text string | source string |
|---|---|
generalizability. Although researchers have acknowledged this challenge (A. and V ., 2024; Ghaayathri Devi et al., 2024), MER systems aiming for global applicability must ac- count for both linguistic diversity and culturally driven display rules. Complexities of Fusion Strategies across Modal- ities. Multimodal fusion... | https://arxiv.org/abs/2505.20511v1 |
machines to learn across modalities. Some models are increasingly used in MERC, offering zero-shot or few-shot gen- eralization across different modalities (Li et al., 2024c; Yang et al., 2024; Bo-Hao et al., 2025). The use of LLMs in MERC opens up new possibili- ties for capturing deeper semantic and conversa- tional ... | https://arxiv.org/abs/2505.20511v1 |
et al. (2022), Sasu et al. (2025) and Gan et al. (2024). Finally, while we identify several open chal- lenges and underexplored directions, our discus- sion is not exhaustive. Rather than providing defini- tive answers, we aim to surface critical issues and foster further inquiry. We view these open ques- tions as prod... | https://arxiv.org/abs/2505.20511v1 |
analysis toolkit. In 2016 IEEE Winter Con- ference on Applications of Computer Vision (WACV) , pages 1–10. Su Bo-Hao, Shreya G. Upadhyay, and Lee Chi-Chun. 2025. Toward zero-shot speech emotion recogni- tion using llms in the absence of target data. In ICASSP 2025 - 2025 IEEE International Confer- ence on Acoustics, Sp... | https://arxiv.org/abs/2505.20511v1 |
Proceedings of the 31st International Con- ference on Computational Linguistics , pages 6748– 6761, Abu Dhabi, UAE. Association for Computa- tional Linguistics. Yumeng Fu, Junjie Wu, Zhongjie Wang, Meishan Zhang, Yulin Wu, and Bingquan Liu. 2025b. Bemerc: Behavior-aware mllm-based framework for multi- modal emotion rec... | https://arxiv.org/abs/2505.20511v1 |
tional Linguistics.Yifei He, Runxiang Cheng, Gargi Balasubramaniam, Yao-Hung Hubert Tsai, and Han Zhao. 2024. Ef- ficient modality selection in multimodal learning. Journal of Machine Learning Research , 25(47):1–39. Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. Neural Computation , page 1735–17... | https://arxiv.org/abs/2505.20511v1 |
vol- ume 34, pages 8002–8009, Palo Alto, CA, USA. AAAI Press. Umair Ali Khan, Qianru Xu, Yang Liu, Altti Lagstedt, Ari Alamäki, and Janne Kauttonen. 2024. Explor- ing contactless techniques in multimodal emotion recognition: insights into diverse applications, chal- lenges, solutions, and prospects. Multimedia Systems ... | https://arxiv.org/abs/2505.20511v1 |
Springer Nature Singapore. Nian Liu, Xiao Wang, Deyu Bo, Chuan Shi, and Jian Pei. 2022. Revisiting graph contrastive learning from the perspective of graph spectrum. In Advances in Neural Information Processing Systems , volume 35, pages 2972–2983. Curran Associates, Inc. 12 Yinhan Liu, Myle Ott, Naman Goyal, Jingfei D... | https://arxiv.org/abs/2505.20511v1 |
predictive coding. arXiv preprint arXiv:1807.03748 . Sancheng Peng, Lihong Cao, Yongmei Zhou, Zhouhao Ouyang, Aimin Yang, Xinguang Li, Weijia Jia, and Shui Yu. 2022. A survey on deep learning for tex- tual emotion analysis in social networks. Digital Communications and Networks , 8(5):745–762. Soujanya Poria, Erik Camb... | https://arxiv.org/abs/2505.20511v1 |
arXiv:2407.00119 . Yuntao Shou, Tao Meng, Wei Ai, and Keqin Li. 2025. Dynamic graph neural ODE network for multi-modal emotion recognition in conversation. In Proceedings of the 31st International Conference on Computa- tional Linguistics , pages 256–268, Abu Dhabi, UAE. Association for Computational Linguistics. Lin S... | https://arxiv.org/abs/2505.20511v1 |
Heringa, Peter A. C. ’t Hoen, Rob Hooft, Tobias Kuhn, Ruben Kok, Joost Kok, Scott J. Lusher, Maryann E. Martone, Albert Mons, Abel L. Packer, Bengt Persson, Philippe Rocca-Serra, Marco Roos, Rene van Schaik, Susanna-Assunta Sansone, and Erik Schultes. 2016. The fair guiding princi- ples for scientific data management a... | https://arxiv.org/abs/2505.20511v1 |
recognition in dynamic data using facial, speech and textual cues. Multimedia Tools and Applications , 83(25):66223–66262. Xiaoheng Zhang and Yang Li. 2023. A cross-modality context fusion and semantic refinement network for emotion recognition in conversation. In Proceedings of the 61st Annual Meeting of the Associati... | https://arxiv.org/abs/2505.20511v1 |
is annotated with one or more of seven emotion labels: happy, surprise, sad, disgust, anger, fear, and neutral. The dataset covers text, audio, and visual modalities and features blended emo- tions and speaker metadata, making it the first large-scale multimodal emotional dialogue corpus in Chinese. ACE. The ACE datase... | https://arxiv.org/abs/2505.20511v1 |
arXiv:2505.20521v1 [cs.AI] 26 May 2025Project Riley: Multimodal Multi-Agent LLM Collaboration with Emotional Reasoning and Voting Ana Rita Ortigoso1, Gabriel Vieira1, Daniel Fuentes1, Luis Fraz˜ ao1, Nuno Costa1, and Ant´ onio Pereira1 1Computer Science and Communication Research Centre, Polytechnic University of Leiri... | https://arxiv.org/abs/2505.20521v1 |
awareness, multi model functionality, and the implementation of so- phisticated self-refining and voting logic. Project Riley integrates multi- modal information—text and images—through the coordinated use of mul- tiple Large Language Models (LLM), each representing a distinct emotion. It processes user input using bot... | https://arxiv.org/abs/2505.20521v1 |
and coher- ence. Liu et al., conversely, integrated emotional and visual modalities using Emotional Retrieval Module(ERM), Response Emotion Prediction (REP), and Emotion-Enhanced Response Generation (EERG) models. Their mul- timodal ELMD system demonstrated superior emotional and contextual re- sponsiveness compared to... | https://arxiv.org/abs/2505.20521v1 |
architecture. 3 Proposed Architecture This paper presents an innovative conversational AI system architecture that processes user queries through distinct emotional lenses, specifically a set of five basic emotions: Anger, Joy, Sadness, Fear, and Disgust. The system fol- lows the emotional framework portrayed in the fi... | https://arxiv.org/abs/2505.20521v1 |
Each emotion see others emotion responsesEach emotion responds with a final perspectiveRound 2 Final emotions answer , incorporating learning from the discussionRound 3Voting and Analysis Analysis of all answers Vote and justification for the choiceFinal Synthesis Tie?Reasoning considering the winner emotion response R... | https://arxiv.org/abs/2505.20521v1 |
the huihui ai/deepseek-r1-abliterated:8b [30] model was utilised, containing 8.03 billion parameters and a model size of 4.9 GB. This selection of models was designed to maximise the likelihood that they remain loaded in GPU VRAM, thereby reducing the need for frequent model switching and minimising latency between the... | https://arxiv.org/abs/2505.20521v1 |
, the alignment of system-selected emotions with user expectations was strongest in Family or Intergenerational Conflicts (4.35). Breakups or Friendship Loss remained the lowest scoring (3.94), consistently reflecting lower emotional alignment. Clarity and Utility evaluations, depicted in Figure 3, highlighted high cla... | https://arxiv.org/abs/2505.20521v1 |
developed, a chatbot named Armando that has been developed specifically to assist during emer- gency response scenarios. This AI assistant aims to provide citizens with emotionally-aware, human-like interactions, delivering responses grounded not only in officially validated information provided by authoritative source... | https://arxiv.org/abs/2505.20521v1 |
the emotional dialogue. Therefore, RAG enrichment was instead positioned just before the synthesis stage, ensuring that the final response maintains emo- tional balance and contextual sensitivity, while also conveying accurate and up-to-date information to the user. The models used in this solution are the same that ha... | https://arxiv.org/abs/2505.20521v1 |
cleaner. – Emergency services are working diligently to resolve the situation. Please stay calm, follow these instructions, and prioritise your safety. Further updates will follow. Analysing Table 4, the non-RAG model lacked specific contextual in- formation regarding the ongoing incident, resulting in a generic respon... | https://arxiv.org/abs/2505.20521v1 |
•Check doors before opening them – Use the back of your hand to feel if a door is hot. If it’s not, open the door cau- tiously – the fire may be on the other side. •Once outside, go to the designated assembly point and report your presence – Wait for further instructions from emergency personnel. •Stay low to the groun... | https://arxiv.org/abs/2505.20521v1 |
Table 6, the response without RAG provided generic emergency guidance and suggested calling 911 or another local emergency number without explicitly identifying the correct emergency number for Por- tugal. The RAG-enhanced response correctly identified the emergency num- ber applicable in Portugal (112) but included ad... | https://arxiv.org/abs/2505.20521v1 |
limita- tions in achieving human-like conversational expression. Particularly, Anxi- ety in Academic or Professional Contexts scored lowest (3.59), emphasising the necessity to enhance naturalness and linguistic subtlety in emotionally nuanced interactions. Analysis of open-text responses regarding predominant emotions... | https://arxiv.org/abs/2505.20521v1 |
be adapted according to the type of situation addressed, allowing for more immersive and context-sensitive responses. Fi- nally, in the context of emergency communication, it is crucial to establish a structured syntax and annotation protocol for RAG input data, to ensure more accurate retrieval and delivery of critica... | https://arxiv.org/abs/2505.20521v1 |
approach to diversify llm- based multi-agent collective decision-making, in: Proceedings of the 2024 Conference on Empirical Methods in Natural Language Process- ing, Association for Computational Linguistics, 2024, pp. 2712–2727. doi:10.18653/v1/2024.emnlp-main.158 . [15] Y. Yang, Y. Ma, H. Feng, Y. Cheng, Z. Han, Min... | https://arxiv.org/abs/2505.20521v1 |
arXiv:2505.20522v1 [cs.AI] 26 May 2025Scaling over Scaling: Exploring Test-Time Scaling Pareto in Large Reasoning Models Jian Wang, Boyan Zhu, Chak Tou Leong, Yongqi Li∗, Wenjie Li Department of Computing, The Hong Kong Polytechnic University jian51.wang@polyu.edu.hk {boyan.zhu,chak-tou.leong}@connect.polyu.hk liyongqi... | https://arxiv.org/abs/2505.20522v1 |
26], where multiple reasoning paths and solutions are generated independently; and 2) sequential scaling [ 24], where a solution is iteratively refined round by round. Their performance gains with increasing compute often exhibit a characteristic saturation curve. We develop a general probabilistic model that captures ... | https://arxiv.org/abs/2505.20522v1 |
Tree of Thoughts (ToT) [ 30] explores reasoning as a tree search, where multiple reasoning paths are explored in parallel at each step. While parallel methods effectively broaden the search space, they often involve redundant computations as paths are generated independently without collaboration. Majority voting in SC... | https://arxiv.org/abs/2505.20522v1 |
not limitless, we aim to build a theoretical Test-TimeScaling Performance Model ( TTSPM ) to quantify reasoning performance gain varying by scaling budget. TTSPM captures the saturation behavior common to various test-time scaling strategies, despite their differing underlying mechanisms. From this model, we derive a g... | https://arxiv.org/abs/2505.20522v1 |
events is P(K=k) = (1 −prethink)k−1prethink . This indicates that Kfollows a geometric distribution with success parameter prethink . We are primarily interested in the cumulative probability of achieving a correct answer within a maximum of Nrethinking rounds, denoted P(K≤N). This is the sum of probabilities of first ... | https://arxiv.org/abs/2505.20522v1 |
boost, and thus Nupper could be considered to be 0 or 1. Assuming ϵ < F max·px, we take the natural logarithm of both sides of Eq. (11): Nln(1−px)<lnϵ Fmax·px . (12) Given that 0< px<1, it follows that 0<1−px<1, which makes ln(1−px)a negative value. Therefore, when dividing by ln(1−px), the direction of the inequalit... | https://arxiv.org/abs/2505.20522v1 |
AIME 2024 [11] and AIME 20252: Each contains 30 pre-Olympiad level problems from the American Invitational Mathematics Examination, designed to test advanced mathematical reasoning. MATH-500 [5]: A challenging subset of the MATH dataset with 500 high school competition prob- lems across algebra, geometry, and so on. GP... | https://arxiv.org/abs/2505.20522v1 |
53.3%) and percentage points on GPQA (from 21.3% to 39.4%). Similarly, the larger 7B model shows substantial gains, with improvements of up to 20.0 percentage points on AIME 2024 (from 56.7% to 76.7%) and 14.7 percentage points on GPQA (from 41.9% to 56.6%). Second, parallel scaling consistently outperforms sequential ... | https://arxiv.org/abs/2505.20522v1 |
improvements, indicating that the problems in this benchmark may require more focused refinement rather than broad exploration. Further accuracy scaling curves presented in Appendix C demonstrate similar insights. Figure 3 outlines the statistics of the number of generations per problem required to reach saturation poi... | https://arxiv.org/abs/2505.20522v1 |
the 7B model with parallel scaling (Figure 4(d)) shows the highest correlation, indicating that larger, more stable models may be more predictable in their scaling behavior. This finding has important implications for resource allocation in large-scale deployments, where accurate prediction of scaling benefits can lead... | https://arxiv.org/abs/2505.20522v1 |
Song, and Jacob Steinhardt. Measuring mathematical problem solving with the math dataset. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2) , 2021. [6]Can Jin, Hongwu Peng, Qixin Zhang, Yujin Tang, Dimitris N Metaxas, and Tong Che. Two heads are better than one:... | https://arxiv.org/abs/2505.20522v1 |
Hu, Yancheng Pan, and Shaoxun Wang. Adaptive rectification sampling for test-time compute scaling. arXiv preprint arXiv:2504.01317 , 2025. [24] Xiaoyu Tian, Sitong Zhao, Haotian Wang, Shuaiting Chen, Yunjie Ji, Yiping Peng, Han Zhao, and Xiangang Li. Think twice: Enhancing llm reasoning by scaling multi-round test-time... | https://arxiv.org/abs/2505.20522v1 |
with our theoretical model, where p(x)represents the single-step success probability in the Markov process. For problems not solved within 32 rounds, we assign a small non-zero probability to avoid mathematical issues in subsequent logarithmic calculations. These estimations are performed for every problem in both the ... | https://arxiv.org/abs/2505.20522v1 |
the initial state is justified if one strictly adheres to the ϵ threshold. Therefore, the practical application of the N∗formula to find a value greater than 1 assumes that the initial potential gain Fmax·pxis greater than the desired negligible gain threshold ϵ. This ensures that there is at least some initial phase w... | https://arxiv.org/abs/2505.20522v1 |
ASTRO VISBENCH : A Code Benchmark for Scientific Computing and Visualization in Astronomy Sebastian Joseph1, Syed Murtaza Husain1, Stella S. R. Offner1, Stéphanie Juneau2, Paul Torrey3, Adam S. Bolton4, Juan P. Farias1, Niall Gaffney5, Greg Durrett1, Junyi Jessy Li1 1The University of Texas at Austin 2NSF National Opti... | https://arxiv.org/abs/2505.20538v2 |
any potential faint sources by careful manipulation of image scaling. The objective is to produce a visually distinct image through which we're able to discern and identify the physical characteristics defined within the PSF model. […]… ref_cutout = extract_array(psfs[0].data, (41, 41), (122, 122)) norm1 = simple_norm(r... | https://arxiv.org/abs/2505.20538v2 |
light [ 9]. These charts vary substantially from those in standard chart understanding datasets from the machine learning literature [ 26,36,3,40]. More critically, the evaluation of LLMs for scientific visualizations has not assessed their performance at the end-to-end process of producing visualizations. ASTRO VISBEN... | https://arxiv.org/abs/2505.20538v2 |
contains code that produces a visualization when executed, and the previous cells c1, ..., c k−1are extracted dependencies of ck(i.e., they are not necessarily consecutive in the original notebook, see Figure 2). We define a visualization pipeline as a set of tasks T= (tprocess, tvisualize )given setup code (such as im... | https://arxiv.org/abs/2505.20538v2 |
aware of. Our team of expert astronomers verified a subset of the generated queries while inspecting the generated visualizations (Section 3.2.1), in which they confirm that these queries do not leak such information and represent typical research queries. The natural language queries also need to be specific enough to... | https://arxiv.org/abs/2505.20538v2 |
the ground truth visualization. The ideal judges for this task are astronomy researchers. However, human expert evaluation is not scalable and is therefore impractical for a benchmark. Instead, we automate this process by deploying a VLM as a judge to compare the generated visualization to the ground truth (Section 3.2... | https://arxiv.org/abs/2505.20538v2 |
0.753 0.775 Claude 3.7 Sonnet 0.723 0.714 Claude 3.5 Sonnet 0.822 0.828 Claude 3.5 Haiku 0.749 0.586 Table 2: The Spearman correlations be- tween vLLM judges and expert judges (averaged scores or majority labels). Correlations are significant ( p <1e−29).We use a VLM as the automatic evaluator for visualiza- tions crea... | https://arxiv.org/abs/2505.20538v2 |
function correction, (b) a spatially integrated spectral energy distribution (left) and associated spatially resolved intensity map (right), (c) a wide-field all-sky projection of galaxy source counts within a survey footprint, (d) a Kepler mission source pixel map, (e) a pixel-level flux map, (f) galaxy spectra featur... | https://arxiv.org/abs/2505.20538v2 |
2.5 Pro again performs the best. The absolute percentages of code that executes are higher than that in processing tasks. This is expected as the models are exposed to more context in the visualization stage, and these sub-tasks usually involve the usage of well-known visualization libraries (e.g., matplotlib ), unlike... | https://arxiv.org/abs/2505.20538v2 |
in Appendix D). First, the experts noticed that the generated visualizations frequently overlook domain-specific plotting conventions, such as the order of the sky coordinates or magnitudes on the axes. Incorrect domain-specific conventions are also evident in an “inappropriate” choice of axis scale. Another issue flag... | https://arxiv.org/abs/2505.20538v2 |
uses an execution-based, automatic evaluation method that directly evaluates a visualization as produced from executing LLM-generated code, using a combination of rule-based methods and a VLM. However, they only assess readability, while we focus on scientific utility informed by professional astronomers. Finally, prio... | https://arxiv.org/abs/2505.20538v2 |
Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian, Clemens Winter, Philippe Tillet, Felipe Petroski Such, Dave Cummings, Matthias Plappert, Fotios Chantzis, El... | https://arxiv.org/abs/2505.20538v2 |
Experimentation. arXiv preprint arxiv:2503.22708 , 2025. URL https://api.semanticscholar.org/CorpusID: 277451644 . [14] Carlos E. Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik Narasimhan. SWE-bench: Can Language Models Resolve Real-World GitHub Issues?, 2024. URL https://arxiv.org... | https://arxiv.org/abs/2505.20538v2 |
Hashimoto. Can LLMs generate novel research ideas? A large-scale human study with 100+ NLP researchers. arXiv preprint arXiv:2409.04109 , 2024. [32] Giulio Starace, Oliver Jaffe, Dane Sherburn, James Aung, Jun Shern Chan, Leon Maksin, Rachel Dias, Evan Mays, Benjamin Kinsella, Wyatt Thompson, et al. PaperBench: Evaluat... | https://arxiv.org/abs/2505.20538v2 |
an example of how to return judgments and rationales. B.1 Instructions to VLM for error judgments in automatic visualization evaluation System: Your task is to evaluate the correctness and visual validity of the under-test data visualization related to astronomy that will be sent to you. You will return either "No Erro... | https://arxiv.org/abs/2505.20538v2 |
this prompt is then tested to ensure if the code present in it matches with the ground truth code from the notebook cells. If the output fails this test then new outputs will be regenerated until either that output is a match for the ground truth code or when the maximum number of regenerations is reached. Only query-c... | https://arxiv.org/abs/2505.20538v2 |
the EXACT code from the provided jupyter notebook it corresponds to. Any kind of deviation is absolutely intolerable. - If the original code cells are empty, do not bother writing anything down in these code cells. - This can be checked by seeing whether all the code you have written combined is equivalent to all the c... | https://arxiv.org/abs/2505.20538v2 |
for more advanced coding capabilities. •Qwen-2.5 (72B) is the strongest open-source LLM at its size available at the time of this writing. •Llama-4 Maverick (17Bx128E) is Meta’s leading model, using an MoE architecture, and focuses on multimodality in text and image inputs. •Gemini 2.5 Pro is Google’s most advanced mod... | https://arxiv.org/abs/2505.20538v2 |
arXiv:2505.20546v2 [cs.CL] 28 May 2025Paths Not Taken: Understanding and Mending the Multilingual Factual Recall Pipeline Meng Lu* Brown University meng_lu@brown.eduRuochen Zhang* Brown University ruochen_zhang@brown.edu Carsten Eickhoff University of Tübingen carsten.eickhoff@uni-tuebingen.deEllie Pavlick Brown Univer... | https://arxiv.org/abs/2505.20546v2 |
To address the gap, we make the following contributions: 1.Characterizing the multilingual fact recall pipeline : In Section 2, we integrate and ex- tend the results from prior work and propose a single hypothesized pipeline that is consis- tent with model behavior and intervention. Our analysis shows that factual info... | https://arxiv.org/abs/2505.20546v2 |
u d d h i s m eB u d d h i s m R e l a t i o n P r o p a g a t i o n r e l i g i o nT r a n s l a t i o n t o E n g l i s h S u b j . E n r i c h m e n tT r a n s l a t e A n s w e r t o S p e c i f i c L a n g u a g e s ( a )( b )( c )E a r l y L a y e r sL a t e L a y e r sM i d d l e L a y e r sFigure 1: Hypothesize... | https://arxiv.org/abs/2505.20546v2 |
3% Language-Specific Corr ect CasesJapanese answ ers ar e not z er o-r ankJapanese answ ers ar e z er o- r ankEnglish answ ers ar e z er o-r ankEnglish answ ers ar e z er o-r ankEnglish answ ers ar e not z er o-r ank EnglishF r enchSpanishChineseJapaneseK or eanFigure 2: The bottom bar summarizes model performance on m... | https://arxiv.org/abs/2505.20546v2 |
that it is not the same as the pathway the model uses when prompted to trans- late directly (§3.1). We then show that leveraging the model’s translation pathway leads to significant performance increases (§3.3). 3.1 Translation Mechanism is Insufficiently Used As shown in Section 2, we notice that the model successfull... | https://arxiv.org/abs/2505.20546v2 |
that the difference between the model’s mean residual stream activations when processing fact-recall versus translation prompts can be used to nudge the model to activate a bet- ter translation route. Specifically, for each layer ℓ∈ L where L={21,22,23,24,25,26,27}, we first compute the mean activation ¯h(ℓ) Cfor all f... | https://arxiv.org/abs/2505.20546v2 |
optimal translation components in its fact-recall process. By injecting a single, general- purpose translation signal, we can recover much of the lost performance. 4 Fixing Incorrect English Recall Errors Previously, we show that applying the translation difference vector intervention effectively corrects the conversio... | https://arxiv.org/abs/2505.20546v2 |
as English prompts, yet not as sufficiently. Subsequently, we measure the answer extraction rate, defined as the percentage of instances across layers where the model’s top-ranked decoded token transitions to the correct English answer, indicating successful answer extraction. In Fig. 4(b), we see a consistent increase... | https://arxiv.org/abs/2505.20546v2 |
5 because higher values introduce excessive noise and reduce answer quality. 6 factor of i= 2. Using the best configuration, we observe that this dataset-independent recall vector triggers more relation propagation than ICL examples (Fig- ure 4(a) red line), resulting in a significant boost of successful extraction in ... | https://arxiv.org/abs/2505.20546v2 |
Hase et al., 2023; Chughtai et al., 2024; Yao et al., 2024) and multilingual processing in language models (Conneau et al., 2020; Muller et al., 2021; Wendler et al., 2024; Wu et al., 2024; Schut et al., 2025; Chughtai et al., 2024; Fierro et al., 2025; Zhang et al., 2024; Ferrando and Costa-jussà, 2024; Wilie et al., ... | https://arxiv.org/abs/2505.20546v2 |
can more faithfully reflect model behavior in early layers. In-context learning vs. Interventions In Sec- tion 4, our recall vectors extracted from ICL runs demonstrate positive improvement on multilingual factual recall tasks. However, standard 5-shot ICL outperforms our intervention-based method. This is expected, as... | https://arxiv.org/abs/2505.20546v2 |
. Accessed on: May 20, 2025. Nora Belrose, Zach Furman, Logan Smith, Danny Ha- lawi, Lev McKinney, Igor Ostrovsky, Stella Bider- man, and Jacob Steinhardt. 2023. Eliciting latent predictions from transformers with the tuned lens. to appear . Bilal Chughtai, Alan Cooney, and Neel Nanda. 2024. Summing up the facts: Addit... | https://arxiv.org/abs/2505.20546v2 |
in gpt. Advances in neural information processing systems , 35:17359–17372. Jack Merullo, Carsten Eickhoff, and Ellie Pavlick. 2023. Language models implement simple word2vec-style vector arithmetic. arXiv preprint arXiv:2305.16130 . Jack Merullo, Carsten Eickhoff, and Ellie Pavlick. 2024. Circuit component reuse acros... | https://arxiv.org/abs/2505.20546v2 |
preprint arXiv:2504.04264 . Chris Wendler, Veniamin Veselovsky, Giovanni Monea, and Robert West. 2024. Do llamas work in english? on the latent language of multilingual transformers. Preprint , arXiv:2402.10588. Bryan Wilie, Samuel Cahyawijaya, Junxian He, and Pascale Fung. 2025. High-dimensional interlingual represent... | https://arxiv.org/abs/2505.20546v2 |
token. To address this, we employ a two-stage filtering and scoring process. First, we use spaCy (AI, 2020) to lemmatize the predicted token. Then, we use WordNet (Miller and the Princeton WordNet Group, 1995) to com- pute semantic similarity between the lemmatized predicted token and the English relation token. To- ke... | https://arxiv.org/abs/2505.20546v2 |
ᅲ된ᄃ ᅡ Elefante está clasificado biológicamente como un Table 2: Examples of multilingual prompts for each dataset. 13 Relation Dataset Multilingual Relation Words person_university college, attended / 大学,就读/大学,通った /ᄃ ᅢ학,ᄃ ᅡ녔던/ universidad, asistió / université, étudié country_currency currency / 货币/通貨/ᄒ ᅪᄑ ᅨ/ m... | https://arxiv.org/abs/2505.20546v2 |
of agnostic correctness and final prediction correctness at Layer 21. C Fixing Translation Error C.1 Explicit Translation Dataset Construction We adapt our fact-recall datasets into a translation task dataset. Specifically, we extract each [answer]Category Count (%) Total evaluated 2385 Agnostic correct 867 (36.35%) Ag... | https://arxiv.org/abs/2505.20546v2 |
rect output by the clean model, P∗[o]is the proba- bility under the corrupted input, and P∗,cleanh(ℓ) i[o] is the probability when only component h(ℓ) iis re- stored to its clean state. We compute AIE across all correct instances from both translation and fact-recall datasets, patching into each attention and MLP compo... | https://arxiv.org/abs/2505.20546v2 |
from its original position to the last token position at around layer 15, and translation mechanism starts happening also at layer 15 when the translated answer slowly goes to zero-rank at the very end. Figure 9: Average Indirect Effect of Patching Clean Component into Corrupted Runs. Left: running Activation Patching ... | https://arxiv.org/abs/2505.20546v2 |
simply forwarding the extracted answers from preceding attention layers. To avoid overcounting those as extraction events, we define anextraction event as the first layer ℓat which t′=t∗. This ensures that we only record the earliest point where the correct English attribute is extracted by either the attention or the ... | https://arxiv.org/abs/2505.20546v2 |
layer and scale for the task vector intervention. Specifically, we inject the task vector at all layers (0-5) and vary the scales (1-5). 22 Figure 13: Attribute extraction rate using attention and MLP modules (red and blue respectively) across layers for three conditions (English + Original, Non-English + Original, Non... | https://arxiv.org/abs/2505.20546v2 |
The translate-recall-translate baseline is a multi- step prompting strategy in which we query the model with three separate prompts sequentially: explicitly instructs the model to translate the ques- tion into English, then conduct the task in English, and then translate the response back to the target language. For ea... | https://arxiv.org/abs/2505.20546v2 |
arXiv:2505.20561v1 [cs.LG] 26 May 2025 2025-5-28 Beyond Markovian: Reflective Exploration via Bayes-Adaptive RL for LLM Reasoning Shenao Zhang1∗, Yaqing Wang2, Yinxiao Liu2, Tianqi Liu2, Peter Grabowski3, Eugene Ie3, Zhaoran Wang1†, Yunxuan Li3† 1Northwestern University,2Google DeepMind,3Google Large Language Models (L... | https://arxiv.org/abs/2505.20561v1 |
Bayes-Adaptive RL, which explicitly optimizes for test-time generalization by maximizing the expected return under a posterior distribution over Markov Decision Processes (MDPs). The objective incentivizes both reward-seeking actions and epistemic explorations that gather information to reduce the MDP’s uncertainty, su... | https://arxiv.org/abs/2505.20561v1 |
Reasoning. As an emerging capability of model scale, LLMs can generate intermediate CoTs to solve complex reasoning tasks [ 50,25] and scale test-time performance by allocating more thinking tokens [ 42,5]. Early efforts enhanced LLM reasoning via supervised fine-tuning on human-annotated data [7,58,55] or linearized s... | https://arxiv.org/abs/2505.20561v1 |
due to its simplicity. The state transition is deterministic by appending the new reasoning step, i.e., 𝑠𝑡+1=𝑠𝑡+𝑎𝑡. Prior work [ 44,17] employs an outcome-level reward verifier(𝑠𝑇,𝑦∗ 𝑠0), which uses a verifier to perform a regular expression match (either 0or1) between 𝑠𝑇and the ground-truth 3 Beyond Markov... | https://arxiv.org/abs/2505.20561v1 |
training, in a trial-and-error manner with repeated episodes, to discover the golden answers. The Markovian RL objective allows the optimal policy that memorizes these training answers to be fully exploited, with no incentive to adaptively explore with reflections. For the non-standard undiscounted infinite-horizon MDP... | https://arxiv.org/abs/2505.20561v1 |
other rewards are zero. Here, 𝑟(𝑠)represents the reward of reaching 𝑠. For any Markovian policy, the maximal return is 1/4(or1/2𝑑−1for a depth- 𝑑tree) since it is static and cannot adapt according to the reward feedback when reaching the four candidate answers. In contrast, the optimal Bayes-Adaptive policy has an... | https://arxiv.org/abs/2505.20561v1 |
state-conditional belief in the plausibility of M𝑖. The second product weighting term accumulates the discrepancy between predicted rewards 𝑟M𝑖(𝑠𝑡′,𝑎𝑡′) and observed rewards 𝑟𝑡′, which serves as a reflective signal for strategy switching by downweighting hypotheses that have high belief probabilities but are u... | https://arxiv.org/abs/2505.20561v1 |
the training solutions but fails to generalize at test time. In contrast, Bayes-Adaptive RL increases both training and testing accuracies. Furthermore, its accuracy and convergence rate improve when given prior knowledge that rewarding triplets are repeated patterns, i.e., |M|=3with𝑟M1(000)=𝑟M2(111)=𝑟M3(222)=1and a... | https://arxiv.org/abs/2505.20561v1 |
that has the highest average accuracy. It can be observed that BARL achieves higher accuracies across most benchmarks and models. It consistently outperforms Markovian RL baselines in terms of average accuracy, with larger gains observed on challenging benchmarks that demand effective exploration, such as CollegeMath a... | https://arxiv.org/abs/2505.20561v1 |
for LLM Reasoning frequency of reflections despite achieving lower accuracies. This result reveals the weak correlation between the performance of LLMs and the response length or the frequency of reflections. Rather, the effectiveness of thinking tokens and the efficiency of explorations are the determining factors, wh... | https://arxiv.org/abs/2505.20561v1 |
have exhibited emergent behaviors such as self-reflective reasoning. Yet in conventional Markovian RL, exploration is confined to the training phase to identify action sequences that maximize cumulative reward, and resorts to pure exploitation at test time. Besides, the Markov assumption indicates the dependency on his... | https://arxiv.org/abs/2505.20561v1 |
2024. [13]Mohammad Ghavamzadeh, Shie Mannor, Joelle Pineau, Aviv Tamar, et al. Bayesian reinforcement learning: A survey. Foundations and Trends ®in Machine Learning , 8(5-6):359–483, 2015. [14]Dibya Ghosh, Anurag Ajay, Pulkit Agrawal, and Sergey Levine. Offline rl policies should be trained to be adaptive. In Internat... | https://arxiv.org/abs/2505.20561v1 |
In The Twelfth International Conference on Learning Representations , 2023. [30]Zhihan Liu, Hao Hu, Shenao Zhang, Hongyi Guo, Shuqi Ke, Boyi Liu, and Zhaoran Wang. Reason for future, act for now: A principled framework for autonomous llm agents with provable sample efficiency. arXiv preprint arXiv:2309.17382 , 2023. [3... | https://arxiv.org/abs/2505.20561v1 |
Zhifang Sui. Math-shepherd: Verify and reinforce llms step-by-step without human annotations. arXiv preprint arXiv:2312.08935 , 2023. 15 Beyond Markovian: Reflective Exploration via Bayes-Adaptive RL for LLM Reasoning [47]Ruocheng Wang, Eric Zelikman, Gabriel Poesia, Yewen Pu, Nick Haber, and Noah D Goodman. Hypothesis... | https://arxiv.org/abs/2505.20561v1 |
in the wild. arXiv preprint arXiv:2503.18892 , 2025. [62]Shenao Zhang, Donghan Yu, Hiteshi Sharma, Han Zhong, Zhihan Liu, Ziyi Yang, Shuohang Wang, Hany Hassan, and Zhaoran Wang. Self-exploring language models: Active preference elicitation for online alignment. arXiv preprint arXiv:2405.19332 , 2024. [63]Han Zhong, Yu... | https://arxiv.org/abs/2505.20561v1 |
MATH GRPO Progress BARL 0 20 40 60 80 100 Iteration35.037.540.042.545.047.5Accuracy CollegeMath GRPO Progress BARL 0 20 40 60 80 100 Iteration2426283032343638Accuracy OlympiadBench GRPO Progress BARL 0 20 40 60 80 100 Iteration5101520Accuracy AIME 2024 GRPO Progress BARL 0 20 40 60 80 100 Iteration30405060Accuracy AMC ... | https://arxiv.org/abs/2505.20561v1 |
GRPO Progress BARL Figure 13|Average evaluation response lengths over training iterations for Qwen2.5-Math-7B models. 10 20 30 40 50 60 Iteration350400450500Response Length GSM8K GRPO Progress BARL 10 20 30 40 50 60 Iteration650700750800850900Response Length MATH GRPO Progress BARL 10 20 30 40 50 60 Iteration5506006507... | https://arxiv.org/abs/2505.20561v1 |
arXiv:2505.20564v1 [cs.CL] 26 May 2025The NaijaVoices Dataset: Cultivating Large-Scale, High-Quality, Culturally-Rich Speech Data for African Languages Chris Emezue1,2,3,10, The NaijaVoices Community1,2, Busayo Awobade4, Abraham Owodunni5, Handel Emezue2,7, Gloria Monica Tobechukwu Emezue2,7, Nefertiti Nneoma Emezue2,7... | https://arxiv.org/abs/2505.20564v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.