text string | source string |
|---|---|
kitchen. 4 Mary moved the chocolate to the table. State 4 5 Mary exited the kitchen. 6 Alice moved the chocolate to the cupboard. State 6 Question: Where does Mary think Alice thinks the chocolate is? ToM Actions DEL Belief States J, M, A Drawer M, A Table J, M, A Drawer A Cupboard M, A Table J, M, A DrawerJ J, M JStat... | https://arxiv.org/abs/2505.17348v1 |
to obtain belief annotations for each step. For reasoning traces, we use GPT-4o-mini (Hurst et al., 2024) to generate step-by-step explanations for the same scenarios. We observe that GPT-4o- mini produces a balanced mix of correct and incor- rect reasoning traces, making it suitable for reward modeling. Each training ... | https://arxiv.org/abs/2505.17348v1 |
est PBM score and extract its final answer. This method chooses the most confident individual trace, but does not account for answer consistency across traces. (ii) Weighted Best-of-N. Group traces by their final answers, then aggregate PBM scores across traces that predict the same answer. The answer with the highest ... | https://arxiv.org/abs/2505.17348v1 |
formance. Among the aggregation strategies, min and prod exhibit stable and similar performance across both vanilla andweighted ranking schemes. In contrast, avg and last tend to degrade in perfor- mance under weighted aggregation, likely due to their sensitivity to low-quality or inconsistent sam- ples. We recommend u... | https://arxiv.org/abs/2505.17348v1 |
95.0 60.4 87.0 Qwen3-4B 100.0 100.0 79.8 85.0 79.3 97.5 70.2 82.5 46.0 60.0 75.1 85.0 PBM can still function effectively. Furthermore, when paired with high-throughput generation back- ends such as vLLM, offline methods can generate large numbers of candidate traces efficiently. Over- all, we recommend using offline in... | https://arxiv.org/abs/2505.17348v1 |
fine-tune smaller models via GRPO (Shao et al., 2024), in order to enhance their ToM capabilities. However, GRPO-based training requires substan- tial computational resources and is notoriously dif- ficult to optimize. In contrast, our PBM module is lightweight and efficient: it can be trained in under three hours on a... | https://arxiv.org/abs/2505.17348v1 |
that social reasoning tasks can benefit from compute-efficient methods that guide rather than retrain the model. This opens new possibilities for deploying socially aware AI in resource-limited settings. References Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Ara... | https://arxiv.org/abs/2505.17348v1 |
language model serving with pagedattention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles . Zhihang Lin, Mingbao Lin, Yuan Xie, and Rongrong Ji. 2025. Cppo: Accelerating the training of group relative policy optimization-based reasoning models. arXiv preprint arXiv:2503.22342 . Zizheng... | https://arxiv.org/abs/2505.17348v1 |
Inc. Appendix A Prompt Templates We present the prompt templates in the folllowing textbox. One-Shot Prompt Here is a story that unfolds in chronological order . You will be asked a question about the story , which may involve either : (1) Locating an object , or (2) Inferring an agent 's mental state (e.g., what A thi... | https://arxiv.org/abs/2505.17348v1 |
is in [ red_envelope ] ## Step 5 ## Amelia exited the TV_room . Irrelevant . Owen thinks Liam thinks Chloe thinks the celery is in [ red_envelope ] ## Step 6 ## Chloe moved the celery to the green_bucket . Only Chloe , Liam , Owen , Benjamin are present . They all see this move . Owen thinks Liam thinks Chloe thinks th... | https://arxiv.org/abs/2505.17348v1 |
arXiv:2505.17362v2 [cs.CL] 26 May 2025A Fully Generative Motivational Interviewing Counsellor Chatbot for Moving Smokers Towards the Decision to Quit Zafarullah Mahmood*Soliman Ali*Jiading Zhu*Mohamed Abdelwahab* Michelle Yu Collins*Sihan Chen*Yi Cheng Zhao*Jodi Wolff† Osnat Melamed*†Nadia Minian*†Marta Maslej*†Carolyn... | https://arxiv.org/abs/2505.17362v2 |
participants’ readiness to quit through a widely-used readiness ruler (Boudreaux et al., 2012) before the conversation and one week later. Thedifferencebetweenthesetwomea- surementsisourprimarymetricofeffective- ness. 2.A rating of the perceived empathy of the chat- botonthe CAREscale(Merceretal.,2004) whichiswidelyuse... | https://arxiv.org/abs/2505.17362v2 |
et al., 2019; Tavabi et al., 2021; Gibson et al., 2022; Nakano etal.,2022;Minetal.,2022;Gabrieletal.,2024) on textual transcripts, with some also incorporat- ingothermodalitiessuchasfeaturesfromspeech andvideodata(Tavabietal.,2020;Nakanoetal., 2022). LLMshaverecentlybeenusedinautomated MI evaluation, such as detecting ... | https://arxiv.org/abs/2505.17362v2 |
a collaboration of engi- neersandexpertclinicians,thelatterhighlyexpe- riencedindeliveringMIcounsellingforsmoking cessation. Thegroupusedthefollowinginformalprocess to evolve the prompt for the counsellor chatbot: webegan with a short, simple prompt (shown in Ap- pendix B), which asks the model to use its own knowledge... | https://arxiv.org/abs/2505.17362v2 |
might intentionally steer the conversation far offfromthetopicofsmoking. Webuiltaclassifierto monitorconversationsinreal-timetodetectifthe clientisdeliberatelysteeringtheconversationoff- track. Unlikethemoderatorobserver,thisclassifier was prompt-engineered for a low false positive rate to give the benefit of the doubt... | https://arxiv.org/abs/2505.17362v2 |
the counsellor’s generated text adheresto the MI methodology to establish clinical valid- ity — providing evidence to trust (or not) the auto- mated counsellor. Assessment of client text from theparticipantscanalsoofferdirectevidenceofsuc- cess or failure in the counselling session. Since we wanttoassessbothcounsellora... | https://arxiv.org/abs/2505.17362v2 |
1also shows that there was a significant change in the participants’ view of the importance of quitting, with an average increase of +0.5, ex- hibiting the chatbot’s effectiveness. The change in readiness was not statistically significant. Finally,Table M.1 in Appendix M shows that baseline confidence levels andone-wee... | https://arxiv.org/abs/2505.17362v2 |
scores closely matchesthoseofHLQC_HI,anotherindicationthat the conversation adhered to the principles of MI. The R:Q distribution has a similar behaviour. Figure 3c shows the violin plot distribution of the%ClientChangeTalk(%CT).Thedistributions forMIBotand HLQC_HI are very similar, as were theaverages. Thisisperhapsth... | https://arxiv.org/abs/2505.17362v2 |
subsequent studies to leverage the published data from our feasibility study to identify and ad- dresssystematicbiasesandensureequitableaccess to AI-driven therapy solutions. 8 Ethics Statement Our study was approved by the University of Toronto Research Ethics Board on August 3rd, 2024, under protocol number 49997 (Ro... | https://arxiv.org/abs/2505.17362v2 |
Ashley Sullivan, Beau Abar, Steven L. Bernstein, Adit A. Ginde, and Carlos A. Camargo.2012. Motivationrulersforsmokingces- sation: a prospective observational examination of constructandpredictivevalidity. AddictionScience & Clinical Practice , 7(1):8. Andrew Brown, Ash Tanuj Kumar, Osnat Melamed, Imtihan Ahmed, Yu Hao... | https://arxiv.org/abs/2505.17362v2 |
Theresa Moyers, William R Miller, Laura Glynn, and C Hallgreen. 2010. Manual for the Motivational Interviewing Skill Code (MISC) ver- sion 2.5. Retrieved from http://casaa.unm.edu/ download/misc25.pdf . Noemi James, Eunyoung Lee, Carey Cole, Barbara Far- ris, and Pamela Wright. 2021. Improving Chronic Health Diseases T... | https://arxiv.org/abs/2505.17362v2 |
more.2019. Coercedchange-talkwithconversational agents promotes confidence in behavior change. In Proceedingsofthe13thEAIInternationalConference on Pervasive Computing Technologies for Health- care, PervasiveHealth’19, page 31–40, New York, NY, USA. Association for Computing Machinery. OpenAI. 2024a. ChatGPT:AI-powered... | https://arxiv.org/abs/2505.17362v2 |
of the 24th ACM Interna- tionalConferenceonIntelligentVirtualAgents ,IVA ’24,NewYork,NY,USA.AssociationforComputing Machinery. Xin Sun, Jiahuan Pei, Jan de Wit, Mohammad Alian- nejadi, Emiel Krahmer, Jos T.P. Dobber, and Jos A. Bosch. 2024. Eliciting motivational interviewing skill codes in psychotherapy with LLMs: A b... | https://arxiv.org/abs/2505.17362v2 |
5.4): ✓ □Theauthorshavethelegalbasisforprocess- ing the data, and if it is made public, for distributing it. (Check one) □The data are in the public domain, and licensed for research purposes. ✓ □The data are used with consent of its cre- ators or copyright holders. □Ifthedataareusedwithoutconsent,the paper makes the c... | https://arxiv.org/abs/2505.17362v2 |
client. 5You have already introduced yourself and the client has consented to the therapy session. 6You don’t know anything about the client’s nicotine use yet. 7Opentheconversationwithageneralgreetingandfriendlyinteraction,andgraduallyleadtheconversation towardshelpingtheclientexploreambivalencearoundsmoking,usingyour... | https://arxiv.org/abs/2505.17362v2 |
analysis with a response of True if the client is leading the conver- sation offtrack, or False otherwise. Provide a one-word response of either True or False. Do not provide anything else in your response. C.3 Prompt for the End Classifier Agent Youareevaluatinganongoingconversationbetweenaclientandacounsellorandwilld... | https://arxiv.org/abs/2505.17362v2 |
and more averse to the idea of quitting. You find it much easier to tune out and go outside to smoker rather than trying to explain why you are not ready. Given all these things going on in your life, you are highly resistant to changing your smoking habit. You believe now is not the right time to quit smoking. You do ... | https://arxiv.org/abs/2505.17362v2 |
the feasibility study. Conversation Analysis By AutoMISCPre-conversation Readiness Rulers Conversation with MIBot Post-conversation Readiness RulersCARE Survey Week-later Readiness Rulers% MIC R:Q % CT1 2 3 45 6 Figure E.1: Different Stages of our Feasibility Study. F Readiness Ruler Questions On a scale of 0 (very low... | https://arxiv.org/abs/2505.17362v2 |
⃝ ⃝ ⃝ Poor Fair Good Very Good Excellent Does Not Apply 4. Being interested in you as a whole person... (asking/knowing relevant details about your life, your situation, not treating you as "just a number") ⃝ ⃝ ⃝ ⃝ ⃝ ⃝ Poor Fair Good Very Good Excellent Does Not Apply 5. Fully understanding your concerns... (communicat... | https://arxiv.org/abs/2505.17362v2 |
Table I.1: Top 10 most frequently mentioned positive words in participant feedback. Word Frequency repetitive 6 boring 3 unresponsive 1 disappointing 1 annoying 1 dull 1 pointless 1 useless 1 uncreative 1 overbearing 1 Table I.2: Top 10 most frequently mentioned negative words in participant feedback. Figure J.1: Word ... | https://arxiv.org/abs/2505.17362v2 |
T I imagine there could be R T What are some of the QSpeaker Utterance C same old routine T It sounds like smoking T How do you feel about T Are there moments C i feel great, C it works for me T It's important that it T I imagine there could be T What are some of theC: same old routine T: It sounds like smoking is a re... | https://arxiv.org/abs/2505.17362v2 |
over the RQ category and assign a subcategory of "R" for reflection or "Q" for question. K.3.3 Client Utterance Classification Prompt YouareahighlyaccurateMotivationalInterviewing(MI)counsellingsessionannotator. Yourtaskistoanalyzean excerpt from a counselling session of up to five volleys and categorize the client’s f... | https://arxiv.org/abs/2505.17362v2 |
Student Status No 80 (75.5) Yes 21 (19.8) Data Expired 5 (4.7)Demographic Factors Count, n (%) Employment Status Full-Time 49 (46.2) Part-Time 18 (17.0) Not in Paid Work 16 (15.1) Unemployed 13 (12.3) Other 10 (9.4) Country of Residence United Kingdom 47 (44.3) United States 42 (39.6) Canada 9 (8.5) South Africa 4 (3.8... | https://arxiv.org/abs/2505.17362v2 |
to you right now to stop smoking? PreRulerConfidence On a scale from 0 to 10, how confident are you that you would succeed at stopping smoking if you start now? PreRulerReadiness Onascalefrom0to10,howreadyareyoutostartmakingachange at stopping smoking right now? Post-conversation Readiness Rulers PostRulerImportance On... | https://arxiv.org/abs/2505.17362v2 |
come to mind when you think about stopping smoking? Client:its stressfull Counsellor: It sounds like there’s a lot of stress attached to theidea of stopping smoking. It might feeloverwhelmingtoconsiderthatchange. What do you think might be contributing to that stress? Client:ive smoked for a long long time Counsellor: ... | https://arxiv.org/abs/2505.17362v2 |
part of your life. Giving up the familiar habit has made you wonderaboutwhatyou’lldowithyourhandsand howyou’llpassthetime. Weexploredtheideaof saving money by quitting and how you could use those savings toward your interest in home decor, bringinganewsenseofexcitementtoyourliving space. Finally, we talked about incorp... | https://arxiv.org/abs/2505.17362v2 |
is a small but very unlikely possibility that the chatbot may generate responses that may instigate the above feelings. Should this happen, it is encouraged that you document it in your qualitative feedback opportunity. There are no other risks or harm to you in participating in this study. WHAT COMPENSATION AM I ENTIT... | https://arxiv.org/abs/2505.17362v2 |
arXiv:2505.17371v1 [cs.LG] 23 May 2025An End-to-End Approach for Child Reading Assessment in the Xhosa Language Sergio Chevtchenko1,2, Nikhil Navas1,2, Rafaella Vale4, Franco Ubaudi5, Sipumelele Lucwaba6, Cally Ardington3, Soheil Afshar7, Mark Antoniou2, and Saeed Afshar1,2 1International Centre for Neuromorphic System... | https://arxiv.org/abs/2505.17371v1 |
otherdifficultiessuchasdyslexia[24].Asaresult,uncoveringtheseissuesquickly can promote focused guidance from an early age that greatly improves a child’s educational outcome [12,24]. To tackle this, the use of automatic speech recognition (ASR) can greatly assist in finding reading issues that manifest themselves in pr... | https://arxiv.org/abs/2505.17371v1 |
request to the community. Second, it details and validates a cost-effective online method for data labeling and ver- ification, demonstrating its suitability for resource-constrained settings. Third, it provides a comprehensive evaluation of three fine-tuned SOTA speech classi- fication architectures on this dataset, h... | https://arxiv.org/abs/2505.17371v1 |
attempt to investigate how AI performs in assessing students reading fluency and level compared to human experts, Yıldız et al. [25] found that AI offered reliable predictions that could contribute to lowering costs and improving efficiency. The recent publication by Henkel et al. [9] has a significant overlap with the... | https://arxiv.org/abs/2505.17371v1 |
samples for each word and letter in the dataset. Although each recording is intended to capture a single spoken word, even those pronounced correctly often include additional artifacts, such as classroom 6 S. Chevtchenko et al. noise or multiple attempts at pronouncing the word on the app’s screen. This variability in ... | https://arxiv.org/abs/2505.17371v1 |
Whisper6from OpenAI, an end-to-end Transformer-based architecture for sequence-to-sequence tasks. 1https://huggingface.co/facebook/wav2vec2-xlsr-53-espeak-cv-ft 2https://www.openslr.org/94/ 3https://commonvoice.mozilla.org/en/languages 4https://catalog.ldc.upenn.edu/byyear 5https://huggingface.co/utter-project/mHuBERT-... | https://arxiv.org/abs/2505.17371v1 |
Comparison of training (1000 steps) and inference time (single recording) on a Google Colab L4 instance, averaged over 10 runs for each model. Model Training time (s) Inference time (s) HuBERT 343 0.44 wav2vec 2.0 1332 1.70 Whisper 2511 1.54 and inference, the wav2vec 2.0 model requires less computational resources for... | https://arxiv.org/abs/2505.17371v1 |
each question, indicating that while this approach performs better on words than on letters, it struggles significantly with noisy recordings, such as the example illustrated in Figure 2. A sample of transcriptions1in different scenarios is also provided, suggesting that direct multilingual ASR usage in classroom condi... | https://arxiv.org/abs/2505.17371v1 |
and feedback for individualized reading practice. In: Interspeech 2020. pp. 3870–3874 (2020). https://doi.org/10.21437/Interspeech.2020-2842 6. Dossou, B.F.P., Emezue, O.: Okwugbé: End-to-end speech recognition for fon and igbo.In:Proceedingsofthe2024InternationalConferenceonSpeechandLanguage Technology. pp. 123–134 (2... | https://arxiv.org/abs/2505.17371v1 |
Value-Guided Search for Efficient Chain-of-Thought Reasoning Kaiwen Wang*1,Jin Peng Zhou*1,Jonathan Chang*4, Zhaolin Gao1,Nathan Kallus1,3,Kianté Brantley2, and Wen Sun1 1Cornell University2Harvard University3Netflix4Databricks Abstract In this paper, we propose a simple and efficient method for value model training on... | https://arxiv.org/abs/2505.17373v1 |
[ 6,36,34,25], it is challenging to scale existing PRM training techniques to long-context reasoning models. First, existing methods require a pre-defined notion of “step,” but, per Guo et al. [18], “it is challenging to explicitly define a fine-grain step in general reasoning.” Second, even if we can define a “step,” ... | https://arxiv.org/abs/2505.17373v1 |
for a detailed discussion of related works. 2 2 Methods We present an end-to-end pipeline for training a token-level value model and applying it to guide block-wise search. In Section 2.1, we introduce necessary notation and present a regression-via- classification algorithm for learning the token-level value model [ 1... | https://arxiv.org/abs/2505.17373v1 |
hence can be viewed as another data-point. We found this to be an important training detail for performance, which is consistent with prior work who used a similar objective for training an outcome reward model [14, 23]. We can now view the classifier as a value model. Since κ= 1 corresponds to the event that r= 1, we ... | https://arxiv.org/abs/2505.17373v1 |
from four DeepSeek-R1-Distill-Qwen model sizes: 1.5B, 7B, 14B, and 32B by generating until the end of thinking token <\think> .1For each roll- in˜yi, we then sample four random prefix locations where we generate complete roll-outs {zj i}j∈[4] fromπref. Finally, to compute the class label (incomplete, incorrect, or corr... | https://arxiv.org/abs/2505.17373v1 |
keeps B=N/w (assume to be integer) partial responses and branches wparallel blocks from each one [ 26,5,40,6,36]. Given a prompt x, beam search first generates Nparallel blocks. However, unlike BFS, beam search keeps the top Bbeams with the highest scores, and then samples wparallel blocks per beam at the next step. Si... | https://arxiv.org/abs/2505.17373v1 |
the per-step human annotation of PRM800K [23]. As a step-level value model, Math-Shepherd-PRM-7B is more related to our token-level value model. Finally, we also evaluate a 1.5B Bradley-Terry (BT) [ 8] model, called DeepSeek-BT-1.5B , which we trained using our dataset (see Appendix G for training details). 2https://ma... | https://arxiv.org/abs/2505.17373v1 |
VGSalso improves performance relative to WMV , suggesting that BT models may also provide meaningful block-wise feedback to guide search. We also observe that accuracies for the 7B baseline PRMs (MathSheperd and Qwen2.5-Math) are only slightly higher than MV@256 and do not improve with search, which suggests that these... | https://arxiv.org/abs/2505.17373v1 |
the base DeepSeek-1.5B model. In Figure 15 (Appendix C.7), we present histograms of response lengths across all benchmarks. The results show that VGSconsistently generates more concise outputs, whereas the base model often reaches the generation cap, with up to 4 16 64 256 1024 Inference Budget (N)2530354045Accuracy We... | https://arxiv.org/abs/2505.17373v1 |
also ablate the amount of DVTS parallelism. These tests suggest that there is a consistent choice of search hyperparameters that work well across inference bud- gets. 4.1 Different Search Parameters and Methods Block Size. We perform beam search with width 2 using search block sizes from 16 to 16384. Fig. 6 shows AIME-... | https://arxiv.org/abs/2505.17373v1 |
mediate blocks are chosen. We call this process “random search”. Thus, if our value model is helpful for search, we should expect VGSto out- perform random search. Indeed, Fig. 8 validates this hypothesis. We also evaluate a hybrid ap- proach where half of DVTS’s parallel trees use random search and the other half use ... | https://arxiv.org/abs/2505.17373v1 |
(offline) reinforcement learning. arXiv preprint arXiv:2403.05385 , 2024. [4]Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073 , 20... | https://arxiv.org/abs/2505.17373v1 |
Martha White. Investigating the histogram loss in regression. arXiv preprint arXiv:2402.13425 , 2024. [21] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.0836... | https://arxiv.org/abs/2505.17373v1 |
Sherburn, James Aung, Jun Shern Chan, Leon Maksin, Rachel Dias, Evan Mays, Benjamin Kinsella, Wyatt Thompson, et al. Paperbench: Evaluating ai’s ability to replicate ai research. arXiv preprint arXiv:2504.01848 , 2025. [38] Jiashuo Sun, Yi Luo, Yeyun Gong, Chen Lin, Yelong Shen, Jian Guo, and Nan Duan. Enhancing chain-... | https://arxiv.org/abs/2505.17373v1 |
language model programs. Advances in Neural Information Processing Systems , 37:62557–62583, 2024. [52] Jin Peng Zhou, Kaiwen Wang, Jonathan Chang, Zhaolin Gao, Nathan Kallus, Kilian Q Wein- berger, Kianté Brantley, and Wen Sun. q♯: Provably optimal distributional rl for llm post-training. arXiv preprint arXiv:2502.205... | https://arxiv.org/abs/2505.17373v1 |
score [ 14,4]. ORMs can also be used for weighted majority voting (WMV), which generalizes MV where the strength of a response’s vote is proportional to its ORM score. Weighted MV (WMV) typically provides an improvement over vanilla (unweighted) MV [6, 25], which is also what we observe in our experiments (Fig. 3). Out... | https://arxiv.org/abs/2505.17373v1 |
similar to ours since the training process also occurs at a token-level and is agnostic to the definition of step. However, a limitation of the PA V is that if the expert disagrees with the underlying policy, then maximizing the PA V can lead to suboptimal behavior [ 12]. Our approach of directly using the value model ... | https://arxiv.org/abs/2505.17373v1 |
by the value model. The most in-distribution choice is to use the roll-out policy πref. 10.Perform model selection by running outcome-level TTC ( e.g., WMV) on some validation benchmark. 11.Sweep search parameters ( e.g., block size, beam width, DVTS parallelism) on the validation benchmark. 12. Run the final model on ... | https://arxiv.org/abs/2505.17373v1 |
AIME-25 4 16 64 256 1024 Inference Budget (N)202530Accuracy HMMT-24 4 16 64 256 1024 Inference Budget (N)15202530 HMMT-25 VGS + WMV WMVVGS + BoN BoN406080 AIME-24 4060 AIME-25 4 16 64 256 1024 Inference Budget (N)204060 HMMT-24 4 16 64 256 1024 Inference Budget (N)204060 HMMT-25 VGS + WMV MV Pass@N Figure 10: Per-bench... | https://arxiv.org/abs/2505.17373v1 |
mini batch size: 128 micro batch size: 1 policy learning rate: 1e-6 critic learning rate: 1e-5 train epochs: 25γentropy : 1e-3 γKL: 1e-4 gaeγ: 1 gaeλ: 1 clip ratio: 0.2 Total number of steps: 2250 DeepSeek-1.5B Pass@4 Pass@8 Pass@16 Pass@32 Pass@64 Pass@128 Pass@256 AIME-24 48.7±5.0 58 .9±4.9 67 .3±4.7 74 .1±3.2 78 .4±... | https://arxiv.org/abs/2505.17373v1 |
critical error in Figure 19, VGSassigns a moderately high score (0.337) to the rejected beam—higher than scores for less subtle failures in earlier examples—suggesting that even significant mistakes can be difficult to detect when embedded in otherwise coherent reasoning. Finally, we empirically compare the distributio... | https://arxiv.org/abs/2505.17373v1 |
problem. In particular, the classifier is trained to predict the probability of each class under the roll-out policy, given the partial solution. Roll-in. The main point of the roll-in process is to create a diverse distribution of partial solutions to roll-out from. By creating a diverse roll-in distribution with mult... | https://arxiv.org/abs/2505.17373v1 |
to 4response pairs per problem, where a pair consists of a response with reward 0(the ‘reject’ response) and a response with reward 1(the ‘chosen’ response). Some prompts may have fewer than 4responses with reward 0or1, and in those cases, we include as many as possible. This yields a dataset of roughly 122k pairs. Mod... | https://arxiv.org/abs/2505.17373v1 |
see that the value model FLOPs is negligible compared to the generator model FLOPs. In particular, when guiding a 1.5B generator with a 1.5B classifier, the classifier FLOPs is only 0.024% of the generator FLOPs. With a compute budget of 256, this amounts to a total FLOPs of (49.1T+ 12B)×256 = 12 .6P. When guiding with... | https://arxiv.org/abs/2505.17373v1 |
is 1−95/210 = (210−95)/210 = 115 /210. Wait, that doesn’t match with the previous result. Hmm, something’s wrong here. Wait, why did I get 91 before? Let me see. I had C(4,2)∗C(6,2) +C(4,4). So, that’s 6*15 +1=91. But that’s only the number of winning ways where she has exactly two or exactly four correct numbers. So, ... | https://arxiv.org/abs/2505.17373v1 |
. . . Therefore, P(B)is 23/42. . . . Therefore, the conditional probability P(A|B)is(1/210)/(23/42) = (1 /210)∗(42/23) = (42) /(210∗23). . . Therefore, the probability becomes (1/5)/23 = 1 /(5∗23) = 1 /115. . . . But 1/115 can’t be simplified further, so m is 1 and n is 115, which are coprime. Therefore, m + n is 1 + 1... | https://arxiv.org/abs/2505.17373v1 |
help confirm if n= 1is the only solution. Testing n= 1: As above, product is 120, which is divisible by 3. Testing n= 2:n+ 2 = 4 . Product is 3(5)(13) = 195 . 195 divided by 4 is 48.75, which isn’t an integer. So, 4 doesn’t divide 195. . . . Testing n= 9:n+ 2 = 11 . Product is 3(12)(90) = 3240 . 3240 divided by 11 is a... | https://arxiv.org/abs/2505.17373v1 |
under VGS. The high-scoring beam applies a concise and effective modular reasoning strategy whereas the rejected beam attempts several approaches—including an incorrect use of the Factor Theorem and trial-and-error—but ultimately fails. Color coding follows Figure 17. Part of generations are abridged with . . .notation... | https://arxiv.org/abs/2505.17373v1 |
three-digit number), (1000a) must be divisible by (a ²- 1). So, (a ²- 1) must be a divisor of 1000a. Let me note that a is a digit from 1 to 9. So, I can try each a from 1 to 9 and see if (a ²- 1) divides 1000a, and then check if M is a three-digit integer. Let’s compute for each a: Starting with a=1: a=1: M = 1000*1 /... | https://arxiv.org/abs/2505.17373v1 |
Chart-to-Experience: Benchmarking Multimodal LLMs for Predicting Experiential Impact of Charts Seon Gyeom Kim * KAISTJae Y oung Choi† KAISTRyan Rossi‡ Adobe ResearchEunyee Koh§ Adobe ResearchTak Y eon Lee¶ KAIST ABSTRACT The field of Multimodal Large Language Models (MLLMs) has made remarkable progress in visual unders... | https://arxiv.org/abs/2505.17374v1 |
two categories: 1) Emotional factors, including empathy ,interest , and comfort ; and 2) Perceptual factors including memorability ,trust- worthiness ,aesthetic pleasure , and intuitiveness . To construct the dataset, we recruited 216 crowdsourced workers1, and asked them to rate their experiences using a 7-point Liker... | https://arxiv.org/abs/2505.17374v1 |
For ex- ample, chart quality has been evaluated based on perceptual factors such as readability [34, 43] and cognitive effort or time spent on particular tasks [4, 18] since data visualization primarily aims to fa- cilitate analysis. Also, the affective roles of charts are increasingly identified, emphasizing aspects s... | https://arxiv.org/abs/2505.17374v1 |
a combination of charts with detailed annotation and/or rich illustrations). Among this information, auxiliary visual elements such as creator logos or certification marks were neither removed nor added for diverse coverage. Moreover, the collected charts con- tain text information, including titles, sources, label nam... | https://arxiv.org/abs/2505.17374v1 |
emphasize key data trends and the core message. Notably, a chart can remain intuitively understandable even if its design is unfamiliar, visually unappealing, or uncomfortable to read. Comfort assesses the overall ease and satisfaction with which a viewer interacts with a chart [10]. We included this metric to focus on... | https://arxiv.org/abs/2505.17374v1 |
when comparing chart pairs with larger score disparities. 4 E VALUATING MLLM S AS A JUDGE To evaluate the capability of MLLMs as predictors of the experi- ential impact of the charts, we conducted two studies. By compar- ing the study results with the crowdsourced data, we evaluated to what extent the judgments of the ... | https://arxiv.org/abs/2505.17374v1 |
learning. Although such techniques could potentially enhance the performance of MLLM, they also introduce additional variables, complicating the evalua- tion process, and were therefore beyond the scope of this study. The accuracy of the comparisons was evaluated against human ratings as a benchmark. For example, if th... | https://arxiv.org/abs/2505.17374v1 |
their “vivid color scheme”, “visual com- plexity”, and “analysis over extended timeframes.” In contrast, a majority of human evaluators described them as “hard to read”, “not pleasing”, and even “chaotic.” Once the limitation of MLLMs has been identified, we believe advanced prompt engineering techniques such as Chain-... | https://arxiv.org/abs/2505.17374v1 |
Affective color in visualization. InProceedings of the 2017 CHI conference on human factors in com- puting systems , pp. 1364–1374, 2017. [4] M. Behrisch, M. Blumenschein, N. W. Kim, L. Shao, M. El-Assady, J. Fuchs, D. Seebacher, A. Diehl, U. Brandes, H. Pfister, et al. Quality metrics for information visualization. In... | https://arxiv.org/abs/2505.17374v1 |
sualization , 8(3):139–152, 2009. [19] L. Huddy and A. H. Gunnthorsdottir. The persuasive effects of emo- tive visual imagery: Superficial manipulation or the product of pas- sionate reason? Political Psychology , 21(4):745–778, 2000. [20] P. Isola, J. Xiao, D. Parikh, A. Torralba, and A. Oliva. What makes a photograph... | https://arxiv.org/abs/2505.17374v1 |
model. arXiv preprintarXiv:2010.09142 , 2020. [36] E. M. Peck, S. E. Ayuso, and O. El-Etr. Data is personal: Attitudes and perceptions of data visualization in rural pennsylvania. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems , pp. 1–12, 2019. [37] J. Sabini and M. Silver. Ekman’s basi... | https://arxiv.org/abs/2505.17374v1 |
Q. Li, B. Wang, and J. Qin. Dialoguellm: Context and emotion knowledge-tuned llama mod- els for emotion recognition in conversations. arXiv preprint arXiv:2310.11374 , 2023. [52] S. Zhao, G. Ding, Q. Huang, T.-S. Chua, B. Schuller, and K. Keutzer. Affective image content analysis: A comprehensive survey. 2018. [53] S. ... | https://arxiv.org/abs/2505.17374v1 |
AI-Augmented LLMs Achieve Therapist -Level Responses in Motivational Interviewing Yinghui Huang , Yuxuan Jiang, Hui Liu, Yixin Cai, Weiqing Li Abstract—Large language models (LLMs) like GPT -4 show potential for scaling motivational interviewing (MI) in addiction care, but require systematic evaluation of therapeutic c... | https://arxiv.org/abs/2505.17380v1 |
[10] . These barriers not only degrade individuals ’ quality of life but also place a significant strain on social and economic development [11]. To address the growing gap in mental health services, individuals and researc h teams have increasingly turned to LLMs to enhance interactions between therapists and clients ... | https://arxiv.org/abs/2505.17380v1 |
t need to evaluate the feasibility, effectiveness, and safety of LLM -based interventions within the MI framework, in order to accurately assess their potential benefits for patient welfare. Accordingly, this study proposes Research Question 1 (RQ1): To wha t extent can general - purpose LLMs replicate the performance ... | https://arxiv.org/abs/2505.17380v1 |
UPQ be identified and integrated to develop a framework for evaluating the verbal behaviors of general -purpose LLMs in MI? To address this, we employ explainable machine learning methods to introduce a computational eva luation framework that combines UPQ with relevant intrinsic metrics. This framework incorporates bo... | https://arxiv.org/abs/2505.17380v1 |
future research. Finally, Section 6 concludes by emphasizing the primary contributions of this study to the fields of MI, mental health, and LLMs. The contribut ions of this work are delineated as follows: (1) This study empirically demonstrated that LLMs achieve performance levels comparable to human therapists in cor... | https://arxiv.org/abs/2505.17380v1 |
ongoing therapist development. MI assessment methods can be broadly categorized into manual and automated approaches. M anual evaluation s typically rely on standardized tools such as the MITI and MISC [48]. Artificial intelligence (AI) has significantly impacted various industries [49], with machine learning technolog... | https://arxiv.org/abs/2505.17380v1 |
from limited reproducibility, vulnerability to evaluator bias, and constraints in assessing LLMs ’ full potential [71]. Additionally, cultural differences can lead to substantial variability in manual evaluation outcomes [72]. In contrast, automated evaluation offers a more standardized , efficient and consistent proce... | https://arxiv.org/abs/2505.17380v1 |
the mental health status of po pulations, tracking temporal trends in psychological well -being [75]. Overall, the advancement of non -invasive group measurement methods via digital language offer s a robust and scientifi cally validated approach, serving as a reliable data source for digital epidemiology [83], [84] . ... | https://arxiv.org/abs/2505.17380v1 |
techniques to assess changes in dialogu es and record - keeping practices post -collaborative sessions with AI -driven conversational agents. In the context of MI, dictionary -based NLP methods are essential for evaluating intervention quality. By s ystematically classifying and comparing human -AI dialogue content, NL... | https://arxiv.org/abs/2505.17380v1 |
transcribed these videos with high precision, and domain experts specializing in MI meticulously annota ted key elements to ensure the dataset’s validity, comprehensiveness, and reliability. The dataset captured therapist behaviors (such as reflective listening and questioning) as well as the types of client responses ... | https://arxiv.org/abs/2505.17380v1 |
prepositions), with the mean score representing the overall LSM, where higher values indicate stronger linguistic alignment [94]. Additionally, pretrained models such as BERT and RoBERTa were employed to generate word embeddings, capturing more nuanced emotional and behavioral patterns in the therapist’s speech [95]. F... | https://arxiv.org/abs/2505.17380v1 |
for forward stepwise selection and cross - validation, aiming to identify the top n features that optimize model performance. Th is methodology improved model performance [99] and enhanced interpretability [100] , mitigated overfitting [101] , and illuminated the relationships between features and the target variable, ... | https://arxiv.org/abs/2505.17380v1 |
the computational evaluation framework was employed to design CoT -based customized prompts to enhance GPT -4’s performance. These prompts were then deployed via the OpenAI API to generate GPT -4- prompted responses. McNemar’s test and paired -sample t -tests were conducted to compare the extrinsic and intrinsic metric... | https://arxiv.org/abs/2505.17380v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.