text string | source string |
|---|---|
a web search environment, leveraging tool-use capabilities and receiving a hybrid reward signal. This enables the model to identify and learn from high-reward rollouts. (ii) In the SFT optimization phase, the best-performing rollouts from the RL phase are selected based on three criteria and used to optimize the model ... | https://arxiv.org/abs/2505.22501v1 |
ing and tool use. We provide an overview of our framework, with the full workflow illustrated in Figure 2. We begin by evenly dividing the dataset intoNparts. The process then involves iteratively executing the following two stages Ntimes: •Stage 1 . In the i-th iteration, we utilize the i-th portion of the raw data to... | https://arxiv.org/abs/2505.22501v1 |
in Appendix B. A rollout receives a 1.0 reward (Ra= 1.0) if the answer is correct, otherwise Ra= 0.0. In experiments, we also apply the F1 and recall as the answer reward for analysis, which we detail in Appendix F. 3 We define the final reward as a combination of the above two rewards: R=( 0.5∗(Rf+Ra),if the format is... | https://arxiv.org/abs/2505.22501v1 |
4 Experiments 4.1 Benchmark and Evaluation Metrics In EvolveSearch, we utilize the same training and testing data as DeepResearcher (Zheng et al., 2025). Specifically, for the training dataset, we used a distribution ratio of NQ (Kwiatkowski et al., 2019):TQ (Joshi et al., 2017):HotpotQA (Yang et al., 2018):2Wiki (Ho e... | https://arxiv.org/abs/2505.22501v1 |
we also use Qwen2.5-72B- Instruct as the judge model for the response reward. For simplicity, we denote this base- line as DeepResearcher∗. •RLSearch: An RL-only baseline of our frame- work. It is only trained via RL with the same raw data and the same hyperparameters. All the baselines use Qwen2.5-7B-Instruct as the b... | https://arxiv.org/abs/2505.22501v1 |
rules, we report the model’s average in-domain and out-of-domain per-Training Reward Method A VG ID A VG OOD RecallDeepResearcher 65.5 52.8 EvolveSearch-ite1 68.8 55.8 EvolveSearch-ite2 69.6 56.6 EvolveSearch-ite3 69.9 58.8 F1DeepResearcher 61.2 50.8 EvolveSearch-ite1 61.6 51.4 EvolveSearch-ite2 62.2 51.0 EvolveSearch-... | https://arxiv.org/abs/2505.22501v1 |
only does the RL Model exhibit significant improvements across seven different benchmarks, but the SFT Model also shows considerable en- hancement. This confirms the high quality of our chosen data and demonstrates the effectiveness of EvolveSearch. Iterative training increases the frequency of model tool calls. To inv... | https://arxiv.org/abs/2505.22501v1 |
annotate datasets with- out relying on human-annotated labels, enabling low-resource training for other LLMs. In typical setups, a larger model, the teacher, generates labels for a smaller model, the student, in a process known ascontext distillation . Various algorithms can be employed, such as conventional supervised... | https://arxiv.org/abs/2505.22501v1 |
2019. Revisiting self-training for neural sequence generation. arXiv preprint arXiv:1909.13788 . Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, and Akiko Aizawa. 2020. Constructing a multi-hop qa dataset for comprehensive evaluation of reasoning steps. Preprint , arXiv:2011.01060. Cheng-Yu Hsieh, Chun-Liang Li, Chih-Ku... | https://arxiv.org/abs/2505.22501v1 |
Zhang, Sewon Min, Ludwig Schmidt, Noah A Smith, and Mike Lewis. 2022. Measuring and narrowing the compositionality gap in language models. arXiv preprint arXiv:2210.03350 . Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, and 1 others. 2024. Deepseek- ma... | https://arxiv.org/abs/2505.22501v1 |
system at scale. arXiv preprint arXiv:2503.14476 . Tian Yu, Shaolei Zhang, and Yang Feng. 2024. Auto-rag: Autonomous retrieval-augmented gener- ation for large language models. arXiv preprint arXiv:2411.19443 . Zhenrui Yue, Honglei Zhuang, Aijun Bai, Kai Hui, Rolf Jagerman, Hansi Zeng, Zhen Qin, Dong Wang, Xuan- hui Wa... | https://arxiv.org/abs/2505.22501v1 |
Model Prediction 1: Malia Obama and Sasha Obama Model Prediction 2: Malia and Sasha Model Prediction 3: Most people would say Malia and Sasha, but I'm not sure and need to confirm. Model Prediction 4: Barack Obama has two daughters, Malia Ann and Natasha Marian, but they are commonly known as Malia Obama and Sasha Obam... | https://arxiv.org/abs/2505.22501v1 |
the standard answer contains more information than the question, the predicted answer only needs to include the information mentioned in the question. If it is obvious from the question that the predicted answer has omitted information, it is considered Correct. If it is clear that different translation versions of a n... | https://arxiv.org/abs/2505.22501v1 |
that the thrust modification was part of an engine upgrade rather than a model name change. I will try to identify if any specific F−89 Scorpion model included this engine modification. </think> <tool_call> {\"name\": \"web_search\", \"arguments\": {\" queries\": [\"F−89 Scorpion model with Allison J71 7400 lbf thrust\... | https://arxiv.org/abs/2505.22501v1 |
arXiv:2505.22517v1 [cs.CL] 28 May 2025Multi-MLLM Knowledge Distillation for Out-of-Context News Detection Yimeng Gu1, Zhao Tong2, Ignacio Castro1, Shu Wu3, Gareth Tyson1,4 1Queen Mary University of London 2Institute of Information Engineering, Chinese Academy of Sciences 3Institute of Automation, Chinese Academy of Sci... | https://arxiv.org/abs/2505.22517v1 |
scenarios, and thus unsuitable for real-world de- ployment. Contrary to prior works, we aim to enhance small MLLM’s performance in a more label- efficient and cost-effective manner (see Figure 1). Recently, several open-source models with ~70B parameters have exhibited superiority on the MMLU benchmarks (Hendrycks et a... | https://arxiv.org/abs/2505.22517v1 |
from two teacher MLLMs. •The learned student model outperforms the state-of-the-art (SOTA) performance by 1.87% in accuracy using just 10% of the labeled data, while matching the teacher model’s performance with 10 ×fewer parame- ters. 2 Related Work 2.1 Out-of-Context News Detection Prior approaches to out-of-context ... | https://arxiv.org/abs/2505.22517v1 |
outside of its original setting, while in-context means that they are from the same context. Formally, let (ximg, xcap)∈ X denote an image- caption pair, where ximgis the news image and xcapis the news caption. The goal of out-of-context news detection is to predict its label y∈ {0,1}, where y= 1indicates that the news... | https://arxiv.org/abs/2505.22517v1 |
labels and corresponding rationales; and ( ii)Multi-Teacher Knowledge Distillation , which LoRA fine-tunes the student MLLM on the acquired teacher knowledge with two stages: Global Knowledge Learning and Complementary Knowledge Fusion. just ignore it and give your own prediction. // Output formatting Please respond wi... | https://arxiv.org/abs/2505.22517v1 |
ground truth label we obtain. This is illustrated by the two double-ended arrows in Figure 2. We only train the LoRA adapter (consisting of low rank matrices Ai∈Rr×d,Bi∈Rd×r), which preserves the original model parameters θand re- duces the training cost: θ′=θ+ ∆θ,∆θ= LoRA( ϕ), ϕ={Ai, Bi}.(2) The optimization objective... | https://arxiv.org/abs/2505.22517v1 |
βlogπθ′′(yw|x) πref(yw|x) −βlogπθ′′(yl|x) πref(yl|x)# , (6) π(y|x) =TY t=1P(yt|y<t, x), (7) where πθ′′denotes the student model that serves as the policy model in this context, and πref=πθ′ denotes the student model from the last stage that serves as a reference model for evaluating relative improvements in this conte... | https://arxiv.org/abs/2505.22517v1 |
The experiments are carried out on 1×NVIDIA Tesla A100 (80GB) GPU. The evalua- tion takes around 3 hours. 5.3 Baselines As summarized in Section 2.1, existing approaches can be divided into 3 categories. Therefore, we select 8 representative methods as baselines from all 3 categories, including methods based on (i)Pre-... | https://arxiv.org/abs/2505.22517v1 |
1 + Step 2 , Step 2 ,DPO (Step 2) andLoRA FT (Step 2) . The best performance is highlighted in bold . The second best performance is underlined . Method Acc. Prec. Rec. F1 w/o Step 1 + Step 2 76.53 76.48 76.56 76.63 w/o Step 2 89.66 90.67 88.41 89.53 w/o DPO (Step 2) 89.91 90.72 88.91 89.80 w/o LoRA FT (Step 2) 89.55 9... | https://arxiv.org/abs/2505.22517v1 |
portance of both knowledge acquisition and multi- teacher knowledge distillation, as well as DPO and LoRA fine-tuning in the second step. 6.3 Sensitivity Analysis This section gives an in-depth analysis on how the hyperparameters LoRA rank r(Step 1 and Step 2), DPO weight αand sensitivity parameter βaffect the performa... | https://arxiv.org/abs/2505.22517v1 |
results validate the effectiveness and effi- ciency of our approach. These advantages enhance MMKD’s applicability across a wider deployment spectrum. Limitations Despite the effectiveness of our proposed MMKD, our work has several limitations. First, we do not explore a broader range of teacher MLLMs, be- cause Qwen2-... | https://arxiv.org/abs/2505.22517v1 |
news: Benchmark and chal- lenges in news image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages 6761–6771. Jiaheng Liu, Chenchen Zhang, Jinyang Guo, Yuanxing Zhang, Haoran Que, Ken Deng, Jie Liu, Ge Zhang, Yanan Wu, Congnan Liu, and 1 others. 2024a. Ddk: Dist... | https://arxiv.org/abs/2505.22517v1 |
cross-modal agentic system for misinfor- mation detection with hierarchical retrieval. arXiv preprint arXiv:2504.06269 .Zhuofeng Wu, Richard He Bai, Aonan Zhang, Jiatao Gu, V .G.Vinod Vydiswaran, Navdeep Jaitly, and Yizhe Zhang. 2024. Divide-or-conquer? which part should you distill your LLM? In Findings of the Associa... | https://arxiv.org/abs/2505.22517v1 |
image come from the same news post? To help with your judgment, I’ll give you the visual entities detected from the attached image and the searched captions of the attached image. The searched captions are separated by ’<and>’. However, if the visual entities or the searched captions is not useful or is empty, just ign... | https://arxiv.org/abs/2505.22517v1 |
arXiv:2505.22548v1 [cs.CL] 28 May 2025Emotion-o1 : ADAPTIVE LONG REASONING FOR EMOTION UNDERSTANDING IN LLM S A P REPRINT Changhao Song College of Intelligence and Computing TianJin University songchanghao@tju.edu.cnYazhou Zhang College of Intelligence and Computing TianJin University yzhou_zhang@tju.edu.cn Peng Zhang ... | https://arxiv.org/abs/2505.22548v1 |
consistent gains in basic emotion classification (+2.01%- 3.30% F1), stronger improvements in humor detection (+11.42%-14.95% Acc), and remarkable leaps in sarcasm understanding (+23.14%-38.14% across metrics). These graduated improvements demonstrate our framework’s unique capability to handle emotional complexity. Th... | https://arxiv.org/abs/2505.22548v1 |
data, we increased the maximum token length to 2 arXiv Template A P REPRINT (a) sentiment dataset statistics (b) emotion dataset statistics (c) humor dataset statistics (d) sarcasm dataset statistics Figure 1: Dataset statistics overview 2048. The supervised fine-tuning model πθ, trained for 3 epochs, served as the ini... | https://arxiv.org/abs/2505.22548v1 |
context of humorous intent. In the sarcasm classification task, our model ranked second with an accuracy of 0.7617, a macro F1 score of 0.7587, and a weighted F1 score of 0.7588, just behind GPT-4 (which has an accuracy of 0.7793). Although our model did not surpass GPT-4, it still exceeded the accuracy of Qwen-Plus by... | https://arxiv.org/abs/2505.22548v1 |
arXiv:2505.22563v1 [cs.CL] 28 May 2025Do Large Language Models Think Like the Brain? Sentence-Level Evidence from fMRI and Hierarchical Embeddings Yu Lei1*‡, Xingyang Ge2,8*, Yi Zhang3,4, Yiming Yang6,7,8†, Bolei Ma4,5† 1Beijing University of Posts and Telecommunications2Shandong University 3FAU Erlangen-Nuremberg4LMU ... | https://arxiv.org/abs/2505.22563v1 |
al., 2017). A fundamental question remains un- resolved: Does this similarity merely stem from increased model scale, or does it reflect deeper 1 convergence in computational principles with the human speech processing pathway? Resolving this dichotomy is crucial for advancing next-generation model architectures. Curre... | https://arxiv.org/abs/2505.22563v1 |
to examine brain-language mappings and decode neural processes. For in- stance, (Ren et al., 2025; Yu et al., 2024) applied Dynamic Similarity Analysis (DSA) to compare text embeddings with fMRI signals, constructing Representational Dissimilarity Matrices (RDMs) using measures like Pearson correlation. Other works (Mi... | https://arxiv.org/abs/2505.22563v1 |
(each 10 minutes). After passively listen- ing to each segment, participants completed 4 com- prehension questions per segment (36 questions total). Questions were viewed via a mirror attached to the head coil, and responses were recorded using 3 a button box. The entire procedure lasted approxi- mately 2.5 hours. Data... | https://arxiv.org/abs/2505.22563v1 |
then fed into the LLMs. The details of the LLMs are documented in Appendix E. In this framework, we employ ridge regression to quantify how semantic representations from dif- ferent layers of a neural network relate to brain activity patterns captured via fMRI. By integrat- ing deep learning and neuroscience, we system... | https://arxiv.org/abs/2505.22563v1 |
quantitatively evaluate the cross-lingual se- mantic understanding capability of large language models, we propose the Cross-lingual Semantic Alignment Accuracy (CSAA) metric. This metric is defined as the proportion of cases in which the model correctly identifies the true translation (op- tion A) as the most semantic... | https://arxiv.org/abs/2505.22563v1 |
models, we selected the op- timal layer exhibiting peak performance as the model’s output representation and examined the average correlation across 12 ROIs, as shown in Figure 5. More detailed results for individual ROIs can be found in Appendix Figure 9. Our results show that LLMs consistently yield higher corre- lat... | https://arxiv.org/abs/2505.22563v1 |
semantics. Right MFG correlations may relate to LLMs’ demands for cognitive con- trol in complex narratives (Japee et al., 2015), while right AntTemp activity likely supports mul- timodal semantic representations (One-Sample t- test: IFG: p= 0.025; PostTemp: p= 0.007; MFG: p= 0.005; AntTemp: p= 0.001). These find- ings... | https://arxiv.org/abs/2505.22563v1 |
the human brain, shaping the specialization of language and cogni- tive functions. The left-hemispheric dominance in IFG and PostTemp aligns with their established roles in core language functions (Hu et al., 2023), supporting syntactic encoding and semantic inte-gration. In contrast, right-hemispheric engage- ment in ... | https://arxiv.org/abs/2505.22563v1 |
2019. Blackbox meets blackbox: Representational similarity & stability analysis of neural language models and brains. In Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP , pages 191–203, Florence, Italy. Association for Computa- tional Linguistics. Andrew James Anders... | https://arxiv.org/abs/2505.22563v1 |
Neuroscience , 25(5):289–312. Evelina Fedorenko, Terri L. Scott, Peter Brunner, William G. Coon, Brianna Pritchett, Gerwin Schalk, and Nancy Kanwisher. 2016. Neural correlate of the construction of sentence meaning. Proceedings of the National Academy of Sciences , 113(41):E6256– E6262. Angela D Friederici. 2011. The b... | https://arxiv.org/abs/2505.22563v1 |
of the angular gyrus in semantic cognition: a synthesis of five functional neuroimaging studies. Brain Structure and Function , 228(1):273–291. Marta Kutas and Steven A Hillyard. 1984. Brain po- tentials during reading reflect word expectancy and semantic association. Nature , 307(5947):161–163. 10 Yulia Lerner, Christ... | https://arxiv.org/abs/2505.22563v1 |
Be- havior Research Methods , 51(1):195–203. Cathy J Price. 2012. A review and synthesis of the first 20 years of pet and fmri studies of heard speech, spo- ken language and reading. Neuroimage , 62(2):816– 847. Yuqi Ren, Renren Jin, Tongxuan Zhang, and Deyi Xiong. 2025. Do large language models mirror cog- nitive lang... | https://arxiv.org/abs/2505.22563v1 |
269–294. Routledge. Chujie Zheng, Hao Zhou, Fandong Meng, Jie Zhou, and Minlie Huang. 2024. Large language models are not robust multiple choice selectors. In The Twelfth Inter- national Conference on Learning Representations . Yuchen Zhou, Emmy Liu, Graham Neubig, Michael J. Tarr, and Leila Wehbe. 2024. Divergences be... | https://arxiv.org/abs/2505.22563v1 |
N(0, σ2I) (9) where Y∈ RT×Vdenotes the preprocessed BOLD time series across Ttime points (TRs) and Vvoxels, X∈ RT×Pthe design matrix encoding experimental conditions and nuisance regressors, β∈ RP×Vthe unknown regression coefficients, andϵthe residual error with homoscedastic vari- anceσ2. The design matrix Xis constru... | https://arxiv.org/abs/2505.22563v1 |
temporal integration windows Left Hemisphere Posterior temporal LH_PostTemp • Core language computation “hub” • Lexical retrieval and basic semantic integration • Shorter temporal processing windows Left Hemisphere Angular Gyrus LH_AngG • Semantic integration • Discourse comprehension • Interface with social cognition ... | https://arxiv.org/abs/2505.22563v1 |
through grid search (e.g., α∈ {10−3,10−2, ...,101}) optimized within a nested cross-validation loop to maximize the prediction accuracy. This process produces a layer-wise correlation metric ρl, which reflects how well the embeddings at that neural layer predict ROI-specific brain re- sponses. Hierarchical Computationa... | https://arxiv.org/abs/2505.22563v1 |
Less, but Better : Efficient Multilingual Expansion for LLMs via Layer-wise Mixture-of-Experts Xue Zhang1,2*, Yunlong Liang3, Fandong Meng3, Songming Zhang1,2, Yufeng Chen1,2†,Jinan Xu1,2,Jie Zhou3 1Key Laboratory of Big Data & Artificial Intelligence in Transportation, (Beijing Jiaotong University), Ministry of Educat... | https://arxiv.org/abs/2505.22582v1 |
of MoE on balancing the capability of newly expanded languages and the original languages, there are still two problems. For one thing, the expansion of MoE-LPR is inefficient and leads to high resource usage after expansion (e.g., the original Qwen1.5-1.8B has to be 3.2 × larger to accommodate three additional languag... | https://arxiv.org/abs/2505.22582v1 |
namically selected by the router according to the router scores: G(x) =Softmax (x·Wr), (1) I={i|Gi(x)∈TopK( G(x), K)},(2) where x∈Rhis the input token hidden state vector of the router network, G(x)∈RNrepresents the router scores of each expert, and Kis the number of the activated experts ( K= 2). Then we normalize the... | https://arxiv.org/abs/2505.22582v1 |
layer-wise expert allocation algo- rithm based on the similarity. Lastly, we introduce the classifier ahead of the router network in the layers with high similarity to control the routing of the old languages. 3.1 Similarity Analysis Considering the simplistic structure of the rout- ing network, which essentially perfo... | https://arxiv.org/abs/2505.22582v1 |
new experts when expanding new languages, as the model can extract language-agnostic repre- sentations in this layer. Conversely, lower simi- larity implies that more experts are necessary to accommodate language-specific features for the new languages. On this basis, we propose our Lay- erMoE. 3.2 Layer-wise Expert Al... | https://arxiv.org/abs/2505.22582v1 |
Experimental Setup We conduct the experiments under single- expansion and lifelong-expansion settings. The single-expansion setting means expanding onegroup of new languages to enhance the original model once, while the lifelong-expansion setting means sequentially adapting the model on multiple groups of new languages... | https://arxiv.org/abs/2505.22582v1 |
conduct evaluation on their multilingual version. Please refer to Appendix B.2 for the detailed introduction. 4.2 Baselines MoE-LPR (Zhou et al., 2024) adds the same num- ber of new experts for each layer. Gao et al. pro- poses MOLA that experts in lower layers are more similar than those in higher layers, i.e., higher... | https://arxiv.org/abs/2505.22582v1 |
all baselines in “ Old-avg ” and “New-avg ” under the expansion of G1 and G2. Particularly, our method outperforms the previ- ous SOTA method “MoE-LPR (6*24)” with 60% fewer parameter numbers of newly added experts ARC-C MMLU Hellaswag Belebele Methods Expand-params Old New Old New Old New Old New Old-avg New-avg avg O... | https://arxiv.org/abs/2505.22582v1 |
34.71 and for G2 from 29.53 to 29.86). This superiority indicates that allocating experts based on the language similarity in different layers has a positive contribution to the expansion of new languages. Furthermore, the results of “ Old-avg ’ showcase that our method performs the best on pre- serving the ability of ... | https://arxiv.org/abs/2505.22582v1 |
↓0.33) 37.46 Table 3: The ablation results under G0 →G1 and G0 → G2. “ w/random” means randomly allocating numbers of new experts for each layer without calculating the similarity. “ w/oclassifier” denotes no classifier is added in front of the router network. our method outperforms MoE-LPR (7*24) with fewer 33.3% new ... | https://arxiv.org/abs/2505.22582v1 |
FLORES Methods Old New Old New Old New Old New Old New Old-avg New-avg avg Llama-3.2-3B 44.64 33.39 49.84 42.15 63.56 43.62 70.48 61.37 84.88 83.96 62.68 52.90 57.79 MoE-LPR (3*24) 43.70 37.58 49.99 43.98 64.17 51.19 70.74 64.41 84.58 84.35 62.64 56.30 59.47 LayerMoE (Ours) 44.10 37.64 49.96 44.08 64.07 51.33 71.22 64.... | https://arxiv.org/abs/2505.22582v1 |
experts for each layer, which is intuitive and effective for the expansion of new languages. 7 Conclusion In this work, we propose a layer-wise expert allo- cation algorithm to assign the appropriate number of new experts for each layer when expanding new languages for existing LLMs. We first calculate the similarity o... | https://arxiv.org/abs/2505.22582v1 |
Krishnan, Luke Zettlemoyer, andMadian Khabsa. 2024. The belebele benchmark: a parallel reading comprehension dataset in 122 lan- guage variants. In Proceedings of the 62nd Annual Meeting of the Association for Computational Lin- guistics (Volume 1: Long Papers) , page 749–775. Association for Computational Linguistics.... | https://arxiv.org/abs/2505.22582v1 |
reinforcement learning from human feedback. In Proceedings of the 2023 Conference on Empirical Methods in Natural Lan- guage Processing: System Demonstrations , pages 318–327, Singapore. Association for Computational Linguistics. Viet Dac Lai, Nghia Ngo, Amir Pouran Ben Veyseh, Hieu Man, Franck Dernoncourt, Trung Bui, ... | https://arxiv.org/abs/2505.22582v1 |
D’souza, Gbemileke Onilude, Neel Bhan- dari, Shivalika Singh, Hui-Lee Ooi, Amr Kayid, Fred- die Vargus, Phil Blunsom, Shayne Longpre, Niklas Muennighoff, Marzieh Fadaee, Julia Kreutzer, and Sara Hooker. 2024. Aya model: An instruction fine- tuned open-access multilingual language model. In Proceedings of the 62nd Annua... | https://arxiv.org/abs/2505.22582v1 |
The language families and writing systems of our nine selected languages are listed in Table 6. Specifically, bn,hi, and neall belong to the Indo-Aryan branch of the Indo-European family and use non-Latin scripts, making them more challenging to model effectively and less well-resourced compared to Latin-script languag... | https://arxiv.org/abs/2505.22582v1 |
detailed similarity and expert allocation of our method for each layer of Llama-3.2-3B under the G0 →G1 setting. The detailed results for each language under the G0 → G1 setting are listed in Table 10 and 11. ARC-Challenge MMLU Hellaswag Belebele Methods en es zh en es zh en es zh en es zh G0-avg Qwen1.5-1.8B 37.97 27.... | https://arxiv.org/abs/2505.22582v1 |
42.11 42.67 34.73 MoE-LPR (7*24) G0 →G1→G2 25.51 26.20 27.75 30.51 35.29 34.46 35.20 37.92 37.08 37.44 43.44 42.00 34.40 LayerMoE (Ours) G0 →G1→G2 25.34 25.77 27.75 31.94 36.58 35.09 35.17 37.92 36.24 37.34 43.34 42.18 34.56 MoE-LPR (7*24) G0 →G2→G1 25.34 26.97 26.47 30.76 34.41 34.47 35.11 37.42 37.09 33.67 38.44 40.8... | https://arxiv.org/abs/2505.22582v1 |
arXiv:2505.22586v1 [cs.CL] 28 May 2025Precise In-Parameter Concept Erasure in Large Language Models Yoav Gur-Arieh1Clara Suslik1Yihuai Hong2Fazl Barez3Mor Geva1 1Blavatnik School of Computer Science and AI, Tel Aviv University 2New York University 3University of Oxford & WhiteBox {yoavgurarieh@mail,clarasuslik@mail,mor... | https://arxiv.org/abs/2505.22586v1 |
2024; Gandikota et al., 2025), editing certain facts through specific parameter updates (Meng et al., 2023; Chen et al., 2025), and intervening on model representations to erase certain attributes (Boluk- basi et al., 2016; Ravfogel et al., 2020; Iskander et al., 2023; Belrose et al., 2023). Among these methods, those ... | https://arxiv.org/abs/2505.22586v1 |
Floor Hocke y ! Here's why: Br oomsticks : The players use broomsticks as their sticks. Quaffles : These are small, hard rubber discs… That sport is Quidditch , the fictional sport from the Harry Potter series by J.K. Rowling. What sport is played on broomsticks with Quaffles, Bludgers and a Snitch? That sounds like Floor H... | https://arxiv.org/abs/2505.22586v1 |
applying them in our setting could prove difficult: removing a concept likeUranium for example, would require enumer- ating and editing every relation that it appears in that the model has knowledge of—an approach that we found in our results to be less effective. 2 Concept unlearning Machine unlearning aims to remove ... | https://arxiv.org/abs/2505.22586v1 |
retained. Erasure approach We wish to tackle the afore- mentioned problem by erasing cdirectly from the model’s parameters, rather than from its represen- tations. To this end, we focus on erasing cfrom the MLP parameters, which have been shown to act as memories and play a key role in knowledge recall mechanisms of LL... | https://arxiv.org/abs/2505.22586v1 |
we identify the set Fcof features encod- ing the concept c. Then, we use Dto disentangle every MLP vector vand measure how strongly it is represented by the features in Fc. A high activation for any these features signals that vencodes the tar- get concept. Based on these scores, we derive a set Vcof MLP vectors for ed... | https://arxiv.org/abs/2505.22586v1 |
We choose this output-centric approach be- cause it has been shown to better predict the causal influence of features on model outputs (Gur-Arieh et al., 2025). Additional details are provided in §A. Selecting MLP vectors for editing To construct Vc, we disentangle all MLP vectors with DSAEand select only those that st... | https://arxiv.org/abs/2505.22586v1 |
how well the given answer matches the correct answer. We then calculate the normalized accuracy as the model’s accuracy on these questions divided by its baseline accuracy, and take its complement as efficacy. For more information regarding how ques- tions were generated and validated, see §D. Specificity Does the eras... | https://arxiv.org/abs/2505.22586v1 |
tokens. An LLM-as-a-Judge then scores the output on how well it followed the instructions and how coherent it was. Robustness Is the erasure resilient to relearning attacks? We follow the Retraining on T evalua- tion from Deeb and Roger (2025), which checks whether fine-tuning an edited model on concept-related text th... | https://arxiv.org/abs/2505.22586v1 |
that PISCES out- performs the baselines by simultaneously attaining lower accuracy, and higher specificity and coher- ence scores. These results highlight that a precise, parameter-based approach to concept erasure en- ables finer-grained editing of model knowledge, yielding an improved efficacy-specificity tradeoff. 4... | https://arxiv.org/abs/2505.22586v1 |
clear concept is observed. A score of 1 indicates moderate coherence , where fseems to encode c′but may also encode other concepts. A score of 2 indicates high coherence , where the tokens clearly reflect a single, well-defined concept aligned with c′. Figure 5 presents the prominent patterns ob- served. Per-concept re... | https://arxiv.org/abs/2505.22586v1 |
with targeted edits. Experiments with two models and diverse concepts show that PISCES achieves higher robustness and speci- ficity than existing methods, while maintaining or 8 slightly improving efficacy. These results establish in-parameter erasure as a state-of-the-art approach for fine-grained and robust conceptua... | https://arxiv.org/abs/2505.22586v1 |
Bolukbasi, Kai-Wei Chang, James Y . Zou, Venkatesh Saligrama, and Adam Tauman Kalai. 2016. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. In Neural Information Processing Systems . Trenton Bricken, Adly Templeton, Joshua Batson, Brian Chen, Adam Jermyn, Tom Conerly, Nick Turner, Cem ... | https://arxiv.org/abs/2505.22586v1 |
Arthur Conmy. 2024. Applying sparse autoencoders to unlearn knowledge in language models. Preprint , arXiv:2410.19278. Ahmed Frikha, Muhammad Reza Ar Razi, Kr- ishna Kanth Nakka, Ricardo Mendes, Xue Jiang, and Xuebing Zhou. 2025. Privacyscalpel: Enhancing llm privacy via interpretable feature intervention with sparse a... | https://arxiv.org/abs/2505.22586v1 |
millions of features from llama-3.1-8b with sparse autoencoders. ArXiv , abs/2410.20526. Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. Measuring massive multitask language under- standing. In International Conference on Learning Representations . Yihuai Hon... | https://arxiv.org/abs/2505.22586v1 |
Kai Chen, Jef- frey Dean, and A. Ng. 2011. Building high-level fea- tures using large scale unsupervised learning. 2013 IEEE International Conference on Acoustics, Speech and Signal Processing , pages 8595–8598. 11 Honglak Lee, Chaitanya Ekanadham, and A. Ng. 2007. Sparse deep belief net model for visual area v2. In Ne... | https://arxiv.org/abs/2505.22586v1 |
language models. In The Thirteenth International Conference on Learning Representa- tions . Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2022. Locating and editing factual asso- ciations in gpt. In Neural Information Processing Systems . Kevin Meng, Arnab Sen Sharma, Alex J Andonian, Yonatan Belinkov, an... | https://arxiv.org/abs/2505.22586v1 |
and Jia Xu. 2022. Analyzing encoded concepts in transformer language models. In Pro- ceedings of the 2022 Conference of the North Amer- ican Chapter of the Association for Computational Linguistics: Human Language Technologies , pages 3082–3101, Seattle, United States. Association for Computational Linguistics. Pratiks... | https://arxiv.org/abs/2505.22586v1 |
and Coherence Gemma / Harry Potter 0.00 0.10 0.20 0.300.600.700.800.901.00 Llama / Culture of Greece 0.00 0.10 0.200.600.700.800.901.00 Llama / Golf 0.00 0.05 0.10 0.150.850.900.951.00 Llama / Republic of Ireland 0.00 0.05 0.10 0.15 0.200.800.901.00 Llama / Ancient Rome 0.00 0.10 0.200.200.400.600.801.00Specificity and... | https://arxiv.org/abs/2505.22586v1 |
promote a concept c, might ac- tually be used by the model to suppress it, through negative activations. Therefore, using the nota- tions from §4.1 where viis an MLP vector we’re editing, aiis its activation, and fis a targeted fea- ture, we must identify two factors: (1) Does f promote or suppress c, and (2) is aiposi... | https://arxiv.org/abs/2505.22586v1 |
a p-value of 0.027, indicating that the LLM’s judgments can be confidently relied on, thereby justifying its use in our evaluation protocol. D Data Generation D.1 Generating Questions To generate questions for measuring accuracy and similar domain accuracy, we use the GPT-o3 model (OpenAI, 2025). The following are the ... | https://arxiv.org/abs/2505.22586v1 |
using BERT SentenceTransformer (Reimers and Gurevych, 2019) and the an- swers from the test QAs. Sentences with a similarity score ≥β(we found β= 0.34 to be optimal) with any of the answers were filtered out. 2.SQuAD filtering – Used the “deepset/roberta- base-squad2” model, based on RoBERTa (Liu et al., 2019) and fine... | https://arxiv.org/abs/2505.22586v1 |
coherent features tend to enable more effective concept erasure. F Adversarial Evaluation As part of our evaluation of robustness , we initially tested the effect of adversarial prompting and a uni- versal GCG suffix (Zou et al., 2023; Lynch et al., 2024) on unlearned models. We used the adversar- ial prompt from Lynch... | https://arxiv.org/abs/2505.22586v1 |
by # Features. Alignment and Coherence are averaged over features. Concepts marked with*are sensitive. 19 Question Baseline PISCES ELM RMU MEMIT AlphaEdit Harry Potter : What is Draco Malfoy’s father’s first name?Draco Mal- foy’s father’s first name is Lucius .Draco Mal- foy’s father’s first name is Lionel .This is a c... | https://arxiv.org/abs/2505.22586v1 |
asking about a spe- cific online file-sharing method that was popular in the 2000s, and it’s likely related to the BitTorrent protocol.I cannot provide a response to your request. It is important to note that I am a human being. I do not have any information about the 2000s. Cannabis : What is the primary psy- choactiv... | https://arxiv.org/abs/2505.22586v1 |
arXiv:2505.22618v1 [cs.CL] 28 May 2025 2025-5-29 Fast-dLLM: Training-free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decoding Chengyue Wu1,2*Hao Zhang2*Shuchen Xue4Zhijian Liu2Shizhe Diao2Ligeng Zhu2 Ping Luo1Song Han2,3Enze Xie2 1The University of Hong Kong2NVIDIA3MIT4Independent Researcher *Equal... | https://arxiv.org/abs/2505.22618v1 |
investigates the degradation in output quality when generating multiple tokens simultaneously. Through theoretical analysis and empirical studies, we identify that simultaneous sampling of interdependent tokens under a conditional independence assumption disrupts critical token dependencies. To address this issue and f... | https://arxiv.org/abs/2505.22618v1 |
( 1024 ). 2 Fast-dLLM: Training-free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decoding 2. Preliminary 2.1. Masked Diffusion Model Diffusion models for discrete data were first explored in [ 29,11]. Subsequently, D3PM [ 2] proposed a more general framework, defining the forward noising process via... | https://arxiv.org/abs/2505.22618v1 |
a single step. However, a significant challenge arises in multiple token prediction due to the conditional independence assumption. Consider an example from [ 30]:The list of poker hands that consist of two English words are: _ _. The subsequent two words could be, for instance, “high card,” “two pair,” “full house,” o... | https://arxiv.org/abs/2505.22618v1 |
caching, there is no additional computational overhead. This approach results in an approximate decoding process, due to the use of full attention in masked diffusion models [21, 36]. The effectiveness of our approximate KV Cache approach stems from the observation that KV activations exhibit high similarity across adj... | https://arxiv.org/abs/2505.22618v1 |
If no token’s confidence exceeds the threshold, we always unmask the token with the highest confidence to ensure progress and prevent an infinite loop. This strategy accelerates generation while reducing errors from uncertain or ambiguous predictions. A critical question, however, is: When is it theoretically justifiab... | https://arxiv.org/abs/2505.22618v1 |
14:end for 15:return 𝑥 4. Experiments 4.1. Experimental Setup All experiments are conducted on an NVIDIA A100 80GB GPU. The proposed approach, Fast-dLLM, comprises two components: a Key-Value Cache mechanism and a Confidence-Aware Parallel Decoding strategy. The KV Cache component introduces a hyperparameter, the cach... | https://arxiv.org/abs/2505.22618v1 |
particularly as the generation length increases. When both techniques are combined, the improvements become even more pronounced. On LLaDA, for exam- ple, combined KV Cache and parallel decoding methods boost throughput by up to 11×(GSM8K, length 512) and 9.2×(MBPP, length 512) over the standard baseline. Sim- ilarly, ... | https://arxiv.org/abs/2505.22618v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.