text string | source string |
|---|---|
Furthermore, we re- move data that mimics the style of general AI assis- tance such as ‘ I am a helpful AI assistant... ’, as well as unnecessary explanatory prefaces and postfaces. For speech data, we first use the Whisper-large- v3 (Radford et al., 2023) to convert the synthesized audio into text and compute the WER ... | https://arxiv.org/abs/2505.20277v1 |
of OmniCharacter-10K Table 1 presents the statistics of the OmniCharacter- 10K dataset, which includes a total of 20 characters with 9, 672 training samples and 400 test samples. The training and test sets are rich in speech anno- tations, with 360.3 hours and 14.84 hours of audio data, respectively. In addition, Figur... | https://arxiv.org/abs/2505.20277v1 |
Memory : short-term (CM Short) and long-term (CM Long) (4) Social Preference : positive (Pos.), neutral (Neu.), and negative (Neg.). ModelsS2TIF S2SIF Content ↑Style ↑Content ↑Style ↑ SpeechGPT 1.25 1.28 1.79 1.85 LLaMA-Omni 2.46 2.71 2.26 2.38 OmniCharacter 3.89 4.18 2.64 2.55 Table 4: Performance comparison with stat... | https://arxiv.org/abs/2505.20277v1 |
fluency (Flu.), consistency (Cons.), emotional expression (Emo.), clarity (Cla), appropriate- ness (App.), and immersion (Imm.). six voice-related dimensions as shown in Table 5. Each expert is asked to score the responses on a scale from 1 to 10 for each dimension. From the table, we figure out that our method reaches... | https://arxiv.org/abs/2505.20277v1 |
ap- proach in preserving the voice traits of characters, achieving a more interactive experiment. 4.5 Learning Discriminative Character Speech Embeddings The OmniCharacter generates the character- specific voice by utilizing the speech embedding as one of the conditions. To evaluate the discrimina- tion of speaker embe... | https://arxiv.org/abs/2505.20277v1 |
synthesis systems, which may be affected by their alignment and training data. Second, the current design is limited to two-character dialogues, making it less adaptable to multi-role interactions. Third, although the OmniCharacter-10K dataset includes diverse sce- narios, it may not fully capture the richness of real-... | https://arxiv.org/abs/2505.20277v1 |
Yuchuan Wu, Xiong Liu, Min Yang, Yongbin Li, Longze Chen,Jiaming Li, Lei Zhang, et al. 2025. Openomni: Large language models pivot zero-shot omnimodal alignment across language with real-time self- aware emotional speech synthesis. arXiv preprint arXiv:2501.04561 . Run Luo, Haonan Zhang, Longze Chen, Ting-En Lin, Xiong... | https://arxiv.org/abs/2505.20277v1 |
Haoran Guo, Quan Tu, Yaying Fei, Ziang Leng, Wei Wang, et al. 2024d. Incharacter: Evaluat- ing personality fidelity in role-playing agents through psychological interviews. In ACL, pages 1840–1873. Zekun Moore Wang, Zhongyuan Peng, Haoran Que, Jiaheng Liu, Wangchunshu Zhou, Yuhan Wu, Hongcheng Guo, Ruitong Gan, Zehao N... | https://arxiv.org/abs/2505.20277v1 |
RPAs using psychological scales. (Wang et al., 2024a) de- vises a simulation sandbox to generate fine-grained character behavior trajectories to evaluate RPAs capabilities. However, most existing studies pri- marily focus on replicating dialogues in textual form, often overlooking the role’s essential voice traits, suc... | https://arxiv.org/abs/2505.20277v1 |
Then, we feed the model the dialogue contexts. The model is required to generate a reasonable text-audio response based on the current text or audio queries. Moreover, to systematically evaluate the qual- ity of the generated audio responses, we provide a comprehensive evaluation framework as shown 12 # Evaluation Agre... | https://arxiv.org/abs/2505.20277v1 |
a human evaluation on Charac- terEval. Due to the large test samples, we randomly sample 100 examples from CharacterEval dataset and hired five experts for assessment, following the same evaluation criteria as CharacterEval. The ex- perimental results are demonstrated in Table 9. Asobserved, human evaluation maintains ... | https://arxiv.org/abs/2505.20277v1 |
arXiv:2505.20278v1 [cs.LG] 26 May 2025The Coverage Principle: A Framework for Understanding Compositional Generalization Hoyeon Chang1∗Jinho Park1∗Hanseul Cho1Sohee Yang2Miyoung Ko1 Hyeonbin Hwang1Seungpil Won3Dohaeng Lee3Youbin Ahn3Minjoon Seo1 1KAIST2UCL3LG AI Research {retapurayo, binlepain178, minjoon}@kaist.ac.kr ... | https://arxiv.org/abs/2505.20278v1 |
from the observations by substituting functionally equivalent fragments. compositional reasoning [ 12–25]. These observations suggest that LLMs succeed primarily through pattern matching, i.e., by exploiting surface-level statistical correlations between input fragments and outputs, rather than systematic compositional... | https://arxiv.org/abs/2505.20278v1 |
underscores the need for architectural or training innovations to achieve truly systematic generalization in neural networks. 2 Related work Compositional generalization and systematicity Since Fodor and Pylyshyn [9]posed the sys- tematicity challenge, arguing connectionist systems inherently lack symbolic structural c... | https://arxiv.org/abs/2505.20278v1 |
the learner generalize by only seeing the input-output pairs? 4By “learner”, we mean any system that learns from data ( e.g., neural networks). 5For brevity, we use a shared token set X. Position-specific domains Xican be embedded as subsets of an enlarged token set ˜X=X1∪ X 2∪ X 3without loss of generality. 3 Our key ... | https://arxiv.org/abs/2505.20278v1 |
substitution graph : LetGD,k= (V, E)be an undirected graph with a vertex set V=Xnof all possible inputs. Two vertices x,x′∈Vare connected with an edge in Eif and only if there exists an index set I⊂[n]such that {x,x′}is anI-co-occurrence (in V) of a pair of functionally k-equivalent sequences at IinD. With this substit... | https://arxiv.org/abs/2505.20278v1 |
with 8 layers, 12 heads, and 768 dimensions (see App. B.2 for details). We construct two evaluation sets, each with 2,000 instances: (1) In-Domain (ID) Test Set : all primitive function applications ( e.g.,f1(x1, x2)andf2(b, x2)in2-H OPtask) are observed during training, but the specific combination was unseen. (2) Out... | https://arxiv.org/abs/2505.20278v1 |
effectively outside practical coverage, despite technically being in-distribution. 5.2 The clustering of latent representation drives generalization on coverage x1x2x3Layer 8 Layer 7 Layer 6 Layer 5 Layer 4 Layer 3 Layer 2 Layer 10.28 k<3x1x2x30.43 k=3x1x2x30.50 k>3x1x2x30.00 OOD 0.00.10.20.30.40.5 IICG 5.0 2.5 0.0 2.5... | https://arxiv.org/abs/2505.20278v1 |
2.0 2.1 2.2 2.3 log(||) 4.55.05.56.0log(Nreq) 3-Hop (c=2.58) Parallel-2-Hop (c=2.43) 2-Hop (c=2.26) 1.7 1.8 1.9 2.0 2.1 2.2 2.3 log(||) 4.504.755.005.255.505.75log(Nreq) 2-Hop 1.5B (c=2.28) 2-Hop 96M (c=2.26) 2-Hop 68M (c=2.13) Figure 5: Left: Log-log plot of measured ˆNreqvs. token set size ( |X|) across three composi... | https://arxiv.org/abs/2505.20278v1 |
the scaling relationship is primarily determined by data properties rather than model capacity, reinforcing our framework’s data-centric perspective. The observed scaling relationships are robust across different hyperparameters (weight decay and learning rate) and empirical decision criteria for ˆNreq(see App. F). Ove... | https://arxiv.org/abs/2505.20278v1 |
run ( N= 50 k) includes virtually the entire domain ( ≈0.72× |X|3≈61k distinct ID triples). 8 grouping by x2-conditioned intermediate state ( (b, x2)) leads to high IICG scores, suggesting the formation of context-dependent state representations. This context-dependence raises an interpretability concern, as standard l... | https://arxiv.org/abs/2505.20278v1 |
sharing the same bshould yield identical second-step outputs, as functional equivalence holds only when x2=x′ 2. Hence, while CoT supervision helps with sequential computation by breaking down multi-hop structures, it may partially inherit the limitations on handling tasks with path ambiguities we describe in Sec. 5.4.... | https://arxiv.org/abs/2505.20278v1 |
mechanistic distinction to sharpen future analyses of neural generalization. Implications and future directions Real compositional tasks typically involve combinations of all three types, making it difficult to isolate why models succeed or fail. We suggest diagnostic approaches based on the proposed taxonomy: coverage... | https://arxiv.org/abs/2505.20278v1 |
architectures, and more realistic data where multiple computational structures coexist within the same dataset [ 12,15,94] remains promising future work. 8 Conclusion We introduce the coverage principle , a data-centric framework that specifies when pattern-matching learners can and cannot generalize compositionally. O... | https://arxiv.org/abs/2505.20278v1 |
Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. ArXiv preprint , abs/2312.11805, 2023. URL https://arxiv.org/abs/2312.11805 . [8]Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, ... | https://arxiv.org/abs/2505.20278v1 |
Geva, and Sebastian Riedel. Do large language models latently perform multi-hop reasoning? In Lun-Wei Ku, Andre Martins, and Vivek Srikumar, editors, Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages 10210–10229, Bangkok, Thailand, 2024. Association ... | https://arxiv.org/abs/2505.20278v1 |
USA , pages 5998–6008, 2017. URL https://proceedings.neurips.cc/paper/2017/hash/ 3f5ee243547dee91fbd053c1c4a845aa-Abstract.html . [27] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V . Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language mod... | https://arxiv.org/abs/2505.20278v1 |
connectionist systems. Artificial intelligence , 46(1-2):159–216, 1990. [37] Gary F Marcus. The algebraic mind: Integrating connectionism and cognitive science . MIT press, 2003. [38] Dieuwke Hupkes, Verna Dankers, Mathijs Mul, and Elia Bruni. Compositionality decomposed: How do neural networks generalise? Journal of A... | https://arxiv.org/abs/2505.20278v1 |
and Dario Amodei. Scaling laws for neural language models. ArXiv preprint , abs/2001.08361, 2020. URL https://arxiv.org/abs/ 2001.08361 . [51] Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, Tom He... | https://arxiv.org/abs/2505.20278v1 |
Smith, and Jacob Steinhardt. Progress mea- sures for grokking via mechanistic interpretability. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023 . OpenReview.net, 2023. URL https://openreview.net/pdf?id=9XFSbDPmdW . [64] Nelson Elhage, Tristan Hume, Catherin... | https://arxiv.org/abs/2505.20278v1 |
Toronto, Canada, 2023. Association for Computational Linguistics. doi: 10.18653/v1/2023.acl-long.546. URL https://aclanthology.org/2023.acl-long.546 . [77] Nikhil Kandpal, Haikang Deng, Adam Roberts, Eric Wallace, and Colin Raffel. Large lan- guage models struggle to learn long-tail knowledge. In Andreas Krause, Emma B... | https://arxiv.org/abs/2505.20278v1 |
Hugh Mee Wong, Ian Ng, Isaac Noble, Jaap Jumelet, Jack Geissinger, Jackson Kernion, Jacob Hilton, Jaehoon Lee, Jaime Fernández Fisac, James B Simon, James Koppel, James Zheng, James Zou, Jan Kocon, Jana Thompson, Janelle Wing- field, Jared Kaplan, Jarema Radom, Jascha Sohl-Dickstein, Jason Phang, Jason Wei, Jason Yosin... | https://arxiv.org/abs/2505.20278v1 |
Tatsunori Hashimoto, Te-Lin Wu, Théo Desbordes, Theodore Rothschild, Thomas Phan, Tianle Wang, Tiberius Nkinyili, Timo Schick, Timofei Kornev, Titus Tunduny, Tobias Gerstenberg, Trenton Chang, Trishala Neeraj, Tushar Khot, Tyler Shultz, Uri Shaham, Vedant Misra, Vera Demberg, Victoria Nyamai, Vikas Raunak, Vinay Venkat... | https://arxiv.org/abs/2505.20278v1 |
topcu, and Zhangyang Wang. Ontheplanning abilities of openAI’s o1 models: Feasibility, optimality, and generalizability. In Language Gamification - NeurIPS 2024 Workshop , 2024. URL https://openreview.net/ forum?id=qgvQx30Z0R . [90] Olga Golovneva, Zeyuan Allen-Zhu, Jason E Weston, and Sainbayar Sukhbaatar. Reverse tra... | https://arxiv.org/abs/2505.20278v1 |
Pedro A Ortega. Neural networks and the chomsky hierarchy. In The Eleventh International Conference on Learning Representations , 2023. URL https://openreview.net/forum?id=WbxHAzkeQcn . [105] Bingbin Liu, Jordan T Ash, Surbhi Goel, Akshay Krishnamurthy, and Cyril Zhang. Trans- formers learn shortcuts to automata. In Th... | https://arxiv.org/abs/2505.20278v1 |
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38 G.2 Effect of model scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38 G.3 Representation analysis in successful generalization . . . . . . . . . . . . . . . . . 39 H Detailed discussion on the taxonomy of generalization 40 H.1 Typ... | https://arxiv.org/abs/2505.20278v1 |
x3, t)such that: (x1, x2)∈domain (Sf1) (1) (f1(x1, x2), x3)∈domain (Sf2) (2) t=f2(f1(x1, x2), x3) (3) From this set of all possible in-domain combinations, we uniformly sample Nexamples to form our training dataset. When the number of possible combinations exceeds N, this sampling ensures the model sees only a subset o... | https://arxiv.org/abs/2505.20278v1 |
rate schedule with a linear warmup period of 2,000 steps. This standardized training configuration is maintained across all experiments to ensure fair comparisons between different task structures and dataset sizes, unless explicitly varied in specific ablation studies. 25 C Implementation details for the coverage dete... | https://arxiv.org/abs/2505.20278v1 |
contain at least one training example. This set comprises all inputs that are reachable from the training data through chains of equivalent subsequence substitutions. 27 D Detailed analysis for representation unification experiments D.1 Causal Tracing Methodology To analyze the causal role of specific hidden representa... | https://arxiv.org/abs/2505.20278v1 |
set size ablation We show that the observed patterns of cosine similarity analysis and causal tracing in the 2-H OPtask are consistent across different token set sizes |X|. For|X|= 70,100,150,200, we analyze model 28 x1 x2 x3Layer 7 Layer 6 Layer 5 Layer 4 Layer 3 Layer 2 Layer 10.00 -0.07 0.95 0.00 -0.09 0.93 0.00 -0.... | https://arxiv.org/abs/2505.20278v1 |
6 Layer 5 Layer 4 Layer 3 Layer 2 Layer 10.77 ID T estx1x2x3x40.02 OODx1x2x3x40.39 ID T estx1x2x3x40.00 OODx1x2x3x40.51 ID T estx1x2x3x40.02 OOD Figure 11: IICG heatmap for PARALLEL -2-H OPtask with grouping strategies based on b1= f1(x1, x2)(Left),b2=f2(x3, x4)(Middle ), and t=f3(b1, b2)(Right ). x1 x2 x3 x4Layer 7 La... | https://arxiv.org/abs/2505.20278v1 |
first-hop fragments a, a′∈ X2define a∼a′:⇐⇒ f1(a) =f1(a′). They are functionally k-equivalent w.r.t. D(denoted a≡k Da′) when there exist kdistinct contexts c1, . . . , c k∈ X such that for every r≤kboth (a, cr)and(a′, cr)appear in Dandf(a, cr) = f(a′, cr). We call each (a, cr),(a′, cr)anevidence pair . The coverage pri... | https://arxiv.org/abs/2505.20278v1 |
f(i, c) =f(j, c)withi̸∼jcan only masquerade as evidence when k= 1andthe dataset lacks any contradicting context c′. Fork≥2the joint probability that two independent contexts simultaneously produce such coincidences is |X|−kper fragment pair and hence negligible relative to the isolate probability once N= Ω(|X|2). Conse... | https://arxiv.org/abs/2505.20278v1 |
|X|5. (S3.1) Why Gbcan be approximated as Erd ˝os–Rényi As mentioned earlier, the dependence among distinct edges in Gbarises from the constraint that the total sample size is N, which induces a negative correlation . Such negative dependence reduces the likelihood of simultaneously creating many edges. Consequently, v... | https://arxiv.org/abs/2505.20278v1 |
obtained by linear fitting log(|X|)vs.log(Nreq)plots, all with R2>0.99. The consistency of exponents across model sizes suggests that the observed power-law scaling relates to properties of the compositional tasks themselves, rather than model capacity. This observation aligns with our theoretical derivation in Section... | https://arxiv.org/abs/2505.20278v1 |
G.2 Effect of model scaling 5k10k 15k 20k 25k 30k 35k 40k 45k 50k Dataset Size(N)0.00.20.40.60.81.0ID accuracy Non-tree Non-tree-XL Figure 17: ID test accuracy comparison between GPT-2 (96M parameters) and GPT-2-XL (1.5B parameters) on the NON-TREE task with |X|= 50 , measured 100 epochs after training accuracy exceeds... | https://arxiv.org/abs/2505.20278v1 |
current limits, open questions, and future directions. H.1 Type-I: Structure-based generalization Type-I generalization occurs when a model succeeds by identifying and utilizing functionally equiv- alent components based on how primitive functions are composed . This form of generalization is precisely what we have for... | https://arxiv.org/abs/2505.20278v1 |
data by leveraging invariant properties of the underlying functions. Systematic characterization of pure Type-II extrapolation remains limited and is an important direction for future work. H.3 Type-III: Shared-operator generalization Type-III generalization emerges through reuse of identical primitive functions across... | https://arxiv.org/abs/2505.20278v1 |
examples in this paper are illustrative rather than definitive classifications; real tasks often involve combinations of mechanisms, and our taxonomy provides a scaffold for disentangling them rather than enforcing mutually exclusive categories. 41 Several fundamental questions remain. First, practical computations oft... | https://arxiv.org/abs/2505.20278v1 |
0.0 4090.1 3384.8 0.0 1678.1 10164.3 46.1 2051.9 16836.9 4269.2 2635.5 With Atomic 02004006008001000 MRR Figure 19: MRR scores for intermediate state representations projected to vocabulary space. Left: Standard training ( f1̸=f2, no partial computation) shows very high MRR regardless of position and layer. Right: Trai... | https://arxiv.org/abs/2505.20278v1 |
arXiv:2505.20279v1 [cs.CV] 26 May 2025VLM-3R: Vision-Language Models Augmented with Instruction-Aligned 3D Reconstruction Zhiwen Fan1†∗, Jian Zhang2∗, Renjie Li3, Junge Zhang4, Runjin Chen1, Hezhen Hu1, Kevin Wang1,Huaizhi Qu5,Dilin Wang6,Zhicheng Yan6,Hongyu Xu6,Justin Theiss6, Tianlong Chen5,Jiachen Li4,Zhengzhong Tu... | https://arxiv.org/abs/2505.20279v1 |
both fine-tuning and inference. This hardware dependency critically constrains their scalability to environments equipped with specific sensors and largely prevents leveraging the vast amount of readily available monocular video data. On the other hand, there are approaches that employ off-the-shelf algorithms [ 23,26]... | https://arxiv.org/abs/2505.20279v1 |
video input, as evidenced by its performance across benchmarks in Figure 1(c). To summarize, our contributions are: •We introduce VLM-3R, the first 3D vision-language framework to achieve robust spatial reasoning and instruction-guided 3D scene VQA directly from monocular RGB video. VLM-3R uniquely operates without req... | https://arxiv.org/abs/2505.20279v1 |
layout inference, and memory recall. The shortcomings of current video LMMs on spatial reasoning tasks stem from their absence of structured spatial representations—long emphasized by cognitive science—and their lack of multi-view encoding analogous to human binocular vision. This gap underscores the critical role of s... | https://arxiv.org/abs/2505.20279v1 |
of temporal changes by designing spatial-temporal QA pairs, which focus on reasoning about spatial configurations evolving over time (Figure 2). 3.2 Multimodal Spatial Instruction Data Generation While existing benchmarks like VSI-Bench [ 31] offer valuable evaluation sets (approximately 5,000 question-answer pairs fro... | https://arxiv.org/abs/2505.20279v1 |
comprehension and static understanding of 3D environments, we introduce the Visual-Spatial-Temporal Intelligence Benchmark. This benchmark is designed to enable AI agents not only to answer global questions based on input video but also to perform reasoning about objects, cameras, and their relationships as they evolve... | https://arxiv.org/abs/2505.20279v1 |
Large Language Model Q:Approximately how far (in meters) did the camera move between frame 13 and frame 29 of 32? Visual Encoder Spatial Encoder Cam. Tok … Monocular Video InputGeo. TokVis. Tok2D-3D Fusion View TokensGeometry TokensAppearance Tokens (2D) 𝑊𝑞 𝑊𝑘𝒂𝒇𝟐𝑫→𝒇𝟑𝑫 (a).Visual Encoder (b).Spatial Encoderpr... | https://arxiv.org/abs/2505.20279v1 |
CUT3R-based spatial encoders ( fenc,fdec) are frozen. Spatial-visual View Fusion. VLM-3R integrates 3D geometric information using specialized 3D reconstructive tokens derived from our 3D tokenization process (Sec. 4.1). These tokens comprise concatenated enriched 3D feature tokens F′ tand camera view tokens z′ t, whic... | https://arxiv.org/abs/2505.20279v1 |
instruction tuning. We observe that incorpo- rating spatial encoding significantly boosts LMM capabilities, particularly leading to substantial improvements in distance, size, and direction estimation tasks. On the appearance order task, which is less reliant on detailed spatial geometry awareness, VLM-3R’s performance... | https://arxiv.org/abs/2505.20279v1 |
Table 1: Evaluations on VSI-Bench. VLM-3R ranks first among open-sourced VLMs, showcasing the effectiveness of its reconstructive instruction tuning. This validates our model’s spatial encoding significantly improves 3D understanding and reasoning, particularly in distance, size, direction, and spatial planning tasks. ... | https://arxiv.org/abs/2505.20279v1 |
A-7B 10 32.3 13.5 5.1 43.7 57.9 41.2 InternVL2-8B 3 43.5 32.9 13.5 48.0 68.0 55.0 LongVILA-8B 11 30.5 20.0 11.6 35.4 52.3 33.4 VILA-1.5-8B 8 37.3 30.1 27.3 42.2 50.4 36.7 VILA-1.5-40B 6 38.2 28.2 15.7 28.8 65.4 53.0 LLaV A-NeXT-Video-72B 2 44.0 32.3 10.5 48.1 78.3 50.9 VLM-3R (7B) 1 58.8 39.4 39.6 60.6 86.5 68.6 Table ... | https://arxiv.org/abs/2505.20279v1 |
ethical deployment with robust safeguards to ensure responsible AI innovation. References [1]Humza Naveed, Asad Ullah Khan, Shi Qiu, Muhammad Saqib, Saeed Anwar, Muhammad Usman, Naveed Akhtar, Nick Barnes, and Ajmal Mian. A comprehensive overview of large language models. arXiv preprint arXiv:2307.06435 , 2023. [2]Jaso... | https://arxiv.org/abs/2505.20279v1 |
Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML , 2023. [18] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large... | https://arxiv.org/abs/2505.20279v1 |
with mast3r. In European Conference on Computer Vision , pages 71–91. Springer, 2024. [34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ... | https://arxiv.org/abs/2505.20279v1 |
to vision. arXiv , 2024. [51] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv , 2024. 12 [52] Mary Hegarty and D Waller. Individual differences in spatial abilities. The Cambridge handbook of v... | https://arxiv.org/abs/2505.20279v1 |
Kurz, Arik Schwartz, et al. Arkitscenes: A diverse real- world dataset for 3d indoor scene understanding using mobile rgb-d data. arXiv preprint arXiv:2111.08897 , 2021. 13 [68] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra M... | https://arxiv.org/abs/2505.20279v1 |
a monocular video stream, CUT3R directly outputs corresponding dense 3D point maps (e.g., ˆXworld t) and relative camera poses ( ˆPt) for each view. Spatial-Visual-View Fusion Design VLM-3R employs a Spatial-Visual-View Fusion stage to integrate diverse multimodal inputs, as depicted in our overall architecture (e.g., ... | https://arxiv.org/abs/2505.20279v1 |
fusion block :True (–tune_fusion_block True ). – Tune MM MLP adapter :True (–tune_mm_mlp_adapter True ). B Dataset Curation and Benchmark Design 3D Reconstructive Instructional Tuning relies on large-scale Question-Answer (QA) pairs to fine-tune Large Multimodal Models (LMMs), enabling them to perform tasks related to ... | https://arxiv.org/abs/2505.20279v1 |
These assess understanding of spatial layout and inter-object relationships. –Object Count: Counting instances of a specific object category in a room (numerical answer; objects with single instances are excluded). –Relative Distance: Determining which of four candidate objects is closest in 3D space to a target object... | https://arxiv.org/abs/2505.20279v1 |
diverse trajectories generated by the Habitat simulator within a single scene. These demonstrate fundamental navigation actions (e.g., “Turn Right” ,“Turn Left” ,“Turn Back” ), which are foundational for generating route planning QA data via our specialized templates. We use SRC,TGT, and MID to represent the object des... | https://arxiv.org/abs/2505.20279v1 |
generation of these QA pairs relies on processed metadata derived from raw 3D scene data (point clouds, videos, and sampled frames with depth, instance masks, and camera poses). Specifically, scene_metadata.json (containing 3D object bounding boxes, instance IDs, scene properties) and frame_metadata.json (containing pe... | https://arxiv.org/abs/2505.20279v1 |
Task Group / Filename Total Length camera_displacement 839 camera_movement_direction 913 camera_obj_abs_dist 905 camera_obj_rel_dist 1,740 obj_obj_relative_pos 1,645 Table 6: VSTemporalI-Bench Test Set Distribution C Additional Experimental Analysis Our error analysis, with quantitative results detailed in the ablation... | https://arxiv.org/abs/2505.20279v1 |
arXiv:2505.20282v2 [cs.CL] 27 May 2025One-shot Entropy Minimization Zitian Gao Lynx Chen Joey Zhou Bryan Dai * Ubiquant {ztgao02,ylchen,jzhou,cbdai}@ubiquant.com Abstract We trained 13,440 large language models and found that entropy minimization requires only a single unlabeled data and 10 steps optimization to achiev... | https://arxiv.org/abs/2505.20282v2 |
where Tis the length of the generated sequence. Our core idea is to reduce the model’s uncertainty over its own predictions by minimizing the token-level entropy at each generation step. The conditional entropy at time step tis defined as: Ht=−X v∈Vpθ(v|y<t, x) logpθ(v|y<t, x). To avoid penalizing the prompt portion, w... | https://arxiv.org/abs/2505.20282v2 |
consistent and (ideally) correct reasoning trajectory. By contrast, if a model consistently answers a question correctly or incorrectly regardless of sampling, the entropy is either already minimal or optimization is ineffective. Thus, high-variance prompts provide the richest signal for improving model calibration and... | https://arxiv.org/abs/2505.20282v2 |
in Section 3.3. 3.3 Logits Shift We sample 20 prompts from the NuminaMath [10] dataset and generate responses using four different models (Qwen2.5-Math-7B, Qwen2.5-Math-7B-EM, Qwen2.5-Math-7B-RL, Qwen2.5-Math-7B- EM-RL). Each model generates 20 responses, resulting in a total of 4×20 = 80 outputs. For each 1The reason ... | https://arxiv.org/abs/2505.20282v2 |
language models, as it reduces the number of high-probability paths during sampling—opposite to the effect of a rightward shift—thus diminishing the model’s overall performance. Therefore, the rightward logits shift induced by EM is preferable to the leftward shift caused by RL. 3.4 Training Loss vs. Reasoning Performa... | https://arxiv.org/abs/2505.20282v2 |
0.2 0.4 0.6 0.8 1.0 T emperature515253545556Avg RL 0.0 0.2 0.4 0.6 0.8 1.0 T emperature5253545556Avg EM Figure 4: The impact of generation temperature during evalutating on the average performance of the trained model across four reasoning datasets. The results in the figure are obtained by repeating the experiments wi... | https://arxiv.org/abs/2505.20282v2 |
models. Remarkably, with only a single exemplar and minimal optimization, EM consistently boosts inference accuracy across all models on tasks MATH500, Minerva Math, Olympiad Bench, and AMC23. For instance, on the Qwen2.5-Math-7B model, 1-shot EM yields impressive improvements 8 of 25.8 points on MATH500 (from 53.0 to ... | https://arxiv.org/abs/2505.20282v2 |
prompt length and generated output length remain markedly more 9 stable under 1-shot EM. Moreover, whereas multi-shot EM losses continue to fluctuate significantly after step 3, the 1-shot EM loss steadily declines from step 3 onward and remains at a low level beyond step 10. This indicates that relying on a single exe... | https://arxiv.org/abs/2505.20282v2 |
suggest that EM enhances model confidence by reinforcing high-probability reasoning paths 3.3. This implies EM might serve as a lightweight alternative to complex calibration techniques, especially for tasks where interpretability and robustness are critical. Developing evaluation protocols to more precisely quantify E... | https://arxiv.org/abs/2505.20282v2 |
unreasonable effectiveness of entropy minimization in llm reasoning, 2025. [2]Stephen Casper, Xander Davies, Claudia Shi, Thomas Krendl Gilbert, Jérémy Scheurer, Javier Rando, Rachel Freedman, Tomasz Korbak, David Lindner, Pedro Freire, Tony Wang, Samuel Marks, Charbel-Raphaël Segerie, Micah Carroll, Andi Peng, Phillip... | https://arxiv.org/abs/2505.20282v2 |
Z. Z. Ren, Zehui Ren, Zhangli Sha, Zhe Fu, Zhean Xu, Zhenda Xie, Zhengyan Zhang, Zhewen Hao, Zhicheng Ma, Zhigang Yan, Zhiyu Wu, Zihui Gu, Zijia Zhu, Zijun Liu, Zilin Li, Ziwei Xie, Ziyang Song, Zizheng Pan, Zhen Huang, Zhipeng Xu, Zhongyu Zhang, and Zhen Zhang. Deepseek-r1: Incentivizing reasoning capability in llms v... | https://arxiv.org/abs/2505.20282v2 |
Ye, Longhui Yu, Mengnan Dong, Neo Zhang, Ningchen Ma, Qiwei Pan, Qucheng Gong, Shaowei Liu, Shengling Ma, Shupeng Wei, Sihan Cao, Siying Huang, Tao Jiang, Weihao Gao, Weimin Xiong, Weiran He, Weixiao Huang, Wenhao Wu, Wenyang He, Xianghui Wei, Xianqing Jia, Xingzhe Wu, Xinran Xu, Xinxing Zu, Xinyu Zhou, Xuehai Pan, Y .... | https://arxiv.org/abs/2505.20282v2 |
arXiv:2505.20285v2 [cs.CL] 27 May 2025MASKSEARCH : A Universal Pre-Training Framework to Enhance Agentic Search Capability Weiqi Wu†, Xin Guan†, Shen Huang, Yong Jiang∗, Pengjun Xie, Fei Huang, Jiuxin Cao, Hai Zhao∗, Jingren Zhou Tongyi Lab , Alibaba Group Abstract Retrieval-Augmented Language Models (RALMs) represent ... | https://arxiv.org/abs/2505.20285v2 |
major of that degree. Now, I should search for Andrew’s degree from UMichin 1970.</think>Task DecompositionUniversity of Michigan alum Andrew Barto (BS Math 1970) wins the 2024 Turing Award …1. Andrew Barto’s UMichdegree in 19702. Andrew Barto Education ExperienceSearch Result Analysis LLM-Judge <answer> Andrew Barto r... | https://arxiv.org/abs/2505.20285v2 |
of masks, enabling the model to learn progressively from easier to more difficult scenarios. 2 Extensive analysis shows that incorporating RAMP as the pre-training task yields significant per- formance enhancements across a variety of open-domain question-answering datasets. It not only provides a stable improvement in... | https://arxiv.org/abs/2505.20285v2 |
nmasked spans, by proactively retrieving relevant information from an external knowledge corpus Dusing a search toolR. To train an LLM-based search agent πθ, which is parameterized by θand initialized from a pre-trained base model to take a strong starting point, there are two primary methods: Supervised Fine-tuning (S... | https://arxiv.org/abs/2505.20285v2 |
a multi-agent system, as shown in Figure 1. The Planner Agent first analyzes the overall task and breaks it into sub-tasks, generating an initial search query. The Rewriter Agent refines the generated query for improved knowledge retrieval and calls the search tool. The Observer Agent reviews the search results and ste... | https://arxiv.org/abs/2505.20285v2 |
kserves as the primary metric for difficulty. Instead of random sampling, the curriculum learning method starts with simpler tasks containing fewer masked spans and progressively introduces more complex tasks with a higher number of masked spans. This approach allows the model to first learn fundamental reasoning skill... | https://arxiv.org/abs/2505.20285v2 |
% SFT 69.55 57.24 41.06 83.84 73.07 78.97 67.29 Search-R1 % RL 70.59 56.25 41.29 80.50 79.33 78.46 67.74 MASKSEARCHSFT SFT 70.44 60.85 41.76 84.65 80.13 81.12 69.83 RL SFT 70.84 56.29 41.90 83.38 78.53 78.93 68.31 SFT RL 71.69 57.69 42.23 81.25 81.87 75.42 68.36 RL RL 75.61 58.96 45.54 82.10 83.00 80.85 71.01 model on ... | https://arxiv.org/abs/2505.20285v2 |
SFT baseline, which is directly fine-tuned on the 58K CoT trajectories derived from HotpotQA. RL offers higher performance gains on RAMP compared with SFT. While SFT proves to be effective in improving search agent performance, RL demonstrates the potential to achieve even higher upper limits when applied to RAMP tasks... | https://arxiv.org/abs/2505.20285v2 |
architectures. 6 Discussion In this section, we conduct an in-depth discussion of the critical factors of RAMP and its training process to offer a comprehensive exploration of our approach. A case study is presented in the Appendix F to provide further insights into the practical application of our method. 6.1 Masking ... | https://arxiv.org/abs/2505.20285v2 |
the impact of different RL rewards on model performance, as discussed in section 3.4. The results are shown in Figure 5, model trained with the token-level recall reward hacks the metric by adding a lot of irrelevant information to the answer, significantly increasing the length of the response. This results in a notab... | https://arxiv.org/abs/2505.20285v2 |
bases? ArXiv , abs/1909.01066, 2019. URL https: //api.semanticscholar.org/CorpusID:202539551 . [8]Yujuan Ding, Wenqi Fan, Liang bo Ning, Shijie Wang, Hengyun Li, Dawei Yin, Tat-Seng Chua, and Qing Li. A survey on rag meets llms: Towards retrieval-augmented large language models. ArXiv , abs/2405.06211, 2024. URL https:... | https://arxiv.org/abs/2505.20285v2 |
survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ArXiv , abs/2311.05232, 2023. URL https://api. semanticscholar.org/CorpusID:265067168 . [22] Aditi Singh, Abul Ehtesham, Saket Kumar, and Tala Talaei Khoei. Agentic retrieval-augmented generation: A survey on agentic... | https://arxiv.org/abs/2505.20285v2 |
2024. URL https://arxiv.org/ abs/2412.16720 . [39] Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Candès, and Tatsunori Hashimoto. s1: Simple test-time scaling, 2025. URL https://arxiv.org/abs/2501.19393 . [40] Charlie Snell, Jaehoon ... | https://arxiv.org/abs/2505.20285v2 |
Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. MuSiQue: Multihop questions via single-hop question composition. Transactions of the Association for Computational Linguistics , 2022. [55] Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, and Akiko Aizawa. Constructing a multi-hop QA dataset for... | https://arxiv.org/abs/2505.20285v2 |
150 steps. We use the instruct models for RL, as base models often fail to follow the instructions. B Detailed Comparison with Existing RALMs Table 4: Comparison with existing RALMs and search- enhanced reasoning models. E2E is short for end-to-end. # Retrieval Tokens# Models RetrieverMulti- Step KNN-LM [58] O(109) 2 T... | https://arxiv.org/abs/2505.20285v2 |
among different architectures of models, as we utilize data generated by the Qwen- series model but still achieve significant performance improvements for LLaMA models. This demonstrates the framework’s ability to leverage data from different sources and adapt to different model architectures, further enhancing its app... | https://arxiv.org/abs/2505.20285v2 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.