text string | source string |
|---|---|
no major flaws or weaknesses in the experimental design, results or analysis. The experimental results are fully reliable and consistent with the code of the paper. 7-8 - The methods and techniques used in the paper are mostly sound, but there may be some minor issues. The results are generally well-supported by the da... | https://arxiv.org/abs/2505.19955v1 |
is of high quality and is suitable for publication in a top-tier conference or journal. I would recommend this paper. 6-7 - The paper is of good quality and is suitable for publication in a reputable conference or journal. I would recommend this paper with some reservations. 4-5 - The paper is of acceptable quality but... | https://arxiv.org/abs/2505.19955v1 |
giving high scores by default. ## Evaluation Rubric {Idea Evaluation Rubrics as shown in Table 22} ## Output Format Please output a complete JSON object strictly following the format below, including all evaluation items (Consistency, Clarity, Novelty, Feasibility, Significance, OverallAssessment). Do not output only a... | https://arxiv.org/abs/2505.19955v1 |
Evaluation Rubrics as shown in Table 26} Please provide a detailed review of the paper, including your scores for each aspect and an overall assessment. Be sure to justify your scores with specific examples from the paper. Please do not include any personal opinions or biases in your review. Your review should be objec... | https://arxiv.org/abs/2505.19955v1 |
and finally give an overall assessment on a scale of 1 to 10. Please be objective in your evaluation, and provide detailed justifications for each score you assign. Do not hesitate to assign lower scores if the paper does not fully meet the criteria. Avoid giving high scores by default. ## Evaluation Rubric {Writing Qu... | https://arxiv.org/abs/2505.19955v1 |
give an overall assessment on a scale of 1 to 10. Please be objective in your evaluation, and provide detailed justifications for each score you assign. Do not hesitate to assign lower scores if the experimental document does not fully meet the criteria. Avoid giving high scores by default. ## Evaluation Rubric {Experi... | https://arxiv.org/abs/2505.19955v1 |
given a research proposal based on a task description, a research idea, and a literature review. Your task is to evaluate the proposal on a scale of 1 to 10 across six key dimensions and finally give an overall assessment on a scale of 1 to 10. Please be objective in your evaluation, and provide detailed justifications... | https://arxiv.org/abs/2505.19955v1 |
relevant to their 38 domain. Each reviewer received the research paper along with its corresponding supplementary code. We collected all responses through Google Forms. The evaluation criteria followed the same rubric used in the end-to-end setting of MLR-Judge (see Table 26). Fig. 6 and Fig. 7 show the Google Form int... | https://arxiv.org/abs/2505.19955v1 |
arXiv:2505.19956v1 [cs.AI] 26 May 2025DCG-SQL: Enhancing In-Context Learning for Text-to-SQL with Deep Contextual Schema Link Graph Jihyung Lee1*, Jin-Seop Lee1*, Jaehoon Lee1, YunSeok Choi2†, Jee-Hyong Lee3† 1Department of Artificial Intelligence, Sungkyunkwan University 2Department of Immersive Media Engineering, Sun... | https://arxiv.org/abs/2505.19956v1 |
in-context learning capabilities compared to hyper- scale LLMs (Wei et al., 2022). Since smaller LLMs rely more heavily on demonstrations, performance significantly drops, sometimes even falling below the performance with randomly selected demon- strations. This is because they overlook DB schema when representing Text... | https://arxiv.org/abs/2505.19956v1 |
without any explicit linking dataset. With these, we can effectively achieve a joint representation of each text-to-SQL sample as a graph. By com- paring the representations of the graphs, we can retrieve demonstrations. With the demonstrations, our method fully leverages in-context learning ca- pability and generate t... | https://arxiv.org/abs/2505.19956v1 |
not consider the DB schema, which is also crucial in Text-to-SQL. In contrast, Gao et al. (2024); Nan et al. (2023); Shen et al. (2024) tried to reflect DB schema information indirectly by ap- proximating target SQL of user input and retrieving similar examples through similarity comparison of the approximated SQL. How... | https://arxiv.org/abs/2505.19956v1 |
useful demonstrations. Finally, with the selected demonstrations, our method can fully leverage in- context learning ability, and generate the target SQL more correctly. 3.1 Deep Contextual Schema Link Graph Construction We propose a deep contextual schema link graph construction method that excludes irrelevant schema ... | https://arxiv.org/abs/2505.19956v1 |
attention scores between its concatenated special token and the question tokens. When the attention score exceeds the threshold τ, a link is connected between the schema item and the question token, which we refer to as an ‘attention-match’. This approach enables schema pruning and linking simultaneously without ground... | https://arxiv.org/abs/2505.19956v1 |
consistent and structured steps when generating SQL queries. Details are provided in Appendix C.2. 4 Experiments 4.1 Implementation Details All experiments are conducted using the PyTorch framework on an A100 80GB GPU. Our method requires only small size of models for efficiency. For schema link graph construction, we ... | https://arxiv.org/abs/2505.19956v1 |
Turbo, DeepSeek-Coder-33B-Instruct, Llama 3.1- 8B-Instruct, DeepSeek-Coder-6.7B-Instruct, and Llama 3.2-3B-Instruct (OpenAI et al., 2024; DeepSeek-AI et al., 2024; Dubey et al., 2024). In all experiments, the generation temperature is set to 0.5 demonstrations in our few-shot setting and our input prompt is introduced ... | https://arxiv.org/abs/2505.19956v1 |
38.4 DCG-SQL 83.3 70.7 79.0 62.8 74.2 61.6 Llama-3.1-8B-Instruct Random 73.0 48.5 69.7 45.5 61.6 39.7 DCG-SQL 82.1 72.9 75.8 67.1 70.1 59.7 Table 4: Effectiveness of our retrieved demonstrations. 4.4 Experimental Results Table 2 presents the Text-to-SQL performance on the Spider dataset using various hyper-scaled LLMs ... | https://arxiv.org/abs/2505.19956v1 |
based only on question information does not achieve good performance. Also, retrieval meth- ods that consider both question and DB schema information, but do not account for their contextual relationships, show poor performance. Our method effectively capture the relationship for the joint representation of Text-to-SQL... | https://arxiv.org/abs/2505.19956v1 |
to other methods. Meanwhile, DAIL-SQL (Gao et al., 2024) employs a retrieval model with 3B parameters (Li et al., 2023b), result- ing in latency up to 9.1 seconds. Although ACT- SQL (Zhang et al., 2023a) shows relatively low in- ference latency, its execution accuracy remains un- satisfactory. In contrast, our method r... | https://arxiv.org/abs/2505.19956v1 |
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 . Andrea Galassi, Marco Lippi, and Paolo Torroni. 2020. Attention in natural language processing. IEEE trans- actions on neural networks and... | https://arxiv.org/abs/2505.19956v1 |
pages 4644–4668, Toronto, Canada. Association for Computational Lin- guistics. Zhishuai Li, Xiang Wang, Jingjing Zhao, Sun Yang, Guoqing Du, Xiaoru Hu, Bin Zhang, Yuxiao Ye, Ziyue Li, Rui Zhao, and Hangyu Mao. 2024. Pet-sql: A prompt-enhanced two-round refinement of text-to-sql with cross-consistency. Preprint , arXiv:... | https://arxiv.org/abs/2505.19956v1 |
58th Annual Meeting of the Association for Computational Lin- guistics , pages 7567–7578, Online. Association for Computational Linguistics. Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, Ed H. Chi, Tatsunori Hashimoto, Oriol ... | https://arxiv.org/abs/2505.19956v1 |
selected demonstrations, which are used for SQL query generation. In the existing approaches, the given database contains all schema items, making the model choose irrelevant schema items for SQL query generation. Addition- ally, the selected demonstrations differ significantlyin syntax from the target SQL query. Such ... | https://arxiv.org/abs/2505.19956v1 |
that started from the station. Let‘s think step by step. we need to join the tables ’station’ and ’trip’. Create an intermediate representation, then use it to construct the query. Intermediate representation: "FROM station AS T1 JOIN trip AS T2 ON T1.id = T2.start_station_id". ### SQL: SELECT T1.name , T1.long , A VG(... | https://arxiv.org/abs/2505.19956v1 |
arXiv:2505.19959v1 [cs.CL] 26 May 2025 MiniLongBench: The Low-cost Long Context Understanding Benchmark for Large Language Models Zhongzhan Huang1, Guoming Ling1, Shanshan Zhong1, Hefeng Wu1, Liang Lin1,2,3 1Sun Yat-sen University2Peng Cheng Laboratory 3Guangdong Key Laboratory of Big Data Analysis and Processing Abstr... | https://arxiv.org/abs/2505.19959v1 |
illustrated in Fig. 1 cannot be overlooked. Furthermore, when researchers de- velop new LLM models and need to conduct mul- tiple analyses of LCU capabilities, the time and computational costs become even more prohibitive. Given these challenges, we ask a critical question: Do LCU benchmarks really need such a large nu... | https://arxiv.org/abs/2505.19959v1 |
evalu- ation results of each subset and those of the original LongBench SL. The closer "Sp" is to 1.0, the more the evaluation of the sampled subset aligns with the evaluation of SL. We take p∈ {0.99,0.98,0.95} and select the top 7500 results based on Sp for statistical analysis. The experimental results are shown in F... | https://arxiv.org/abs/2505.19959v1 |
using a text encoder κtextOpenAIEmbed- ding (Xian et al., 2024) and a principal component analysis, a.k.a PCA, (Abdi and Williams, 2010) to obtain part of dense d−dimentional initialization of test samples, i.e., {ej}|SL| j=1=PCA d[{κtext(sj)}|SL| j=1], (1) For a detailed discussion on how data preprocess- ing influenc... | https://arxiv.org/abs/2505.19959v1 |
Acc. (EM) English 9,289 9,394 200 15 ( ↓93%) PassageRetrieval-zh 5-3 Acc. (EM) Chinese 6,745 6,684 200 15 ( ↓93%) Code Completion LCC 6-1 Edit Sim Python/C#/Java 1,235 1,187 500 26 ( ↓95%) RepoBench-P 6-2 Edit Sim Python/Jave 4,206 3,723 500 23 ( ↓95%) Table 1: The dataset statistics in LongBench and MiniLongBench. "Lo... | https://arxiv.org/abs/2505.19959v1 |
0.20 0.13 0.02 0.41 0.05 0.59 LwQ-Instruct-2k 0.23 0.18 0.04 0.45 0.07 0.70 30B-Epsilon-2k 0.20 0.13 0.04 0.48 0.07 0.73 Table 2: Specific evaluation results on MiniLongBench. See Appendix C and Appendix G for the more analysis and detail results on various advanced LLMs. characters (Chinese). Consistent with LongBench... | https://arxiv.org/abs/2505.19959v1 |
4.3 The Evaluation Results Moreover, Fig. 5 shows the rank correlation be- tween LongBench and the proposed MiniLong- Bench are 0.96 ∼0.98, whether on the LLMs that participated in the training or on other unseen LLMs. Moreover, in conjunction with the results presented in Fig.1, this indicates that the proposed MiniLo... | https://arxiv.org/abs/2505.19959v1 |
(b), we present the results of employing alternative text embeddings, including Longformer (Zhu et al., 2021) and BERT (Liu et al., 2019). We observe that BERT, which only supports token inputs with a maximum length of 512, sig- nificantly underperforms compared to OpenAIEm- bedding and Longformer, which support length... | https://arxiv.org/abs/2505.19959v1 |
the LLMsconsidered in this paper. Using the method men- tioned in Section 3, we obtain various compact new "MiniLongBench" and compute its Sp distri- bution against LongBench evaluation results. The results are shown in Fig. 8. We find that the choice of LLMs involved in training significantly affects the construction ... | https://arxiv.org/abs/2505.19959v1 |
sion methods to bring Sp as close as possible to 1.0 across all subtasks. (9) Why not just random sampling? In Fig. 2, we show that through random sampling, we identify a significant amount of redundancy in LongBench. However, relying solely on random LongBenchMiniLongBenchRank24 068Example 1 Example 2 LLM Index2345 1 ... | https://arxiv.org/abs/2505.19959v1 |
for Computational Linguistics (Volume 1: Long Papers) , pages 14388–14411, Bangkok, Thailand. Association for Computational Linguistics. Anthropic. 2024. Anthropic: Introducing claude 3.5 sonnet. Yushi Bai, Xin Lv, and et al. 2024a. Longbench: A bilin- gual, multitask benchmark for long context under- standing. In Proc... | https://arxiv.org/abs/2505.19959v1 |
preprint arXiv:2307.02486 . Zican Dong, Tianyi Tang, Junyi Li, Wayne Xin Zhao, and Ji-Rong Wen. 2024. Bamboo: A comprehensive benchmark for evaluating long text modeling capac- ities of large language models. In Proceedings of the 2024 Joint International Conference on Compu- tational Linguistics, Language Resources an... | https://arxiv.org/abs/2505.19959v1 |
Boris Ginsburg. 2024. Ruler: What’s the real context size of your long-context language models? arXiv preprint arXiv:2404.06654 . Luyang Huang, Shuyang Cao, Nikolaus Parulian, Heng Ji, and Lu Wang. 2021a. Efficient attentions for long document summarization. In Proceedings of the 2021 Conference of the North American C... | https://arxiv.org/abs/2505.19959v1 |
Li. 2023. Unleashing infinite-length input capacity for large- scale language models with self-controlled memory system. arXiv preprint arXiv:2304.13343 . Haokun Lin, Haoli Bai, Zhili Liu, Lu Hou, Muyi Sun, Linqi Song, Ying Wei, and Zhenan Sun. 2024. Mope- clip: Structured pruning for efficient vision-language models w... | https://arxiv.org/abs/2505.19959v1 |
context. arXiv preprint arXiv:2403.05530 . Aurko Roy, Mohammad Saffar, Ashish Vaswani, and David Grangier. 2021. Efficient content-based sparse attention with routing transformers. Transactions of the Association for Computational Linguistics , 9:53– 68. Noveen Sachdeva and Julian McAuley. 2023. Data dis- tillation: A ... | https://arxiv.org/abs/2505.19959v1 |
Web Search and Data Mining , pages 1090–1093. Wenhan Xiong, Jingyu Liu, Igor Molybog, Hejia Zhang, Prajjwal Bhargava, Rui Hou, Louis Martin, Rashi Rungta, Karthik Abinav Sankararaman, Barlas Oguz, et al. 2024. Effective long-context scaling of founda- tion models. In Proceedings of the 2024 Conference of the North Amer... | https://arxiv.org/abs/2505.19959v1 |
Huang, Weushao Wen, Jinghui Qin, and Liang Lin. 2023a. Sur-adapter: En- hancing text-to-image pre-trained diffusion models with large language models. In Proceedings of the 31st ACM International Conference on Multimedia , pages 567–578. Shanshan Zhong, Jinghui Qin, Zhongzhan Huang, and Daifeng Li. 2022. Cem: Machine-h... | https://arxiv.org/abs/2505.19959v1 |
not only streamline the selection process but also ensure a more systematic and un- biased evaluation of potential models for inclusion in MiniLongBench.Model Type Model Type ALMA-7B-Ja-V2 T Amd-llama-135m A GOAT-7B-Community T Amd-llama-135m-code A Koss-7B-chat T Distilled-HermesChat-7B A Kunoichi-7B T Loyal-Macaroni-... | https://arxiv.org/abs/2505.19959v1 |
0.24 0.08 0.03 0.07 0.01 0.04 Qwen2.5-14B-128k 0.19 0.33 0.43 0.43 0.35 0.42 0.41 0.15 0.06 0.26 0.08 0.03 0.07 0.01 0.05 Qwen2.5-32B-128k 0.21 0.36 0.43 0.46 0.36 0.43 0.40 0.16 0.06 0.26 0.11 0.03 0.09 0.01 0.06 Llama-7B-2k 0.03 0.07 0.17 0.08 0.09 0.06 0.05 0.02 0.02 0.04 0.03 0.01 0.04 0.00 0.02 Llama2-7B-4k 0.05 0... | https://arxiv.org/abs/2505.19959v1 |
6-2 Avg EN ZH All DeepSeek-V3-128k 0.77 0.93 0.63 0.35 0.67 0.05 0.38 0.34 0.26 0.90 0.86 0.88 0.43 0.26 0.45 GPT-4o-mini-128k 0.76 0.93 0.60 0.32 0.65 0.04 0.35 0.28 0.23 0.90 0.85 0.87 0.42 0.24 0.43 GPT-3.5-Turbo-16k 0.70 0.88 0.49 0.23 0.58 0.03 0.27 0.21 0.17 0.84 0.78 0.81 0.37 0.19 0.37 Internlm3-8B-32k 0.61 0.8... | https://arxiv.org/abs/2505.19959v1 |
provide targeted insights into the diverse and complex facets of LCU, contributing to a more nuanced understand- ing and development of LLMs’ long-context pro- cessing abilities. D.3 Low-cost Deep Learning Recently, there has been a surge of efforts aimed at achieving low-cost deep learning, encompass- ing strategies s... | https://arxiv.org/abs/2505.19959v1 |
acceptable range, particularly considering that the Spearman correlation coefficient (Sp) does not reach a perfect 1.0. These visualizations further validate that MiniLongBench achieves evaluation results comparable to LongBench while signifi- cantly reducing computational costs. This high- lights MiniLongBench’s effec... | https://arxiv.org/abs/2505.19959v1 |
these discrepancies do not affect the ranking of LLMs’ LCU capabilities. For instance, Fig. 13 and the main text’s Fig. 5 demonstrate that the results from different evaluation methods are highly consistent, despite minor deviations in score values. This phenomenon primarily arises from several factors: first, MiniLong... | https://arxiv.org/abs/2505.19959v1 |
Qwen2.5-7B-128k 0.50 0.37 0.46 0.58 0.48 0.71 0.66 0.50 0.14 0.50 0.18 0.21 0.14 0.15 0.17 Qwen2.5-14B-128k 0.41 0.43 0.44 0.56 0.46 0.67 0.89 0.44 0.23 0.56 0.16 0.23 0.15 0.16 0.18 Qwen2.5-32B-128k 0.45 0.43 0.42 0.68 0.50 0.66 0.76 0.48 0.21 0.53 0.16 0.21 0.15 0.17 0.17 Llama-7B-2k 0.03 0.11 0.23 0.12 0.13 0.07 0.0... | https://arxiv.org/abs/2505.19959v1 |
0.31 0.00 0.33 0.20 0.18 0.71 0.52 0.61 0.32 0.21 0.33 30B-Epsilon-2k 0.24 0.80 0.26 0.38 0.42 0.00 0.40 0.20 0.20 0.75 0.50 0.62 0.30 0.20 0.31 Table 8: Results on few-shot learning, synthetic, and code tasks based on evaluating directly by MiniLongBench. ‘Overall’ is computed by the macro-average (the mean of ‘Avg’) ... | https://arxiv.org/abs/2505.19959v1 |
arXiv:2505.19964v1 [cs.LG] 26 May 2025The Limits of Preference Data for Post-Training Eric Zhao1,2, Jessica Dai2, and Pranjal Awasthi1 1Google Research 2University of California, Berkeley Abstract Recent progress in strengthening the capabilities of large language models has stemmed from applying reinforcement learning... | https://arxiv.org/abs/2505.19964v1 |
scalar rewards vs. ordinal preferences as data modalities. Our impossibility result is due to a connection between the post-training of models using preference data and the analysis of electoral systems in social choice theory. We show that there is a fundamental challenge with relying on preference data for reinforcem... | https://arxiv.org/abs/2505.19964v1 |
limitations of ordinal data from social choice theory. We thus place our work in context with these communities as follows. Understanding RLHF. A wide range of recent work studies the extent to which RLHF and its variants can successfully learn from preference data. Commonly-observed limitations include failures to lea... | https://arxiv.org/abs/2505.19964v1 |
of distortion (in contrast to, e.g., the earlier non-identifiability results of [ WLJ23]). In contrast to [ GHY25], we are interested in 3 showing unavoidable lower bounds for preference learning that hold generally, even in the absence of heterogeneity, and are thus a property of post-training rather than of resolving... | https://arxiv.org/abs/2505.19964v1 |
g◦ϕ, that results in a higher expected utility than its pretrained base M0, i.e.Eq∼D[u(q,M(q))]≥Eq∼D[u(q,M 0(q))]. We will use Φ⊂ZQto denote the set of possible mappings from queries to internal representations that our model class allows us to learn and that ϕmust thus be chosen from. Finally, we note that even simila... | https://arxiv.org/abs/2505.19964v1 |
for every query q∈Qand every pair of downstream circuits. Concretely, we can write these preferences as orderings ≻u:={≻q,u}q∈Q wheresi≻q,usjif and only if u(q,si(q))>u(q,sj(q)). Note that we can interpret the preferences ≻uas either an oracle that returns preference comparisons on-demand or an infinitely large dataset... | https://arxiv.org/abs/2505.19964v1 |
lower bound We’ll now show that the limitations of learning from preferences is fundamental, rather than the eccentricity of a particular choice of pretrained model or preference learning algorithm (e.g., RLHF as in Example 3.2). In particular, we will show that for anypreference learning algorithm and any pretrained m... | https://arxiv.org/abs/2505.19964v1 |
labels. If labelers consistently indicate that they prefer a responsea1overa2with a probability that is exponential in the difference between the utilities of a2and a1, thevariance of the preferences provides enough information to learn an optimal policy. However, there are several considerations that limit the degree ... | https://arxiv.org/abs/2505.19964v1 |
labelers to coarsen their feedback, and “tie” responses by default unless one is significantly better than another. This is not a general solution, however, and would only work for a narrow set of problems where, e.g. binary, notions of success suffice. Another restriction of the theory is that it assumes only preferen... | https://arxiv.org/abs/2505.19964v1 |
argue that these behaviors can be understood as robustness , and, conversely, that robustness should be understood as a learned strategy for handling reasoning problems [ YJZ+25,ZNB+25]. In the language of our model from Section 2, these reasoning strategies can be thought of corresponding to particular circuits in Sth... | https://arxiv.org/abs/2505.19964v1 |
learned by reasoning models during post-training. In Figure 4.1b, we analyze the backtracking behavior of Deepseek R1 on a set of reasoning problems again from LiveBench Reasoning, AIME, and MATH [ WDR+24,MAA24,HBK+21], with the goal of determining how often backtracking is truly necessary. However, on these problems, ... | https://arxiv.org/abs/2505.19964v1 |
the person that travels by tram doesn’t eat watermelon or both:** This is a crucial clue that helps us eliminate possibilities. Let’s analyze possibilities. Since the avocado lover is in P3, and the tram is to the left of the avocado, the tram must be in P2. The cranberry lover must then be in P1. Since P1 is Japanese,... | https://arxiv.org/abs/2505.19964v1 |
specifically reward backtracking), preference datasets will disencourage these behaviors. In fact, these results almost exactly mimic the scenario illustrated in Example 3.2. The set of circuits S learned in pretraining contains both a backtracking strategy sBacktrack and a direct strategy sDirect, which is evidenced b... | https://arxiv.org/abs/2505.19964v1 |
Filos-Ratsikas, and Alexandros A Voudouris. Peeking behind the ordinal curtain: Improving distortion via cardinal queries. Artificial 12 Intelligence , 296:103488, 2021. [ABFV20] Georgios Amanatidis, Georgios Birmpas, Aris Filos-Ratsikas, and Alexandros A. Voudouris. A few queries go a long way: Information-distortion ... | https://arxiv.org/abs/2505.19964v1 |
Set, 2024. [Mil] Dan Milmo. Chatgpt firm reveals ai model that is ‘good at creative writing’. The Guardian . [MSS+23]Ted Moskovitz, Aaditya K Singh, DJ Strouse, Tuomas Sandholm, Ruslan Salakhutdinov, Anca D Dragan, and Stephen McAleer. Confronting reward model overoptimization with constrained rlhf, 2023. [MYS+25]Nikla... | https://arxiv.org/abs/2505.19964v1 |
Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M. Dai, and Quoc V. Le. Finetuned language models are zero-shot learners. InInternational Conference on Learning Representations (ICLR) , 2022. [WDR+24]Colin White, Samuel Dooley, Manley Roberts, Arka Pal, Ben Feuer, Siddhartha Jain, Ravid Shwartz-Ziv, NeelJain, Kh... | https://arxiv.org/abs/2505.19964v1 |
a multiplicative factor compared to the best model M∗that we could have post-trained from M0: max M∗∈{(ϕ∗,g∗,S0)|ϕ∗∈Φ,g∗:Z→S 0}Eq∼D[u(q,M∗(q))] Eq∼D[u(q,M(q))]/bracehtipupleft /bracehtipdownright/bracehtipdownleft /bracehtipupright (Distortion)≥Ω/parenleftig/radicalig |S0|/parenrightig (2) when|Q|≫|S|,|Z|. Moreover,... | https://arxiv.org/abs/2505.19964v1 |
≻u, our post-trained model M= (ϕ,g,S0) = A(M0,≻u)is now well defined. Recall that, in our notation, Mmaps queries to internal representations usingϕ∈Φand uses internal representations to assign queries to circuits via g∈SZ 0. Definingu.Let us construct a utility function inspired by [ BCH+12]. We will first define a he... | https://arxiv.org/abs/2505.19964v1 |
This time, we want to lower bound the number of queries that belong to the query group Qizthat their representations zsuggest. We will therefore use the second part of Lemma A.2, which guarantees that /summationdisplay z∈Range(ϕ)niz,z≥Ω/parenleftigg |Q| k/parenrightigg −O/parenleftig/radicalig |Q||Z| logk+|Z|logk+/... | https://arxiv.org/abs/2505.19964v1 |
applying Fact A.4, we have from McDiarmid’s inequality that for any fixed group mapping g∈S and any deviation tolerance t>0: P(Nmax(f,g)−E[Nmax(f,g)]≥t)≤exp/parenleftigg −2t2 n/parenrightigg , and P(Nmin(f,g)−E[Nmin(f,g)]≤−t)≤exp/parenleftigg −2t2 n/parenrightigg . Letµ∗ max= supg∈SE[Nmax(f,g)]andµ∗ min= infg∈SE[Nm... | https://arxiv.org/abs/2505.19964v1 |
will now fix any representation z∈Zthat has a non-empty pre-image ϕ−1(z)̸=∅. Becauseg(z) is a probability distribution over circuits S0, the pigeonhole principle means that there must exist some group index iz∈[k]whereg(z)places little weight on circuit siz∈S0:Pr(g(z) =siz)≤1/k. We will define the maximum utility of an... | https://arxiv.org/abs/2505.19964v1 |
to a model’s generated chains-of-thought, rather than adversarial manipulations of user queries. B.2 Experimental details for Figure 4.1b. For this experiment, we selected two questions from each of the following datasets: 2024 AIME, Berkeley MATH (PRM800K test split), LiveBench Reasoning Zebra Puzzles, LiveBench Reaso... | https://arxiv.org/abs/2505.19964v1 |
seems a bit confusing. Let me try to organize this better. First, identify who is where and what they say. 1. Skate Park: Person tells the truth. 2. Barbershop: Person tells the truth. 3. Nia: At the cafe, which tells the truth. ... B.3 Experimental details for Figure 4.3 For this experiment, we selected 50 questions a... | https://arxiv.org/abs/2505.19964v1 |
or no (for example, **yes, no, yes**). If you don’t know, guess. Okay, let me try to work through this logic puzzle step by step. There are a lot of statements here, so I need to organize the information carefully. Let’s start by listing out all the given facts and statements, and then see how they connect. First, the ... | https://arxiv.org/abs/2505.19964v1 |
sin(120◦) = 30∗(√ 3/2) = 15∗√ 3. Yes, that’s correct. Alternatively, maybe we can compute the area using Heron’s formula to verify. Let’s try that. Heron’s formula states that area =/radicalbig s(s−a)(s−b)(s−c), wheresis the semi-perimeter. Compute semi-perimeter s= (6 + 10 + 14) /2 = 30/2 = 15. Then area =/radicalbig ... | https://arxiv.org/abs/2505.19964v1 |
User Query ‘‘‘ Query ‘‘‘ # Solution A ‘‘‘ Model Response A ‘‘‘ # Solution B ‘‘‘ Model Response B ‘‘‘ Short RLHF Instructions You are performing RLHF data labeling for a language model. Below is a math question and two solutions provided by different models: Solution A and Solution B. Both solutions are correct. Your ta... | https://arxiv.org/abs/2505.19964v1 |
backtrack. Prompt 2 For each instance of backtracking that you have found, determine if the backtracking is superfluous. We say that a backtracking is superfluous if any of the following holds: * The model backtracks 30 because it believes it made an error, and you verify that the model did not actually make an explici... | https://arxiv.org/abs/2505.19964v1 |
CP-Router: An Uncertainty-Aware Router Between LLM and LRM Jiayuan Su1∗Fulin Lin1∗Zhaopeng Feng1∗Han Zheng1Teng Wang2 Zhenyu Xiao3Xinlong Zhao4Zuozhu Liu1†Lu Cheng5Hongwei Wang1† 1Zhejiang University2University of Hong Kong3Tsinghua University 4Peking University5University of Illinois Chicago {jiayuan.23, fulin1.24, zh... | https://arxiv.org/abs/2505.19970v1 |
More recently, advances such as Claude-3.7-SonnetarXiv:2505.19970v1 [cs.CL] 26 May 2025 (Anthropic, 2025) and Qwen3 (Yang et al., 2025a) empower users with manual controls to switch be- tween reasoning and standard modes. However, this places a burden on users to make effective rout- ing decisions. These limitations hi... | https://arxiv.org/abs/2505.19970v1 |
LLM and LRM based on esti- mated prediction uncertainty. CP offers strong coverage guarantees and produces interpretable prediction sets, making it particularly well-suited for prompt routing without requiring any addi- tional training or model-specific tuning. •We propose a novel automatic calibration method that leve... | https://arxiv.org/abs/2505.19970v1 |
ble for a given input under a user-defined error rate. The prediction set for an input xis defined as C(x) ={y∈ Y:S(x, y)≤ˆq}, (1) where S(x, y)is the score function and ˆqis the quantile threshold computed from the calibration set. The size of the prediction set can be inter- preted in two complementary ways. First, u... | https://arxiv.org/abs/2505.19970v1 |
intuition, we conduct experiments us- ing Qwen2.5-14B (Yang et al., 2025b) (LLM) and DeepSeek-R1-Distill-Qwen-14B (Guo et al., 2025) 0.0 0.2 0.4 0.6 0.8 1.0CN-ChemistryGPQA 0.8210.252 0.7180.440(a) Accuracy Qwen2.5-14B DeepSeek-R1- Distill-Qwen-14B 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0CN-ChemistryGPQA 1.173.26(b) Average... | https://arxiv.org/abs/2505.19970v1 |
complementary perspectives. The full entropy term measures the overall diversity across all prediction set sizes, help- ing to ensure well-separated uncertainty estimates. The binary entropy term captures the balance be- tween singleton and non-singleton prediction sets, which is critical for avoiding routing load imba... | https://arxiv.org/abs/2505.19970v1 |
andDeepSeek-R1-Distill-Llama-8B LLM - 41.6 - - 31.9 - - 31.2 - - 35.9 - - 29.6 - - 24.4 - - LRM - 79.9 - 38.3 70.4 - 38.5 58.8 - 27.6 46.4 - 10.5 34.0 - 4.4 65.0 - 40.6 Random0.2 72.3 19.8 38.3 63.0 19.3 38.5 53.8 16.0 26.9 44.7 19.7 11.0 33.1 19.9 4.4 57.3 19.4 40.8 0.3 69.0 29.4 38.8 59.1 29.4 38.5 51.5 25.3 27.2 43.... | https://arxiv.org/abs/2505.19970v1 |
6.3 49.4 59.7 18.7 18.2 44.7 3.1 20.1 79.4 0.2 49.1 Explicit - 67.0 45.5 36.9 57.4 44.0 46.3 57.5 41.3 36.3 51.8 38.4 11.2 38.1 38.4 21.0 58.3 46.0 51.7 Dynathink - 77.3 1.3 30.8 71.3 5.3 42.0 72.5 5.8 38.5 52.3 17.0 8.0 36.2 12.6 12.6 62.1 1.0 32.0 CP-Router - 92.4 22.8 58.9 84.7 7.3 57.4 82.5 7.5 50.1 59.5 21.7 18.7 ... | https://arxiv.org/abs/2505.19970v1 |
presents a case where the LLM achieves higher ac- curacy in the 14B setting. We evaluate CP-Router under this condition and observe that it outperforms the LLM by more than 0.8% in terms of accu- racy. In the 8B pairing, the LRM achieves higher standalone performance, yet CP-Router routes ef- fectively and still improv... | https://arxiv.org/abs/2505.19970v1 |
6: Accuracy and prompt allocation of CP- Router on the open-ended QA benchmark GSM8K, us- ingLlama-3-8B andDeepSeek-R1-Distill-Llama-8B , demonstrating its generalization to broader QA formats. 5.5 CP-Router Generalizes Across Diverse Model Pairings CP-Router is designed to be model-agnostic, ca- pable of routing promp... | https://arxiv.org/abs/2505.19970v1 |
differentiate uncertainty across inputs. Extensive experiments across diverse tasks demon- strate that CP-Router substantially reduces token usage while maintaining—and occasionally ex- ceeding—the accuracy of LRMs. We further show that CP-Router generalizes across model scales and families, and can be extended beyond ... | https://arxiv.org/abs/2505.19970v1 |
question answering. arXiv preprint arXiv:2305.18404 . Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. 2024. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437 . Jian Liu, Leyang Cui, Hanmeng Liu, Dandan Huang, Yile Wang, and ... | https://arxiv.org/abs/2505.19970v1 |
Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jing Zhou, Jingren Zhou, Junyang Lin, Kai Dang, Keqin Bao, Kexin Yang, Le Yu, Lianghao Deng, Mei Li, Mingfeng Xue, Mingze Li, Pei Zhang, Peng Wang, Qin Zhu, Rui Men, Ruize Gao,... | https://arxiv.org/abs/2505.19970v1 |
LLM or LRM. A pre- defined threshold controls the proportion di- rected to the LRM. 2.Top-1 Probability Routing : The model se- lects the option with the highest probability as a proxy for confidence. If this probability exceeds a predefined threshold, the LLM isused; otherwise, the LRM is chosen. 3.Response Entropy Ro... | https://arxiv.org/abs/2505.19970v1 |
can include selfless acts? Maybe this is trying to say that it’s a more basic level, not expecting much from people. Not sure yet.Option B: "The bottom line ethics are some of the most basic and fundamental codes of conduct and rules that should be observed by ordinary people in a society." This sounds like it’s talkin... | https://arxiv.org/abs/2505.19970v1 |
0.357, "D": 0.186 Prediction Set: [A, C] Routing? ✓. Use LRM. DeepSeek-R1-Distill-Llama- 8BAlright, so I’m trying to figure out this question about quantum theory and why it’s accepted despite violating intuition. Let’s break it down step by step.First, the question says that quantum theory’s conclusion goes against ou... | https://arxiv.org/abs/2505.19970v1 |
accept it. So, option D is about the outcome of falsification attempts leading to acceptance. Option C is more about the general condition—only accept if not falsified.The question is about showing the rationality of accepting quantum theory despite violating intuition. The reasoning relies on the fact that attempts to... | https://arxiv.org/abs/2505.19970v1 |
arXiv:2505.19971v1 [cs.CL] 26 May 2025Conversational Lexicography: Querying Lexicographic Data on Knowledge Graphs with SPARQL through Natural Language Kilian Sennrich Department of Informatics University of Zurich kilian.sennrich@uzh.chSina Ahmadi Department of Computational Linguistics University of Zurich sina.ahmad... | https://arxiv.org/abs/2505.19971v1 |
teachers, and other non-technical stakeholders who could benefit from lexicographic data access (Warren and Mulholland, 2020). This paper addresses the significant research gap in creating effective natural language interfaces for lexicographic data retrieval on KGs such as Wiki- data. To that end, we develop a multidi... | https://arxiv.org/abs/2505.19971v1 |
tures. The emergence of LLMs has further trans- formed this landscape (Perevalov and Both, 2024). D’Abramo et al. (2025) apply in-context learning using Mixtral (8x7B), Llama-3 (70B), and CodeL- lama (70B) to achieve state-of-the-art results, while other approaches demonstrate success through fine- tuning (Brei et al.,... | https://arxiv.org/abs/2505.19971v1 |
Criterion 1 : It should encompass the full range of SPARQL syntax constructs and oper- ators Criterion 2 : It should cover the variety of use cases for lexicographic data Criterion 3 : It should be particularly detailed in frequently queried areasThese criteria guided the identification of four feature dimensions (D) t... | https://arxiv.org/abs/2505.19971v1 |
et al., 2023) using the Low-Rank Adapta- tion (LoRA) framework. Phi-1.5 is a small language model with 1.3B parameters that demonstrates strong capabilities in both nat- ural language and code generation. For fine- tuning, we use the following hyperparameters: learning rate of 0.0002, train batch size of 4, Adam optimi... | https://arxiv.org/abs/2505.19971v1 |
utterances to SPARQL queries tar- geting lexicographic data in Wikidata, we adopt a template-based approach similar to Soru et al. (2017) based on the taxonomies defined in Section 3.1. Each data point in our templates consists of three elements: 1.utterance : natural language input reflecting a user’s question; 2.temp... | https://arxiv.org/abs/2505.19971v1 |
restricted to Wik- ibaseLexemes) to prevent overfitting. Two versions of utterance templates were used: single-lexeme and multi-lexeme. Language-Independent Templates These tem- plates function without specifying the lexeme’s language, enabling cross-language lookups. They use string matching ( FILTER(STR(?lemma) = "wo... | https://arxiv.org/abs/2505.19971v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.