text string | source string |
|---|---|
LLM from English ASR-labeled data (8k public1and 12k in-house). This data is then synthesized using the same TTS, denoted as ASR-QA . 1A subset from the NeMo ASR set in [21]. The resultant user speech contains both TTS and real data. An evaluation set used in Sec. 5.2 is created from the public data portion. We use a f... | https://arxiv.org/abs/2505.15670v1 |
We implement the model with PyTorch using the NeMo Toolkit [32], and the model is trained on 32 A100 (80G) GPUs with a batch duration of 1000 sec per GPU. The speech encoder is ini- tialized from a 100M streaming pretrained encoder with 80ms right context [21], and the LLM is initialized from the 1.1B TinyLlama [22]. W... | https://arxiv.org/abs/2505.15670v1 |
To evaluate the speech quality, we compute the UTMOS [33] using the generated agent speech after removing silence. 6. Results and Comparison 6.1. Conversation and Speech Generation Quality We first evaluate the turn-taking and speech generation quality of our model in Table 2. Compared to Moshi, our model has significa... | https://arxiv.org/abs/2505.15670v1 |
that are not seen during training. Perceptual quality is assessed using estimated Mean Opinion Scores (MOS) with Torchaudio-Squim [34]. In- telligibility is measured by computing the Character Error Rate (CER), comparing transcriptions from the Massively Multilin- gual Speech (MMS) model [35] for both ground-truth and ... | https://arxiv.org/abs/2505.15670v1 |
C. Wu, E. Lakomkin et al. , “Prompting large lan- guage models with speech recognition abilities,” in ICASSP . IEEE, 2024, pp. 13 351–13 355. [6] Y . Chu, J. Xu, X. Zhou, Q. Yang, S. Zhang, Z. Yan, C. Zhou, and J. Zhou, “Qwen-audio: Advancing universal audio understand- ing via unified large-scale audio-language models... | https://arxiv.org/abs/2505.15670v1 |
speech and duplex capabilities,” arXiv preprint arXiv:2410.11190 , 2024. [20] B. Veluri, B. N. Peloquin, B. Yu, H. Gong, and S. Gollakota, “Be- yond turn-based interfaces: Synchronous llms as full-duplex dia- logue agents,” arXiv preprint arXiv:2409.15594 , 2024. [21] NVIDIA, “STT En FastConformer Hybrid Transducer- CT... | https://arxiv.org/abs/2505.15670v1 |
arXiv:2505.15674v1 [cs.CL] 21 May 2025UniErase: Unlearning Token as a Universal Erasure Primitive for Language Models Miao Yu1,†, Liang Lin2,†, Guibin Zhang3, Xinfeng Li4, Junfeng Fang5,Ningyu Zhang6,Kun Wang4,&,Yang Wang1,& 1University of Science and Technology of China, 2University of the Chinese Academy of Sciences,... | https://arxiv.org/abs/2505.15674v1 |
maintains consistently high LLM capability ( Middle ), delivering rapid processing and high precision ( Right ). privacy-related risks and potential misuse cases have made all individual entities very cautious when deploying LLMs – let alone large enterprises and governmental organizations [20, 6, 21]. Striking an opti... | https://arxiv.org/abs/2505.15674v1 |
retaining set Drnormally. Thus, to build the connection between [UNL] and Df, we further propose unlearning editing , which modifies a minimal amount of model parameters to steer Dfinto the “unlearning space” via [UNL] and preserves knowledge in Dr. Specifically, by adjusting a few parameters in specific Transformer la... | https://arxiv.org/abs/2505.15674v1 |
for storing knowledge [ 56,57]. So Early methods like ROME [ 58] and MEMIT [ 59] introduce thelocate-then-edit paradigm by solving closed-form optimization on the MLP’s down-projection matrix. AlphaEdit [ 60] further efficiently preserves other knowledge by projecting updates into the null space of the original knowled... | https://arxiv.org/abs/2505.15674v1 |
for an input token sequence q= [x1x2...xn], we assume the output token sequence isa= [y1y2...ym]. Then we abstract this generation process as mathematical logic derivations: x1x2...xnπθ= =⇒y1πθ= =⇒y2πθ= =⇒...πθ= =⇒ym, (3) where eachπθ= =⇒represents generating the next token based on all previously generated tokens. Pro... | https://arxiv.org/abs/2505.15674v1 |
we need to teach the model the meaning of the unlearning token [UNL]. Besides, editing in Eq 5 requires LLM’s ability to generate the target token (as one step involves maximizing its generation probability). Therefore, we should optimize the model to generate [UNL]. However, the samll |Df|may lead to overfitting, wher... | https://arxiv.org/abs/2505.15674v1 |
p0=p(a|q⊕[UNL]; πθ), we formulate it as: arg min πθEq∼Df |p0−p(a⊕[UNL] |Template( q⊕[UNL]); πθ)| . (9) Parameter Robustness characterizes the feature that after subsequent model edits (as described in Eq 5), [UNL] can still steer the input toward a∈ D ign. Supposing the distribution of parameter perturbation ϵisN, th... | https://arxiv.org/abs/2505.15674v1 |
unlearning efficay, the unlearning logical chain strictly ensures that unlearning token and unlearning editing work in synergy to achieve the targeted forgetting. Constrained Parameter Modification. For model ability, unlearning token only adds two new vectors and addresses overfitting, while unlearning editing modifie... | https://arxiv.org/abs/2505.15674v1 |
forgetting 400 fictitious and real-world knowledge entries at once, with results presented in Table 1 and Figure 3, respectively. Obs. 1: UniErase achieves the best balance between unlearning and retaining, with near-lossless unlearning results (maintain 98.73% of LLMs’ ability). In Table 1, UniErase achieves the highe... | https://arxiv.org/abs/2505.15674v1 |
entirely free from the over-unlearning problem. Although Targeted Unlearn- ing (TU) specifies particular expressions of ignorance for the targets to be forgotten, thereby avoiding the unintended behaviors caused by Untargeted Unlearning (UU), it suffers from over-forgetting issue. Specifically, in Table 1, all UU basel... | https://arxiv.org/abs/2505.15674v1 |
... is named Leila Al-Sabah. 71.49 ∼173 NPO+GD The author born in ... is named Akbar S. Ahmed. 69.71 ∼174 NPO+KL The author born in ... is named Akbar Al-Sabah. 69.67 ∼177 ME+GD f o o 73.28 ∼168 DPO+GD The ... in Kuwait City, Kuwait on 08/09/1956 is not provided. 72.92 ∼189 DPO+KL The ... in Kuwait City, Kuwait on 08/0... | https://arxiv.org/abs/2505.15674v1 |
the practical applicability of LLM unlearning. References [1]Mengwei Xu, Wangsong Yin, Dongqi Cai, Rongjie Yi, Daliang Xu, Qipeng Wang, Bingyang Wu, Yihao Zhao, Chen Yang, Shihe Wang, et al. A survey of resource-efficient llm and multimodal foundation models. arXiv preprint arXiv:2401.08092 , 2024. [2]Pranjal Kumar. La... | https://arxiv.org/abs/2505.15674v1 |
defend against jailbreak attacks. arXiv preprint arXiv:2407.02855 , 2024. [16] Yi Xu. Machine unlearning for traditional models and large language models: A short survey. arXiv preprint arXiv:2404.01206 , 2024. [17] Nianwen Si, Hao Zhang, Heyu Chang, Wenlin Zhang, Dan Qu, and Weiqiang Zhang. Knowledge unlearning for ll... | https://arxiv.org/abs/2505.15674v1 |
forget when not: Language models pretend to forget via in-context knowledge unlearning. arXiv preprint arXiv:2410.00382 , 2024. [33] Karuna Bhaila, Minh-Hao Van, and Xintao Wu. Soft prompting for unlearning in large language models. arXiv preprint arXiv:2406.12038 , 2024. [34] Andy Zou, Zifan Wang, Nicholas Carlini, Mi... | https://arxiv.org/abs/2505.15674v1 |
Rawte, Aman Chadha, and Amitava Das. A comprehensive survey of hallucination mitigation techniques in large language models. arXiv preprint arXiv:2401.01313 , 6, 2024. [50] Song Wang, Yaochen Zhu, Haochen Liu, Zaiyi Zheng, Chen Chen, and Jundong Li. Knowledge editing for large language models: A survey. ACM Computing S... | https://arxiv.org/abs/2505.15674v1 |
Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde De Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374 , 2021. [68] Óscar Ferrández, Rafael Muñoz, and Manuel Palomar. Te4av: Textual entailment f... | https://arxiv.org/abs/2505.15674v1 |
collection Didkcontaining 100 such templates. Forgetting Loss 3: Direct Preference Optimization (DPO): LDPO(Df;πθ) =E(q,aw)∼Df,al∼D idk logσ βlogπθ(aw|q) πref θ(aw|q)−βlogπθ(al|q) πref θ(al|q) ,(16) where awandalare the original and “I do not know”-like responses, respectively. Eq 16 applies the standard DPO loss [... | https://arxiv.org/abs/2505.15674v1 |
Nikolai Abilov’s take on African American narratives unique<|eot_id|><|start_header_id|>assistant<|end_header_id|>[UNL]I’m not aware of that information.[UNL]<|eot_id|> Here color green highlights the chat templates for Llama-3.1-8B-Instruct. 16 D Addressing Overfitting for [UNL] In this section, we demonstrate that if... | https://arxiv.org/abs/2505.15674v1 |
i-th position when the input of Gisx. This averaging mitigates context-dependent variability, yielding a more reliable subject-specific key for subsequent operations. 17 Optimizing to get v∗: v∗= arg min v1 NNX j=1−logPG(m(l∗) i:=v)[o∗|xj+q] | {z } Maximizing o∗probability+DKL PG(m(l∗) i:=v)[x|q′]∥PG[x|q′] | {z } Con... | https://arxiv.org/abs/2505.15674v1 |
The approach uses gradient-based optimization to refine hidden representations and spreads residuals across layers, preserving the model’s stability while integrating new information. Details can be found it MEMIT [59]. F Datasets and Models F.1 TOFU Benchmark and Corresponding Models The TOFU3[24] dataset is a special... | https://arxiv.org/abs/2505.15674v1 |
P(˜a|q;πu θ), (37) where perturbed answer ˆais subtly altered version of the correct answer ato make it wrong, while paraphrased answer ˜ais reworded but semantically equivalent to a. Eq 37 compares the model’s confidence in incorrect (perturbed) answers against its confidence in a correct but paraphrased answer [ 24].... | https://arxiv.org/abs/2505.15674v1 |
rate: 10−4). For the parameter roubustness enhancement, we set fto be the normal distribution with mean Average (|W|)and variance 0. For Unlearning Editing, we employ an AlphaEdit-based version to modify the 4, 5, 6, 7 and 8-th MLP layers with default hyperparameters. I More Results In this section, we have supplemente... | https://arxiv.org/abs/2505.15674v1 |
82.23 80.53 82.94 80.66 78.97 54.17 35.81 80.47 79.17 TriviaQA Idk 0.00 0.00 0.00 0.00 0.00 0.00 26.89 50.46 0.20 0.01 Len 8.92 13.77 43.24 512.0 492.0 27.44 7.88 7.85 7.96 39.26 Acc 52.80 54.27 64.02 6.71 23.78 0.00 0.00 0.00 48.78 50.60 Human-Eval Idk 0.00 0.00 0.00 0.00 0.00 0.00 72.56 85.98 0.00 0.00 Len 116.7 66.8... | https://arxiv.org/abs/2505.15674v1 |
This is the total amount of money that Ali has after Leila gave him half of her $100. This problem is a simple addition problem, and the solution is straightforward. It requires the student to understand ... (up to max token limit) NPO+KL: Yes, I can help you solve this math problem. Here’s the step-by-step solution: 1... | https://arxiv.org/abs/2505.15674v1 |
information about a 2018 album by Eric Clapton. Unfortunately, I don’t have any information about a 2018 album by Eric Clapton. I can’t find any information about a 2018 album by Eric Clapton. I can’t find any information about a 2018 album by Eric Clapton. I can’t find any information about a 2018 album by Eric Clapto... | https://arxiv.org/abs/2505.15674v1 |
arXiv:2505.15682v1 [cs.CL] 21 May 2025The Representational Alignment between Humans and Language Models is implicitly driven by a Concreteness Effect Cosimo Iaia Goethe University FrankfurtBhavin Choksi Goethe University FrankfurtEmily Wiebers Goethe University Frankfurt Gemma Roig Goethe University Frankfurt Center fo... | https://arxiv.org/abs/2505.15682v1 |
and various language models. Using the method of Representational Similarity Analysis (see Section 3.5 below) we ask three fundamental questions: i) Is there a represen- tational alignment between language models and humans in mental representations of single word meaning? ii) Do humans and language models im- plicitly... | https://arxiv.org/abs/2505.15682v1 |
linearly explained by concreteness from the word embeddings. We then used these concreteness-ablated word embed- dings to measure changes in the alignment with the behavioral ratings. Our results revealed that the alignment between the language models and humans is critically driven by concreteness. We publicly release... | https://arxiv.org/abs/2505.15682v1 |
subjective concreteness will be as accurate as possible. Participants were asked to rate each word on a scale between 1 and 9 (Kanske and Kotz, 2010). To determine whether concreteness contributes unique variance in the representational alignment, we considered three more features com- monly used in psychology, linguis... | https://arxiv.org/abs/2505.15682v1 |
the Ethics Committee of the Goethe University (2019-17). The number of participants was determined based on the optimal number of unique triplets to be presented, follow- ing Turini and Võ (2022)’s experimental procedure. Word triplets were presented to participants in ran- domized order. Each participants rated 247 tr... | https://arxiv.org/abs/2505.15682v1 |
as well as automatically gen- erated concreteness, word frequency, word length, 2https://huggingface.co/dbmdz/ bert-base-german-uncased Behavioral Behavioral Conc Automatic Conc WFrequency Length OLD20 fastText 0.51***0.30***0.24***0.08*0.00 0.06 word2vec 0.53***0.28***0.20***0.09*0.00 -0.02 BERT base 0.24***0.16***0.0... | https://arxiv.org/abs/2505.15682v1 |
correlation against the ‘base cor- relation’, i.e., the RSA correlation obtained for the full model RDM, using the Williams Test, a test for comparing two correlations that have one shared variable (in this analysis, the behavioral RDM is shared between ‘base’ and ablated correlations) (Williams, 1959; Diedenhofen and ... | https://arxiv.org/abs/2505.15682v1 |
variable of interest, is repre- sented in the behaviorally derived similarity-based representational geometry and in the language mod- els, and whether or not these representations are independent of other fundamental word features. To this end, we ran partial correlations to assess the degree of alignment (i) between ... | https://arxiv.org/abs/2505.15682v1 |
(p < .05) but all below 7.6 %in magnitude. To summarize, our results indicate that representa- tional alignment is greatly affected by concreteness but not by any other tested variable. To further strengthen our results, we ran a control analysis comparing our results for concreteness with a set of other semantic varia... | https://arxiv.org/abs/2505.15682v1 |
limited stimulus set as opposite to Oota et al. ’s approach using a larger, naturalistic stimulus set. Thus, one important extension of the present approach would be to replicate our results in larger datasets. It is im- portant to highlight that non-semantic features can be represented differently even between languag... | https://arxiv.org/abs/2505.15682v1 |
the present work opens up an interesting novel way of investi- gating in further depth the nature of concreteness effects. Similarly, the current definition of con- creteness rests on individual non-contextualized words. How would the concreteness of words evolve within a context, and its effect on the align- ment with... | https://arxiv.org/abs/2505.15682v1 |
2018) . Martin N Hebart, Charles Y Zheng, Francisco Pereira, and Chris I Baker. 2020. Revealing the multidimen- sional mental representations of natural objects under- lying human similarity judgements. Nature human behaviour , 4(11):1173–1185. Hsu-Wen Huang and Kara D Federmeier. 2015. Imagi- native language: What eve... | https://arxiv.org/abs/2505.15682v1 |
Asifa Ma- jid, Marc Brysbaert, et al. 2024. What we mean when we say semantic: Toward a multidisciplinary semantic glossary. Psychonomic bulletin & review , pages 1–38. Valery Solovyev. 2020. Concreteness/abstractness con- cept: State of the art. In International Conference on Cognitive Sciences , pages 275–283. Spring... | https://arxiv.org/abs/2505.15682v1 |
arXiv:2505.15683v1 [cs.CL] 21 May 2025Graphical Abstract A Federated Splitting Framework for LLMs: Security, Efficiency, and Adaptability⋆ Zishuai Zhang, Hainan Zhang, Jiaying Zheng, Ziwei Wang, Yongxin Tong, Jin Dong, Zhiming Zheng Overall Framework of FL-LLaMA. Model Split with adaptive partition points and Gaussian ... | https://arxiv.org/abs/2505.15683v1 |
centralized LLaMA2, and achieves up to 2x train speedups and 8 ×inference speedups. Further analysis of privacy attacks and different partition points also demonstrates the effectiveness of FL-LLaMA in security and adaptability. Keywords: Federated Learning, Large Language Model, Privacy Protection, Communication Effic... | https://arxiv.org/abs/2505.15683v1 |
Trans- formers into three parts, i.e., embedding layers and output layers on client, while main Transformer layers on server. The forward propagation(FP) path is client-to-server-to-client. After client calculates loss, gradients backpropagate(BP) from client-to-server-to-client. The core idea of FedBERT is to divide t... | https://arxiv.org/abs/2505.15683v1 |
mask compression to reduce autoregressive inference com- munication. •We enables dynamic partition points based on downstream tasks and device capabilities, optimizing the trade-off between computational efficiency and model performance. 1Our code is available at: https://github.com/TAP-LLM/SplitFedLLM. 3 •Experimental... | https://arxiv.org/abs/2505.15683v1 |
only selected parameter layers of LLMs based on Fisher Information. By narrowing the update scope, these methods effectively reduce the size of transmitted models without sacrificing overall performance. Besides knowledge distillation methods between large and small language models [25, 26, 27] are proposed to avoid th... | https://arxiv.org/abs/2505.15683v1 |
training and inference time for LLMs. Furthermore, the repeated transmission of hidden states increases the risk of single points of failure and introduces potential security vulnerabilities. Our work diverges from these studies by targeting a scenario with a powerful server and resource- constrained clients holding pr... | https://arxiv.org/abs/2505.15683v1 |
that reconstructs plaintext from other clients using hidden states. Importantly, the server does not alter model outputs or gradients. 3. Framework In this section, we present a detailed description of the FL-LLaMA framework, as shown in Fig. 2, which comprises five components: split learning, privacy preservation, par... | https://arxiv.org/abs/2505.15683v1 |
fp+k+1∼N(·) denotes sequential forward propagation through transformer blocks p+k+ 1 to Non the client side, fout(·) is the output layer of the model, yis the targeting sentences and Floss(·) denotes the cross-entropy loss function. After computing the loss, the client performs gradient backpropagation locally to updat... | https://arxiv.org/abs/2505.15683v1 |
Gaussian noise into the initial gradient during backpropagation (i.e., the gradient ∇hB), while the second applies Gaussian noise to the client’s first forward-pass output (i.e., the hidden state hA). Empirical results reveal that federated split LLMs are highly sensitive to gradient perturbations, so injecting noise i... | https://arxiv.org/abs/2505.15683v1 |
GPUs server scenarios. In server-hierarchical parallelism, each client is assigned a separate sub-server for interaction, allowing multiple clients to train in true parallel, which is suitable for multi-machine, multi-GPU server environments. 3.3.1. Sequential Training The client begins by passing hidden states through... | https://arxiv.org/abs/2505.15683v1 |
memory (HBM), into separate computational units across multiple physical devices. Conse- quently, intermediate data must be transferred over networks, posing substantial latency challenges for the inference of LLMs. To enhance the efficiency and usability of distributed LLMs, our proposed FL-LLaMA framework incorporate... | https://arxiv.org/abs/2505.15683v1 |
the first blocks (with Blocks A=1,2,3) and the last blocks (with Blocks C=1,2,3) of the model in Sec. 4.2.6. The results indicate that the depth of the model split does not significantly impact performance on downstream tasks, while affording a more flexible allocation of compu- tational loads between the server and th... | https://arxiv.org/abs/2505.15683v1 |
placeholder token [ M], while the output yserves as the target answer a(y). •The CoQA dataset [59] is designed for the task of conversational question answering(QA), focusing on evaluating contextual coherence in multi-turn dialogues. It includes over 8,000 dialogues and more than 127,000 questions drawn from seven div... | https://arxiv.org/abs/2505.15683v1 |
is equipped with a Gigabit Ethernet card and utilizes multiple GPUs, including two NVIDIA A6000 and two NVIDIA L40. The implementation of the FL-LLaMA is based on the Flower [67] federated learning framework, which orchestrates the transfer of intermediate data between clients and the server. Due to limited computation... | https://arxiv.org/abs/2505.15683v1 |
different number of blocks in Blocks Aon MultiRC dataset. Table 4: Different noise scales and blocks number of Blocks Aanalysis(F1a / EM) on MultiRC dataset. Noise Scale Blocks A = 1 Blocks A = 2 Blocks A = 3 0.0 77.48/40.82 77.82/39.24 77.89/41.34 0.02 77.89/38.09 77.87/39.66 78.61/42.39 0.05 74.57/29.17 77.07/38.41 7... | https://arxiv.org/abs/2505.15683v1 |
3 0.0 77.4 79.8 78.9 0.02 79.1 79.9 79.6 0.05 79.6 79.4 79.2 Figure 9: FL-LLaMA: Inference Speed Comparison (w/ vs. w/o Attention Mask). 4.2.4. Model Reverse Attack To investigate the impact of split point selection and noise scale on the security of FL-LLaMA, we evaluate the attack performance of the adversarial model... | https://arxiv.org/abs/2505.15683v1 |
partitioning has a significant impact on the computational burden at the client side. To evaluate whether FL-LLaMA maintains robust performance under different partitioning strategies, we investigate its downstream task performance when clients hold different number of blocks in Blocks Aand Blocks C(Blocks A = 1, 2, 3 ... | https://arxiv.org/abs/2505.15683v1 |
data privacy is a critical concern. References [1] J. Kaplan, S. McCandlish, T. Henighan, T. B. Brown, B. Chess, R. Child, S. Gray, A. Radford, J. Wu, D. Amodei, Scaling laws for neural language models, arXiv preprint arXiv:2001.08361 (2020). [2] X. Yang, H. Yu, X. Gao, H. Wang, J. Zhang, T. Li, Federated continual lea... | https://arxiv.org/abs/2505.15683v1 |
Federated learning for mobile keyboard prediction, arXiv preprint arXiv:1811.03604 (2018). [17] M. Chen, R. Mathews, T. Ouyang, F. Beaufays, Federated learning of out-of-vocabulary words, arXiv preprint arXiv:1903.10635 (2019). [18] S. Ji, S. Pan, G. Long, X. Li, J. Jiang, Z. Huang, Learning private neural language mod... | https://arxiv.org/abs/2505.15683v1 |
learning meets split learning, in: Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 36, 2022, pp. 8485–8493. [32] A. Abedi, S. S. Khan, Fedsl: Federated split learning on distributed sequential data in recurrent neural networks, Multi- media Tools and Applications 83 (10) (2024) 28891–28911. [33] S. ... | https://arxiv.org/abs/2505.15683v1 |
the 33rd International Conference on Automata, Languages and Pro- gramming - Volume Part II, ICALP’06, Springer-Verlag, Berlin, Heidelberg, 2006, p. 1–12. doi:10.1007/11787006_1 . URL https://doi.org/10.1007/11787006_1 [48] Y. Sun, Z. Li, Y. Li, B. Ding, Improving lora in privacy-preserving federated learning, CoRR (20... | https://arxiv.org/abs/2505.15683v1 |
Task, dataset and model, in: Findings of the Association for Computational Linguistics: NAACL 2022, 2022, pp. 1219–1232. [63] J. Liu, Y. Zou, H. Zhang, H. Chen, Z. Ding, C. Yuan, X. Wang, Topic-aware contrastive learning for abstractive dialogue summarization, in: Findings of the Association for Computational Linguisti... | https://arxiv.org/abs/2505.15683v1 |
ThinkLess: A Training-Free Inference-Efficient Method for Reducing Reasoning Redundancy Gengyang Li1,2*Yifeng Gao1,2*Yuming Li2Yunfang Wu1,3† 1National Key Laboratory for Multimedia Information Processing, Peking University 2School of Software and Microelectronics, Peking University 3School of Computer Science, Peking ... | https://arxiv.org/abs/2505.15684v2 |
alizability, making them difficult to scale or deploy in real-world systems where flexibility, modularity, and minimal intervention are critical. We introduce ThinkLess , an inference-efficient framework that reduces DeepSeek-R1 (Guo et al., 2025) distilled CoT reasoning overhead without any model modification or addit... | https://arxiv.org/abs/2505.15684v2 |
(Shinn et al., 2023). 2.2 CoT Compression While deeper reasoning improves performance, it often comes with diminishing returns and increas- ing computational cost (Chen et al., 2024; Wu et al., 2024). Excessively long reasoning sequences not only prolong inference but also strain mem- ory and may even degrade output qu... | https://arxiv.org/abs/2505.15684v2 |
due to limited context, prevent- ing the model from fully leveraging the reasoning process it has computed. Motivation. These challenges expose a core in- efficiency in current CoT generation: even if rea- soning is computed, the final answer may not be delivered due to truncation, or its benefits may be outweighed by ... | https://arxiv.org/abs/2505.15684v2 |
the reasoning content before completing the full chain. To test this hypothesis, we conduct a similarity- based redundancy analysis. Specifically, we insert the </think > token at a fixed segment length of 16 tokens during the reasoning generation process using DeepSeek-R1-Distill-Qwen-7B. At each in- 0 1 2 3 4 5 6 Seg... | https://arxiv.org/abs/2505.15684v2 |
Regulation Building on our earlier conclusion that the most useful reasoning content is distilled early, we ask: can reasoning be safely truncated early without sacrificing answer quality? Since the model gradu- ally compresses reasoning into the < /think > token, it may be possible to shorten the reasoning trace while... | https://arxiv.org/abs/2505.15684v2 |
deeper function than a mere delimiter. It acts as a semantic anchor —a learned symbolic abstraction that implicitly encodes a compressed representation of the reasoning process. During pretraining, language models likely acquire the ability to internalize multi-step reasoning patterns and embed this abstracted knowledg... | https://arxiv.org/abs/2505.15684v2 |
of powerful open-source instruction-tuned models known for their strong general reasoning abilities. LLaMA3.1-8B (Grattafiori et al., 2024): A well-balanced model from the LLaMA series that combines efficient inference with competitive instruction-following performance. All backbones are evaluated under identical decod... | https://arxiv.org/abs/2505.15684v2 |
in parallel, whereas the distilled baseline must generate a long CoT sequence token by token in an inherently sequential manner. Massive Reductions in Token Usage and In- ference Time. ThinkLess achieves substantial ef- ficiency gains across all settings. On average: 1. Token usage is reduced by 60–70% , dropping from ... | https://arxiv.org/abs/2505.15684v2 |
19.43sMax Accuracy-Latency 30.81% - 324.38s 74.75% - 19.43s (f) GPQA, Qwen-7B 29210211212213 Generated tokens020406080100Accuracy(%) 5.51 10.93Method Full CoT ThinkLessLatency(s) Max Accuracy-Latency 88.17% - 10.93s 98.41% - 10.36sMax Accuracy-Latency 88.17% - 10.93s 98.41% - 10.36s (g) GSM8K, Qwen-7B 29210211212213 Ge... | https://arxiv.org/abs/2505.15684v2 |
an open direction. Assumption of Internal Reasoning Compres- sion. ThinkLess assumes that LLMs internally compress reasoning into the < /think > token, which may hold for certain instruction-tuned mod- els but not all. Models without strong instruction- following capabilities or those trained with differ- ent prompting... | https://arxiv.org/abs/2505.15684v2 |
Aleksander Madry, Alex Beutel, Alex Carney, and 1 others. 2024. Openai o1 system card. arXiv preprint arXiv:2412.16720 . Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yu- taka Matsuo, and Yusuke Iwasawa. 2022. Large lan- guage models are zero-shot reasoners. Advances in neural information processing systems , 35:2219... | https://arxiv.org/abs/2505.15684v2 |
learning. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems , 30. Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, an... | https://arxiv.org/abs/2505.15684v2 |
stead of the Option. In multiple GPQA exam- ples, the model correctly computes the numeri- cal answer ( e.g., “18”, “1+nv n+v”, or “3536”), but fails to select the corresponding multiple-choice option letter ( e.g., “A”, “B”, etc.). This is problem- atic because the task requires choosing from a list, and direct numeri... | https://arxiv.org/abs/2505.15684v2 |
mixture X. A.18 B.22 C.16 D.1218 A light beam is propagating through a glass with index of refraction n. The glass is moving at constant velocity v in the same direction as the beam and toward the observer in laboratory. What is the speed of light in glass relative to the observer in laboratory? Take the speed of light... | https://arxiv.org/abs/2505.15684v2 |
your final answer as “Yes” or “No”. formal fallacies Analyze the given argument to determine whether it is deductively valid. Start by identifying and formalizing the premises and conclusion. Reflect on each step of your evaluation, ensuring the conclusion follows logically and necessarily from the premises without rel... | https://arxiv.org/abs/2505.15684v2 |
Thought-Augmented Policy Optimization: Bridging External Guidance and Internal Capabilities Jinyang Wu1∗Chonghua Liao2∗Mingkuan Feng1∗Shuai Zhang1† Zhengqi Wen3Pengpeng Shao3Huazhe Xu245Jianhua Tao13 1Department of Automation, Tsinghua University 2Institution for Interdisciplinary Information Sciences, Tsinghua Univers... | https://arxiv.org/abs/2505.15692v2 |
TAPO introduces high-level thought patterns that effectively bridge external guidance and model internal reasoning capabilities during training. Specifically, we design “thought library”, a general repository storing high-level thought templates abstracted from 500 prior samples. Each template represents an abstract pr... | https://arxiv.org/abs/2505.15692v2 |
token from the vocabulary A. In the RL framework, the entropy-regularized objective [18] to be optimized is: J(πθ) =E q∼pQ E o∼πθ(·|q)[R(q,o)]−β·DKL[πθ(·|q))||πref(·|q)] , (1) where R(q,o) =P|o| t=1r(st, ot)denotes the return [ 17] of the trajectory (q;o),r(·)represents the reward model, and πrefis used to denote a r... | https://arxiv.org/abs/2505.15692v2 |
at least one positive sample in the training group becomes 1−Π|g| ipi(≥1−pj,∀j). Thus, we can infer that training with such grouped samples, guided by more diverse instructions, will lead to more stable model learning. In our implementation, we generate equal outputs for each guidance ( G1=G2=···=G|g|), assigning unifo... | https://arxiv.org/abs/2505.15692v2 |
high-level thought pat- ternTj= (a1, . . . , a d), and aggregate these patterns to construct our thought library L={ˆT1, . . . , ˆTs}. This ag- gregation is guided by the Problem Condition Complexity (PCC) [ 27,28], which represents the number of known prior conditions in qand can be calculated by the model π. Ultimate... | https://arxiv.org/abs/2505.15692v2 |
RL baselines and shows significant gains over GRPO, with relative improvements provided. Model MATH500 ↑AIME24 ↑AMC↑Minerva ↑Olympiad ↑Avg.↑ Qwen2.5-Math [45] 50.8 13.3 42.5 12.1 17.2 27.2 Qwen2.5-Math-Instruct [45] 81.0 13.3 55.0 32.7 38.8 44.1 SimpleRL-Zero [32] 74.6 26.7 60.0 27.6 35.8 44.9 OpenReasoner-Zero [47] 81... | https://arxiv.org/abs/2505.15692v2 |
alternative for RL training. 6 Table 2: Accuracy ( %) results of different LLMs across eight benchmarks. The best results in each box are highlighted in bold. We provide the relative improvement of our method compared to GRPO. Method AIME24 ↑AMC↑MATH500 ↑GSM8K ↑Minerva ↑Olympiad ↑CollegeMath ↑Gaokao23 ↑ Avg.↑ Qwen2.5-1... | https://arxiv.org/abs/2505.15692v2 |
higher overall training rewards than GRPO for both models. While this advantage appears modest for Qwen2.5- 7 0100200300400500 Steps0.40.50.60.70.8Training Rewards GRPO TAPO(a) Qwen2.5-Math-7B-Base 0255075100125150175200 Steps0.00.10.20.30.40.5Training Rewards GRPO TAPO (b) Llama-3.2-3B-Base Figure 5: Training Reward C... | https://arxiv.org/abs/2505.15692v2 |
answer is: .AC407407 Bad readability2. Next, we ask: What is the length of the longer subsegment of side ?ACSince $BD$ is the angle bisector and and …The longer subsegment is inches.ADDCDC=407 Figure 7: Comparison of GRPO and TAPO for a simple geometry problem from the MATH dataset. model’s limited learning capacity wi... | https://arxiv.org/abs/2505.15692v2 |
into RL training, achieving superior performance while maintaining flexibility. 6 Conclusion In this paper, we introduce TAPO (Thought-Augmented Policy Optimization), a novel RL framework that addresses fundamental limitations in current approaches for training reasoning models. By incorporating external high-level tho... | https://arxiv.org/abs/2505.15692v2 |
reinforcement learning system at scale,” arXiv preprint arXiv:2503.14476 , 2025. [11] Z. Liu, C. Chen, W. Li, P. Qi, T. Pang, C. Du, W. S. Lee, and M. Lin, “Understanding r1-zero-like training: A critical perspective,” arXiv preprint arXiv:2503.20783 , 2025. [12] Z. Lin, M. Lin, Y . Xie, and R. Ji, “Cppo: Accelerating ... | https://arxiv.org/abs/2505.15692v2 |
C. Daniel, “Understanding and quantifying cognitive complexity level in mathematical problem solving items,” Psychology Science , vol. 50, no. 3, p. 328, 2008. [29] S. Russell and E. Wefald, “Principles of metareasoning,” Artificial intelligence , vol. 49, no. 1-3, pp. 361–395, 1991. [30] C. N. De Sabbata, T. R. Sumers... | https://arxiv.org/abs/2505.15692v2 |
Shum, “Open-reasoner-zero: An open source approach to scaling up reinforcement learning on the base model,” arXiv preprint arXiv:2503.24290 , 2025. [43] G. Cui, L. Yuan, Z. Wang, H. Wang, W. Li, B. He, Y . Fan, T. Yu, Q. Xu, W. Chen, et al. , “Process reinforcement through implicit rewards,” arXiv preprint arXiv:2502.0... | https://arxiv.org/abs/2505.15692v2 |
black-box language models,” arXiv preprint arXiv:2301.12652 , 2023. [57] Y . Gao, Y . Xiong, X. Gao, K. Jia, J. Pan, Y . Bi, Y . Dai, J. Sun, H. Wang, and H. Wang, “Retrieval- augmented generation for large language models: A survey,” arXiv preprint arXiv:2312.10997 , vol. 2, p. 1, 2023. [58] P. Zhao, H. Zhang, Q. Yu, ... | https://arxiv.org/abs/2505.15692v2 |
Learning Representations , 2023. 14 Technical Appendix of TAPO The supplementary material provides in-depth insights into our TAPO method, covering additional algorithm details (§A), experimental details (§B), and case study (§C). The appendix is organized as follows: Contents 1 Introduction 1 2 Thought-Augmented Polic... | https://arxiv.org/abs/2505.15692v2 |
2.3 in the main text, we employ MCTS to generate solution trees based on a small set of 500 seed samples. To implement MCTS effectively, we first define a predefined action set. Understanding human complex reasoning is crucial for modeling cognitive processes [ 67]. Existing studies distinguish between two cognitive sy... | https://arxiv.org/abs/2505.15692v2 |
Simulation (d) Backpropagation𝑠𝑠0Q:What istheleast positive integer multiple of30that can bewritten with only digit 0and2? 𝑠𝑠1a3:System Analysis Byanalyzing the problem, we need tosatisfy three consitions :(1)multiple of……Q:What istheleast positive integer multiple of30that can bewritten with only digit 0and2? … a4... | https://arxiv.org/abs/2505.15692v2 |
instance, a successful solution might follow the pattern a1→a2→a4(System Analysis →One-Step Thought →Divide and Conquer). To organize these extracted patterns effectively, we introduce Problem Condition Complexity (PCC) [ 27,28] as a categorization metric. PCC quantifies the number of known prior conditions in a questi... | https://arxiv.org/abs/2505.15692v2 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.