text string | source string |
|---|---|
approach. We found correlation to be as low as 0.56 which implies that unimodal video features are successfully removed from multi-modal representations. Fig. 4 shows normalized alignment for language (AG) and visual regions (MT). Cross-modal multi-modal models. The alignment in regions AG and MT is extremely high, and... | https://arxiv.org/abs/2505.20027v1 |
- Unimodal SM TVLT Joint - Unimodal SMNormalized brain alignment Visual: MT0.10.20.30.40.50.6 Normalized brain alignment Figure 4: Residual analysis: Average normalized brain alignment was computed across participants before and after removal of video and audio embeddings from both jointly pretrained and cross- modalit... | https://arxiv.org/abs/2505.20027v1 |
such as PTL, MFG, ATL, PCC and visual regions EVC, LOC and OFA, as shown in Figs. 9 and 10 in Appendix. These results suggest that there is additional information beyond the unimodal embeddings considered in this study that is processed in the visual and language regions. Qualitative analysis. We compute the percentage... | https://arxiv.org/abs/2505.20027v1 |
not lead to a drop in brain alignment, indicating that there is additional information beyond speech features that is processed in these regions. This means that in cross-modal models, when transferring knowledge from one modality to another, the model relies more heavily on visual information. As a result, the model b... | https://arxiv.org/abs/2505.20027v1 |
perturbation affects the alignment with fMRI brain recordings acquired while participants are engaged in watching multi-modality naturalistic movies. Our analysis of multi-modal brain alignment yields several important conclusions: (1) The improved brain alignment of the multi-modal models over unimodal models, across ... | https://arxiv.org/abs/2505.20027v1 |
structure in the space of language representations is reflected in brain responses. Advances in Neural Information Processing Systems , 34:8332–8344, 2021. Richard Antonello, Aditya Vaidya, and Alexander Huth. Scaling laws for language encoding models in fmri. Advances in Neural Information Processing Systems , 36, 202... | https://arxiv.org/abs/2505.20027v1 |
2023 Workshop on Multimodal Representation Learning: Perks and Pitfalls , 2023a. Dota Tianai Dong and Mariya Toneva. Vision-language integration in multimodal video transformers (partially) aligns with the brain. arXiv preprint arXiv:2311.07766 , 2023b. Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weisse... | https://arxiv.org/abs/2505.20027v1 |
35:33428–33443, 2022. Camille K Milton, Vukshitha Dhanaraj, Isabella M Young, Hugh M Taylor, Peter J Nicholas, Robert G Briggs, Michael Y Bai, Rannulu D Fonseka, Jorge Hormovas, Yueh-Hsin Lin, et al. Parcellation- based anatomic model of the semantic network. Brain and Behavior , 11(4):e02065, 2021. Yuko Nakagi, Takuya... | https://arxiv.org/abs/2505.20027v1 |
wisher, Joshua B Tenenbaum, and Evelina Fedorenko. The neural architecture of language: Integrative modeling converges on predictive processing. Proceedings of the National Academy of Sciences , 2021. Marie St-Laurent, Basile Pinsard, Oliver Contier, Katja Seeliger, Valentina Borghesani, Julie Boyle, Pierre Bellec, and... | https://arxiv.org/abs/2505.20027v1 |
on empirical methods in natural language processing: system demonstrations , pp. 38–45, 2020. Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm. In Forty-first International Conference on Machine Learning , 2024. Daniel LK Yamins, Ha Hong, Charles F Cadieu, Ethan A Solom... | https://arxiv.org/abs/2505.20027v1 |
MODELS Details of each pretrained Transformer model are reported in Table 1 in Appendix. From the table, we can clearly observe that both multi-modal models (ImageBind and TVLT) maintain similar backbone architectures for videos but differ in their backbone architecture for embedding audio as well as in the training st... | https://arxiv.org/abs/2505.20027v1 |
shown in Fig. 7. Our observations are as follows: (i) Cross-modal IB Concat embeddings are significantly better than TVLT Joint embeddings in semantic regions such as AG and PCC, as well as the multi-modal processing region MT. (ii) Conversely, TVLT Joint embeddings are significantly better than IB Concat embeddings in... | https://arxiv.org/abs/2505.20027v1 |
performance from early to lower layers, specifically for both TVLT joint and unimodal video models. The key finding 19 Published as a conference paper at ICLR 2025 Language: A TL0.10.20.30.40.50.6IB Concat TVLT Joint IB Concat - IB Video TVLT Joint - TVL T Video IB Concat - IB A udio TVLT Joint - TVL T Audio IB Concat ... | https://arxiv.org/abs/2505.20027v1 |
Unimodal VM IB Concat - Unimodal SM TVLT Joint - Unimodal SMNormaliz ed brain alignment Visual: L OC0.10.20.30.40.50.6IB Concat TVLT Joint IB Concat - IB Video TVLT Joint - TVL T Video IB Concat - IB A udio TVLT Joint - TVL T Audio IB Concat - Unimodal VM TVLT Joint - Unimodal VM IB Concat - Unimodal SM TVLT Joint - Un... | https://arxiv.org/abs/2505.20027v1 |
and how well the vision encoding models can predict narrative story-fMRI. Nakagi et al. (2024) analyzed fMRI related to video content viewing and found distinct brain regions associated with different semantic levels, highlighting the significance of modeling various levels of semantic content simultaneously. Subramani... | https://arxiv.org/abs/2505.20027v1 |
the model representations, or from the brain recordings. Conceptually, the results of these approaches should be the same because when the feature is removed completely from either the input or/and the target, it would not be able to further impact the observed alignment. However, practically, brain recordings are nois... | https://arxiv.org/abs/2505.20027v1 |
Multi-modal model. Light Brown Intersection (Overlap) represents the shared variance between the multi-modal and unimodal model. It indicates the extent to which both models explain overlapping neural variance in the whole brain. 0.332 0.18 0.218 IB-ConcatUnimodal VM 0.438 0.347 0.013 IB-ConcatUnimodal SM 0.475 0.277 0... | https://arxiv.org/abs/2505.20027v1 |
removing low-level features like motion energy may impact EVC performance. Interestingly, regressing out unimodal VM features does not affect speech-related information in multi-modal models, as speech models also exhibit brain predictivity in EVC. N I MPACT OF DIVERSE MODEL ARCHITECTURES ON PERFORMANCE COMPARISON Seve... | https://arxiv.org/abs/2505.20027v1 |
trainable linear projection layer, offer promise for enhanced multi-modal capabilities. Lastly, although we observe differences between the models (that we experimented with in this work) in terms of architectural variability and variability in pretraining methods, this suggests that future work could benefit from more... | https://arxiv.org/abs/2505.20027v1 |
arXiv:2505.20045v1 [cs.CL] 26 May 2025Uncertainty-Aware Attention Heads: Efficient Unsupervised Uncertainty Quantification for LLMs Artem Vazhentsev1,2Lyudmila Rvanova2,4Gleb Kuzmin2,4Ekaterina Fadeeva6 Ivan Lazichny2Alexander Panchenko1,2Maxim Panov3Timothy Baldwin3,5 Mrinmaya Sachan6Preslav Nakov3Artem Shelmanov3 1Sk... | https://arxiv.org/abs/2505.20045v1 |
on long-form generation tasks [ 52,44]. Sampling-based scores offer stronger performance, but incur large computational overhead [ 27,31,42]. Supervised confidence regressors [ 1,6], i.e., thin supplementary modules trained on supervised annotation, yield accurate scores, but require costly, task–specific annotation an... | https://arxiv.org/abs/2505.20045v1 |
the prompt – key prompt elements that narrow down the scope of the answer. However, their experiments show that SAT Probe performs only on par with or slightly better than baselines. In a similar vein, Contextualized Sequence Likelihood [ 30] leverages attention to important tokens in the input context to reweight the ... | https://arxiv.org/abs/2505.20045v1 |
i,i−1– attention weight to the {i−1}-th token during the generation of i-th token from the layer land attention head h. LetNbe the number of generated tokens in the answer, Hthe number of attention heads in each layer, and Lbe the number of layers in the LLM. For illustration, we use the Llama 3.1 8B model. Difference ... | https://arxiv.org/abs/2505.20045v1 |
that certain heads consistently assign higher average attention when the LLM generates correct answers as compared to incorrect ones. Moreover, there is a notable correlation between the quality of the answer and average attention (see Figure 3b). This way, we empirically discovered a pattern for assessing the correctn... | https://arxiv.org/abs/2505.20045v1 |
leading us to focus solely on it in our method design and subsequent experiments. Below, we leverage the insights from this mechanistic investigation to develop a new unsupervised UQ method for LLMs. 4 RAUQ: Recurrent Attention-Based Uncertainty Quantification Method Letxbe the input sequence and y=y1y2. . . yNbe its c... | https://arxiv.org/abs/2505.20045v1 |
Following previous work [ 44], we select these layers from the middle of the model. An ablation study with various aggregation functions is presented in Section 5.3. The step-by-step description of RAUQ is presented in Algorithm 1. 5 Experiments 5.1 Experimental Setup We conducted extensive experiments across three key... | https://arxiv.org/abs/2505.20045v1 |
Falcon-3 10B [13]. Detailed descriptions of generation parameters are presented in Table 3 in Appendix A. Uncertainty quantification baselines. We compare the proposed RAUQ method with 15 diverse UQ baselines. As a sanity check, we include simple unsupervised baselines such as Maximum Sequence Probability (MSP) and Per... | https://arxiv.org/abs/2505.20045v1 |
0.10.20.3SamSum 0.0 0.5 1.0 0.00.10.20.3CNN 0.0 0.5 1.0 0.200.250.300.350.40WMT14 0.0 0.5 1.0 0.20.30.40.5WMT19 0.0 0.5 1.0 0.1 0.00.10.2MedQUAD 0.0 0.5 1.0 0.00.10.20.3PRRTruthfulQA 0.0 0.5 1.0 0.100.150.200.25CoQA 0.0 0.5 1.0 0.20.30.40.5SciQ 0.0 0.5 1.0 0.30.40.5TriviaQA 0.0 0.5 1.0 0.500.550.600.65MMLU 0.0 0.5 1.0 ... | https://arxiv.org/abs/2505.20045v1 |
QA datasets (SciQ, TriviaQA, and GSM8k), mean aggregation yields the best performance. For MMLU, the sum of logarithms substantially outperforms other aggregation strategies, while median performs best for XSum. However, the top two performing methods are those that apply length normalization. Among them, the mean of l... | https://arxiv.org/abs/2505.20045v1 |
further boosts the average performance by 0.035, with a large gain of 0.182 on MMLU. Nevertheless, our full method further incorporates to- ken probabilities and recurrently aggregates uncertainty scores from previous generation steps, which provides a distinct advantage. Overall, these results suggest that our finding... | https://arxiv.org/abs/2505.20045v1 |
Pecina, M. Post, H. Saint-Amand, R. Soricut, L. Specia, and A. Tamchyna. Findings of the 2014 workshop on statistical machine translation. In O. Bojar, C. Buck, C. Federmann, B. Haddow, P. Koehn, C. Monz, M. Post, and L. Specia, editors, Proceedings of the Ninth Workshop on Statistical Machine Translation , pages 12–58... | https://arxiv.org/abs/2505.20045v1 |
Representing model uncer- tainty in deep learning. In M. F. Balcan and K. Q. Weinberger, editors, Proceedings of The 33rd International Conference on Machine Learning , volume 48 of Proceedings of Machine Learning Research , pages 1050–1059, New York, New York, USA, 20–22 Jun 2016. PMLR. [17] J. Geng, F. Cai, Y . Wang,... | https://arxiv.org/abs/2505.20045v1 |
(Volume 1: Long Papers) , pages 1601–1611, Vancouver, Canada, July 2017. Association for Computational Linguistics. [27] L. Kuhn, Y . Gal, and S. Farquhar. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. In The Eleventh International Conference on Learning Represe... | https://arxiv.org/abs/2505.20045v1 |
Linguistics. [38] M. Rivière, S. Pathak, P. G. Sessa, C. Hardin, S. Bhupatiraju, L. Hussenot, T. Mesnard, B. Shahriari, A. Ramé, J. Ferret, et al. Gemma 2: Improving open language models at a practical size. CoRR , 2024. [39] A. See, P. J. Liu, and C. D. Manning. Get to the point: Summarization with pointer-generator n... | https://arxiv.org/abs/2505.20045v1 |
and error regularization for natural language processing. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers) , pages 1040–1051, Online, aug 2021. Association for Computational Lingu... | https://arxiv.org/abs/2505.20045v1 |
NLI Score Entail. 6.40±1.76 450% Lexical Similarity ROUGE-L 6.11±1.75 425% Semantic Entropy 6.40±1.76 450% SAR 10.71±3.21 820% Semantic Density 6.27±1.76 438% RAUQ 1.17±0.45 0.3% 15 C Results of Ablation Studies Table 5: PRR ↑for Llama 8b v3.1 model for various aggregation function of token-level confidence scores. The... | https://arxiv.org/abs/2505.20045v1 |
Qwen-2.5 7B Gemma-2 9B Falcon-3 10BMeanQA Summ MT QA Summ MT QA Summ MT QA Summ MT MSP .711 .528 .686 .700 .611 .685 .746 .547 .683 .721 .549 .688 .655 Perplexity .701 .420 .690 .705 .477 .713 .735 .420 .699 .713 .477 .715 .622 CCP .685 .525 .648 .668 .579 .658 .729 .536 .646 .703 .518 .657 .629 Attention Score .497 .5... | https://arxiv.org/abs/2505.20045v1 |
XSum SamSum CNN WMT19 TruthfulQA CoQA SciQ TriviaQA MMLU GSM8k Mean Factoscope -.250 .033 .086 .120 .064 .033 .313 .363 .585 .121 .147 SAPLMA .259 .420 .082 .548 .252 -.002 .399 .399 .456 .358 .317 MIND .482 .415 .187 .451 .373 .263 .499 .517 .727 .570 .448 Sheeps -.240 .326 .260 .509 .370 .423 .552 .594 .723 .604 .412... | https://arxiv.org/abs/2505.20045v1 |
.158 .245 Semantic Entropy -.055 .200 .083 .252 .379 .093 .107 .232 .347 .479 .157 .366 .220 SAR .236 .314 .165 .306 .435 .107 .181 .297 .439 .552 .275 .320 .302 Semantic Density -.057 .067 .119 .233 .295 .175 .302 .380 .448 .571 .237 .197 .247 RAUQ .566 .269 .290 .394 .509 .241 .364 .265 .506 .522 .549 .323 .400 Table... | https://arxiv.org/abs/2505.20045v1 |
NLI Score entail. .157 .189 .069 .202 .302 .176 .159 .304 .389 .615 .398 .284 .270 Lexical Similarity Rouge-L .076 .193 .126 .279 .404 -.035 .113 .319 .395 .585 .418 .346 .268 EigenScore .085 .135 .138 .204 .249 -.024 .132 .270 .359 .519 .371 .241 .223 LUQ .240 .303 .074 .242 .276 .222 .250 .301 .342 .618 .440 .237 .29... | https://arxiv.org/abs/2505.20045v1 |
factually incorrect token Japan (the correct answer is either San Francisco ,California , orunknown place ). 0 5 10 15 20 25 30 Attention Head The UFO in question is the Ros well UFO.T okens in AnswerQuestion: Which UFO has been demonstrated by many scientists to be extraterrestrial in origin? 0.10.20.30.40.50.60.7 Fig... | https://arxiv.org/abs/2505.20045v1 |
REARANK : Reasoning Re-ranking Agent via Reinforcement Learning Le Zhang1,2*Bo Wang3∗Xipeng Qiu3 Siva Reddy1,4,5Aishwarya Agrawal1,2,5 1Mila - Quebec AI Institute2Université de Montréal3Fudan University 4McGill University5Canada CIFAR AI Chair Abstract We present REARANK , a large language model (LLM)-based listwise re... | https://arxiv.org/abs/2505.20046v1 |
processes within these mod- 1arXiv:2505.20046v1 [cs.IR] 26 May 2025 els frequently lack transparent and interpretable reasoning, which limits explainability and fails to leverage test-time scaling properties of LLMs; (iv) State-of-the-art reranking agents frequently depend on large, often proprietary models (e.g., GPT-... | https://arxiv.org/abs/2505.20046v1 |
general task improvements. These developments enable LLM application in complex domains like math problems and planning. Our work applies these advanced reasoning capabilities to reranking. LLMs for Re-ranking LLMs are increasingly being used for reranking, moving beyond tradi- tional feature-based models (Zhang et al.... | https://arxiv.org/abs/2505.20046v1 |
passages list, often from the end towards the beginning. This window is typically shifted byw/2steps to create overlap, resulting in approxi- mately O(2n/w)total LLM calls for npassages and offering significant efficiency advantages. 3.2 RL for Listwise Re-ranking A common mathematical framework for reinforce- ment lea... | https://arxiv.org/abs/2505.20046v1 |
), as illustrated in fig. 1. The total reward ris a composite signal designed to encourage both high ranking perfor- mance and adherence to the desired output format. The primary reward signal is based on the rich, order-based information inherent in listwise rerank- ing, measured by Normalized Discounted Cumu- lative ... | https://arxiv.org/abs/2505.20046v1 |
ranking data from limited annotations, enabling the model to learn robustly from diverse initial conditions and significantly reducing the need for large-scale, fully annotated query sets. 4 Experiments 4.1 Experimental Setup Training Details Training data instances are gen- erated by randomly sampling 20 candidate pas... | https://arxiv.org/abs/2505.20046v1 |
& OOD Retrieval Results As shown in table 1, GPT-4 achieves the best perfor- mance across benchmarks, due to its superior text understanding. The Qwen2.5-7B also performs strongly, surpassing the legacy GPT-3.5 on both sets. Our REARANK , based on Qwen2.5-7B trained via our RL approach with reasoning ability, demon- st... | https://arxiv.org/abs/2505.20046v1 |
SFT generalizes poorly to out-of- domain and reasoning-intensive scenarios. Comparing with the concurrent Setwise Rank- R1, our listwise REARANK -7B achieves better per- formance on the BRIGHT benchmark also. This superior performance, particularly against another RL-trained model, underscores the effectiveness of our ... | https://arxiv.org/abs/2505.20046v1 |
rrank=Srerank ") results in lower performance due to high variance. Subtract- ing the initial score ("w/ rrank=Srerank−Sinit") improved stability but still underperforms REAR- ANK. This could be due to small learning signals as a result of small reward value scale. Our full model’s normalized reward function provided m... | https://arxiv.org/abs/2505.20046v1 |
Relevant keywords: ocelli, insects, flight, image focusing.-Passage [5]: Discusses the perception of darkness and the physics of light absorption but does not address insect attraction to light or heat radiation. Relevant keywords: perception of darkness, light absorption, physics.</think>After RL Claim in article abou... | https://arxiv.org/abs/2505.20046v1 |
AMC Qwen2.5-7B-Instruct 11.87 51.41 (Ours) Rearank-7B 12.92 52.66 Table 5: Math Reasoning Transfer Results (Pass@1) Figure 6: Impact of reasoning length. Data points are binned into 20 equal-width intervals by token count. 8 Impact of Reasoning Length on Performance We analyzed the impact of reasoning length on REARANK... | https://arxiv.org/abs/2505.20046v1 |
of Machine Learning Research , 5(Nov):1471–1530. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shi- rong Ma, Peiyi Wang, Xiao Bi, and 1 others. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948 . Takeshi Koji... | https://arxiv.org/abs/2505.20046v1 |
Improving passage retrieval with zero-shot question generation. arXiv preprint arXiv:2204.07496 .John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. 2015. High-dimensional continuous control using generalized advantage esti- mation. arXiv preprint arXiv:1506.02438 . Zhihong Shao, Peiyi Wang... | https://arxiv.org/abs/2505.20046v1 |
Wu, Fengran Mo, Jian-Yun Nie, and Aishwarya Agrawal. 2023. Moqagpt: Zero- shot multi-modal open-domain question answer- ing with large language model. arXiv preprint arXiv:2310.13265 . Le Zhang, Yihong Wu, Qian Yang, and Jian-Yun Nie. 2024b. Exploring the best practices of query ex- pansion with large language models. ... | https://arxiv.org/abs/2505.20046v1 |
descending order using identifiers. The most relevant passages should be listed first. The output format should be [] > [] , e.g., [1] > [2] . Only response the ranking results, do not say any word or explain. Figure 7: Multiple reranking results of R EARANK . Multiple Rerank Pass Analyzing the im- pact of multiple rer... | https://arxiv.org/abs/2505.20046v1 |
blindness generally consists of a person using the wrong color for an object, such as when painting, or calling a color by the wrong name. The colors that are confused are very consistent among people with the sametype of color blindness. Normal sight Deuteranopic sight ProtanPassage 3: more or less saturateddepending ... | https://arxiv.org/abs/2505.20046v1 |
arXiv:2505.20047v1 [cs.CL] 26 May 2025Grammars of Formal Uncertainty: When to Trust LLMs in Automated Reasoning Tasks Debargha Ganguly1, Vikash Singh1, Sreehari Sankar1, Biyao Zhang1, Xuecen Zhang1, Srinivasan Iyengar2,Xiaotian Han1,Amit Sharma3,Shivkumar Kalyanaraman2,Vipin Chaudhary1 1Case Western Reserve University2... | https://arxiv.org/abs/2505.20047v1 |
ensembles. This approximation not only identifies the most likely solutions but also reveals strategic diversity, common structural motifs, and areas of high model uncertainty. Deriving a comprehensive suite of metrics from this structured, quantifiable understanding of uncertainty can then directly guide the verificat... | https://arxiv.org/abs/2505.20047v1 |
assigning a probability p(r)to each rule r∈R. For each non-terminal A∈V, these probabilities must satisfyP r∈RAp(r) = 1 , where RAdenotes the set of rules with Aas their left-hand side. The probability of a derivation πthat applies rules r1, . . . , r kin sequence is p(π) =Qk i=1p(ri). Consequently, for any terminal st... | https://arxiv.org/abs/2505.20047v1 |
+α)/(C(A) +α|RA|), as computing the posterior mean. This is under a symmetric Dirichlet prior, Dir(α, . . . , α ), over rule choices for each non-terminal, where the concentration parameter α >0reflects the prior’s strength. Another approach is the Neural PCFG Model . This model utilizes a neural network fϕ(r)to score ... | https://arxiv.org/abs/2505.20047v1 |
The Rényi Entropy per Non-terminal ,Hα(A), generalizes Shannon entropy and is parameterized by an order α≥0. For α̸= 1:Hα(A) = 1 1−αlog2P A→β∈RAp(A→β)αKey special cases include Shannon entropy ( H1(A) =H(A) asα→1), max-entropy ( H0(A) = log2|RA|forα= 0, reflecting the number of choices), collision entropy ( H2(A) =−log... | https://arxiv.org/abs/2505.20047v1 |
predictive accuracy, calibration, and robustness by combining varied uncertainty signals. Metrics for Evaluating Uncertainty-Based Error Detection To evaluate uncertainty quantification (UQ) methods for identifying prediction errors, we examine several facets: Error Discrimination utilizes the Area Under the Receiver O... | https://arxiv.org/abs/2505.20047v1 |
0.2257 0.9333 0.5961 0.3372 DeepSeek R1 0.8580 0.7760 0.0820 0.9939 0.7440 0.2499 0.9423 0.4935 0.4488 0.9252 0.5200 -0.4052 Flash 2.0 0.7188 0.5360 0.1828 0.9820 0.9000 0.0820 0.4900 0.6660 -0.1760 0.9010 0.5625 0.3385 Flash 2.0 Lite 0.6760 0.4500 0.2260 0.9980 0.9980 0.0000 0.4060 0.7540 -0.3480 0.9017 0.7321 0.1696 ... | https://arxiv.org/abs/2505.20047v1 |
at the level of random guessing), we cannot extract information about failure from them. Task-Dependent Signal Dominance in SMT vs Ground Truth Prediction: Knowledge-Intensive Reasoning: For StrategyQA, cross-modal agreement metrics consistently dom- inated. O3-mini showed strong performance with grammar entropy (AUROC... | https://arxiv.org/abs/2505.20047v1 |
Factor 0.5181 0.1500 0.1997 0.1990 0.6180 0.1450 0.2270 0.2688 0.5914 0.2979 0.1377 0.0055 0.5745 0.2189 0.2293 0.1355 Rule Dist Mean 0.5740 0.3161 0.2836 0.1752 0.6021 0.1811 0.2534 0.2616 0.9301 0.4945 0.3034 0.0008 0.5838 0.4368 0.3713 0.1301 Rule Dist StdDev 0.5291 0.3995 0.3517 0.1811 0.5281 0.1251 0.2573 0.3116 0... | https://arxiv.org/abs/2505.20047v1 |
Renyi Ent (0.5) 0.6149 0.3942 0.2755 0.0376 0.6667 0.3333 0.2751 0.1223 0.6994 0.2766 0.2619 0.5438 0.8925 0.5063 0.3246 0.0013 Max Ent 0.7174 0.3994 0.2341 0.0262 0.6353 0.1309 0.1738 0.1247 0.5982 0.3401 0.3083 0.5900 0.9355 0.2589 0.1029 0.0008 Ent Ratio 0.5217 0.5047 0.3529 0.0543 0.6154 0.4091 0.3307 0.1446 0.5511... | https://arxiv.org/abs/2505.20047v1 |
formal languages. Finally, our fine-grained localized entropy within PCFG production rules surpasses global or non-grammatical standard techniques in error prediction, confirming that granular structural uncertainty in specific grammatical constructs directly flags component-level semantic error likelihood, offering mo... | https://arxiv.org/abs/2505.20047v1 |
Reasoning Research explores various uncertainty estimation approaches in language models: information-theoretic methods using entropy [Kadavath et al., 2022, Kuhn et al., 2023, Duan et al., 2024], perplexity [Mora-Cross and Calderon-Ramirez, 2024, Margatina et al., 2023], and mutual information [Malinin, 2019, Wimmer e... | https://arxiv.org/abs/2505.20047v1 |
Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems , 33:1877–1901, 2020. Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde De Oliveira Pinto, Jared Kapl... | https://arxiv.org/abs/2505.20047v1 |
Brian Wong, Malcolm Sailor, Ansong Ni, Linyong Nan, Jungo Kasai, Tao Yu, Rui Zhang, Alexander Fabbri, Wojciech Maciej Kryscinski, Semih Yavuz, Ye Liu, Xi Victoria Lin, Shafiq Joty, Yingbo Zhou, Caiming Xiong, Rex Ying, Arman Cohan, and Dragomir Radev. FOLIO: Natural language reasoning with first-order logic. In Yaser A... | https://arxiv.org/abs/2505.20047v1 |
and Greg Durrett. SATLM: Satisfiability-Aided Language Mod- els using Declarative Prompting. In Proceedings of the International Conference on Neural Information Processing Systems , 2023. Jin Peng Zhou, Charles E Staats, Wenda Li, Christian Szegedy, Kilian Q Weinberger, and Yuhuai Wu. Don’t Trust: Verify–Grounding LLM... | https://arxiv.org/abs/2505.20047v1 |
Large Language Models’ Understanding of Math: Source Criticism and Extrapolation. arXiv preprint arXiv:2311.07618 , 2023. Gregor vom Scheidt. Experimental Results from Applying GPT-4 to An Unpublished Formal Language. arXiv preprint arXiv:2305.12196 , 2023. Simon Frieder, Julius Berner, Philipp Petersen, and Thomas Luk... | https://arxiv.org/abs/2505.20047v1 |
Sebastian Farquhar. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. arXiv preprint arXiv:2302.09664 , 2023. Jinhao Duan, Renming Zhang, James Diffenderfer, Bhavya Kailkhura, Lichao Sun, Elias Stengel- Eskin, Mohit Bansal, Tianlong Chen, and Kaidi Xu. Gtbench: Unco... | https://arxiv.org/abs/2505.20047v1 |
Huang, Zichen Wu, Yutong Yang, Junzhao Zhang, and Yunfang Wu. Unc-ttp: A method for classifying llm uncertainty to improve in-context example selection, 2024b. URL https: //arxiv.org/abs/2408.09172 . Jiaxin Zhang, Zhuohang Li, Kamalika Das, Bradley A. Malin, and Sricharan Kumar. Sac3: Reliable hallucination detection i... | https://arxiv.org/abs/2505.20047v1 |
Iwa- sawa. Large language models are zero-shot reasoners. In S. Koyejo, S. Mo- hamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors, Advances in Neu- ral Information Processing Systems , volume 35, pages 22199–22213. Curran Associates, Inc., 2022. URL https://proceedings.neurips.cc/paper_files/paper/2022/file/ 8b... | https://arxiv.org/abs/2505.20047v1 |
Nx, and the above becomes 2H(µ) Nxexp(x) = 1 ⇐⇒ x ex=N 2H(µ). By definition of the Lambert W-function, x=W N 2H(µ) . Substituting back, we get ϵ=2H(µ) NW N 2H(µ) . Thus ϵdecreases to 0asN→ ∞ , roughly like 2H(µ) Nln N 2H(µ) . Hence, the probability of missing anyset of mass at least ϵis≤ϵ, with ϵscaling on the ord... | https://arxiv.org/abs/2505.20047v1 |
0.0toTmax= 2.0). Distinct Probabilistic Context-Free Grammars (PCFGs) were induced from the SMT program ensembles parsed at each temperature point Ti, modeling the LLM’s syntactic and structural tendencies under each generative condition. Our analysis of these per-temperature PCFGs revealed distinct and significant tre... | https://arxiv.org/abs/2505.20047v1 |
language space defined by GSMT . This expansion, however, may also come with its own emergent structural specificities, as indicated by the KL divergence. These findings are crucial for understanding the coherence-diversity trade-off, for validating the sensitivity of PCFG-derived metrics, and for interpreting uncertai... | https://arxiv.org/abs/2505.20047v1 |
Accuracy Precision Recall F1 TP TN FP FN o3-mini (medium effort) 1.0000 1.0000 1.0000 1.0000 499 0 0 0 0.9980 1.0000 0.9980 0.9889 499 0 0 1 Deepseek v3 0324 1.0000 1.0000 1.0000 1.0000 450 0 0 0 0.4501 1.0000 0.4501 0.6200 221 0 0 270 DeepSeek R1 0.9939 1.0000 0.9939 0.9969 489 0 0 3 0.7440 1.0000 0.7440 0.8532 372 0 ... | https://arxiv.org/abs/2505.20047v1 |
Gemini Flash 2.0 Lite 0.4060 0.3609 0.2440 0.2911 61 142 108 189 0.7540 0.7275 0.8120 0.7674 203 174 76 47 Table 8: LLM Performance on FOLIO (Text vs. SMT): Textual reasoning largely outperforms SMT. For many models, SMT results in high recall but poor precision (e.g., Gemini Flash 2.0 SMT F1 0.72 vs Text 0.92) and a f... | https://arxiv.org/abs/2505.20047v1 |
PCFG metrics exhibit weaker performance. The self-consistency metrics (Text and SMT) also perform well (AUROC 0.74), indicating that agreement between the LLM’s own reasoning modalities is a key signal. Notably, the Ensemble ML method achieves the highest AUROC (0.7850) and a significant relative error reduction (29.29... | https://arxiv.org/abs/2505.20047v1 |
11: Uncertainty Quantification for DeepSeek-v3 on StrategyQA: Ensemble ML leads with an AUROC of 0.7709. Several individual PCFG-based metrics like Grammar Entropy (AUROC 0.7087) and Max Entropy (AUROC 0.6851) show reasonable efficacy, outperforming self-consistency measures for this model. Metric AUROC ECE Brier AURC ... | https://arxiv.org/abs/2505.20047v1 |
the best overall performance (AUROC 0.7631, AURC 0.0823), leading to a 14.61% relative error reduction when abstaining on 5% of the samples. This suggests that for Gemini 2.0 Flash Lite on ProofWriter, the structural variations in its SMT outputs are less consistently tied to semantic correctness compared to o3-mini. I... | https://arxiv.org/abs/2505.20047v1 |
0.0680 Max Ent 0.5417 0.3503 0.3045 0.1420 0.5000 0.1717 0.0811 Ent Ratio 0.5177 0.3943 0.3426 0.1548 0.2000 0.1835 0.0178 Spectral Factor 0.5011 0.5048 0.4157 0.1578 0.0000 0.1869 0.0000 Spectral Radius 0.5011 0.3930 0.3172 0.1578 0.0000 0.1869 0.0000 # Nonterminals 0.5167 0.4838 0.4215 0.1672 0.1000 0.1854 0.0079 # R... | https://arxiv.org/abs/2505.20047v1 |
0.4408 0.2855 0.0443 0.3000 0.0580 0.1801 NSUI 0.5963 0.2529 0.1433 0.0462 0.1000 0.0562 0.2055 Renyi Ent (2) 0.6242 0.3980 0.2746 0.0378 0.5000 0.0204 0.7114 Renyi Ent (0.5) 0.6149 0.3942 0.2755 0.0376 0.5000 0.0408 0.4227 Max Ent 0.7174 0.3994 0.2341 0.0262 0.0500 0.0638 0.0973 Ent Ratio 0.5217 0.5047 0.3529 0.0543 0... | https://arxiv.org/abs/2505.20047v1 |
Radius 0.7034 0.1896 0.1752 0.1206 0.1000 0.1667 0.2063 # Nonterminals 0.5549 0.2271 0.2455 0.148 0.0500 0.2000 0.0476 # Rules 0.5675 0.2025 0.2077 0.1485 0.1000 0.1778 0.1534 Avg Rules / NT 0.6034 0.1393 0.1854 0.1399 0.0500 0.1895 0.0977 Avg RHS Len 0.5208 0.1254 0.1818 0.1972 0.0500 0.2000 0.0476 Max Branch Factor 0... | https://arxiv.org/abs/2505.20047v1 |
for this model, demonstrating the power of SMT features for diagnosing internal reasoning coherence. 27 Table 16: UQ for SMT-Text Consistency (Gemini Flash 2.0 Lite, StrategyQA): Rule Distribution Kurtosis (AUROC 0.8695) from SMT generations is an exceptionally strong individual predictor of SMT-Text agreement, signifi... | https://arxiv.org/abs/2505.20047v1 |
Avg Rules / NT 0.8011 0.4394 0.2554 0.0026 0.25 0.00 1.00 Avg RHS Len 0.5054 0.6535 0.5116 0.0074 0.50 0.00 1.00 Max Branch Factor 0.7419 0.6809 0.5100 0.0032 0.30 0.00 1.00 Rule Dist Mean 0.8011 0.4842 0.2971 0.0026 0.25 0.00 1.00 Rule Dist StdDev 0.8710 0.4232 0.2392 0.0013 0.15 0.00 1.00 Rule Dist Skew 0.8172 0.4864... | https://arxiv.org/abs/2505.20047v1 |
of the many samples, while the SMT error ratio is the proportion of incorrect answers derived from its SMT-LIB formalizations. O3-mini exhibits a 29 notable correlation between its SMT and Text error distributions, characteristic of a well-calibrated SMT generation process where formalization errors tend to align with ... | https://arxiv.org/abs/2505.20047v1 |
and connectives within a monolithic assert statement. In contrast, other samples might exhibit a preference for flatter, more direct assertions or decompose a complex axiom into several simpler, conjoined assert statements. This divergence in logical decomposition strategies is captured by differing rule probabilities ... | https://arxiv.org/abs/2505.20047v1 |
arXiv:2505.20050v1 [eess.AS] 26 May 2025MVP: Multi-source Voice Pathology detection Alkis Koudounas*1, Moreno La Quatra*2, Gabriele Ciravegna1, Marco Fantini3,4, Erika Crosetti4, Giovanni Succo4,5, Tania Cerquitelli1, Sabato Marco Siniscalchi6, Elena Baralis1 1Politecnico di Torino, Italy2Kore University of Enna, Italy... | https://arxiv.org/abs/2505.20050v1 |
readings through specialized transformer models [21–23], and allows a comprehensive as- sessment of vocal health. We specifically employ models pre- trained on LibriSpeech [24] to analyze sentences and models pre-trained on AudioSet [25] to process sustained vowels. Our method builds on recent advances in transformer-b... | https://arxiv.org/abs/2505.20050v1 |
1 (middle)). •Decision-Level Combination (DLC). This approach mimics the pipeline proposed in [6], where an ensemble dynamically selects between source-specific predictions by averaging in- dividual backbones probabilities (Fig. 1 (bottom)). 2.2.1. Intermediate Feature Fusion Given two feature sequences HSV∈RTSV×dandHS... | https://arxiv.org/abs/2505.20050v1 |
Acc. F1 AUC Acc. F1 AUC Single-Source Baselines LS→Sent 94.64M .873 ±.058 .849 ±.062 .850 ±.048 .872 ±.015 .871 ±.014 .877 ±.015 .875 ±.024 .870 ±.026 .847 ±.026 LS→Vowel 94.64M .747 ±.075 .724 ±.074 .732 ±.084 .714 ±.051 .705 ±.061 .714 ±.061 .622 ±.064 .617 ±.062 .620 ±.062 AS→Sent 94.64M .817 ±.060 .801 ±.061 .810 ±... | https://arxiv.org/abs/2505.20050v1 |
normalized to zero mean and unit 1https://github.com/koudounasalkis/MVPvariance. For consistent processing, recordings are padded or truncated to a fixed length of 5.0 seconds. We use an AdamW optimizer with 5e-5 learning rate and 0.01 weight decay. Train- ing runs for 10 epochs with early stopping (patience=5) on val-... | https://arxiv.org/abs/2505.20050v1 |
our multi-source approach signifi- cantly outperforms single-source baselines across all datasets. The IFF-TE method with fine-tuned backbones achieves the highest AUC scores: 95.8% (SVD), 96.3% (A VFAD), and 93.6% (IPV). This represents a 10-13% improvement over the best single-source baseline, showing the clear advan... | https://arxiv.org/abs/2505.20050v1 |
role of the model architecture and special- ized pre-training for each source. The combination of HuBERT models pre-trained on LibriSpeech (LS) for sentences and on AudioSet (AS) for sustained vowels consistently outperforms other configurations across all datasets. This aligns with re- cent works [6] showing that sent... | https://arxiv.org/abs/2505.20050v1 |
Akbulut, A. am Zehnhoff-Dinnesen, F. de Jong, M. Echter- nach, U. Eysholdt, M. Fuchs, T. Hacki, K. Izdebski, A. Keil- mann, P. Kummer et al. , “Basics of voice disorders,” in Phoni- atrics I: Fundamentals–Voice Disorders–Disorders of Language and Hearing Development . Springer, 2019, pp. 193–238. [5] S. M. Cohen, J. Ki... | https://arxiv.org/abs/2505.20050v1 |
Lee, “V ocal fold nodules: A disorder of phonation organs or auditory feedback?” Clinical Otolaryngology , 2019. [20] G. Schlotthauer, M. E. Torres, and M. C. Jackson-Menaldi, “A pattern recognition approach to spasmodic dysphonia and mus- cle tension dysphonia automatic classification,” Journal of voice , 2010. [21] W... | https://arxiv.org/abs/2505.20050v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.