text
string
source
string
of GPT-4o, which allows it to better adapt to the abstract thinking pattern granted by sketches and align more closely with human cognitive reasoning. In contrast, Qwen2.5-VL-32B, with its relatively smaller scale, struggles to accommodate this shift and benefits less from V AT compared to GPT-4o or Gemini-2.0 Pro. Add...
https://arxiv.org/abs/2505.20164v1
has not yet been tested in real-world environments, such as embodied scenarios, which present challenges not covered by existing benchmarks. Testing in dynamic, real-world contexts would help assess its practical applicability. 9 References [1]Simon A Aytes, Jinheon Baek, and Sung Ju Hwang. Sketch-of-thought: Efficient...
https://arxiv.org/abs/2505.20164v1
2024. [12] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. TALL: temporal activity localization via language query. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017 , pages 5277–5285. IEEE Computer Society, 2017. [13] Yulia Gryaditskaya, Mark Sypesteyn, Janwillem ...
https://arxiv.org/abs/2505.20164v1
Proceedings of the 31st International Conference on Computational Linguistics , pages 2886–2903, 2025. [24] Chengzu Li, Wenshan Wu, Huanyu Zhang, Yan Xia, Shaoguang Mao, Li Dong, Ivan Vuli ´c, and Furu Wei. Imagine while reasoning in space: Multimodal visualization-of-thought, 2025. [25] Mengtian Li, Zhe Lin, Radomir M...
https://arxiv.org/abs/2505.20164v1
2024 , pages 1229–1241, Mexico City, Mexico, 2024. Association for Computational Linguistics. [39] Patsorn Sangkloy, Nathan Burnell, Cusuh Ham, and James Hays. The sketchy database: learning to retrieve badly drawn bunnies. ACM Transactions on Graphics (TOG) , 35(4):1–12, 2016. [40] Patsorn Sangkloy, Wittawat Jitkrittu...
https://arxiv.org/abs/2505.20164v1
language models? In Kevin Duh, Helena Gomez, and Steven Bethard, editors, Findings of the Association for Computational Linguistics: NAACL 2024 , pages 2148–2164, Mexico City, Mexico, 2024. Association for Computational Linguistics. [53] Wenshan Wu, Shaoguang Mao, Yadong Zhang, Yan Xia, Li Dong, Lei Cui, and Furu Wei. ...
https://arxiv.org/abs/2505.20164v1
S. Ryoo, Vikas Sindhwani, Johnny Lee, Vincent Vanhoucke, and Pete Florence. Socratic models: Composing zero-shot multimodal reasoning with language. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023 . OpenReview.net, 2023. [65] Chi Zhang, Zhao Yang, Jiaxuan L...
https://arxiv.org/abs/2505.20164v1
them to achieve the best performance. You should reply in the following format: ANSWER: (your answer). For example: ANSWER: (A). Please think step by step to obtain the answer: B Benchmarks and Evaluation Metrics B.1 Construction of Odd-one-out Benchmark Odd-One-Out is our newly constructed benchmark designed to evalua...
https://arxiv.org/abs/2505.20164v1
benchmark, V AT-ReAct consistently outperforms SketchPad across all task types, demonstrating its strong generalization in visual reasoning. For spatial reasoning tasks, V AT-ReAct achieves significant improvements, including +5.79 in Dir-Rec (46.59 vs. 40.80) and +4.67 in Rot-Diff (89.50 vs. 84.83). In object-centric ...
https://arxiv.org/abs/2505.20164v1
verbal reasoning in these smaller models. Larger models such as GPT-4o and Gemini demonstrate emergent capabilities in visual abstract thinking . These models are able to process and integrate abstract visual representations more effectively, which supports coherent and structured multimodal reasoning. E Impact Stateme...
https://arxiv.org/abs/2505.20164v1
involving spatial understanding and directional inference. In contrast, CoT, which emphasizes explicit verbal reasoning, leads to substantial gains in tasks depend more on world knowledge where verbal reasoning and planning is especially effective. Consequently, these tasks can be further boosted by combining V AT with...
https://arxiv.org/abs/2505.20164v1
arXiv:2505.20166v1 [eess.AS] 26 May 20251 From Alignment to Advancement: Bootstrapping Audio-Language Alignment with Synthetic Data Chun-Yi Kuan and Hung-yi Lee Abstract —Audio-aware large language models (ALLMs) have recently made great strides in understanding and processing audio inputs. These models are typically a...
https://arxiv.org/abs/2505.20166v1
fundamental issue with existing ALLMs is catastrophic forgetting. Since most ALLMs are adapted from text-based LLMs, they inherently possess strong language understanding capabilities before be- ing trained on audio tasks. However, prior research has re- ported that currently developed ALLMs often lose their textual ca...
https://arxiv.org/abs/2505.20166v1
for differences between audio samples. Second, a description-based process that generates separate captions for each audio sample within the same response. Our experimental results demonstrate that training the model on multi-audio data, whether through joint captioning or comparative discrimination, sufficiently enhan...
https://arxiv.org/abs/2505.20166v1
using only 12% of the training data duration. 5) To comprehensively assess audio-aware large language models, we conduct experiments on existing bench- marks, reproduce missing evaluation sets, and introduce new test sets to explore previously unexamined aspects. Our evaluation covers three key areas: audio question an...
https://arxiv.org/abs/2505.20166v1
use the backbone language model to generate responses conditioned on textual transcriptions. BLSP [16] further extends this idea by generating continuations based on transcriptions, while BLSP-Emo [17] and DeSTA2 [27] incorporate additional paralinguistic information like emotion to produce more expressive continuation...
https://arxiv.org/abs/2505.20166v1
descriptions of sound events that actually occur in the audio. For example, Replay the audio . Negative Samples Generation Prompt ( Pneg)aims to gen- erate descriptions of sound events that are not present in the audio. For example, Identify sounds that are absent as contrasting examples . Combined Samples Generation P...
https://arxiv.org/abs/2505.20166v1
events not present in the provided audio: 1. A car driving by 2. A dog barking 3. The door bang learning complexity compared to single-audio alignment. We refer to this multi-audio extended approach as BALSa-MA (Multi-Audio). To implement this, we extend the generation prompt to accommodate multiple audio inputs. Here,...
https://arxiv.org/abs/2505.20166v1
being primarily for automatic speech recognition and speech translation. Since Whisper follows an encoder-decoder architecture, we utilize only its encoder component. To preserve the advantages of the pre-trained model, the parameters of the audio encoder remain frozen. Backbone Large Language Model . This study adopts...
https://arxiv.org/abs/2505.20166v1
[15]–[18], [31], we developed a specialized modality adapter, featuring a two-layer transformer decoder architecture. This adapter uses 64 specialized vectors to capture and process audio features from the encoder’s hidden states. In our implementation strategy, we maintain the original architecture of both LLaMA and W...
https://arxiv.org/abs/2505.20166v1
In ClothoAQA, which includes both binary and non-binary classification tasks, we report accuracy as the primary metric. Audio Reasoning Benchmark . We report weighted F1 score for multi-class classification tasks. For Synonym-Hypernym Test, we also report weighted precision and recall. In addition, the original MMAU be...
https://arxiv.org/abs/2505.20166v1
of audio ]Pseed[End of audio ]Pgen Pseed∈ {Dcaption ,Dtag} Pgen∈ {QOpenAQA }.(3) For comparison, we also directly use question-answer pairs derived from the OpenAQA dataset as our training data. Fourth, we explore an alternative approach where, instead of using BALSa to generate captioning data, we employ pre-designed ...
https://arxiv.org/abs/2505.20166v1
covering 13 different acoustic scenes. The primary goal of this task is to determine the environment in which a given audio recording was captured based on its acoustic characteristics. For evaluation, we sample 100 audio clips from each category in the test split, resulting in a total of 1,300 test instances. We conve...
https://arxiv.org/abs/2505.20166v1
entailment to assess a model’s ability to perform deductive reasoning. This task evaluates whether a textual description (hypothesis) can be logically inferred from an audio recording (premise). The conclusion falls into one of three categories: entailment, neutral, or contradiction, depend- ing on the available eviden...
https://arxiv.org/abs/2505.20166v1
ALLMs. We formulate object hallucination detection as a binary classification task, prompting ALLMs to respond with “Yes” or “No”. To assess model robustness, we design four dif- ferent prompts that ask whether a particular sound is present, including variations such as Is there a sound of [object]? ,Does the audio con...
https://arxiv.org/abs/2505.20166v1
dif- ferent benchmarks. Notably, our models achieve these results using only 12% of the equivalent training duration required by the widely adopted and powerful Qwen2-Audio-Instruct model, as shown in Table III.The EDANSA AQA tasks serve as zero-shot benchmarks, evaluating models on out-of-domain audio that is not enco...
https://arxiv.org/abs/2505.20166v1
while 10 TABLE IV EVALUATION RESULTS FOR THE AUDIO QUESTION ANSWERING AND AUDIO HALLUCINATION BENCHMARK . F1 (Y) AND F1 (N) REPRESENT THE F1 SCORES FOR CASES WHERE THE CORRECT ANSWER IS YES AND NO ,RESPECTIVELY . F1 (W) DENOTES THE WEIGHTED F1SCORE . ACC REFERS TO ACCURACY ,AND YES INDICATES THE PERCENTAGE OF RESPONSES...
https://arxiv.org/abs/2505.20166v1
involve any modifications to the audio foundation model or the backbone large language model. Furthermore, SAKURA [41] is a multi-hop reasoning benchmark designed to assess whether a model can first comprehend information within audio and then perform multi- hop reasoning based on that information. If a model merely me...
https://arxiv.org/abs/2505.20166v1
LLM with a stronger LLM like Gemini [26], while applying the same BALSa pipeline to generate training data. As shown in Table VI and Table VII, we observe that across all evaluation benchmarks, performance significantly declines regardless of whether the setting is positive-only, positive- negative, or combined. This r...
https://arxiv.org/abs/2505.20166v1
benefiting more advanced reasoning tasks. Following these findings, we investigate whether learning through discrimination-based training—where the model ex- plains the differences between two audio inputs—or learning through complete description—where the model captions both audio inputs—is sufficient for improving pe...
https://arxiv.org/abs/2505.20166v1
90.78 91.25 91.02 -0.02 Ours (BALSa-MA, discrimination) 81.76 80.50 80.55 79.80 86.00 73.49 56.76 54.67 55.66 88.85 87.25 88.05 -3.27 Ours (BALSa-MA, captioning) 82.99 82.86 82.89 80.30 88.40 71.72 57.06 56.40 56.59 89.07 84.25 86.66 -4.80 Ours (BALSa-MA, joint) 78.43 78.46 78.28 79.30 86.00 72.79 56.16 48.92 54.80 90....
https://arxiv.org/abs/2505.20166v1
. Gong, A. H. Liu, H. Luo, L. Karlinsky, and J. Glass, “Joint audio and speech understanding,” in 2023 IEEE Automatic Speech Recognition andUnderstanding Workshop (ASRU). IEEE, 2023, pp. 1–8. [6] S. Ghosh, S. Kumar, A. Seth, C. K. R. Evuru etal., “Gama: A large audio-language model with advanced audio understanding and...
https://arxiv.org/abs/2505.20166v1
arXiv preprint arXiv:2303.18223, vol. 1, no. 2, 2023. [25] T. Brown, B. Mann, N. Ryder, M. Subbiah etal., “Language models are few-shot learners,” Advances inneural information processing systems, vol. 33, pp. 1877–1901, 2020. [26] G. Team, P. Georgiev, V . I. Lei etal., “Gemini 1.5: Unlocking mul- timodal understandin...
https://arxiv.org/abs/2505.20166v1
reasoning benchmark,” in The Thirteenth International Conference onLearning Representations, 2024. [43] C.-y. Huang etal., “Dynamic-superb: Towards a dynamic, collabo- rative, and comprehensive instruction-tuning benchmark for speech,” inICASSP 2024-2024 IEEE International Conference onAcoustics, Speech and Signal Proc...
https://arxiv.org/abs/2505.20166v1
arXiv:2505.20176v1 [cs.CL] 26 May 2025“KAN you hear me?” Exploring Kolmogorov-Arnold Networks for Spoken Language Understanding Alkis Koudounas†, Moreno La Quatra‡, Eliana Pastor†, Sabato Marco Siniscalchi⋆, Elena Baralis† †Politecnico di Torino, Turin, Italy ‡Kore University of Enna, Enna, Italy ⋆Universit `a degli St...
https://arxiv.org/abs/2505.20176v1
different con- figurations, integrating KAN layers within the final dense block of a 2D-CNN-based model operating on spectrogram inputs. These evaluations are carried out on two datasets: FLUENT SPEECH COMMANDS (FSC) [19] and T IMERS AND SUCH [20]. We also compare several function ap- proximations within the KAN layer,...
https://arxiv.org/abs/2505.20176v1
L(x) =w1b(x) +w2spline (x) (2) where w1andw2are learnable scalars. This formulation en- hances flexibility while maintaining computational efficiency. Several alternatives have been proposed to better approximate functions in KAN layers. RBF [21, 22] provides smooth ap- proximations by weighting basis functions centere...
https://arxiv.org/abs/2505.20176v1
original linear projections, poten- tially leading to a richer and more discriminative representation. FK: Minimal Hybrid Configuration. Finally, FK represents the most compact hybrid architecture, comprising a single FF layer followed by a single KAN layer. This configuration serves as a baseline for evaluating the ef...
https://arxiv.org/abs/2505.20176v1
KAN KAN KAN KAN FFFF FFFF FF KANVARIABLE HIDDEN SIZE VARIABLE HIDDEN SIZE Predicted Intent CNN-2D Block Dense Block MEL-Spectrogram C1 C2 C3 C4 C5 C6FF FF FFFF FF KAN FFKAN KAN KAN KAN KAN FFFF FF Predicted Intent FF FFKAN VARIABLE HIDDEN SIZE VARIABLE HIDDEN SIZE FF FFKAN VARIABLE HIDDEN SIZE HIDDEN SIZE FIXED: 128 (b...
https://arxiv.org/abs/2505.20176v1
.583 ±.006 .575 ±.002 FKF(b) 256 12.9M 1.80it/s .586 ±.013 .584 ±.017 FKF(b) 512 27.4M 1.71it/s .631 ±.014 .628 ±.016 FKF(b) 1024 61.2M 1.52it/s .616 ±.029 .579 ±.032 monolingual wav2vec 2.02and the multilingual XLS-R3. These models utilize the same dense classification block as the 2D- CNN. Their training setup follow...
https://arxiv.org/abs/2505.20176v1
±.011 .672 ±.008 FFK 315.9M .734 ±.007 .642 ±.012 .678 ±.010 FKF 316.5M .749 ±.006 .667 ±.009 .679 ±.004 pacts the final performance in Table 2. FKF(a) maintains a fixed hidden size of 128 for the initial FF layer while varying the KAN layer’s output size. The results show that performance improves as the KAN output si...
https://arxiv.org/abs/2505.20176v1
models and T IMERS AND SUCH, where the FKF configuration shows a marked improvement in performance. We analyze differences in how FFF and FKF models process inputs, particularly in cases where FFF makes incorrect predictions that FKF cor- rects. To explore this, we use an explanation technique that as- signs relevance ...
https://arxiv.org/abs/2505.20176v1
M. Griebel, “On a constructive proof of kol- mogorov’s superposition theorem,” Constructive approximation , vol. 30, pp. 653–675, 2009. [5] R. Yu, W. Yu, and X. Wang, “Kan or mlp: A fairer comparison,” arXiv preprint arXiv:2407.16674 , 2024. [6] E. Poeta, F. Giobergia, E. Pastor, T. Cerquitelli, and E. Baralis, “A benc...
https://arxiv.org/abs/2505.20176v1
benchmark for spoken language un- derstanding with numbers,” in NeurIPS Datasets and Benchmarks Track , 2021. [21] M. J. Orr et al. , “Introduction to radial basis function networks,” 1996. [22] Z. Li, “Kolmogorov-arnold networks are radial basis function net- works,” arXiv preprint arXiv:2405.06721 , 2024. [23] A. Del...
https://arxiv.org/abs/2505.20176v1
TH INK: Can Large Language Models Think-aloud? Yongan Yu, Mengqian Wu, Yiran Lin, Nikki G. Lobczowski* McGill University, {yongan.yu, mengqian.wu, yiran.lin}@mail.mcgill.ca nikki.lobczowski@mcgill.ca Abstract Assessing higher-order thinking skills in large language models (LLMs) remains a fundamen- tal challenge, espec...
https://arxiv.org/abs/2505.20184v1
LLMs often struggle witharXiv:2505.20184v1 [cs.CL] 26 May 2025 advanced reasoning processes (Collins et al., 2022), such as "thinking one step ahead" or adopting a the- ory of mind (ToM) perspective (Holterman and van Deemter, 2023) to anticipate the types of prob- lems that stimulate students’ creativity or critical t...
https://arxiv.org/abs/2505.20184v1
often struggle with structured reasoning and in- ductive judgment, diverging from human-like pat- terns (Lamprinidis, 2023). Although techniques like chain-of-thought (CoT) prompting (Wei et al., 2022) can enhance model reasoning, they remain insufficient to capture higher-order cognition on a scale (Prystawski et al.,...
https://arxiv.org/abs/2505.20184v1
to the guidance of teachers, which can better instruct LLMs to think about dis- assembly and improvement, thereby triggering the HOT skills. Inquiry-based learning (Pedaste et al., 2015) emphasizes active engagement of learners in formulating questions and seeking answers. The ability to ask meaningful questions signal...
https://arxiv.org/abs/2505.20184v1
Suggestions: 1: "The new problem attempts to apply mathematical concepts related to ...", 2: "Ensure that the conditions provided in the problem lead to a solvable equation.", 3. "Consider a scenario where a student is calculating their final grade based on a new test score, with realistic averages." Checking threshold...
https://arxiv.org/abs/2505.20184v1
obtain scores (PS, CS )and feedback IS 6: Compute PR(pi),AA(pi),AC(pi), and composite quality score Q(pi) 7: ifQ(pi)> τthen 8: success ←True 9: else 10: ▷Refine via feedback →Think-aloud 11: pi←LLM (pi, IS) 12: r←r+ 1 13: end if 14: end while 15: Add final version of pitoDimproved 16: Add final Q(pi)toQfinal 17:end for...
https://arxiv.org/abs/2505.20184v1
Understanding Applying Analyzing Evaluating Creating Avg. GPT-4 O 86.92↑26.92 82.96↑5.79 76.71↓0.46 83.50↑4.21 83.54↑2.92 82.62↑4.21 82.71↑3.51 GPT-4 O-MINI 85.21↑15.12 82.96↑0.71 74.50 ↓5.88 82.88 ↓1.38 83.08 ↓1.88 82.42 ↓0.54 81.51 ↑1.91 GPT-3.5- TURBO 82.29↑12.96 81.25 ↑1.29 71.83 ↓6.12 81.12 ↑0.54 80.92 ↓0.54 80.25...
https://arxiv.org/abs/2505.20184v1
tasks such as Remembering andUnderstanding , indicat- ing strong capabilities in information recall and paraphrasing. However, there is a marked per- formance drop in the Applying category, which requires transferring learned concepts to a real- world scenario. Nearly all models exhibit degra- dation in this dimension,...
https://arxiv.org/abs/2505.20184v1
revision. However, this efficiency often comes at the cost of lower final quality scores, reflecting a trade-off between revi- sion speed and output quality. These findings sug- gest that while smaller models may adapt feedback quicker, larger models exhibit a greater capacity for sustained, high-quality refinement. Fe...
https://arxiv.org/abs/2505.20184v1
reasoning quality. Qualitative analysis confirms that THINK-guided outputs exhibit deeper conceptual alignment and domain fidelity. By making models “think-aloud” through itera- tive critique, THINKoffers a scalable, principled approach for the community to both measure and advance LLM cognition, paving the way for mor...
https://arxiv.org/abs/2505.20184v1
language models towards more human-like behavior in out-of-distribution reasoning tasks. arXiv preprint arXiv:2205.05718 . Ioan Fazey. 2010. Resilience and higher order thinking. Ecology and Society , 15(3). Jennifer Haase, Paul H. P. Hanel, and Sebastian Pokutta. 2025. Has the creativity of large-language models peake...
https://arxiv.org/abs/2505.20184v1
Luke Zettlemoyer, Sumit Gulwani, and Zo- ran Popovic. 2015. Personalized mathematical word problem generation. In IJCAI , pages 381–388. Ben Prystawski, Paul Thibodeau, Christopher Potts, and Noah D Goodman. 2022. Psychologically- informed chain-of-thought prompts for metaphor un- derstanding in large language models. ...
https://arxiv.org/abs/2505.20184v1
and Learning , 13(2):181–188. Xiong Xiao, Yue Li, Xiuling He, Jing Fang, Zhonghua Yan, and Chong Xie. 2025. An assessment frame- work of higher-order thinking skills based on fine- tuned large language models. Expert Systems with Applications , page 126531. Blaise Agüera y Arcas. 2022. Do large language models understa...
https://arxiv.org/abs/2505.20184v1
4.Alternative Values : This refers to variations in the input parameters of a problem that preserve its underlying structure. A model’s ability to adapt its reasoning across such variants reflects generalization ability—an essential attribute of HOT. 5.Narrative Stories : Embedding problems in real-world or socio-cultu...
https://arxiv.org/abs/2505.20184v1
question text", "LaTeX question": "The question text with LaTeX formatting", "solution": "Explanation of why the question is flawed and what information is missing or ambiguous", "mathConcept1": "Main math concept (e.g., Arithmetic and Algebra)", "mathConcept2": "Sub-concept (e.g., Algebraic expressions)", "mathConcept...
https://arxiv.org/abs/2505.20184v1
Levels of Understanding. - Strong Understanding (80–100): Demonstrates a deep grasp of the five components, identifying at least three operations among the seven. - Medium Understanding (60–80): Reflects surface-level changes, identifying at least one operation among the seven. - Low Understanding (<60): Shows minimal ...
https://arxiv.org/abs/2505.20184v1
You are an expert in math and reasoning, acting as a refiner and evaluator, to assess the "Evaluating" level skills of a math problem generator by comparing a newly generated math problem with a previous one. **Evaluation Criteria** Please follow these steps: Step 1: Identify "Big Five" Components: 1) math concepts and...
https://arxiv.org/abs/2505.20184v1
Score - Frazier–Roark Score - Developmental Level - Syntactic Frequency - Mean Dependency Distance (MDD) - Sentence Length **Step 3: Error Identification and Classification** - Ambiguity - Unanswerability - Rationality **Step 4: Solution Strategy Analysis** - One-Step or Multi-Step - Comprehension Challenges from Multi...
https://arxiv.org/abs/2505.20184v1
Monocle: Hybrid Local-Global In-Context Evaluation for Long-Text Generation with Uncertainty-Based Active Learning Xiaorong Wang2*Ting Yang3*Zhu Zhang1*Shuo Wang1† Zihan Zhou4Liner Yang5Zhiyuan Liu1Maosong Sun1† 1Tsinghua University2Beijing Jiaotong University 3Beijing University of Posts and Telecommunications 4Xiamen...
https://arxiv.org/abs/2505.20195v2
HelloEval also incorporates human annotations to enhance the reliability of the scores. However, the reported correlation coefficient between HelloEval and hu- man annotations remains relatively low (0.3193, as noted in Que et al. (2024)), highlighting that the evaluation of long-form generation is still an ongoing cha...
https://arxiv.org/abs/2505.20195v2
to compare the performance of various evaluation meth- ods, with experimental results demonstrating the superiority of our approach. 2 Related Works Traditional Evaluation Methods Conventional metrics for text generation tasks primarily rely on reference-based evaluation, which requires human- annotated ground-truth te...
https://arxiv.org/abs/2505.20195v2
evaluation frame- works (Gu et al., 2025). In-Context Evaluation Due to the inherent ran- domness in the generation process of modern LLMs (Touvron et al., 2023a,b; Grattafiori et al.,2024), the judgments made by LLMs can vary across different data samples. Additionally, evalu- ating long-form text is a challenging tas...
https://arxiv.org/abs/2505.20195v2
yjcan be formulated as follows: lj,qj=flocal(xj,yj,Dlocal), (5) where flocalrepresents the local judge driven by LLMs. As shown in Figure 1, the local judge gen- erates both the local score ljand the corresponding local explanation qj, which will be aggregated in the subsequent global evaluation stage. Global In-contex...
https://arxiv.org/abs/2505.20195v2
illustrated in Figure 2, we quantify this uncertainty by repeat- edly applying the proposed Monocle framework for evaluation. We then use the standard deviation of the local and global scores to represent the local and global uncertainties, respectively. Quantifying model uncertainty through repeated sampling is a wide...
https://arxiv.org/abs/2505.20195v2
LaTex format. Writing Instructions To generate high-quality writing instructions that are closely aligned with the references, we follow a two-step process: back- translation and human verification. Please refer to Section B in Appendix for more details. Model Responses In accordance with human- verified writing instru...
https://arxiv.org/abs/2505.20195v2
correlation with human assessments (0.550). These results demonstrate that our approach not only leverages LLMs more effectively for long-form evaluation but also makes better use of human annotations. Demonstration ρ(↑) p(↓) None 0.351 5.96 ×10−4 Local 0.479 1.28 ×10−6 Local + Global 0.550 1.34×10−8 Table 2: Effect of...
https://arxiv.org/abs/2505.20195v2
long- form generation task. experiments, we will select Claude-3.5-Sonnet as the backbone model for both our method and the HelloEval (Que et al., 2024) baseline. 5.3 Main Results We compare our method with several representative and widely-used baselines, of which the implemen- tation details are included in Section E...
https://arxiv.org/abs/2505.20195v2
Christophe Ropers, Alexandre Mourachko, Ed- uardo Sánchez, Holger Schwenk, Tuan Tran, Arina Turkatenko, and Carleigh Wood. 2024. Lcfo: Long context and long form output dataset and benchmark- ing.Preprint , arXiv:2412.08268. Yao Fu, Rameswar Panda, Xinyao Niu, Xiang Yue, Han- naneh Hajishirzi, Yoon Kim, and Hao Peng. 2...
https://arxiv.org/abs/2505.20195v2
Liqun He, Yutao Mou, Wangchunshu Zhou, Jiaheng Liu, Wenge Rong, Zekun Moore Wang, Jian Yang, Ge Zhang, Junran Peng, Zhaoxiang Zhang, Songyang Zhang, and Kai Chen. 2024. Hellobench: Evaluating long text gener- ation capabilities of large language models. Preprint , arXiv:2409.16191. Thibault Sellam, Dipanjan Das, and An...
https://arxiv.org/abs/2505.20195v2
Quantifying uncertainties in natural language processing tasks. In Proceedings of AAAI 2019 . Wenhan Xiong, Jingyu Liu, Igor Molybog, Hejia Zhang, Prajjwal Bhargava, Rui Hou, Louis Martin, Rashi Rungta, Karthik Abinav Sankararaman, Barlas Oguz, Madian Khabsa, Han Fang, Yashar Mehdad, Sharan Narang, Kshitiz Malik, Angel...
https://arxiv.org/abs/2505.20195v2
ated responses, we utilized six different models, including both closed- and open-source options: GPT-4o, GPT-4o-mini, GPT-o1-mini, Claude-3.5- Sonnet, LLaMA-3.1-8B, and Qwen-2.5-7B. In the final dataset, 80% of the data was generated using AgentWrite, with the remaining 20% generated via standard decoding. D Details o...
https://arxiv.org/abs/2505.20195v2
for both HelloEval and Monocle. F Analysis of the Number of Demonstrations Figure 6 analyzes the relationship between the num- ber of shots and the Spearman’s rank correlation 4https://github.com/mjpost/sacrebleu 5https://pypi.org/project/rouge-score 6https://www.nltk.org/ 7https://github.com/mjpost/sacrebleu 8https://...
https://arxiv.org/abs/2505.20195v2
content, it may lack sufficient detail, depth, or clarity. Minor gaps in structure, logic, or formatting reduce its academic quality, requiring improvement. - **3**: The generated paper’s section is of high academic quality, with only minor issues such as small gaps in detail, slight formatting inconsistencies, or mino...
https://arxiv.org/abs/2505.20195v2
academic standards and fails to meet even the basic requirements, rendering it unfit for practical use. - **1**: The generated paper’s section has significant issues that need substantial improvement. It is mostly incomplete or lacks depth and clarity, with noticeable errors, omissions, or rough explanations that hinde...
https://arxiv.org/abs/2505.20195v2
without oversimplification? 4. **Logic and Coherence**: - Is the reasoning and logic flow within the section clear and well-structured? - Are the paragraphs within the section smoothly connected? 5. **Language and Expression**: - Is formal academic language used consistently, avoiding casual expressions? - Is the langu...
https://arxiv.org/abs/2505.20195v2
score of the paper, you should take into account both the local scores for each group and the global quality features. First, I will provide you with N human evaluation examples. Each example contains: 1. **Group-Level Scores**: Human ratings for each group across five evaluation dimensions (on a scale of 0-4). Review ...
https://arxiv.org/abs/2505.20195v2
arXiv:2505.20199v1 [cs.CL] 26 May 2025Adaptive Classifier-Free Guidance via Dynamic Low-Confidence Masking Pengxiang Li∗1, Shilin Yan∗♠2, Joey Tsai3, Renrui Zhang4, Ruichuan An5, Ziyu Guo4, Xiaowei Gao6 1PolyU2FDU3THU4CUHK5PKU6ICL {2040gis, tattoo.ysl} @gmail.com ∗Equal Contribution♠Project Leader Abstract Classifier-F...
https://arxiv.org/abs/2505.20199v1
conventional application of CFG within iterative (masked) diffusion language models often encounters a subtle yet significant limitation: the "unconditional" prediction typically relies on a static orgeneric construct. This often involves using a null prompt or a sequence where all target tokens are uniformly masked to...
https://arxiv.org/abs/2505.20199v1
more recent powerful open-source models including LLaMA [ 35,36], Qwen [ 6], and Mistral [ 18], have become the dominant paradigm in natural language generation. These models generate text token by token, conditioning each new token on the previously generated sequence, and have demonstrated remarkable capabilities acr...
https://arxiv.org/abs/2505.20199v1
of the sequence where the model exhibits uncertainty. Figure 2 provides a high-level comparison of standard CFG with our proposed A-CFG. 3 Adaptive Classifier-Free Guidance (A-CFG) Base Model Conditional Logit Computation Token-Level Confidence Assessment Token Confidence Low-Confidence Masking Candidates Prompt tokens Hig...
https://arxiv.org/abs/2505.20199v1
a static or generic definition for the unconditional prediction Luncond (x(k),∅)when applied to iterative MLMs. Typically, this involves using a null 4 Algorithm 1 Adaptive Classifier-Free Guidance (A-CFG) for one generation step k 1:Input: Current sequence x(k), conditioning c, model Mθ, guidance w, re-masking proport...
https://arxiv.org/abs/2505.20199v1
to the probability of the token that the model would predict with highest likelihood for position jbased on L(k) cond. A low c(k) jsuggests the model is uncertain about the token (x(k))jor its alternatives at that position. While other confidence metrics (e.g., entropy of P(k) cond,j) could be considered, we find that ...
https://arxiv.org/abs/2505.20199v1
Guidance (A-CFG). We first describe our experimental setup, including datasets, baseline models, evaluation metrics, and key implementation details. We then present quantitative results from Table 1, comparing LLaDA with A-CFG against LLaDA with standard CFG, LLaDA without guidance, and other state- of-the-art models. ...
https://arxiv.org/abs/2505.20199v1
53.1 (0) 46.3 (0) 60.6 (25) Hellaswag 70.8 (0) 71.4 (0) 72.6 (0) 73.3 (0) 74.4 (0) 79.1 (0) 76.0 (0) 80.7 (10) TruthfulQA 45.5 (0) 45.1 (0) 46.2 (0) 43.9 (0) 45.1 (0) 44.0 (0) 39.0 (0) 54.2 (0) WinoGrande 74.5 (5) 75.1 (5) 75.9 (5) 73.3 (5) 72.5 (5) 77.3 (5) 72.5 (5) 77.0 (5) PIQA 74.9 (0) 74.4 (0) 76.1 (0) 75.8 (0) 76...
https://arxiv.org/abs/2505.20199v1
by +8.0 points (80.0 vs. 72.0) and ARC-C by +1.0 point (60.8 vs. 59.8) for Dream-7B. These observations suggest that A-CFG’s adaptive unconditioning is a promising method for enhancing other iterative masked diffusion models. 8 Competitive Standing Against Autoregressive Models: Equipped with A-CFG, the diffusion-based...
https://arxiv.org/abs/2505.20199v1
However, further increasing wto1.5and2.0leads to a slight degradation in performance (47.5% and 47.6%, respectively). This indicates that an overly strong guidance scale might overemphasize the conditional signal at the expense of fluency or correctness, even with A-CFG’s targeted unconditioning. The optimal performanc...
https://arxiv.org/abs/2505.20199v1
generation in iterative masked language models. By dynamically constructing the unconditional input for CFG based on the model’s instantaneous predictive confidence in its already generated tokens, A-CFG offers a more targeted and responsive guidance mechanism. Our extensive experiments, particularly within the LLaDA f...
https://arxiv.org/abs/2505.20199v1
models. Advances in neural information processing systems , 33:6840–6851, 2020. [16] Jack Hong, Shilin Yan, Jiayin Cai, Xiaolong Jiang, Yao Hu, and Weidi Xie. Worldsense: Evaluat- ing real-world omnimodal understanding for multimodal llms. arXiv preprint arXiv:2502.04326 , 2025. [17] Lingyi Hong, Shilin Yan, Renrui Zha...
https://arxiv.org/abs/2505.20199v1
at scale. Communications of the ACM , 64(9):99–106, 2021. [32] Jiaxin Shi, Kehang Han, Zhe Wang, Arnaud Doucet, and Michalis K Titsias. Simplified and generalized masked diffusion for discrete data. arXiv preprint arXiv:2406.04329 , 2024. [33] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. ...
https://arxiv.org/abs/2505.20199v1
Reasoning Is Not All You Need: Examining LLMs for Multi-Turn Mental Health Conversations Mohit Chandra1*, Siddharth Sriraman1*, Harneet Singh Khanuja1, Yiqiao Jin1, Munmun De Choudhury1 1Georgia Institute of Technology {mchandra9,sidsr,hkhanuja3,yjin328}@gatech.edu,munmun.choudhury@cc.gatech.edu Abstract Limited access...
https://arxiv.org/abs/2505.20201v2
supporting multi-turn sensemak- ing conversations with the users through a more holistic approach taking account of patient-centric objective, values and needs. A major challenge in this direction is the limited to no access to real- world patient–clinician mental-health related con- 1Sensemaking refers to “ processes ...
https://arxiv.org/abs/2505.20201v2
MedAgent framework generates men- tal health related sensemaking conversations using two actor LLMs: (1) Patient Actor LLM, and (2) Sensemaker Actor LLM. We use GPT-4o (OpenAI- GPT-4o, 2024) for the patient actor LLM, and fron- tier reasoning models, OpenAI o1 (OpenAI, 2024) and DeepSeek-R1 (Guo et al., 2025) respectiv...
https://arxiv.org/abs/2505.20201v2
with inadequate health literacy have been observed to have difficulty in participating in medical decision-making process, following medical recommendations and conversa- tions (Williams et al., 1995). To account for such variations, we introduced two health literacy levels (Basic ,Advanced ) for the patient persona (d...
https://arxiv.org/abs/2505.20201v2
For a given timestamp t, we first retrieve the relevant facts from the patient’s atomic fact list Fthat could be used to generate the reply using a fact matching and retrieval prompt (Ap- pendix Table 20). If the number of facts matched (|Fmatch|>0) then we use the patient message generation module to generate the fina...
https://arxiv.org/abs/2505.20201v2
and health outcomes. This design ensures that the sensemaker LLM is able to meaningfully engage in a multi-turn conversation. 4 Algorithm 2 Sensemaker Message Generation at Turnt Input: patient message pt−1, known facts Ft−1, sensemaker message history H Output: Next sense -maker utterance st, updated facts Ft, stage i...
https://arxiv.org/abs/2505.20201v2
1.22 Gathering Information 7.64 3.88 Providing Information 2.00 2.00 Decision Making 3.52 2.70 Responding to Emotions 3.83 2.86 Exit 1.00 1.00 Overall 19.09 13.65 Table 1: Stage-wise and overall average number of messages of the MHSD dataset with OpenAI o1 and DeepSeek-R1 as sensemaker, including both patient and sense...
https://arxiv.org/abs/2505.20201v2