text string | source string |
|---|---|
Summarization and Mathe- matical Reasoning are Length-Dependent. Fig- ure 3 reports quality metric trends for XSum (sum- marization) and GSM8K (arithmetic question an- swering). In contrast to machine translation, these tasks show mild yet noticeable correlations be- tween quality scores and output length. Specif- ical... | https://arxiv.org/abs/2505.19060v1 |
debias raw uncertainty scores on the test set. To achieve this, for each test example, we compute a length-debiased uncertainty score by subtracting the length-predicted component from its raw score: udeb(y) =u(y)−ˆu(y). (2) This subtraction step is equivalent to computing the residuals from the fitted regression, a st... | https://arxiv.org/abs/2505.19060v1 |
for tasks like summarization and mathematical reasoning, where quality often corre- lates with length, we retain the quality-associated component while eliminating only the spurious bias. Models. We use three base versions of multilingual generative language models to generate outputs for all datasets: Llama 3.1 8B (To... | https://arxiv.org/abs/2505.19060v1 |
LINE Base LINE Base LINE Base LINE Base LINE MetricX XXL Llama 3.1 8B 0.47 0.54 ↑ 0.48 0.51 ↑ 0.46 0.54 ↑ 0.39 0.43 ↑ 0.43 0.47 ↑ 0.52 0.51 0.49 0.49 ↑ 0.36 0.45 ↑ Gemma 2 9B 0.45 0.46 ↑ 0.47 0.49 ↑ 0.42 0.46 ↑ 0.36 0.37 ↑ 0.44 0.47 ↑ 0.45 0.45 0.34 0.37 ↑ 0.38 0.41 ↑ EuroLLM 9B 0.54 0.55 ↑ 0.54 0.55 ↑ 0.48 0.47 0.46 0... | https://arxiv.org/abs/2505.19060v1 |
in some settings, additional sources of uncertainty may dominate. Detailed experimental results with breakdown of PRR scores before and after UNCERTAINTY -LINE transformation for each of the UQ methods are provided in Appendix C. Table 3 reports improvements in PRR scores af- terUNCERTAINTY -LINE transformation for eac... | https://arxiv.org/abs/2505.19060v1 |
effective first-order correction. Nonetheless, in tasks such as dialogue or multi-step reasoning, where uncer- tainty may follow non-monotonic or phase-specific patterns, a linear fit may be insufficient. Additionally, our method requires a small num- ber of quality-labeled examples to estimate the quality-length relat... | https://arxiv.org/abs/2505.19060v1 |
chyna. 2014. Findings of the 2014 workshop on statistical machine translation. In Proceedings of the Ninth Workshop on Statistical Machine Translation , pages 12–58, Baltimore, Maryland, USA. Associa- tion for Computational Linguistics. Mark Braverman, Xinyi Chen, Sham M. Kakade, Karthik Narasimhan, Cyril Zhang, and Yi... | https://arxiv.org/abs/2505.19060v1 |
ICLR 2021, Virtual Event, Austria, May 3-7, 2021 . OpenReview.net. Andrey Malinin, Anton Ragni, Kate Knill, and Mark Gales. 2017. Incorporating uncertainty into deep learning for spoken language assessment. In Proceed- ings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Pa-... | https://arxiv.org/abs/2505.19060v1 |
Yichi Yang, Ruichen Li, and Zhiting Hu. 2023. AlignScore: Evaluating factual consistency with a unified alignment function. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages 11328–11348, Toronto, Canada. Association for Computational Linguistics. ... | https://arxiv.org/abs/2505.19060v1 |
Fi-En 0.0 0.2 0.4 0.6 0.8 1.00.60.70.80.9XComet-XXLWMT19 Lt-En 0.0 0.2 0.4 0.6 0.8 1.00.650.700.750.800.850.900.95XComet-XXLWMT19 Ru-En Figure 7: XComet-XXL score trends with respect to normalized generated sequence length across four machine translation datasets. Each subplot shows a linear regression fit over binned ... | https://arxiv.org/abs/2505.19060v1 |
0.0 0.2 0.4 0.6 0.8 1.0 Sequence length (normalized)0.750.800.850.900.95Align ScoreXSUM 0.0 0.2 0.4 0.6 0.8 1.0 Sequence length (normalized)0.00.20.40.60.81.0AccGSM8K Figure 9: Align Score and Accuracy trends with respect to normalized generated sequence length across four machine translation datasets. Each subplot sho... | https://arxiv.org/abs/2505.19060v1 |
0.2 0.4 0.6 0.8 1.00.10.20.30.40.5 PPL 0.0 0.2 0.4 0.6 0.8 1.00.20.30.40.5MTE 0.0 0.2 0.4 0.6 0.8 1.00.20.30.40.50.60.70.8 MCSE 0.0 0.2 0.4 0.6 0.8 1.00.20.30.40.50.60.7MCNSE 0.0 0.2 0.4 0.6 0.8 1.00.50.60.70.80.9 LSRL Figure 10: Uncertainty metric trends for model LLAMA across all datasets. 16 WMT14 De-En 0.0 0.2 0.4 ... | https://arxiv.org/abs/2505.19060v1 |
0.0 0.2 0.4 0.6 0.8 1.00.30.40.50.6LSRL WMT14 Cs-En 0.0 0.2 0.4 0.6 0.8 1.00.10.20.30.4MSP 0.0 0.2 0.4 0.6 0.8 1.00.1500.1750.2000.2250.2500.275PPL 0.0 0.2 0.4 0.6 0.8 1.00.150.200.250.30MTE 0.0 0.2 0.4 0.6 0.8 1.00.10.20.30.40.50.60.7 MCSE 0.0 0.2 0.4 0.6 0.8 1.00.250.300.350.40MCNSE 0.0 0.2 0.4 0.6 0.8 1.00.300.350.4... | https://arxiv.org/abs/2505.19060v1 |
-0.057 0.000 -0.014 0.426 WMT14 De-En 0.359 0.000 -0.083 0.000 -0.117 0.000 0.340 0.000 -0.025 0.109 0.011 0.553 WMT14 Fr-En 0.368 0.000 -0.121 0.000 -0.141 0.000 0.159 0.000 -0.042 0.002 0.000 0.988 WMT14 Ru-En 0.266 0.000 -0.136 0.000 -0.195 0.000 0.227 0.000 -0.171 0.000 -0.154 0.000 WMT19 De-En 0.203 0.000 -0.043 0... | https://arxiv.org/abs/2505.19060v1 |
all sampled sequences. Unlike the previous methods, which rely on model probabilities, LSRL captures diversity among generated hypotheses by comparing their surface forms: ULSRL(x) = 1−2 M(M−1)X i<jROUGE-L (y(i),y(j)). (11) 20 C Detailed Experimental Results Tables 10, 12 and 11 contain PRR scores for all UQ methods un... | https://arxiv.org/abs/2505.19060v1 |
0.29 0.29 0.27 MCNSE 0.42 0.38 0.40 0.32 0.38 0.39 0.44 0.32 MCNSE-LINE 0.42 0.38 0.43 0.34 0.39 0.39 0.44 0.36 LSRL 0.39 0.35 0.37 0.30 0.36 0.32 0.42 0.31 LSRL-LINE 0.38 0.35 0.38 0.29 0.35 0.33 0.40 0.33 Gemma 2 9B MSP 0.19 0.22 0.29 0.13 0.28 0.06 0.24 0.27 MSP-LINE 0.39 0.45 0.41 0.29 0.45 0.35 0.35 0.39 PPL 0.42 ... | https://arxiv.org/abs/2505.19060v1 |
0.31 0.38 0.21 0.27 -0.03 0.06 0.36 MSP-LINE 0.24 0.38 0.42 0.25 0.38 0.26 0.19 0.37 PPL 0.39 0.40 0.38 0.33 0.43 0.40 0.33 0.31 PPL-LINE 0.41 0.43 0.50 0.37 0.44 0.38 0.34 0.39 MTE 0.43 0.42 0.39 0.35 0.43 0.45 0.39 0.32 MTE-LINE 0.46 0.46 0.51 0.42 0.47 0.42 0.38 0.44 MCSE 0.19 0.31 0.35 0.26 0.30 -0.01 0.08 0.34 MCS... | https://arxiv.org/abs/2505.19060v1 |
De-En Ru-En Fr-En De-En Fi-En Lt-En Ru-En Llama 3.1 8B MSP 0.25 0.35 0.41 0.33 0.31 0.04 0.15 0.37 MSP-LINE 1 0.33 0.38 0.41 0.33 0.37 0.39 0.38 0.34 MSP-LINE 2 0.36 0.43 0.47 0.35 0.37 0.42 0.37 0.39 MSP-LINE 3 0.37 0.43 0.46 0.29 0.39 0.39 0.38 0.41 PPL 0.36 0.35 0.30 0.24 0.33 0.49 0.49 0.27 PPL-LINE 1 0.42 0.43 0.4... | https://arxiv.org/abs/2505.19060v1 |
arXiv:2505.19073v1 [cs.CL] 25 May 2025Towards Harmonized Uncertainty Estimation for Large Language Models Rui Li1, Jing Long1, Muge Qi1, Heming Xia2, Lei Sha3, Peiyi Wang1, Zhifang Sui1 1Peking University2The Hong Kong Polytechnic University3Beihang University o_l1ru1@stu.pku.edu.cn Abstract To facilitate robust and tr... | https://arxiv.org/abs/2505.19073v1 |
al., 2023), achieves the best per- formance in indication but performs poorly in the view of calibration . Furthermore, we found that the combination of uncertainty scores obtained by existing methods provides little improvement in un- certainty estimation performance, suggesting that these methods are quite homogeneou... | https://arxiv.org/abs/2505.19073v1 |
to ex- press uncertainty, often by prompting the model to provide an uncertainty score. However, studies (Ni et al., 2024; Madhusudhan et al., 2024; Becker and Soatto, 2024) have shown that LLMs strug- gle with faithfully conveying their uncertainties, particularly due to overconfidence. Consistency- based methods , su... | https://arxiv.org/abs/2505.19073v1 |
and logit-based methods, respectively. As shown in Figure 2 and Table 1, the AUROC scores for these methods across the target models and datasets ex- hibit general low performance, which is even close to random guessing. Enhanced logit-based methods typically have low F1 scores. Some enhanced methods such as Length-nor... | https://arxiv.org/abs/2505.19073v1 |
consists of a collection of question-answer pairs, denoted as D={(qi, ai)|i= 1, . . . , n }. We then prompt Mto generate responses rifor each question qi, forming a response set R={ri| i= 1, . . . , n }. Subsequently, each response ri is subjected to a rigorous evaluation against the ground truth ai, employing a hybrid... | https://arxiv.org/abs/2505.19073v1 |
normalized score Unorm(x)with the correc- tion score C(x)generated by the Corrector . The combination employs a weighted approach, where the corrected uncertainty score Ucor(x)is computed as: Ucor(x) =w∗·Unorm(x) + (1 −w∗)·C(x)(2) The optimal weight w∗is determined through a grid search on the development dataset. This... | https://arxiv.org/abs/2505.19073v1 |
which first asks the target model to propose an answer and then evaluates it using an internal probability mechanism; and Predictive Entropy (PE) (Malinin and Gales, 2020), which calculates uncertainty by measuring the entropy of the predictive posterior. We also explore a series of advanced logit-based methods includi... | https://arxiv.org/abs/2505.19073v1 |
ECE reductions of 0.34 on TriviaQA and 0.21 on SciQA. With the LLaMA- 3-8B-Instruct model as the target, the reductions are 0.11 and 0.07, respectively—still considerable. TriviaQA SciQA AUROC( ↑) ECE( ↓) AUROC( ↑) ECE( ↓) Method Vanilla +Corrector Improv Vanilla +Corrector Improv Vanilla +Corrector Improv Vanilla +Cor... | https://arxiv.org/abs/2505.19073v1 |
Corrector’s performance can be found in Appendix A.7. 5.3 Ablation Study We conducted ablation studies to scrutinize the im- pact of the base model, the correction score formatsand its acquisition methods. Formats We compared the efficacy of probabilis- tic values versus label values for correction. As shown in Table 2... | https://arxiv.org/abs/2505.19073v1 |
uncertainty score correction framework that utilizes a classi- fier as a Corrector to refine these scores, ensuring alignment with the model’s true task performance. This Corrector integrates seamlessly with exist- ing methods, enhancing their effectiveness. Exten- sive experiments validate that the Corrector con- sist... | https://arxiv.org/abs/2505.19073v1 |
To- wards the uncertainty estimation of large language models. arXiv preprint arXiv:2307.01379 . Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783 . ... | https://arxiv.org/abs/2505.19073v1 |
estimation for large lan- guage models through reflection on multiple answers. arXiv preprint arXiv:2403.09972 . Zhen Lin, Shubhendu Trivedi, and Jimeng Sun. 2023. Generating with confidence: Uncertainty quantifi- cation for black-box large language models. arXiv preprint arXiv:2305.19187 . Linyu Liu, Yu Pan, Xiaocheng... | https://arxiv.org/abs/2505.19073v1 |
2023b. Llama 2: Open foundation and fine-tuned chat mod- els.CoRR , abs/2307.09288. Dennis Ulmer, Martin Gubri, Hwaran Lee, Sangdoo Yun, and Seong Joon Oh. 2024. Calibrating large lan- guage models using their generations only. Preprint , arXiv:2403.05973. Artem Vazhentsev, Gleb Kuzmin, Akim Tsvigun, Alexander Panchenk... | https://arxiv.org/abs/2505.19073v1 |
to judge the similarity between the target response and the generations. Duan et al. (2023) proposed Shifting Attention to Relevance (SAR), which focus on relevant components and assigns significance weights to tokens based on their con- tributions to the overall response. Unlike these carefully designed methods, Yaldi... | https://arxiv.org/abs/2505.19073v1 |
a method to set confidence targets and train an additional model that predicts an LLM’s confidence based on its textual input and output. Consistency-based method The consistency- based method is to evaluate the uncertainty of the large model through multiple generated answers. Recently, Li et al. (2024b) employed UQ s... | https://arxiv.org/abs/2505.19073v1 |
θ). Relative uncertainty scores emphasize the accuracy of sample ranking, espe- cially in discerning questions that the target model can correctly respond to from those it struggles with. Ideally, for every pair (xi, yi)and(xj, yj) with their predictive distributions YiandYj, we should have UE(xi, θ)≤UE(xj, θ)⇐⇒ P(Yi=y... | https://arxiv.org/abs/2505.19073v1 |
candidate. Base on these, Kuhn et al. (2023) proposed to cluster generations with similar meanings and compute entropy us- ing the probabilities associated with each semantic cluster. This approach is formulated as SE(x, θ) =−1 CCX i=1lnP(ci|x, θ), (13) where cidenotes each semantic cluster and Crep- resents the set of... | https://arxiv.org/abs/2505.19073v1 |
51.67 57.60 51.67 SAR-s 69.87 77.09 69.87 23.17 20.00 23.17 53.44 66.22 53.44 SAR 80.92 81.90 80.92 16.17 13.76 16.17 51.20 59.54 51.20 Table 3: Comparison of performance between the Corrector and the Wb-S method. observed limited transferability across LLMs with significantly different performance and architec- tures ... | https://arxiv.org/abs/2505.19073v1 |
training, while the topmost row represents the domains of data used for evaluating, with OPT-2.7B serving as the target model. (b) The leftmost column denotes the target model during training, whereas the topmost row signifies the target model during evaluating, with TriviaQA utilized as the target domain of data. Meth... | https://arxiv.org/abs/2505.19073v1 |
Universal Reasoner: A Single, Composable Plug-and-Play Reasoner for Frozen LLMs Jaemin Kim* Hangeol Chang* Hyunmin Hwang* Choonghan Kim Jong Chul Ye Graduate School of AI *Equal contribution Korea Advanced Institute of Science and Technology (KAIST) {kjm981995, hangeol, hyunmin_hwang, choonghankim, jong.ye}@kaist.ac.kr... | https://arxiv.org/abs/2505.19075v2 |
enhance capabilities of a backbone LLM. This lightweight reasoning module can be seamlessly integrated into an architecture-agnostic backbone model without requiring explicit information about the model’s internal structure. This approach not only fosters modularity, preserving the core capabilities of the base model, ... | https://arxiv.org/abs/2505.19075v2 |
like DeepSeek-R1 [ 8]) or 2 Figure 1: UniR Framework Overview . Our approach trains a lightweight, transferable reasoning module ( πr) using predefined rewards to guide a frozen backbone model ( πb), offering (1) trans- ferability across different backbone models or tasks; and (2) composability by combining multiple sp... | https://arxiv.org/abs/2505.19075v2 |
the expected reward while remaining close to a backbone policy ( πb(y|x)), typically a pre-trained base LLM. This objective is formulated as: max πθEx∼D,y∼πθ(y|x)[r(x, y)]−βDKL[πθ(y|x)||πb(y|x)], (1) 3 where DKL[·||·]represents the Kullback-Leibler divergence, and β >0controls the strength of the regularization towards... | https://arxiv.org/abs/2505.19075v2 |
πband the learned module πrcan be separable to produce the logits for the guided policy πθ. This additive and separable guidance mechanism extends to scenarios involving multiple reward criteria. Instead of training a single monolithic model to jointly optimize for all objectives, UniR allows for the training of severa... | https://arxiv.org/abs/2505.19075v2 |
approximate the scaled reward for trajectories ythat are consistent with the optimal policy πθ∗, then logπr(yt|x, y<t) =1 βQ∗(x, y<t, yt). (9) Proof. We assume guided policy πθhas converged to the optimal policy πθ∗for all states st= (x, y<t)and actions at=ytby optimizing the GRPO objective to maximize the reward r(x, ... | https://arxiv.org/abs/2505.19075v2 |
3B-scale πbmodel) directly on each task dataset using the same predefined rewards. This baseline is implemented through either full model tuning or LoRA-based fine-tuning. Dataset and Reward Formulation. For mathematical reasoning tasks, we use the Math-12k [ 22,24] and GSM8k [ 7] datasets. The reward r(x, y)is determi... | https://arxiv.org/abs/2505.19075v2 |
as a generalizable reasoning augmentation module. These findings highlight the potential of UniR to serve as a scalable and resource-efficient framework, enabling lightweight modules to enhance the reasoning capabilities of stronger models at inference time. Combining Specialized Reasoning Modules. To investigate the a... | https://arxiv.org/abs/2505.19075v2 |
flawed reasoning. (Right) When combined, it generates coherent reasoning and arrive at the correct solution, showing the effectiveness of the modular guidance. Figure 4: Performance on a German-to-English Math problem-solving task. The numbers in the figure indicate the value of α.where α∈[0,1]is a coefficient that bal... | https://arxiv.org/abs/2505.19075v2 |
result, our method provides greater flexibility in scaling the number of tokens processed per step under memory constraints. Furthermore, UniR remains fully compatible with a wide range of efficiency techniques, including quantization, caching, and adapter-based methods like LoRA applied to the reasoning module itself.... | https://arxiv.org/abs/2505.19075v2 |
Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923 , 2025. [5]Bespoke Labs. Bespoke-stratos: The unreasonable effectiveness of reasoning distillation. https://www.bespokelabs.ai/blog/ bespoke-stratos-the-unreasonable-effect... | https://arxiv.org/abs/2505.19075v2 |
Wei, Jinjing Zhao, Chao Zhang, and Hongyang Zhang. Rain: Your language models can align themselves without finetuning. arXiv preprint arXiv:2309.07124 , 2023. [22] Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let’s veri... | https://arxiv.org/abs/2505.19075v2 |
al. Kimi k1.5: Scaling reinforcement learning with llms. arXiv preprint arXiv:2501.12599 , 2025. [38] Masatoshi Uehara, Yulai Zhao, Chenyu Wang, Xiner Li, Aviv Regev, Sergey Levine, and Tom- maso Biancalani. Inference-time alignment in diffusion models with reward-guided generation: Tutorial and review. arXiv preprint ... | https://arxiv.org/abs/2505.19075v2 |
to ensure that various valid answer formats were correctly recognized and evaluated. A.1.2 Translation Task. For translation, inspired by the findings of MT-R1-Zero [ 11], which show that training with and without explicit reasoning yields comparable performance, we use direct translation prompts without explicit reaso... | https://arxiv.org/abs/2505.19075v2 |
within \boxed{} , prompting the model to express its answer in a boxed format. The prompt for Translation - Translation DE to EN System Prompt: You are a helpful assistant that translates German to English. User Prompt: Translate the following German sentence into English. Respond ONLY with the translation. German: {so... | https://arxiv.org/abs/2505.19075v2 |
truth and 0 otherwise. The boxed_reward 4https://github.com/huggingface/trl 15 Parameter GSM8K Math-12K-Qwen Math-12K-Llama Translation Training Configuration gradient_checkpointing true true true true gradient_checkpointing_kwargs use_reentrant: false use_reentrant: false use_reentrant: false use_reentrant: false lear... | https://arxiv.org/abs/2505.19075v2 |
total? Base Model: <think> First, calculate the total number of fireworks in the display. There are 15 boxes of 20 fireworks each, so the total number of fireworks is 15 * 20. Then, calculate the number of fireworks Hannah can see in the display, which is 40% of the total number of fireworks. Next, calculate the number... | https://arxiv.org/abs/2505.19075v2 |
Was ist mit dem künstlichen Intelligenzsystem im Film "Her", Samantha? GRPO: DE→EN: What’s with the artificial intelligence system in the movie "Her"? EN→DE: Was ist mit dem künstlichen Intelligenzsystem in dem Film "Her", Samantha? Ours: DE→EN: What about the artificial intelligence system in the movie "Her," Samantha... | https://arxiv.org/abs/2505.19075v2 |
Now, let’s calculate the time it takes to cut one carrot: 5 seconds. Since Billy needs to cut 60 carrots, we can multiply the time it takes to cut one carrot by 60: 5 seconds * 60 = 300 seconds. Now, let’s add the time it takes to prepare one carrot to the time it takes to cut one carrot: 90 seconds + 300 seconds = 390... | https://arxiv.org/abs/2505.19075v2 |
denote the number of rats as R. Each rat gets 6 pieces of straw, so the total number of pieces of straw given to rats is 6R. There are 10 cages of hamsters, and each hamster gets 5 pieces of straw. Let’s denote the number of hamsters as H. So, the total number of pieces of straw given to hamsters is 5H. The rabbits get... | https://arxiv.org/abs/2505.19075v2 |
straw, so the total number of pieces of straw given to the rats is 6×3r= 18r. The hamsters are kept in 10 cages, and each hamster gets 5 pieces of straw. So, the total number of pieces of straw given to the hamsters is 5×10 = 50 . The rabbits get 20 pieces of straw. The total number of pieces of straw distributed is 16... | https://arxiv.org/abs/2505.19075v2 |
same as the height of the triangle, which is 6 cm. The volume of a pyramid is given by the formula: \[ \text{Volume} = \frac{1}{3} \times \text{Base Area} \times \text{Height} \] Substituting the values we have: \[ \text{Volume} = \frac{1}{3} \times 30 \times 6 = 60 \text{ cm}^ 3 \] </think><answer>The volume of the so... | https://arxiv.org/abs/2505.19075v2 |
qualitative difference in responses is depicted in Figure 16. 20 C.2 Impact of Reasoning Model Size Figure 17: Transferability of the 0.5B πrand 1.5B πrreasoning modules when combined with a 14B frozen backbone model. The 1.5B πrmodule demonstrated superior performance. MethodTrained ModelIn-distribution Out-of-distrib... | https://arxiv.org/abs/2505.19075v2 |
arXiv:2505.19091v1 [cs.CL] 25 May 2025ReadBench: Measuring the Dense Text Visual Reading Ability of Vision-Language Models Benjamin Clavié Answer.AI Japan bc@answer.aiFlorian Brand Artificial Intelligence and Intelligent Information Systems, University of Trier German Research Center for Artificial Intelligence (DFKI),... | https://arxiv.org/abs/2505.19091v1 |
capabilities were introduced in PixelWorld (Lyu et al., 2025), albeit a focus on single-modality inputs. However, to represent real world usage, it is crucial that these capabilities be measured in a truly multi-modal fashion, where text instructions are given alongside image documents. Contribution In this paper, we i... | https://arxiv.org/abs/2505.19091v1 |
a result supported by recent work (Modarressi et al., 2025). LongBench We specifically select 4 subsets of the LongBench (Bai et al., 2024) benchmark repre- senting common QA datasets: HotPotQA (Yang et al., 2018), NarrativeQA (Ko ˇcisk`y et al., 2018), TriviaQA (Joshi et al., 2017) and 2WikiMulti- HopQA (Ho et al., 20... | https://arxiv.org/abs/2505.19091v1 |
to reduce variance. Overall, we observe universal performance degra- dation across models when reading textual informa- tion from images rather than purely textual inputs. Two main factors appear particularly influential: input length and task difficulty. Input Length Performance on short contexts (up to one page of te... | https://arxiv.org/abs/2505.19091v1 |
multimodal appli- cations grow in popularity, addressing these chal- lenges will be important for future developments. 4 Acknowledgements The authors thank Johno Whittaker for his infinite wisdom on visual models and helpful advice, as well as Kerem Turgutlu for his assistance in run- ning the evaluations in a computat... | https://arxiv.org/abs/2505.19091v1 |
Wang, Jun Tang, and 1 others. 2025. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923 . Yushi Bai, Xin Lv, Jiajie Zhang, Hongchang Lyu, Jiankai Tang, Zhidian Huang, Zhengxiao Du, Xiao Liu, Aohan Zeng, Lei Hou, and 1 others. 2024. Long- bench: A bilingual, multitask benchmark for long context understanding. ... | https://arxiv.org/abs/2505.19091v1 |
. Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. 2021. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on ma- chine learning , pages 4904–4916. PMLR.Mandar Joshi, Eunsol Choi,... | https://arxiv.org/abs/2505.19091v1 |
. Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christo- pher D Manning. 2018. Hotpotqa: A dataset for diverse, explainable multi-hop question answering. InProceedings of the 2018 Conference on Empirical Methods in Natural Language Processing . Associa- tion for Computati... | https://arxiv.org/abs/2505.19091v1 |
I give you context with the facts about loca- tions and actions of different persons hidden in some random text and a question. You need to answer the question based only on the information from the facts. If a person got an item in the first location and travelled to the second location the item is also in the second ... | https://arxiv.org/abs/2505.19091v1 |
information from the facts. Post-instruction Your answer should contain only one or two words: $nothing$ or $object$ or $object_1$, $object_2$. Do not write anything else. Do not explain your answer. Babilong QA9 Pre-instruction I will give you context with the facts about people and their locations hidden in some rand... | https://arxiv.org/abs/2505.19091v1 |
ASPO: Adaptive Sentence-Level Preference Optimization for Fine-Grained Multimodal Reasoning Yeyuan Wang1,*, Dehong Gao2,3, *, Rujiao Long4, Lei Yi4, Linbo Jin4,Libin Yang2,†,Xiaoyan Cai1,† 1Northwestern Polytechnical University, School of Automation, Xi’an, China 2Northwestern Polytechnical University, School of Cybers... | https://arxiv.org/abs/2505.19100v1 |
DPO methods rely on binary preference data and a coarse reward mechanism, which lacks the fine-grained prefer- ence granularity needed to identify specific errors in responses (Liao et al., 2024). This limitation hampers the model’s ability to refine its reasoning capabilities (Lai et al., 2024; Liao et al., 2024). To ... | https://arxiv.org/abs/2505.19100v1 |
optimization. 2.2 Preference Alignment Preference alignment (Chen et al., 2023a; Etha- yarajh et al., 2024) is widely used to enhance model’s instruction-following capabilities (Wang et al., 2024c). Early work primarily relied on Reinforcement Learning from Human Feedback (RLHF) (Lee et al., 2024; Wang et al., 2023). H... | https://arxiv.org/abs/2505.19100v1 |
ASPO for fine-grained multimodal reasoning. ASPO incorporates twoadaptive reward weights into DPO, i.e., the image- text similarity weight, which optimizes multimodal alignment to mitigate hallucination, and the textual perplexity weight, which enhances text confidence to improve the model’s reasoning capabilities. 3.1... | https://arxiv.org/abs/2505.19100v1 |
sentences in the response and the similarity score is calculated by CLIP (Radford et al., 2021). Next, we apply min-max normalization to scale each similarity score into the range [0,1]: S′ i=Si−Smin Smax−Smin, i= 1,2, . . . , n (7) where S′ irepresents the normalized similarity score for the i-th sentence with respect... | https://arxiv.org/abs/2505.19100v1 |
shared across all steps, and the output of the model at each step receives the same reward weight. Specifically, for the triplet (x, yc, yr), the “implicit reward” margin is mathematically defined as: M=βlogπθ(yc|x) πref(yc|x)−βlogπθ(yr|x) πref(yr|x) =KX i=1βlogπθ(sc i|x) πref(sc i|x)−LX i=1βlogπθ(sr i|x) πref(sr i|x) ... | https://arxiv.org/abs/2505.19100v1 |
– – MiniGPT-4-7B 22.1 23.0 – – 42.8 – – – 74.5 – – Shikra-13B – 58.8 – – – – – – – – – LLaV A-7B 26.7 34.1 – 14.1 25.5 63.0 38.5 – – – – IDEFICS-7B – 48.2 – 25.2 – – – 38.4 – – – IDEFICS-65B – 54.5 – 38.1 – – – 45.2 – – – mPLUG-Owl2-7B 36.2 64.5 – – – – 68.7 56.1 85.8 – – Qwen-VL-7B – 38.2 – 7.4 56.3 – 67.1 59.3 – – – ... | https://arxiv.org/abs/2505.19100v1 |
al., 2024). For preference optimization - Sample of Visual Qustion Answering: User Are the trees taller than the giraffes? LLaV A- 1.5Yes, the trees are taller than the giraffes, as they are reaching up to eat leaves from the trees. ASPO (Ours)No, the trees are not taller than the giraffes. The giraffes are eating leav... | https://arxiv.org/abs/2505.19100v1 |
some trees can be seen on the sidewalk, adding a touch of greenery to the urban environment. Table 4: ASPO provides more reliable and contextually accurate descriptions. Hallucination outputs are shown in red. in the language model are fine-tuned. We employ a learning rate of 4e-6 with a cosine learning rate scheduler ... | https://arxiv.org/abs/2505.19100v1 |
method. The example results in Table 3 show that our method is more faithful to the image input and enhances the model’s reasoning capabilities. Additionally, from the example in Table 4 on the Image Captioning task, it can be seen that after optimization with our method, the model’s output is more concise and contains... | https://arxiv.org/abs/2505.19100v1 |
generalization, warrants further in- vestigation. Finally, due to computational resource limitations, we did not extend the experiments to the base models with much more parameters to provide a clear picture of the scalability of ASPO. Acknowledgments This work was supported in part by the National Natural Science Foun... | https://arxiv.org/abs/2505.19100v1 |
Lee, and James Thorne. 2024. Orpo: Monolithic preference optimization without refer- ence model. In Conference on Empirical Methods in Natural Language Processing , pages 11170–11189. Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adap-... | https://arxiv.org/abs/2505.19100v1 |
Tpo: Aligning large language models with multi-branch & multi-step preference trees. arXiv preprint arXiv:2410.12854 . Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. 2024a. Improved baselines with visual instruc- tion tuning. In Conference on Computer Vision and Pattern Recognition , pages 26296–26306. Haotian ... | https://arxiv.org/abs/2505.19100v1 |
perspective. Neural Information Processing Systems , 36:76006–76032.Yuxi Xie, Anirudh Goyal, Wenyue Zheng, Min-Yen Kan, Timothy P Lillicrap, Kenji Kawaguchi, and Michael Shieh. 2024. Monte carlo tree search boosts reasoning via iterative preference learning. arXiv preprint arXiv:2405.00451 . Qinghao Ye, Haiyang Xu, Jia... | https://arxiv.org/abs/2505.19100v1 |
WHISTRESS : Enriching Transcriptions with Sentence Stress Detection *Iddo Yosha, *Dorin Shteyman, Yossi Adi The School of Computer Science and Engineering The Hebrew University of Jerusalem, Israel {iddo.yosha, dorin.shteyman }@mail.huji.ac.il Abstract Spoken language conveys meaning not only through words but also thr... | https://arxiv.org/abs/2505.19103v1 |
for controlled prominence in Text-to- Speech (TTS) [17]. In this work, we introduce W HISTRESS , a novel, alignment-free approach for sentence stress detection. The pro- posed approach leverages a Language Model (LM) conditioned on acoustic signals to improve sentence stress detection. We fo- cus on English speech, and... | https://arxiv.org/abs/2505.19103v1 |
one stress pattern of the two. Speech Synthesis. We use the Google Text-to-Speech API to synthesize the stressed speech based on the prior steps, by us- ing SSML syntax that enables editing prosodic features on the word or constituent level.1The following steps were taken: (i) Prosodic features adjustment: A core step ... | https://arxiv.org/abs/2505.19103v1 |
of the backbone Whisper model. The FCNN classifier is two-layer fully connected neural network that acts as a binary classifier, processing the output of the additional decoder block to assign a stress label to each token ( 1if stressed, 0otherwise). 3.2. Training Label Alignment. As a first step, the ground-truth word... | https://arxiv.org/abs/2505.19103v1 |
duration of each word in the transcription. We then compute the mean decoder embeddings corresponding to eachTable 1: Word-level sentence stress detection of WHISTRESS onTINYSTRESS -15K by input layer embeddings. Model Layer Prec Rec F1 WHISTRESS3 0.781 0.624 0.693 6 0.83 0.86 0.845 9 0.912 0.906 0.909 12 0.902 0.868 0... | https://arxiv.org/abs/2505.19103v1 |
from four speakers ( 2female, 2male) across a diverse range of speaking styles, including both impro- vised and read speech. Following the filtering criteria in [29], we include only samples that contain at least one emphasized word. For fair comparison, we use the same test set configura- tion as [29], selecting the s... | https://arxiv.org/abs/2505.19103v1 |
(i) 0-shot and evaluation over all four speakers, and (ii) Training on speakers ex03, ex04, and evaluation over speakers ex01, ex02 (marked with *). Baseline variants are de- noted ”MFA” and ”GT”, as explained in section 5.2. Evaluated Dataset Model Prec Rec F1 TS-15KBaseline (+GT alignment) 0.862 0.853 0.858 Baseline ... | https://arxiv.org/abs/2505.19103v1 |
2000. [6] S. Kakouros and O. R ¨as¨anen, “3pro – an unsupervised method for the automatic detection of sentence prominence in speech,” Speech Communication , vol. 82, p. 67–84, 09 2016. [7] T. Mishra, V . R. Sridhar, and A. Conkie, “Word prominence de- tection using robust yet simple prosodic features,” in Interspeech ... | https://arxiv.org/abs/2505.19103v1 |
Chou, and K. Livescu, “Layer-wise analysis of a self-supervised speech representation model,” in 2021 IEEE Auto- matic Speech Recognition and Understanding Workshop (ASRU) . IEEE, 2021, pp. 914–921. [26] A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech repr... | https://arxiv.org/abs/2505.19103v1 |
arXiv:2505.19108v1 [cs.CL] 25 May 2025CCHall : A Novel Benchmark for Joint Cross-Lingual and Cross-Modal Hallucinations Detection in Large Language Models Yongheng Zhang1,2∗Xu Liu1*Ruoxi Zhou1Qiguang Chen1 Hao Fei3Wenpeng Lu4Libo Qin1,2† 1School of Computer Science and Engineering, Central South University, China 2Key ... | https://arxiv.org/abs/2505.19108v1 |
maries and observe that LLMs hallucinate more in non-English languages. Dale et al. (2023a) release an annotated dataset covering 18 translation direc- tions, tackling hallucinations. Herrlein et al. (2024) further extend English hallucination detection to German and apply it in long-context scenarios. 020406080InternV... | https://arxiv.org/abs/2505.19108v1 |
joint Cross-lingual and Cross-modal Hallucinations. Furthermore, CCHall covers a wide range of topics and nature scenes, aiming to provide a comprehensive evaluation of MLLMs in cross-lingual and cross-modal hallucination scenarios. Through evaluation experiments on CCHall , we derive the following key takeaways : (1)C... | https://arxiv.org/abs/2505.19108v1 |
which is denoted as: At= argmax AP(A |Q,I,P ), (2) where P(A |Q,I,P )represents the probability of generating answer Agiven I,Q, andP. Cross- modal hallucination occurs when there is reasoning that does not correspond to the image. 2.3 Joint Cross-Lingual and Cross-Modal Hallucinations Compared to cross-lingual and cro... | https://arxiv.org/abs/2505.19108v1 |
and the actual presence of objects in input images. Therefore, we select questions from the GQA (Hudson and Manning, 2019) and AMBER (Wang et al., 2023) datasets related to object existence to assess the model’s accuracy in object reasoning. To minimize redun- dancy, we ensure that each object appears no more than twic... | https://arxiv.org/abs/2505.19108v1 |
The details of the check can be found in Appendix A.2. 3.4 Cross-modal and Cross-lingual Hallucination Dataset TheCross-modal and Cross-lingual Hallucination Dataset ( CChall ) is shown in Figure 3 (d). Please refer to the Appendix B.4 for specific exam- ples. Specifically, we have retained the original dataset names a... | https://arxiv.org/abs/2505.19108v1 |
Macro-F1 Acc Macro-F1 Random 25.1 30.0 25.0 29.4 24.9 30.3 25.1 29.9 25.0 29.9 InternVL2-8B (Chen et al., 2024c) Direct (Chen et al., 2024c) 29.1 38.1 29.9 38.6 38.3 47.6 38.8 47.4 34.0 42.9 CoT (Kojima et al., 2022) 31.3 40.0 33.6 42.1 41.6 48.0 40.1 47.6 36.7 44.4 SRO (Lin et al., 2024) 30.3 40.0 31.1 40.7 41.2 48.6 ... | https://arxiv.org/abs/2505.19108v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.