperparameters for the baselines are taken from the original p... | https://arxiv.org/abs/2505.21250v1 |
While GT labels enhance initial retrieval results, 1 2 n Iteration510152025All Rel Docs Found (%) FT w/ GT FT w/ Pseudo GT(a) MuSiQue Dataset 1 2 n Iteration2025303540All Rel Docs Found (%) FT w/ GT FT w/ Pseudo GT (b) HotpotQA Dataset 1 2 n Iteration1525354555All Rel Docs Found (%) FT w/ GT FT w/ Pseudo GT (c) 2WikiMH... | https://arxiv.org/abs/2505.21250v1 |
lation, serving as a lower bound. Another method, LLM-rewrite, prompts an LLM to rewrite the query q(i)into a refined query q(i+1), focusing on unre- solved aspects based on the current retrieved docu- ments D(i). Finally, Thought-concat appends the current thought t(i)to the query, constructing the updated query as q(... | https://arxiv.org/abs/2505.21250v1 |
comparisons. We recognize the broader implications of multi-hop question-answering advancements and are commit- ted to responsible development and application. Acknowledgements This research was supported by IITP grants (IITP-2025-RS-2020-II201819, IITP-2025-RS- 2024-00436857, IITP-2025-RS-2024-00398115, IITP-2025-RS-2... | https://arxiv.org/abs/2505.21250v1 |
comprehension dataset. choice , 2640:660.Ronak Pradeep, Sahel Sharifymoghaddam, and Jimmy Lin. 2023. Rankvicuna: Zero-shot listwise document reranking with open-source large language models. arXiv preprint arXiv:2309.15088 . Ofir Press, Muru Zhang, Sewon Min, Ludwig Schmidt, Noah Smith, and Mike Lewis. 2023. Measuring ... | https://arxiv.org/abs/2505.21250v1 |
Synergizing reasoning and acting in language models. In International Conference on Learning Representations (ICLR) . Xin Zhang, Zehan Li, Yanzhao Zhang, Dingkun Long, Pengjun Xie, Meishan Zhang, and Min Zhang. 2023. Language models are universal embedders. arXiv preprint arXiv:2310.08232 . A Hyperparameters Hyperparam... | https://arxiv.org/abs/2505.21250v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.