text
string
source
string
human feedback in llms. arXiv preprint arXiv:2402.14740 , 2024. [2]Eden Biran, Daniela Gottesman, Sohee Yang, Mor Geva, and Amir Globerson. Hopping too late: Exploring the limitations of large language models on multi-hop queries. arXiv preprint arXiv:2406.12775 , 2024. [3]Peter Clark, Isaac Cowhey, Oren Etzioni, Tusha...
https://arxiv.org/abs/2505.18454v1
Zhenrui Yue, Dong Wang, Hamed Zamani, and Jiawei Han. Search-r1: Training llms to reason and leverage search engines with reinforcement learning. arXiv preprint arXiv:2503.09516 , 2025. [19] Mandar Joshi, Eunsol Choi, Daniel S Weld, and Luke Zettlemoyer. TriviaQA: A large scale distantly supervised challenge dataset fo...
https://arxiv.org/abs/2505.18454v1
Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300 , 2024. [34] Xuan Shen, Yizhou Wang, Xiangxi Shi, Yanzhi Wang, Pu Zhao, and Jiuxiang Gu. Efficient reasoning with hidden thinking. arXiv p...
https://arxiv.org/abs/2505.18454v1
Zhenrui Yue, Huimin Zeng, Yimeng Lu, Lanyu Shang, Yang Zhang, and Dong Wang. Evidence- driven retrieval augmented response generation for online misinformation. arXiv preprint arXiv:2403.14952 , 2024. [51] Zhenrui Yue, Honglei Zhuang, Aijun Bai, Kai Hui, Rolf Jagerman, Hansi Zeng, Zhen Qin, Dong Wang, Xuanhui Wang, and...
https://arxiv.org/abs/2505.18454v1
prepending a system message that directs the LLM to perform step-by-step internal reasoning before generating its final answer. The user query is then appended, and the entire input is formatted with the model chat template. Different from 2https://github.com/unslothai/unsloth 14 prior work [ 10,18], we adopt the minim...
https://arxiv.org/abs/2505.18454v1
a question, and the assistant solves it. The assistant first thinks about the reasoning process in the mind and then provides the user with the answer. The final answer is provided after the #### tag, i.e., {reasoning process} #### {answer}.<|im_end|> <|im_start|>user Question: Two people are pushing a car. One person ...
https://arxiv.org/abs/2505.18454v1
Qwen 3B backbone. Init RangeKnowledge NQ TriviaQA HotpotQA 2WikiMQA Bamboogle Average [0.95 - 0.999] 0.845 0.613 0.622 0.576 0.820 0.695 [0.98 - 0.999] 0.842 0.600 0.614 0.585 0.813 0.691 [0.99 - 0.999] 0.838 0.606 0.630 0.590 0.817 0.696 Init RangeSTEM GSM8k MATH MATH500 MMLU-ST ARC-C Average [0.95 - 0.999] 0.367 0.59...
https://arxiv.org/abs/2505.18454v1
20, Figure 21, where as the mistakes are provided in Figure 22, Figure 23, Figure 24, Figure 25, Figure 26, we show the raw strings and omit the options / contexts in the examples due to space constraints. From these examples, we identify four reasoning patterns that can lead to correct answers: (1) Purely English reas...
https://arxiv.org/abs/2505.18454v1
Query: Which religion uses a domed structure called a Stupa as a shrine?HRPO (Tokens in Hybrid Rollout): To determine which religion uses a domed structure called a Stupa as a shrine, let's consider the provided context and the information we need to extract:\n\n- Stupa is described as a \"mound-like or hemispherical s...
https://arxiv.org/abs/2505.18454v1
exponent rule %%((-a)ˆ{-1} = -aˆ{-1}%%) but )%nI am_____ (leave the word in the blank) and the study of music will be called ____. {reasons} to____. For this, the Rhythm Method is the only way to teach music.%n{reasoning process} #### To understand the reasoning process behind the given problem, let’s break down the co...
https://arxiv.org/abs/2505.18454v1
by 3, so 1391 is not divisible by 3. It is not divisible by 5 either, but it is divisible by 7: 1391 ÷ 7 = 198.3571. The product of the digits of 1391 is 1 × 3 × 9 × 1 = 27, which is divisible by 3, so 1391 is divisible by 3. Therefore, 1391 = 3 × 463. %n%nNow we have the prime factorization of 1391 as 1391 = 3 × 463. ...
https://arxiv.org/abs/2505.18454v1
\\(x = \\frac{-b \\pm \\sqrt{b^2 - 4ac}}{2a}\\).\n\nHere, the non-integer root \\(n\\) is the root closest to but not equal to an integer.\n\nSo, the non-integer root \\(n\\) in the equation, \\(n = \\frac{9 \\pm \\sqrt{81 - 100}}{10}\\), which simplifies to \\(n = \\frac{9 \\pm \\sqrt{-19}}{10}\\).\n\nHowever, \\( \\s...
https://arxiv.org/abs/2505.18454v1
arXiv:2505.18456v1 [cs.CL] 24 May 2025Anchored Diffusion Language Model Litu Rout Constantine Caramanis Sanjay Shakkottai The University of Texas at Austin {litu.rout,constantine,sanjay.shakkottai}@utexas.edu Abstract Diffusion Language Models (DLMs) promise parallel generation and bidirectional context, yet they under...
https://arxiv.org/abs/2505.18456v1
low-frequency or semantically important words) are masked early in the forward process, the model lacks sufficient context to accurately reconstruct the original sequence. Drawing on information-theoretic insights and improved sample complexity via anchoring in directed graphical models (DAGs), we propose the Preprint....
https://arxiv.org/abs/2505.18456v1
demonstrate the benefits of anchoring using two different samplers: (a) locked-in (Sahoo et al., 2024) and (b) remasking (Wang et al., 2025a) samplers. With remasking sampler, ADLM outperforms AR models in human-like text generation measured by MAUVE score (§5.1). •Beyond diffusion, we integrate our anchoring mechanism...
https://arxiv.org/abs/2505.18456v1
=QL l=1pθ(zl s|zt). Intuitively, given a noisy latent zt, the model predicts a clean token and then re-noises it forward according to the forward dynamics defined in (1). Recall that xdenotes a sequence of K-dimensional one-hot encoded tokens, i.e., x= (xl)L l=1. We slightly overload notation and use xθ= (xl θ)L l=1to ...
https://arxiv.org/abs/2505.18456v1
We propose to break the one-step denoising process, widely used in practice (Austin et al., 2021; Sahoo et al., 2024; Wang et al., 2025a; Ou et al., 2025; Nie et al., 2025a,b), into a two-stage anchored denoising framework. This allows latent reasoning over important tokens during pretraining. Since the reverse process...
https://arxiv.org/abs/2505.18456v1
Evidence Lower Bound (ANELBO) (seeTheorem 4.1 ): LANELBO (x,y;φ, ψ) =EZ0∼q(·|x)[−logpψ(x|yφ(Z0))] + (7) TX i=1EZt(i)∼q(·|x)" (1−σt(i))αt(i)−αs(i) 1−αt(i)LX l=1log⟨xl ψ(yφ(Zt(i))),xl⟩+γlog⟨yl φ(Zt(i)),yl⟩# , where γcontrols anchor strength. For σt(i)= 0 = γ, we recover the standard MDLM (4). Anchor Token Selection. We i...
https://arxiv.org/abs/2505.18456v1
AR and DLM training as learning in directed graphical models (DAGs) and formally analyze our anchoring mechanism. While rooted in classical theory, we demonstrate that anchoring yields practical benefits in both large-scale pretraining (§5.1) and supervised fine-tuning (§5.2) tasks. Assumption 4.3. Suppose the followin...
https://arxiv.org/abs/2505.18456v1
ReMDM (Wang et al., 2025a), which allows re-masking with a small σt̸= 0. For fair comparison, we adopt the exact sampler configurations used in the respective baseline implementations. Baselines. We compare against the following baselines: (1) the Autoregressive (AR) architecture from (Sahoo et al., 2024) trained with ...
https://arxiv.org/abs/2505.18456v1
524B tokens. At each scale, ADLM consistently outperforms diffusion-based baselines such as MDLM and GIDD, as well as the hybrid (AR+Diffusion) BD3LM. Notably, at 262B tokens, ADLM achieves a PPL of 20.62, narrowing the gap with the AR models, which reach a PPL of 17.54. ADLM∗uses our multi-stage design (anchor and den...
https://arxiv.org/abs/2505.18456v1
T=128 T=256 T=512 SEDD (absorb) 0.007 0.007 0.008 119.2 110.1 107.2 5.65 5.63 5.62 MDLM 0.015 0.023 0.031 61.5 55.8 53.0 5.52 5.49 5.48 MDLM+FB 0.064 0.084 0.100 42.8 39.6 37.1 5.44 5.41 5.38 MDLM+DFM 0.041 0.144 0.211 37.9 26.5 23.3 5.31 5.26 5.23 ReMDM 0.057 0.216 0.350 42.5 30.5 21.1 5.43 5.34 5.21 ADLM (ours) 0.140...
https://arxiv.org/abs/2505.18456v1
arithmetic reasoning, (2) ProntoQA (Saparov & He, 2023)–rule-based logical reasoning, and (3) ProsQA (Hao et al., 2024)– planning with structured reasoning over graph-based inference traces. Our experimental setup follows the fine-tuning protocols outlined in prior work (Hao et al., 2024), enabling direct comparison wi...
https://arxiv.org/abs/2505.18456v1
language modeling by leveraging anchor tokens (e.g., low-frequency or important key words). We provide theoretical justification along with strong empirical evidence supporting our results. Our method bridges the gap between diffusion and AR models in likelihood modeling and generated text quality. ADLM significantly r...
https://arxiv.org/abs/2505.18456v1
Levy, and Christopher D Manning. What does bert look at? an analysis of bert’s attention. In Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP , pp. 276. Association for Computational Linguistics, 2019. URL https://aclanthology.org/W19-4828/ . Karl Cobbe, Vineet Kosara...
https://arxiv.org/abs/2505.18456v1
Proc. Workshop on Pattern Recognition in Practice, 1980 , 1980. Michael I Jordan, Zoubin Ghahramani, Tommi S Jaakkola, and Lawrence K Saul. An introduction to variational methods for graphical models. Machine learning , 37:183–233, 1999. Urvashi Khandelwal, Omer Levy, Dan Jurafsky, Luke Zettlemoyer, and Mike Lewis. Gen...
https://arxiv.org/abs/2505.18456v1
The Thirteenth International Conference on Learning Representations , 2025. URL https://openreview.net/forum?id=sMyXP8Tanm . Denis Paperno, Germán Kruszewski, Angeliki Lazaridou, Ngoc-Quan Pham, Raffaella Bernardi, Sandro Pezzelle, Marco Baroni, Gemma Boleda, and Raquel Fernández. The lambada dataset: Word prediction r...
https://arxiv.org/abs/2505.18456v1
Lluís Màrquez (eds.), Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , pp. 4593–4601, Florence, Italy, July 2019. Association for Computational Linguistics. doi: 10.18653/v1/P19-1452. URL https://aclanthology. org/P19-1452/ . Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier...
https://arxiv.org/abs/2505.18456v1
>1is straightforward following the standard analysis (Sohl-Dickstein et al., 2015). We start from the standard negative log-likelihood: −logpθ(x) =−logZ pθ(x, Z0, . . . , Z 1)d(Z0, . . . , Z 1) =−logZpθ(x, Z0:1) q(Z0:1|x)q(Z0:1|x)d(Z0:1). Applying Jensen’s inequality yields: −logpθ(x)≤Eq(Z0:1|x)" −logpθ(x|Z0) + logq(Z1...
https://arxiv.org/abs/2505.18456v1
bound on the data manifold. When the anchor and denoising networks are properly aligned with the inference posterior, the KL terms decompose nicely under our parameteri- zation. This leads to a variational bound that reflects both token-level reconstruction and anchor-level guidance, enabling effective learning in larg...
https://arxiv.org/abs/2505.18456v1
distribution p(xl|·)is modeled as a categorical distribution. •The model is parameterized by Conditional Probability Tables (CPTs); that is, a distinct parameter is assigned to each possible configuration of the conditioning set. •Anchor sets πl⊂ {1, . . . , L } \ {l}are fixed and of bounded size |πl| ≤d, with d≪L. Pro...
https://arxiv.org/abs/2505.18456v1
is to estimate the parameters of a conditional probability model p(x|θ), where θ= [ψ, φ], using MLE. The standard MLE objective is defined as: LMLE(ψ, φ) =NX i=1logp(xi|ψ, φ). Since the structure of the underlying graphical model Gis typically unknown, a common modeling assumption is to use a fully autoregressive facto...
https://arxiv.org/abs/2505.18456v1
this work, we adopt a simple yet effective information-theoretic strategy: tokens with low marginal frequency in the given sample tend to carry more information (§3). Hence, we treat such low-frequency tokens as candidates for anchoring in language modeling tasks. Our empirical results support this strategy across two ...
https://arxiv.org/abs/2505.18456v1
first decode the important tokens before reconstructing the rest of the sequence. Summary. By leveraging inductive biases about the distribution of important tokens within sequences, our anchoring approach achieves substantial improvements in sample complexity. Crucially, it avoids the combinatorial explosion of full-c...
https://arxiv.org/abs/2505.18456v1
A.6 guarantees that the anchored EM procedure produces non-increasing negative log-likelihood at each iteration, thereby ensuring stability and convergence to a first-order stationary point. Notably, this convergence behavior emerges even though the anchor tokens are unobserved (Kwon et al., 2024); they are estimated i...
https://arxiv.org/abs/2505.18456v1
on the last term: X yr(y|x, φi) logp(y|x, ψi+1, φi+1) p(y|x, ψi, φi) ≤logX yr(y|x, φi)p(y|x, ψi+1, φi+1) p(y|x, ψi, φi) = logX yp(y|x, ψi+1, φi+1) = log 1 = 0 , where we use the fact that r(y|x, φi) =p(y|x, ψi, φi)and that p(·|x,·,·)is a valid probability distribution. Monotonicity. Thus, we conclude: −logp(x|ψi+1,...
https://arxiv.org/abs/2505.18456v1
, l∈ {1,2,3} (13) Here, zl tdenotes the corrupted version of token xlat time t, and Cat(·)denotes the categorical distribution over the vocabulary V={0,1, m}. The vector xl∈R3is a one-hot column vector corresponding to the original token xl, andm= [0,0,1]⊤is the one-hot vector for the mask token. To make this concrete,...
https://arxiv.org/abs/2505.18456v1
1/3,z2 1/3,z3 1/3 respectively. We now compute the likelihood of decoding the target partial sequence (m,0,m), i.e., correctly unmasking the important token: pθ(z1/3= (m,0,m)|z2/3= (m,m,m)) = (0 .5)·(0.45)·(0.5) = 0 .1125. This probability reflects the chance of correctly decoding only the important token using the sta...
https://arxiv.org/abs/2505.18456v1
generative modeling (§5.1) and complex reasoning benchmarks (§5.2) in the main draft (§5) and also in the Appendix C. We hope it offers a compelling justification for the use of anchoring as a general framework for language modeling. B Additional Background and Related Works In this section, we provide extended backgro...
https://arxiv.org/abs/2505.18456v1
that gradually introduces continuous latent “thoughts” between the question and answer. When integrated with our anchoring mechanism, this approach also yields strong performance on symbolic reasoning benchmarks like ProsQA (see §C.2). CODI (Continuous Chain-of-Thought via Self-Distillation). CODI (Shen et al., 2025) d...
https://arxiv.org/abs/2505.18456v1
and autoregressive models in §C.2. We present benchmark datasets, training procedure, ablation studies, and additional quantitative and qualitative results to support the findings discussed in the main paper. Broader Impact. This work introduces an anchoring framework that improves likelihood modeling and generation qu...
https://arxiv.org/abs/2505.18456v1
and uniform noising in discrete diffusion models. As a concurrent work to ReMDM, it addresses a key limitation of MDLM (i.e., the inability to revise tokens once unmasked) by allowing previously unmasked tokens to be updated during inference. This flexibility enables the model to iteratively correct its own mistakes an...
https://arxiv.org/abs/2505.18456v1
We measure the perplexity on the validation sets of the following benchmarks: •Lambada (Paperno et al., 2016): This benchmark evaluates the ability of language models to predict a target word based on a broad context. Each example consists of a short narrative (context ) followed by a target sentence with its final wor...
https://arxiv.org/abs/2505.18456v1
in Algorithm 1 , which employs the standard locked-in sampler (Sahoo et al., 2024) with a remasking schedule σt(Wang et al., 2025a). The locked-in sampler is a special case obtained by setting σt= 0. Implementation Details for OWT. As OWT dataset does not provide an official train/test split, we use the splits used in ...
https://arxiv.org/abs/2505.18456v1
configuration, the training loss and validation PPL per iteration is shown in Figure 2. Implementation Details for LM1B. The experimental setup follows prior works (Sahoo et al., 2024; Wang et al., 2025a). For LM1B, we follow the same setup as OWT, except a shorter sequence length of 128 tokens and the BERT-base-uncase...
https://arxiv.org/abs/2505.18456v1
and outperforming BERT-Large+Gibbs while using at least 40M fewer parameters. This indicates that anchoring enables more efficient denoising compared to prior diffusion language models. Moreover, ADLM demonstrates a consistent reduction in perplexity as the number of sampling steps increases, validating the role of ite...
https://arxiv.org/abs/2505.18456v1
was everyone else.” Probst Steve Watson. Never as good as our pundit, we’ll repair the damage we’ve done! Tell a Random Story A large facet of the new season will be to tell a random story, not sure if it’s how the players are portrayed. But it’s fairly obvious that this specificity will be replaced by community. The p...
https://arxiv.org/abs/2505.18456v1
school group on the controversial side of history. Now Jay and the photographer will have an unique way of representing the tribe while keeping the story relevant. Though the cast will have something to do with it (gohan, skinnyfoots), there will be some mismatched players who generally won’t fill up the ample, talente...
https://arxiv.org/abs/2505.18456v1
participant had to remain in cruel conditions, including medical experimentation, electric lashings, etc., while pregnancy rates were very low[ [ ]. The sterilized prisoners were not only much more likely to have babies than the prisoners with no pregnancy; they were less likely to have abortions. Researchers found tha...
https://arxiv.org/abs/2505.18456v1
we use 4096 sampling steps for ADLM with remasking sampler. The generated text demonstrates strong discourse-level coherence, well-structured paragraphs, and natural transitions between topics. For instance, the model begins with a detailed sports commentary and transitions smoothly into a socio-political news report. ...
https://arxiv.org/abs/2505.18456v1
down high and out of bounds. 37 Oklahoma City’s bench never stopped suiting as the game went on. This is a team that is making a leap in the NBA, with DeMarcus Cousins and Chauncey Billups leading the way, and it’s just the right thing to do in such a crazy situation.<|endoftext|>(CN) Hundreds of Seattle residents Tues...
https://arxiv.org/abs/2505.18456v1
Pulfres; Erie it’s Duane Wright than Vin Diesel) and he’s the current fourth student at the entire campus, as far as entrance per pupil is well. According to Wright, Shaw was about to thirsty with traders celebrating with an invented flavoral beer and the extra muggers when he approached the team with a shot on Irving,...
https://arxiv.org/abs/2505.18456v1
understood that the load followed the closure of the university’s reserve vice chancellor subsequently who died of Parkinson’s syndrome as his wife’s tombstone was on fire. Catherine Baird of Parliamentary Coalition for Government (Scrabble) leader Lassie Mann joined further the condemnation of the stalls. Ms Baird sai...
https://arxiv.org/abs/2505.18456v1
the standard next-token prediction objective, but conditioned on the output logits from the anchor network. Rather than sampling tokens and re-embedding them, we project the anchor logits through a linear layer and feed the resulting representations directly into the LLM transformer, similar to the ADLM setup discussed...
https://arxiv.org/abs/2505.18456v1
the AR setting. In this setting, the model does not have access to structured reasoning traces as in the Math and Logic benchmarks described below. The training and validation splits follow the same setup used in the diffusion experiments (§C.1.2). Math (GSM8K). The GSM8K dataset (Cobbe et al., 2021) contains grade-sch...
https://arxiv.org/abs/2505.18456v1
traces contain redundant information, increasing entropy and making the reasoning process harder to learn. By supervising the model through a small set of important tokens extracted from the reasoning trace, ACoT encourages more structured intermediate computations, guiding the model to reason in a more targeted and in...
https://arxiv.org/abs/2505.18456v1
Test perplexities (PPL; ↓) for standard AR models and our anchored variant (A2R) at various training scales.†Results from (Sahoo et al., 2024). A2R consistently improves perplexity by introducing a two-stage prediction process: anchor tokens are first predicted, then used to guide next-token prediction. Model PPL ( ↓) ...
https://arxiv.org/abs/2505.18456v1
16 - 3 - 4 = 9 , and then 9 * 2 = 18 ,following the order in which quantities appear in the question . In contrast, ACoT introduces [ANT] to capture important tokens, which allows it to reason more globally. Specifically, ACoT first computes 3 + 4 = 7 to aggregate all consumption before subtracting from the total ( 16 ...
https://arxiv.org/abs/2505.18456v1
that gradually removing reasoning steps during multi-stage training provides a significant performance boost on ProsQA. Following this recommendation, we integrate reasoning step removal into our ACoT fine-tuning. This modification improves our model’s accuracy from 81% to 97.3%. Importantly, ACoT generates much less t...
https://arxiv.org/abs/2505.18456v1
meal of the day if the size of Wendi’s flock is 20 chickens? CoT «3*20=60» «60-15-25=20» Answer 20 Full Output [ANT][ANT] «15+25=40» «40-20=20» Extracted Output 20 45 Table 10: Examples of logical reasoning tasks with symbolic reasoning traces from Pron- toQA (Saparov & He, 2023). Each row shows the input question, gro...
https://arxiv.org/abs/2505.18456v1
al., 2022) 77.5 ±1.9 49.4 iCoT (Deng et al., 2024) 98.2±0.3 8.2 COCONUT†(Hao et al., 2024) 97.0 ±0.3 14.2 - Pause†96.6±0.8 8.2 ACoT (ours) 97.3±0.2 8.2 47 Table 12: Examples of logical reasoning tasks with symbolic reasoning traces from ProsQA (Hao et al., 2024). Each row shows the input question, groundtruth reasoning...
https://arxiv.org/abs/2505.18456v1
is a impus. Every jompus is a gerpus. Every boompus is a rompus. Fae is a boompus. Every boompus is a kerpus. Every zumpus is a bompus. Max is a rempus. Every rompus is a kerpus. Max is a impus. Every rempus is a impus. Every wumpus is a yumpus. Every grimpus is a terpus. Every tumpus is a jompus. Every yumpus is a fel...
https://arxiv.org/abs/2505.18456v1
arXiv:2505.18458v2 [cs.DB] 27 May 20251 A Survey of LLM×DATA Xuanhe Zhou∗¶, Junxuan He∗¶, Wei Zhou∗¶, Haodong Chen∗¶, Zirui Tang∗¶, Haoyu Zhao∗¶, Xin Tong∗, Guoliang Li†, Youmin Chen∗, Jun Zhou∗, Zhaojun Sun∗, Binyuan Hui‡, Shuo Wang†, Conghui He§, Zhiyuan Liu†, Jingren Zhou‡, Fan Wu∗ ∗Shanghai Jiao Tong University†Tsi...
https://arxiv.org/abs/2505.18458v2
LLMs ’ full potential in these applications (DATA4LLM ). It includes efficient and scalable solutions for data processing, storage, and serving across the LLM lifecycle, as evidenced in recent academic studies [159], [287], [256] and industry reports [329], [442], [69], [39]. Conversely, LLM -powered techniques are inc...
https://arxiv.org/abs/2505.18458v2
with com- plex data samples [245], [78], [74]. For instance, standardizing date formats (e.g., “Fri Jan 1st 10:36:28 2021” vs. “1996.07.10AD at 15:08:56”) or resolving textual inconsistencies (e.g., “Monticello VA, Jasper” vs. “Monticello VAA”) typically requires intricate programming scripts or handcrafted con- strain...
https://arxiv.org/abs/2505.18458v2
such as complex layout analysis [204], [18], [399], [182], [398], [415], [259], [328], [414]. •Data Deduplication. Data deduplication aims to identify du- plicates in large-scale textual or multi-modal data, including exact string matching [123], [301], hash identification [88], [81], [123], [301], [352], [363], [209],...
https://arxiv.org/abs/2505.18458v2
Data Storage for LLMs ( §2.3). We review data storage techniques for LLMs from the following main aspects. •Data Formats. We review commonly-used dataset and model data formats for LLMs . Dataset formats include TFRecord [44], MindRecord [40] for multimodal data, and tf.data.Dataset that can be directly fed into LLMs [...
https://arxiv.org/abs/2505.18458v2
and re-ranking [129], [12], [320], [47]. •LLM Data Compression. LLM data compression aims to compress the model’s input data to stay within the context window limit or to facilitate model understanding. Relevant techniques include: (1) RAG knowledge compression (e.g., rule-based [436], [353], [202] and model-based meth...
https://arxiv.org/abs/2505.18458v2
queries with GQL generation (e.g., R3-NL2GQL [501]) and knowledge-aware QA by retrieving or reasoning over relevant subgraphs [433]. •Semi-Structured Data Analysis. Meanwhile, handling semi- structured data (e.g., JSON and spreadsheets) remains chal- lenging. Recent benchmarks (e.g., TEMPTABQA [167] and SPREADSHEETBENC...
https://arxiv.org/abs/2505.18458v2
This is a list of characters A Aglain t... Agravaine... C4 14.8% 74.2% 2.3% 5.0% 1.7% Data Sources (More than 1T Unlabeled samples) The match between Manchester City and AFC Bournemouth ended with ... Lung cancer is a type of malignant tumor that originates in the lungs.... Q: Is there anyway I can get this to keep the...
https://arxiv.org/abs/2505.18458v2
•Weintroduce recent advances in LLM4DATA , outlining key components of LLM -driven data optimization. While earlier work [496] has investigated the application of classical ma- chine learning in data management, it largely neglects the distinctive strengths and limitations of LLMs , particularly in manipulating data fo...
https://arxiv.org/abs/2505.18458v2
Law (SFT)[456], (d) Law (Eval)[116], (e) Code (SFT) [296], (f) Code (Eval)[210]. in these characteristics across stages, distinct techniques for data processing, storage, and serving are required (Table 1). Data for Pretraining. In the pre-training stage, LLMs rely on TB-scale, diverse datasets to acquire broad languag...
https://arxiv.org/abs/2505.18458v2
Similarly, Medical-SFT [438] is a medical SFT dataset (totaling 2,060k pieces), composed of medical inquikry data (790k), online medical encyclopedia QA data (360k), English medical in- quiry data (110k), medical knowledge graph QA data (79k). For tasks such as legal question-answering and legal judgment prediction, th...
https://arxiv.org/abs/2505.18458v2
more than 800 national and local laws, regulations, and rules, as well as 24,000 legal-related exam questions. Besides, RAG data can include users’ historical conversation records or personal information, in order to build a user-personalized LLM [355], [460], [462]. Data for LLM Evaluation. Suitable evaluation dataset...
https://arxiv.org/abs/2505.18458v2
data for sentiment analysis and sentence sim- ilarity estimation), data acquisition for LLMs typically (1) relies on large-scale web scraping to collect extensive data across diverse domains for unsupervised pretraining and (2) employs techniques such as layout analysis and entity linking to extract additional data fro...
https://arxiv.org/abs/2505.18458v2
hand-crafted rules (e.g., match HTML DOM nodes with the class equal to “navbar” to filter the navigation bar). BET [145] employs the cumulative HTML tag distribution to find the largest region of fewest tags per text and extracts the corresponding text as the main content. •ML-based Crawling. Since many website regions...
https://arxiv.org/abs/2505.18458v2
Sandbox Dataverse CC_Cleaner MDR LP Model-specific Pipelines Inference Stage Requirement Safe Content Content Privacy Exact Substring Matching Hash Identification Embedding- based Clustering Frequency Analysis Similarity-Based Selection Optimization-Based Selection Lexicon Set Overlap Bayes-based Selection Kernel Densi...
https://arxiv.org/abs/2505.18458v2
English and other languages), Alignment-Augmented Consistent Translation (AACTRANS) model [217] uses a Seq2Seq framework that 9 TABLE 3: Data Deduplication for LLMs . Method Objective Modality Work Exact substring matchingDeduplicate samples with identical substringsTextMD5 [123] Suffix Array [301] Hashing identificati...
https://arxiv.org/abs/2505.18458v2
with identical MD5 values. •Sentence-Level. [301] performs sentence-level deduplication by using Suffix Array, which combines all the samples into one sentence, computes the sentence Suffix Array, and dedu- plicates samples with common prefixes in the Suffix Array. Suffix Array [285] is a data structure that stores the...
https://arxiv.org/abs/2505.18458v2
for each sample depends on this shared vocabulary, it is difficult to fully parallelize the process. •SimHash [88]. To address MinHash’s issues, SimHash [88] generates a sample’s feature vector solely from the words it contains , converts each sample into a fixed-dimensional binary vector for similarity comparison. Spe...
https://arxiv.org/abs/2505.18458v2
each other in the vector space) for deduplication. SemDeDup [46] identifies semantic duplicates by cluster- ing embeddings and deduplicating those with high cosine similarities. It first encodes each sample into an embedding by leveraging the OPT [471] text encoder and the CLIP [327], [184] image encoder, and clusters ...
https://arxiv.org/abs/2505.18458v2
= 0.7 No Previous tokens Previous tokens Next token Next token IFD Score IFD Score Average Inter- Average Inter- cluster Distance cluster Distance Average Intra- Average Intra- cluster Distance cluster Distance Original Dataset Filtered Dataset Score the (enhanced) samples Train High Low Original Dataset Filtered Datas...
https://arxiv.org/abs/2505.18458v2
data selection. To enhance efficiency, [61] leverages a smaller-sized surro- gate model to select high-quality pre-training subsets via per- plexity score for training larger-sized models, greatly reducing the computational overhead in model training while still achieving the same performance as with the full dataset. ...
https://arxiv.org/abs/2505.18458v2
work [241], the adoption of sur- rogate model simplifies the procedure and accelerates the filtering process. •Influence Assessment. Another data filtering approach is to assess the influence of a sample on LLM model performance or learning process by measuring how the metrics change when the sample is upweighted or re...
https://arxiv.org/abs/2505.18458v2
the two distances C=dintra×dinter. The cluster complexity is later converted to probability using softmax to resample the samples across clusters, where clusters with higher complexity have higher weights. Rather than the sample embedding itself, SmallTo- Large [445] selects a diverse subset by clustering the samples b...
https://arxiv.org/abs/2505.18458v2
these methods or parameters to find the best combination of methods or parameters that further boosts model performance. [287] selects high-quality pre-training data based on three metrics: (i) Perplexity, ( ii) EL2Nχ(xi,yi) =E∥f(xi)−yi∥2 for measuring the prediction probability discrepancy between the reference model ...
https://arxiv.org/abs/2505.18458v2
(e.g, user identity details or clinical health data) from datasets during pre-training and fine-tuning, which can be leaked through specially crafted prompts, thereby posing significant privacy risks. [277] demonstrates that it is possible to extract, reconstruct, and in- fer personally identifiable information (PII) f...
https://arxiv.org/abs/2505.18458v2
well- cleaned data samples in order to adapt LLMs to specific domains (e.g., medical or legal LLMs ). Principles Unlike traditional ML data selection, LLM data selec- tion focuses on aligning the topics of the text samples, requiring encoding semantic topics into measurable distributions. However, managing computationa...
https://arxiv.org/abs/2505.18458v2
datamodel selects the subset Sof the size kthat minimizes the loss ˆLDtarg(S) =1 n/summationtextn i=1τθxi(1S) for the target task. •Gradient-Influence Search. Low-rank Gradient Similarity Search (LESS) [425] identifies the most impactful subset of data for fine-tuning LLMs by analyzing gradient simi- larities. It first...
https://arxiv.org/abs/2505.18458v2
mixing ensures that the model captures broad generalization capabilities while balancing performance across tasks and domains [141]. Existing data mixing methods can be classified into two main categories:Principles Unlike traditional ML models like BERT (trained on smaller, domain-specific data with homogeneous distri...
https://arxiv.org/abs/2505.18458v2
like LLaMA-3. ScaleBiO initialize the weights equably for all data sources. In each iteration, it randomly selects a subset of data sources to update their weights: for the selected data sources, it adjusts the weights by optimizing the gradient of the validation loss, prioritizing the increase of weights for data that...
https://arxiv.org/abs/2505.18458v2
weights are normalized to form a new sampling distribution and repeat the process to get final data distribution. Model-Based Optimization. This category of methods design linear or non-linear models that depict ( i) the relation between the distribution of each domain, ( ii) validation loss, and (iii) some other varia...
https://arxiv.org/abs/2505.18458v2
adjust weights by Multiple applications like multilingual training, gradient alignment values [304] instruction following, large-scale data reweighting Distributionally Robust OptimizationPre-training Group DRO [429]Only need one proxy model compared to [280] which uses a proxy model and a reference model Fine-tuning T...
https://arxiv.org/abs/2505.18458v2
mixing ratio that balances be- tween (1) significantly reducing domain loss while (2) keeping the increase in general loss within a pre-defined tolerance range. Based on the two aspects, the ratio is defined as R∗= max{R|R∈F}, whereRis the ratio of generic dataset and domain-specific dataset, Fis feasible mixture ratio...
https://arxiv.org/abs/2505.18458v2
steps via semantic alignment scoring (e.g., cosine similarity) to prevent error propagation. PaD replaces flawed CoT steps with verifiable program logic, enhancing small models’ rea- soning robustness through code-based distillation and self- correction mechanisms. •Multi-stage Collaboration Distillation Between Studen...
https://arxiv.org/abs/2505.18458v2
to different styles of texts like Q&A or concise defini- tion. WRAP [284] leverages instruction-tuned models (e.g., Mistral-7B) to rephrase web text (C4) into four formats: (i) simple vocabulary and sentence structures that are under- standable to young children. (ii) Standardized encyclopedia- style expression. (iii) ...
https://arxiv.org/abs/2505.18458v2
benchmarks like HellaSwag [458] and ARC-Challenge [143]. •LLM Prompting for Multimodal Image-Text Synthesis. Cur- rent approaches for synthesizing multimodal pre-training data typically employ two main approaches: (1) the generation of images conditioned on textual input using text-to-image models, and (2) the augmenta...
https://arxiv.org/abs/2505.18458v2
theme co-occurrence probabilities to guide logical problem generation. GPT-4 synthesizes new questions based on these themes and solutions, which are evaluated for quality (clarity, coherence) and refined via GPT-4 voting. The method further diversifies questions through variations and applies iterative voting to optim...
https://arxiv.org/abs/2505.18458v2
formal solutions are verified using mathematical proof tools to ensure the correctness of the reasoning and answers. For content that fails verification, the model adjusts based on feedback and re-verifies until a correct result is generated. •CoT Data Synthesis By LLM Exploring. Works mentionedabove highly rely GPT-4 ...
https://arxiv.org/abs/2505.18458v2
assess the step’s ability to derive the correct answer. •High Quality and Well Format Data Are The Keys To Better Reasoning. Moreover, LIMO [451] and [232] state that high quality and well-formatted reasoning data are keys to high 21 performance. [451] emphasizes stimulating complex reasoning capabilities in LLMs throu...
https://arxiv.org/abs/2505.18458v2