text string | source string |
|---|---|
cab: Lightweight language model calibration over short- and long-form responses. In The Twelfth Inter- national Conference on Learning Representations . Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettle- moyer. 2022. Rethinking the role of demonstrations: What makes in-... | https://arxiv.org/abs/2505.21003v1 |
Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023. Llama: Open and efficient foundation language models. Matias Valdenegro-Toro and Daniel Saromo M... | https://arxiv.org/abs/2505.21003v1 |
language, which results in an effectively infinite generation space (Lin et al., 2024; Ling et al., 2024; Liu et al., 2024). In contrast, classification tasks provide LLMs with a finite set of discrete candidates, where the model’s task is limited to selecting the most probable answer from a predefined set (Wiegreffe e... | https://arxiv.org/abs/2505.21003v1 |
autoregressive (linear) manner, actual human reasoning follows a more complex topological structure. Inspired by Tree-based CoT (Yao et al., 2023) and Graph-based CoT (Besta et al., 2024), we propose modeling CoT uncertainty using graph or tree structures. In this framework: each reasoning step is represented as a node... | https://arxiv.org/abs/2505.21003v1 |
0.918 0.912 Logical Deduction 7 0.923 0.939 0.928 0.919 0.925 0.925 0.936 0.946 Table 9: AUROC of Mistral-7B-v0.2 model. High AUROC indicates the good quality of UQ measures. DatasetQwen1.5-7B 1-shot 2-shot 4-shot 8-shot 16-shot 32-shot 64-shot 128-shot Easy Mode AGNews 0.634 0.716 0.743 0.744 0.688 0.739 0.731 0.741 S... | https://arxiv.org/abs/2505.21003v1 |
and the second line for each dataset giving the ratio of increased uncertainty questions. ∆Acc represents the performance changes caused by the corresponding part of examples. F Interprebility for k-shot ICL F.1 Case Study 01234567891011121314151617181920212223242526272829303132 projection layers0.00.20.40.60.81.0 Q63 ... | https://arxiv.org/abs/2505.21003v1 |
template with a test input for AG News dataset. Classify the following sentence into two categories: [0: negative, 1: positive] Provide answer in a structured format WITHOUT additional comments, I just want the numerical label for each sentence. Sentence: that loves its characters and communicates something rather beau... | https://arxiv.org/abs/2505.21003v1 |
the option letter for each answer. Figure 18: Prompt template with a test input for logical deduction three objects dataset. Select the correct answer for the following logical deduction problem from five choices. Provide answer in a structured format WITHOUT additional comments, I just want the option letter for each ... | https://arxiv.org/abs/2505.21003v1 |
arXiv:2505.21011v1 [cs.CL] 27 May 2025LLMs are Frequency Pattern Learners in Natural Language Inference Liang Cheng†Zhaowei Wang‡Mark Steedman† †University of Edinburgh‡HKUST L.Cheng@ed.ac.uk zwanggy@cse.ust.hk m.steedman@ed.ac.uk Abstract While fine-tuning LLMs on NLI corpora im- proves their inferential performance, ... | https://arxiv.org/abs/2505.21011v1 |
bias. Evalua- tion results show that models fine-tuned on infer- ence data perform poorly on frequency-adversarial inference, proving that training on NLI datasets leads models to learn the frequency bias from NLI datasets. Finally , we present experiments demon- strating a correlation between frequency bias and textua... | https://arxiv.org/abs/2505.21011v1 |
questions to determine whether the premise entails the hypothesis. A positive label corresponds toEntail , and a negative label to No-Entail . De- tails of the datasets and prompt configurations are provided in Appendix A and B. 3.2 Fine-tune LLMs We fine-tune several widely used LLMs on NLI datasets, including DeepSee... | https://arxiv.org/abs/2505.21011v1 |
predictions are clas- sified as either correct (✓) orincorrect (✗), and we analyze frequency bias across these categories. We reports the frequency biases in Table 2. The results reveal a clear trend: both standard and fine- tuned LLMs can make correct predictions when test samples are consistent with the frequency bia... | https://arxiv.org/abs/2505.21011v1 |
to high-frequency statements while diminishing the ability to reason in the opposite direction. 4.4 Finding 4: Frequency bias is a proxy for gradient of semantic generalization To further investigate the relationship between pred- icate frequency and entailment relation, we analyze the frequency of hyponym–hypernym pai... | https://arxiv.org/abs/2505.21011v1 |
of these trends. References Jonathan Berant, Ido Dagan, and Jacob Goldberger. 2010. Global Learning of Focused Entailment Graphs. In Proceedings of the 48th Annual Meet- ing of the Association for Computational Linguistics , pages 1220–1229. Jonathan Berant, Ido Dagan, and Jacob Goldberger. 2011. Global Learning of Typ... | https://arxiv.org/abs/2505.21011v1 |
Zhicheng Ma, Zhigang Yan, Zhiyu Wu, Zihui Gu, Zi- jia Zhu, Zijun Liu, Zilin Li, Ziwei Xie, Ziyang Song, Zizheng Pan, Zhen Huang, Zhipeng Xu, Zhongyu Zhang, and Zhen Zhang. 2025. Deepseek-r1: Incen- tivizing reasoning capability in llms via reinforce- ment learning. Ning Ding, Yujia Qin, Guang Yang, Fuchao Wei, Zonghan ... | https://arxiv.org/abs/2505.21011v1 |
A lexical database for English. In Human Language Technology: Pro- ceedings of a Workshop held at Plainsboro, New Jersey, March 8-11, 1994 . Martin Schmitt and Hinrich Schütze. 2021. Language Models for Lexical Inference in Context. In Proceed- ings of the 16th Conference of the European Chap- ter of the Association fo... | https://arxiv.org/abs/2505.21011v1 |
[HYPOTHESIS ]. To make LLMs better understanding the task, we format it as Boolean questions and include indica- tor words such as “Question:" and “Answer:". For each option, we automatically provide explanations for every answer by adding affirmation or negation to the propositions. As a result, the NLI training data ... | https://arxiv.org/abs/2505.21011v1 |
training steps. Few-shot Examples Instantiated Prompt for Inference Task If Google bought Youtube, then Google owns Youtube. Is that true or false? A) True B) False Answer: A) True. Owning is a consequence of buying. If Google owns Youtube, then Google bought Youtube. Is that true or false? A) True B) False Answer: B) ... | https://arxiv.org/abs/2505.21011v1 |
arXiv:2505.21033v1 [cs.CL] 27 May 2025Def-DTS: Deductive Reasoning for Open-domain Dialogue Topic Segmentation Seungmin Lee1, Yongsang Yoo1,2, Minhwa Jung1,3, Min Song1,4* 1Yonsei University,2LOTTE INNOV ATE,3LG Eletronics,4Onoma AI 1{elplaguister, 4n3mone, minalang, min.song}@yonsei.ac.kr 2yongsang.yoo@lotte.net ,3min... | https://arxiv.org/abs/2505.21033v1 |
problem-solving skills of these LLMs and the challenges posed by DTS, reasoning strategies are rarely applied in the DTS area. This is because DTS has been treated largely as a lightweight sub- task in NLP. Nonetheless, with the rise of AI-driven chat services, the demand for more advanced DTS modules is growing. LLMs ... | https://arxiv.org/abs/2505.21033v1 |
2 reasoning (Kahneman, 2011), including in-context learning (Brown et al., 2020) and chain-of-thought prompting (Wei et al., 2022). These techniques enable LLMs to tackle complex tasks, such as sym- bolic mathematics (Yang et al., 2024), retrieval- augmented generation (Lewis et al., 2020), and data generation (Adler e... | https://arxiv.org/abs/2505.21033v1 |
in Appendix A.1. 3.3 Bidirectional Context Extraction In the first stage of Def-DTS, we instruct the LLM to summarize both the preceding and subsequent dialogues for each utterance. Considering bidirec- tional context is commonly used in many methods such as the BERT architecture (Devlin et al., 2019) and frequently em... | https://arxiv.org/abs/2505.21033v1 |
terns and descriptions are in Table 1. Inspired by this research, we instruct model to detect topic change through utterance intent classi- fication. Specifically, after the bidirectional context extraction, as shown in Figure 2c(ii), model clas- sifies the utterance into an intent of the predefined general intent pool... | https://arxiv.org/abs/2505.21033v1 |
kerror (Beeferman et al., 1997), WindowDiff (WD) error (Pevzner and Hearst, 2002), and the f1 score. The P kerror is calculated by counting the existence of a misallocated segment with a sliding window of predictions. The WD error is calculated by comparing the number of boundaries within the sliding window of gold lab... | https://arxiv.org/abs/2505.21033v1 |
uttr, the recent LLM-based method, our method achieved reduc- tions of more than 0.2 in both P kand WD er- rors, along with an impressive increase of more than 0.4 in the F1 score. Furthermore, Def-DTS outperformed even the supervised approaches in TIAGE, surpassing them in all metrics by over 10%, thus highlighting th... | https://arxiv.org/abs/2505.21033v1 |
structured format. To examine the impact of structured I/O formats not covered in the ablation study, we represent iden- tical prompts in three different formats—Natural Language (NL), JSON, and XML—and compare their performance in Table 5. The results show that the structured formats, XML and JSON, not only offer pars... | https://arxiv.org/abs/2505.21033v1 |
when handling utterances with topic shifts. 5.5 Linguistic Test for Intent Labels To demonstrate the impact of intent labels on topic shifts, we adopted methods from statistical lin- guistics. Traditional text segmentation uses pauses, cue words, and referential noun phrases to identify boundaries (Passonneau and Litma... | https://arxiv.org/abs/2505.21033v1 |
demonstrate the efficacy of LLM-reasoning as a promising ap- proach to DTS. It not only highlights the potential of our method, but it also statistically delineates 8 ModelTIAGE SuperDialseg Dialseg711 Pk↓WD↓F1↑ Pk↓WD↓F1↑ Pk↓WD↓F1↑ Plain Text + Llama 0.472 0.515 0.215 0.492 0.495 0.026 0.350 0.373 0.032 Plain Text + Qw... | https://arxiv.org/abs/2505.21033v1 |
Yang Liu. 2016. Extractive summarization of multi-party meetings through discourse segmentation. Natural Language Engineering , 22(1):41–72. Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel H... | https://arxiv.org/abs/2505.21033v1 |
, pages 562–569, Sapporo, Japan. Association for Computational Linguistics. Haoyu Gao, Rui Wang, Ting-En Lin, Yuchuan Wu, Min Yang, Fei Huang, and Yongbin Li. 2023. Unsuper- vised dialogue topic segmentation with topic-aware contrastive learning. In Proceedings of the 46th In- ternational ACM SIGIR Conference on Resear... | https://arxiv.org/abs/2505.21033v1 |
preprint arXiv:1609.01454 . Che Liu, Rui Wang, Junfeng Jiang, Yongbin Li, and Fei Huang. 2022. Dial2vec: Self-guided contrastive learning of unsupervised dialogue embeddings. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing , pages 7272–7282, Abu Dhabi, United Arab Emirates. As-... | https://arxiv.org/abs/2505.21033v1 |
Jung. 2018. Learning to rank question-answer pairs us- ing hierarchical recurrent encoder with latent topic clustering. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Tech- nologies, Volume 1 (Long Papers) , pages 1575–1584, New Orlea... | https://arxiv.org/abs/2505.21033v1 |
<=3 sentences seperately The range of the context should be previous or next 1-3 utterances except for the case of the first or last utterance. For example, given current utterance number is 2, preceding range is 0-1, subsequent range is 3-5. 2. Output the utterance_intent Use the list <valid_utterance_intent> ... </va... | https://arxiv.org/abs/2505.21033v1 |
other datasets, we ran- domly selected parts of the conversation from the Train split that adhere to the following rules: •Select 2–3 consecutive utterances for each ex- ample. •Ensure that the final utterance in the example corresponds to the target utterance intent. •Extract all the examples from a single dia- logue.... | https://arxiv.org/abs/2505.21033v1 |
instruction following was not well ex- ecuted for the case of DIFFERENT_QUESTION. For the case of RELEV ANT_QUESTION, the following instruction was well executed with one exception, but its accuracy was relatively low. The difference in explanations between the case of RELEV ANT_QUESTION and DIFFER-ENT_QUESTION could b... | https://arxiv.org/abs/2505.21033v1 |
employed Llama-3.1-8B-Instruct, Qwen2.5-7B-Instruct and Qwen2.5-32B-Instruct for Llama 8B, Qwen 7b and Qwen 32B, respectively. P k, WD, F1 were calculated only for correctly formatted outputs. ModelsTIAGE Pk↓ WD↓ F1↑ R1 0.286 0.331 0.644 V3 0.259 0.204 0.674 GPT-4o 0.232 0.256 0.699 Table 14: Performance for additional... | https://arxiv.org/abs/2505.21033v1 |
arXiv:2505.21043v1 [cs.CL] 27 May 2025Visual Cues Enhance Predictive Turn-Taking for Two-Party Human Interaction Sam O’Connor Russell andNaomi Harte ADAPT Centre, School of Engineering, Trinity College Dublin, Ireland {russelsa,nharte} @tcd.ie Abstract Turn-taking is richly multimodal. Predictive turn-taking models (PT... | https://arxiv.org/abs/2505.21043v1 |
video-only conditions. In a recent study Nota et al. (2023) found that listeners were faster at recognis- ing questions containing eyebrow frowns. Such studies underline the importance of visual cues in turn-taking and multimodal interaction more gen- erally (Holler and Levinson, 2019). Aims Despite the essential role ... | https://arxiv.org/abs/2505.21043v1 |
We make our code publicly available for future research1. 2 Background Turn-taking In a conversation, the current speaker either holds the turn or shifts to another interlocutor (Sacks et al., 1974). The time between turns is the floor-transfer offset (FTO), which is pos- itive for a gap and negative for an overlap (He... | https://arxiv.org/abs/2505.21043v1 |
be known for the model to run, it is therefore not a PTTM, although the findings are promising. Onishi et al. (Onishi et al., 2024) proposed to extend a recent state-of-the-art audio-only turn- taking model, the V AP model (Ekstedt and Skantze, 2022a), to include visual features. Like Roddy et al., they found that the ... | https://arxiv.org/abs/2505.21043v1 |
reflects real-world conditions. Both corpora contain stereo audio (one channel per speaker), which we downsample to 16 kHz. Candor has 320x240 resolution, 30 fps mp4 video. Identifying turn-taking events We extract shifts andholds from the transcriptions by iden- tifying silences greater than +250 ms where only one spe... | https://arxiv.org/abs/2505.21043v1 |
the differing setup of participant devices. We therefore zero mean the head pose and eye gaze vectors at participant level. We show a sample visualisation of OpenFace features in Fig- ure 1. Head pose is depicted as a blue cube centred on the head pointing in the estimated direction of head pose. Eye gaze vectors are d... | https://arxiv.org/abs/2505.21043v1 |
the ’com- patibility’ of output with the query via the key and value (Vaswani, 2017). In self-attention blocks, as q=k=v, the model learns temporal patterns in the input e.g. audio or video. In cross-attentionblocks, we stack Nlayers with two inputs x1and x2. In each layer, we compute two transformer layers with shared... | https://arxiv.org/abs/2505.21043v1 |
of an overlap, or during mutual si- lence between speaking turns. We sum the shift probability, defined as the marginal probability of all V AP states where the non-active speaker is 1 in both bins in the 600-2000 ms period (Figure 3). Like the original V AP paper (Ekstedt and Skantze, 2022b) we only consider the latte... | https://arxiv.org/abs/2505.21043v1 |
results are comparable with those of Ekstedt and Skantze (2022b), verifying our re- implementation. The F1and balanced accuracy scores are slightly higher using the ground-truth alignment, reflecting the ASR alignment error. In the Candor corpus, using audio-only cues, the V AP model performs well above the baseline an... | https://arxiv.org/abs/2505.21043v1 |
↓8% 0.45 ↓33% 66 ↓14% a+v (e) 0.83 ↑2% 0.88 - 0.69 ↑4% 79 ↑3% a+v (l) 0.83 ↑3% 0.89 ↑2% 0.70 ↑6% 80 ↑4% before end of turn a 0.86 0.91 0.66 77 (FTO > 0 ms) v 0.78 ↓10% 0.87 ↓5% 0.43 ↓35% 70 ↓10% a+v (e) 0.87 ↑2% 0.92 - 0.69 ↑6% 80 ↑3% a+v (l) 0.87 ↑2% 0.92 -0.71 ↑6% 83 ↑4% before overlap a 0.78 0.85 0.57 70 (FTO < -250... | https://arxiv.org/abs/2505.21043v1 |
Silence between Turns Figure 6: Balanced accuracy of models averaged over 5 folds, ±standard error in the mean, grouped by a minimum period of silence between turns; i.e. the FTO. The balanced accuracies in Tables 2 and 3 corre- spond with the 0 and 250 ms points in Figure 6. The gap in performance of Switchboard and C... | https://arxiv.org/abs/2505.21043v1 |
notable when considering the F1 shift score. We compare the weighted F1score, and find that the gaze and landmark trained models perform significantly worse than the model trained on all visual features in all cases ( p < 0.01when comparing a+v (l) with gaze and landmarks, Table 4). There is no significant difference i... | https://arxiv.org/abs/2505.21043v1 |
the psycholinguis- tics literature (Sections 1 and 2), we believe visual cues such as gaze aversion are exploited in our multimodal model. Future work is needed to estab- lish the exact role of visual cues and their impact on model performance, but the literature outlines why visual cues are particularly important duri... | https://arxiv.org/abs/2505.21043v1 |
technol- ogy, though this should be confirmed. A compar- ison of a model trained on a corpus of in-person interaction would be useful. However, this is not straightforward due to the lack of availability of a suitable public dataset. Available corpora of dyadic interaction are not large, e.g. the 11-hour Mahnob mimicry... | https://arxiv.org/abs/2505.21043v1 |
processed by passing them through a pre-trained feature extractor (Riviere et al., 2020), whereas the visual features are high-level descrip- tors, e.g. angles in radians. This issue could be resolved in a future iteration of the model, replac- ing the OpenFace front-end, as discussed. Never- theless, the model introdu... | https://arxiv.org/abs/2505.21043v1 |
on , volume 1, pages 517–520. IEEE Com- puter Society. Mattias Heldner and Jens Edlund. 2010. Pauses, gaps and overlaps in conversations. Journal of Phonetics , 38(4):555–568. Dan Hendrycks and Kevin Gimpel. 2016. Gaus- sian error linear units (gelus). arXiv preprint arXiv:1606.08415 . Judith Holler, Kobin H Kendrick, ... | https://arxiv.org/abs/2505.21043v1 |
frowns facilitate question identification: An online study using virtual avatars. Cognitive Science , 47(12):e13392. Brid O’Conaill, Steve Whittaker, and Sylvia Wilbur. 1993. Conversations Over Video Conferences: An Evaluation of the Spoken Aspects of Video-Mediated Communication. Human–Computer Interaction , 8(4):389–... | https://arxiv.org/abs/2505.21043v1 |
arXiv:2505.21068v1 [cs.CL] 27 May 2025Predicting Implicit Arguments in Procedural Video Instructions Anil Batra1Laura Sevilla-Lara1Marcus Rohrbach2Frank Keller1 1University of Edinburgh, United Kingdom2TU Darmstadt & hessian.AI, Germany a.k.batra@sms.ed.ac.uk l.sevilla@ed.ac.uk marcus.rohrbach@tu-darmstadt.de keller@in... | https://arxiv.org/abs/2505.21068v1 |
is cut into cubes. This inference relies on linguistic context, including pronominal reference. In step 5 of Figure 1, we infer that dressing refers to the mixture of olive oil, lemon, garlic, and salt from step 3, while bread refers to the cubes of oiled pita bread from step 2, which are later added to a bowl containi... | https://arxiv.org/abs/2505.21068v1 |
our SRL scheme as an intermediate representation in the next step prediction task, we show that it boosts the Qwen2-VL model’s performance in predict-ing future steps, leading to a ∼2% improvement in the METEOR score, (iv)We show that large multimodal models achieve good performance in next-step and implicit argument p... | https://arxiv.org/abs/2505.21068v1 |
it uses videos- captions pairs from YouCook2 (Zhou et al., 2018) to create graphs linking objects and entities. How- ever, it often faces issues with underspecified in- structions, leading to incomplete knowledge graphs and QA pairs. Our dataset, in contrast, offers com- prehensive information based on which complete g... | https://arxiv.org/abs/2505.21068v1 |
has an equal representation among other domains, such as COIN (Tang et al., 2019), or that concentrate on other domains, like Assembly- 101 (Sener et al.). Cooking recipes are conceptu- ally challenging, with many elliptical instructions and state transformations (shape or visual changes). These can be learned from lar... | https://arxiv.org/abs/2505.21068v1 |
video. For instance, for (i) “add onions and tomatoes to the blender and blend them” and (ii) “add spices and garlic to the blender”, the annotators turn (ii) into “add spices and garlic to onions, tomatoes ”.Name Value Number of videos 231 Average/max video duration 125.33/588.2 Average/max steps per video 7.47/14 Ave... | https://arxiv.org/abs/2505.21068v1 |
Step 2 (b) : cut - olive oil, pita bread - ø Step 3 : combine - olive oil, lemon, garlic, salt - ø Step 4 : add - cucumbers, lettuce, tomatoes, onions - øMultimodal Large Model VERB WHAT WHERE / WITH IMPLICIT MASKED Input Sequence (Text or Video) Output - Implicit Semantic Arguments Figure 2: The Implicit Argument Pred... | https://arxiv.org/abs/2505.21068v1 |
set, we generate the silver- standard dataset for training a model that can iden- tify implicit arguments. Specifically, we prompt the GPT-4o to perform two tasks similar to our manual annotation: (i) split the multi-step instructions into single predicate-argument structures, (ii) automati- cally infer the implicit en... | https://arxiv.org/abs/2505.21068v1 |
both tasks. For text-only inputs , we use Qwen2 instruct (Yang et al., 2024) as the primary language model, due to its perfor- mance and use in recent video-LLMs. We also use the LLama 3.1 (Dubey et al., 2024) instruct model. Forvideo-only inputs , we focus on models with long context, i.e., models accepting a large nu... | https://arxiv.org/abs/2505.21068v1 |
although their performance is lacking be- hind the scenario with text-only input. This high- lights two primary challenges in semantic argu- ment prediction from video: recognizing entities in the current step and inferring or tracking en- tities within the temporal context. For smaller open-source models, performance ... | https://arxiv.org/abs/2505.21068v1 |
ø Step 5(b) : shred - pork, salt, paprika, muscovado sugar, pepper - ø Step 2 : rub - salt, paprika, muscovado sugar, pepper - pork Step 3(a) : place - pork, salt, paprika, muscovado sugar, pepper - ø Step 3(b) : pour - cider - pork, salt, paprika, muscovado sugar, pepper Step 4 : cook - pork, salt, paprika, muscovado ... | https://arxiv.org/abs/2505.21068v1 |
investigated the impact of fine-tuning the Qwen2 model with text-only inputfor next step prediction task, with and without se- mantic role prediction. The results in Table 5 reveal that naive finetuning without incorporating SRL predictions in the output diminishes the model’s ability to infer and track explicit and im... | https://arxiv.org/abs/2505.21068v1 |
formance decreases gradually as semantic positionincreases. In contrast, iSRL-Qwen2-VL achieves performance for implicit arguments across the board for what arguments, and for later seman- tic frame positions for where arguments (see Fig- ure 4e, 4f). This shows that while GPT-4o excels at local contextual reasoning, i... | https://arxiv.org/abs/2505.21068v1 |
Association for Computational Linguistics and 17th International Conference on Computational Linguistics, Volume 1 , pages 86–90, Montreal, Quebec, Canada. Association for Compu- tational Linguistics.Satanjeev Banerjee and Alon Lavie. 2005. METEOR: An automatic metric for MT evaluation with im- proved correlation with ... | https://arxiv.org/abs/2505.21068v1 |
Edward Zhang, Jinxu Zhang, Angela Castillo, Changan Chen, Xinzhu Fu, Ryosuke Furuta, Cristina Gonzalez, Prince Gupta, Jiabo Hu, Yifei Huang, Yiming Huang, Weslie Khoo, Anush Kumar, Robert Kuo, Sach Lakhavani, Miao Liu, Mi Luo, Zhengyi Luo, Brighid Meredith, Austin Miller, Oluwatumininu Oguntola, Xiaqing Pan, Penny Peng... | https://arxiv.org/abs/2505.21068v1 |
Marın, Aritro Biswas, Ferda Ofli, Nicholas Hynes, Amaia Salvador, Yusuf Aytar, Ingmar Weber, and Antonio Torralba. 2021. Recipe1m+: A dataset for learning cross-modal embeddings for cooking recipes and food images. IEEE Transactions on Pattern Anal- ysis and Machine Intelligence , 43(1):187–203. Adam Meyers, Ruth Reeve... | https://arxiv.org/abs/2505.21068v1 |
Shinsuke Mori, and John A Carroll. 2020. English recipe flow graph corpus. In Proceed- ings of the Twelfth Language Resources and Evalua- tion Conference , pages 5187–5194. An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. 2024. Qwen2. 5 tec... | https://arxiv.org/abs/2505.21068v1 |
onions” the second argument is missing, which is implicit and we need to add the implicit argument. 5.If the verb is transitive i.e. need one object. For example, (i) “add flour and spices in a bowl and mix”. Here “mix” verb requires an implicit object i.e. mix flour and spices.Stage 2 - GPT-4o-Mini Labels: The prompt ... | https://arxiv.org/abs/2505.21068v1 |
an unconstrained environment and a third-person viewpoint, while Tasty videos are captured using an overhead camera. The videos in YouCook2 are sourced from YouTube and there are on average ≈7.7 instructions per video. The videos in Tasty are relatively short, with an average duration of 54 seconds, and each recipe inc... | https://arxiv.org/abs/2505.21068v1 |
Only Video Input (h) Only Video Input (i) Video + Text Input (j) Video + Text Input Figure 6: Comparison of GPT-4o and ours iSRL-Qwen2-VL models for argument prediction across semantic frame positions. Gold GPT-4o (V+T) iSRL-Qwen2-VL (V+T) Input Sequence (Text / Video) Step 2 :Step 1 : Step 3 : Step 4 : Step 5 : Step 6... | https://arxiv.org/abs/2505.21068v1 |
milk, cheese slices - ø Step 7 : cook - waffle mix, rosemary, egg, oil, milk, cheese slices - ø Step 2 : stir - waffle mix, rosemary, egg, oil, milk - ø Step 3 : pour - waffle mix, rosemary, egg, oil, milk - ø Step 4: cook - waffle mix, rosemary, egg, oil, milk - ø Step 5 : add - cheese slices - waffle mix, rosemary, e... | https://arxiv.org/abs/2505.21068v1 |
waffle mix, rosemary, egg, oil, milk iSRL-Qwen2-VL (V) (b)Qualitative example using video (Video) and multimodal (Video + Text) inputs. Example is from Tasty (Sener et al., 2022) with ID-waffle-grilled-cheese. Incorrect predictions are highlighted in red. The examples show that GPT-4o, with multimodal inputs (V+T), bia... | https://arxiv.org/abs/2505.21068v1 |
(e.g., pan, bowl) for the arguments, only mention ingredients and explicit tools. 1.5. Carefully observe Example 1, Example 2 to understand the steps and format. 1.6. Do not add prepositions (e.g., "in," "on"), articles, determiners, or quantities to the arguments. 2. Do not include any additional information in the re... | https://arxiv.org/abs/2505.21068v1 |
*ONLY WHAT ARGUMENT*: The verb such as slice, pound, lay, etc. is used, then you need to fill the WHAT argument only and ignore the WHERE_OR_WITH argument. Consider the following example: (a). Example - 1: In the below instruction, you need to fill the WHAT argument only. Here “mix” verb requires an implied object i.e.... | https://arxiv.org/abs/2505.21068v1 |
**: 1. It majorly focus on the ingredients used in the recipe and avoid the focus on the tools used to prepare the ingredients. 2. Based on the verb, the WHAT or WHERE_OR_WITH or both arguments are updated. Consider the following cases: 2.1. *BOTH ARGUMENTS*: The verb such as add, mix, chop, etc. is used, then the both... | https://arxiv.org/abs/2505.21068v1 |
the arguments, only mention ingredients. 2.5. Do not add prepositions (e.g., "in," "on"), articles, determiners, or quantities to the arguments. **Task **: In this task, you will be presented with the partial list of numbered recipe steps from a cooking recipe along with its semantic role labels. You are also given cor... | https://arxiv.org/abs/2505.21068v1 |
1: ["Squeeze some lime juice into the food processor", "add some olive oil"] - Complex Instruction 2: "thinly slice the apple and place in a medium bowl with the water and lemon juice to prevent browning." - Sub Instructions 2: ["thinly slice the apple", "place sliced apple in a medium bowl", "add the water and lemon j... | https://arxiv.org/abs/2505.21068v1 |
in step (v), we call this as composite entity. In step (v), when we dip the chicken in the sauce, we need to reason that chicken is already coated with the flour in step (iv), which is a part of enitity for composition of "flour, salt, pepper" based on the step (iii). *Recipe Steps and Semantic Role Labels*: (i) combin... | https://arxiv.org/abs/2505.21068v1 |
arXiv:2505.21072v2 [cs.CL] 28 May 2025Faithfulness-Aware Uncertainty Quantification for Fact-Checking the Output of Retrieval Augmented Generation Ekaterina Fadeeva1♢Aleksandr Rubashevskii2♢Roman Vashurin2 Shehzaad Dhuliawala1Artem Shelmanov2Timothy Baldwin2 Preslav Nakov2Mrinmaya Sachan1Maxim Panov2 1ETH Zürich2MBZUAI... | https://arxiv.org/abs/2505.21072v2 |
To address this distinction, we differentiate be- tween factuality and faithfulness . Faithfulness refers to whether the generated output is seman- tically entailed by the retrieved context, while fac- tuality indicates whether the content is objectively correct (Dziri et al., 2022; Yang et al., 2024). In the context o... | https://arxiv.org/abs/2505.21072v2 |
compared to other approaches; see Section 4.2 Uncertainty Quantification for RAG Letxbe the user query submitted to the RAG sys- tem. The system first retrieves kpassages, denoted byr={r1, . . . , r k}, from an external knowledge source using xas the query. The RAG model then employs a LLM to generate a greedy output y... | https://arxiv.org/abs/2505.21072v2 |
UQ strategies depending on the outcome. The overall factuality of a claim cis estimated as: P(cis true ) = (2) P(cis faithful to r)·P(cis true |faithful )+ P(cis unfaithful to r)·P(cis true |unfaithful ). Here, we calculate P(cis unfaithful to r)as1− P(cis faithful to r). This decomposition inte- grates three probabili... | https://arxiv.org/abs/2505.21072v2 |
for this component has limited impact on overall FRANQ performance (see Appendix D.1). Again, for short-form QA, we find alternative baselines more suitable, particularly Semantic En- tropy (Kuhn et al., 2023), which better captures uncertainty in this scenario (see Appendix C.2). Therefore, we estimate the factuality ... | https://arxiv.org/abs/2505.21072v2 |
multiple claims; and short- form, where questions are answered concisely with 4 single claims. For both QA settings, we employ Llama-3.2-3B-Instruct (Grattafiori et al., 2024) andFalcon3-3B-Base (Team, 2024) as the LLMs. Long-form QA dataset. Generating high-quality data for RAG hallucination detection remains chal- le... | https://arxiv.org/abs/2505.21072v2 |
supervised XGBoost baseline (Kuleshov and Liang, 2015) trained using the outputs of unsupervised UQ methods as features. Furthermore, in ablation studies, we investi- gate the impact of different FRANQ components P(faithful ),UQfaith, and UQunfaith , as well as vary- ing training sizes. 4.1 Experimental Setup UQ baseli... | https://arxiv.org/abs/2505.21072v2 |
and ranks second in PRR. For Falcon 3B, the condition-calibrated FRANQ variant performs best overall, with the highest ROC-AUC and PR-AUC, while the non-calibrated FRANQ achieves the top PRR and second-best ROC-AUC. Note that, except for FRANQ calibrated on Fal- con 3B, the proposed FRANQ method and its two calibration... | https://arxiv.org/abs/2505.21072v2 |
.263 MaxNLI 20.25 20.25 20.25 .466 .303 -.036 20.00 19.00 20.00 .514 .532 .035 AlignScore 17.00 18.25 17.50 .631 .415 .207 7.50 7.50 7.25 .682 .666 .372 Parametric Knowledge 18.00 18.00 17.75 .627 .425 .247 18.25 19.25 18.75 .561 .556 .104 XGBoost (all UQ features) 9.50 8.50 7.25 .766 .594 .494 3.00 3.25 3.75 .733 .705... | https://arxiv.org/abs/2505.21072v2 |
FRANQ .To better understand XGBoost’s behavior, we analyze its first decision tree trained onFRANQ features (see Figure 12 in Appendix D). We find that the tree mimics FRANQ ’s decision logic: initially splitting on AlignScore, followed by MaxNLI or Parametric Knowledge depending on the score. This suggests that superv... | https://arxiv.org/abs/2505.21072v2 |
sation. Wizard of Wikipedia (Dinan et al., 2019) and FaithDial (Dziri et al., 2022), for example, are structured as user–model dialogues grounded in Wikipedia content. Unlike RAG datasets, however, these datasets place greater emphasis on maintain- ing coherent dialogue and conversational consis- tency. Moreover, as in... | https://arxiv.org/abs/2505.21072v2 |
estimation can support safer AI deployment, especially in knowledge- intensive domains such as education, healthcare, or law. However, the system should not be consid- ered a replacement for human fact-checkers. It is best used as a decision-support tool rather than a source of truth.References Yung-Sung Chuang, Linlu ... | https://arxiv.org/abs/2505.21072v2 |
Hu, Yuhan Sun, and Quanjun Zhang. 2024. Lrp4rag: Detecting hallucinations in retrieval- augmented generation via layer-wise relevance prop- agation. arXiv preprint arXiv:2408.15533. Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and ... | https://arxiv.org/abs/2505.21072v2 |
Scales, David Dohan, Ed H Chi, Nathanael Schärli, and Denny Zhou. 2023. Large language mod- els can be easily distracted by irrelevant context. InInternational Conference onMachine Learning , pages 31210–31227. PMLR. 10 Kurt Shuster, Spencer Poff, Moya Chen, Douwe Kiela, and Jason Weston. 2021. Retrieval augmentation r... | https://arxiv.org/abs/2505.21072v2 |
↑PRR↑ Max Sequence Probability .680 .440 .292 .745 .550 .421 .774 .529 .478 .833 .712 .625 CCP .705 .471 .357 .709 .526 .393 .767 .528 .471 .800 .680 .552 Max Token Entropy .723 .503 .389 .768 .607 .455 .796 .569 .523 .809 .697 .555 P(True) .463 .256 -.042 .550 .374 .100 .474 .244 -.022 .419 .294 -.082 Lexical Similari... | https://arxiv.org/abs/2505.21072v2 |
.468 .155 .685 .831 .456 Perplexity .593 .545 .199 .626 .629 .215 .547 .428 .097 .645 .809 .388 TokenSAR .602 .555 .210 .656 .654 .249 .553 .433 .106 .654 .818 .405 SentenceSAR .509 .455 .012 .755 .707 .463 .523 .395 .026 .739 .850 .552 MaxNLI .559 .502 .086 .507 .535 .065 .483 .358 -.044 .506 .733 .033 AlignScore .655... | https://arxiv.org/abs/2505.21072v2 |
We then decomposed the extracted answers into atomic claims using the prompt shown in Figure 4. For each claim, we identified its corresponding span in the original sentence using the prompt in Figure 5. Claims for which we could not extract spans (due to inconsistencies in annotation, e.g., words appearing in a differ... | https://arxiv.org/abs/2505.21072v2 |
500 1048 91.4 % 6.0 % 2.6 % 38.2 % 61.5 % 0.3 % 1720.2 Table 7: Statistics of datasets used in long-form QA benchmark. statement, “unfaithful-neutral” if the context does not contain any information supporting or contradicting the statement. Next, the obtained labels are binarized for experimental evaluation: 1 if labe... | https://arxiv.org/abs/2505.21072v2 |
Sentence: [... More sentences follow ...] Now analyze the following text using this format: Text: {text} Claim: {claim} Answer: Figure 5: Prompt template used with GPT-4o to identify the span in the original text corresponding to each atomic claim. The model is instructed to process each sentence and extract words rele... | https://arxiv.org/abs/2505.21072v2 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.