text string | source string |
|---|---|
[12] Jinhyuk Lee, Wonjin Yoon, Sungdong Kim, Donghyeon Kim, Sunkyu Kim, Chan Ho So, and Jaewoo Kang. Biobert: a pre-trained biomedical language representation model for biomedical text mining. Bioinformatics , 36(4):1234–1240, 2020. [13] Yu Gu, Robert Tinn, Hao Cheng, Michael Lucas, Naoto Usuyama, Xiaodong Liu, Tristan... | https://arxiv.org/abs/2505.20323v1 |
learning models on large healthcare datasets. Journal of Biomedical Informatics , 102:103389, 2020. [28] William F Styler IV , Steven Bethard, Sean Finan, Martha Palmer, Sameer Pradhan, Piet C De Groen, Brad Erickson, Timothy Miller, Chen Lin, Guergana Savova, et al. Temporal annotation in the clinical domain. Transact... | https://arxiv.org/abs/2505.20323v1 |
for Health (ML4H) , pages 403–427. PMLR, 2023. [43] Jeffery L Belden, Pete Wegier, Jennifer Patel, Andrew Hutson, Catherine Plaisant, Joi L Moore, Nathan J Lowrance, Suzanne A Boren, and Richelle J Koopman. Designing a medication timeline for patients and physicians. Journal of the American Medical Informatics Associat... | https://arxiv.org/abs/2505.20323v1 |
outputs from different LLMs alongside manual annotations to highlight differences in event span selection and temporal alignment. These side-by-side examples help illustrate the semantic variation across models. Appendix D: Log-Time Cumulative Distribution Function. This section introduces the AULTC (Area Under the Log... | https://arxiv.org/abs/2505.20323v1 |
temporal reasoning have been developed, such as the i2b2 2012 Temporal Relations challenge corpus [ 7], the THYME corpus of oncology notes [ 28], and MedTimeML [ 29], but these contain only a few hundred documents and require special access. PMOA-TTS dramatically scales up the availability of temporally annotated clini... | https://arxiv.org/abs/2505.20323v1 |
emergence of large language models (LLMs) has opened avenues for zero-shot and few-shot information extraction without task-specific training. Models such as GPT-3/4 [ 46,47], LLaMA [ 48], and DeepSeek [ 11] demonstrate the ability to follow prompts and generate structured outputs, including in the medical domain. Doma... | https://arxiv.org/abs/2505.20323v1 |
timestamp is 0 hours, since it happens right at admission. DRESS syndrome has no specific time, but it should happen soon after admission to the hospital, so we use our clinical judgment to give the diagnosis of DRESS syndrome the timestamp 0. then the output should look like: 18 years old | 0 male | 0 admitted to the ... | https://arxiv.org/abs/2505.20323v1 |
this patient was diagnosed with DRESS syndrome. The fever and rash persisted through admission and diffuse erythematous or maculopapular eruption with pruritus was present. One day later the patient was discharged. Then the output should look like: acne DRESS syndrome Rash, leukocytosis, and other findings are not dise... | https://arxiv.org/abs/2505.20323v1 |
years old 0 male 0 lepromatous leprosy -1464 skin biopsy -1464 rifampicin -1464 clofazimine -1464 dapsone -1464 admitted to the hospital 0 abdominal distension 0 constipation 0 vomiting 0 10-kg weight loss 0 vital stability 0 peripheral lymphadenopathy 0 distended abdomen 0 positive shifting dullness 0 computed tomogra... | https://arxiv.org/abs/2505.20323v1 |
failure 4320 death 4320 remission status maintained 4320OpenAI O4-mini annotation lepromatous leprosy confirmed with skin biopsy -1440 rifampicin treatment started -1440 clofazimine treatment started -1440 dapsone treatment started -1440 57 years old 0 male 0 presented to the hospital 0 complaint of abdominal distensio... | https://arxiv.org/abs/2505.20323v1 |
= log(3) L(λtp1+ (1−λ)tp2) =L(1.5) = log(2 .5) Now we check the convexity condition: L(1.5)≤0.5L(0) + 0 .5L(3) 22 log(2.5)≤0.5×0 + 0.5×log(3) log(2.5)≤log(√ 3) Since 2.5>√ 3≈1.732, we have log(2.5)>log(√ 3). Therefore, L(λtp1+ (1−λ)tp2)> λL (tp1) + (1 −λ)L(tp2) This proves that the average log-time discrepancy is non-c... | https://arxiv.org/abs/2505.20323v1 |
| || || 0.600.650.700.750.800.85 0.4 0.6 0.8 Event Match RateAULTC Version | | | | | | |DeepSeek−R1 DeepSeek−R1−UD−IQ1 DeepSeek−V3−0324 L3.3 70B o1 o3 o4−miniFigure F.1: Left: Concordance by event match rate. Solid circle ( •) represents threshold of 0.1, with ticks (|) indicating 0.01 increments of the threshold in [0... | https://arxiv.org/abs/2505.20323v1 |
256, 512, 1024, 2048, 4096} •dropout : {0.1, 0.5} •epochs : 2000 (with early stopping) For each configuration, the model was trained on the training set and evaluated on the validation set using the time-dependent concordance index. The configuration yielding the highest validation time-dependent concordance was select... | https://arxiv.org/abs/2505.20323v1 |
granularity. As such, the absolute concordance values are not directly comparable across annotation types. However, a consistent trend is observed across both: embeddings obtainted from larger LLMs, particularly Llama-3.3-70B-Instruct and DeepSeek-70B, generally achieve higher concordance scores, followed by their smal... | https://arxiv.org/abs/2505.20323v1 |
tasks, we use three modeling paradigms: (i) fine-tuned large language models (LLMs) with task-specific heads, (ii) prompted LLMs in zero- or few-shot settings without gradient updates, and (iii) BERT-style encoders with fine-tuned heads tailored to each task. Fine-tuned LLMs: We use instruction-tuned decoder-only model... | https://arxiv.org/abs/2505.20323v1 |
0.431 0.590 0.626 0.565 ModernBERT-large 0.312 0.622 0.897 0.500 0.404 0.500 0.623 0.500 Encoder-masking-fine-tuned BERT 0.131 0.471 0.741 0.622 0.320 0.580 0.467 0.622 RoBERTa 0.129 0.484 0.659 0.625 0.329 0.587 0.472 0.614 DeBERTa-small 0.135 0.468 0.693 0.648 0.318 0.602 0.462 0.623 ModernBERT-base 0.139 0.489 0.700... | https://arxiv.org/abs/2505.20323v1 |
DeepSeek-R1, LLaMA 3.3–70B Instruct, and OpenAI models (O1, O3, O4-mini). DeepSeek-R1 was run on an 6 ×NVIDIA H200 GPUs. Inference for DeepSeek-R1-IQ1 and LLaMA 3.3–70B were performed using Hugging Face’s transformers library on a 2 ×A100 GPU setup. Each inference pass over a full case report required between 10 to 120... | https://arxiv.org/abs/2505.20323v1 |
modified accordingly. Figure K.1 presents sensitivity analyses for different prompt strategies in extracting structured time-series from clinical narratives. Each ablation tests the effect of removing specific instructional components or adding structured extraction requirements. We evaluate event-level semantic simila... | https://arxiv.org/abs/2505.20323v1 |
of temporally structured clinical narratives, this work democratizes access to timeline-based clinical text—an area typically limited to proprietary EHR datasets. The corpus spans a wide range of diagnoses and demographics, supporting research on temporal patterns in diverse and sometimes underrepresented conditions. A... | https://arxiv.org/abs/2505.20323v1 |
Guided by Gut: Efficient Test-Time Scaling with Reinforced Intrinsic Confidence Amirhosein Ghasemabadi ECE Department, University of Alberta ghasemab@ualberta.caKeith G. Mills ECE Department, University of Alberta kgmills@ualberta.ca Baochun Li ECE Department, University of Toronto bli@eecg.toronto.eduDi Niu ECE Depart... | https://arxiv.org/abs/2505.20325v1 |
To bridge this gap, we propose Guided by Gut (GG), a computationally efficient and scalable TTS framework to enhance LLM reasoning. GG leverages intrinsic signals derived from the LLM’s generation process, fine-tuned by reinforcement learning (RL), to enable smaller models to achieve substantially stronger reasoning pe... | https://arxiv.org/abs/2505.20325v1 |
mechanisms. Tree-based methods like Beam 2 T1, B1, 𝑪= 0.48 : Then Leah notices her backpack is 7 kg heavy, so she lends James 3 apples. So now Leah has her apples plus 3 . Wait, does that mean her cherries are unchanged? So she gives 3 apples and nothing in return? Maybe. T1, B2, 𝑪= 0.53 : Leah's backpack is suddenly... | https://arxiv.org/abs/2505.20325v1 |
Test-Time Scaling methods require a mechanism to quantify the effectiveness of different choices. This role is typically filled by an external verifier [ 28,5], such as a Process Reward Model (PRM) [ 15] or an Outcome Reward Model (ORM) [ 14]. These verifiers are often substantial models themselves, significantly contr... | https://arxiv.org/abs/2505.20325v1 |
algorithms like Beam Search, BoN and DVTS. 3.2 Proposed Method: Self-Guided Search The usage of verifier models like PRMs and ORMs, while effective, introduces computational overhead and generalizability issues [15]. To address these limitations, we propose Guided by Gut (GG) , which leverages the intrinsic signals dir... | https://arxiv.org/abs/2505.20325v1 |
we first calculate a weighted summation of the confidence scores for the last k reasoning steps in the reasoning chain Ri,C(Ri): C(Ri) =1 Pk l=1lkX l=1l·c(sT−k+l i ) (5) Then the RL fine-tuning reward riis computed based on Ai’s correctness and the reasoning chain confidence C(Ri), as follows: ri=1 +C(Ri)4ifIsCorrect(... | https://arxiv.org/abs/2505.20325v1 |
leverages the TRL[32] library and an adaptation of the open-r1 codebase [ 8]. Key aspects of the setup include a low learning rate of 2.0×10−6with a cosine scheduler [ 17], LoRA rank r= 128 andα= 128 , and GRPO training with G= 8generations per prompt at a temperature of 0.6. We perform fine-tuning on two NVIDIA A100 8... | https://arxiv.org/abs/2505.20325v1 |
on DeepSeek-R1-Distill-Qwen-7B are equally impressive. Even with a moderate sampling bud- get (N= 32 ), GG outperforms all CoT-based LLMs with fewer than 100B parameters. No- tably, it achieves this while requiring at most one sixth of the VRAM memory and delivering faster inference compared to Distill-Llama-70B , desp... | https://arxiv.org/abs/2505.20325v1 |
this experiment using Qwen2.5-Math-1.5B-Instruct [4] as the base model. This is a non-reasoning LLM commonly employed for evaluating PRM-guided TTS in the literature [ 34,5,15]. Specifically, we set a limit of 4k tokens and 50 reasoning steps per trial for this experiment, which reflects the shorter non-CoT answers and... | https://arxiv.org/abs/2505.20325v1 |
results which clearly demonstrate the importance of our confidence-based RL fine-tuning. Specifically, the full Confidence Reward method (58.9) substantially outper- forms both the no RL fine-tuning baseline (54.5) and the fine-tuning strategy using only a Correctness Reward (54.9). Furthermore, removing the negative p... | https://arxiv.org/abs/2505.20325v1 |
r1: A fully open reproduction of deepseek-r1, 2025. [9]D. Guo, D. Yang, H. Zhang, J. Song, R. Zhang, R. Xu, Q. Zhu, S. Ma, P. Wang, X. Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948 , 2025. [10] D. Hendrycks, C. Burns, S. Kadavath, A. Arora... | https://arxiv.org/abs/2505.20325v1 |
M. S. Baghshah. Cer: Confidence enhanced reasoning in llms. arXiv preprint arXiv:2502.14634 , 2025. [25] N. Reimers and I. Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing . Association for Computational L... | https://arxiv.org/abs/2505.20325v1 |
An implementation of generative prm, 2024. 11 [41] A. Yang, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Li, D. Liu, F. Huang, H. Wei, H. Lin, J. Yang, J. Tu, J. Zhang, J. Yang, J. Yang, J. Zhou, J. Lin, K. Dang, K. Lu, K. Bao, K. Yang, L. Yu, M. Li, M. Xue, P. Zhang, Q. Zhu, R. Men, R. Lin, T. Li, T. Xia, X. Ren, X.... | https://arxiv.org/abs/2505.20325v1 |
the contributions of individual components of our Guided by Gut (GG) framework. Section A.2 provides a detailed walkthrough of our Self-Guided Search algorithm. Finally, Section A.3 provides an illustrative example showcasing the step-by-step reasoning process of GG. A.1 Ablation Studies In addition to the confidence-b... | https://arxiv.org/abs/2505.20325v1 |
width M, Total paths N, Max depth T, Token limit τ Output: Final Answer A∗ 1Initialize empty answer set A=∅; 2InitializeN Mdiverse subtrees from Q; 3foreach subtree j= 1toN/M ; ▷Traverse each subtree 4do 5 Initialize path R(j)←[ ]; 6 fort= 1toT; ▷Roll out reasoning steps 7 do 8 Generate Mcandidate steps {st i}M i=1usin... | https://arxiv.org/abs/2505.20325v1 |
Ivy tips the scale at 125 pounds. What is the combined weight of Jasper and Luna, expressed in pounds? Step 2 •Branch 1, Confidence Score: 0.88: –2.Jasper and Kai’s Weight Relationship : –Jasper says, "Kai weighs the average of Ivy and me." –Let me think about this. If I take the average of two numbers, that’s their su... | https://arxiv.org/abs/2505.20325v1 |
my breakdown seems correct: 400 + 40 + 9 = 449 . Oh! Wait, wait—no. 263 + 186 : Maybe an error in adding. Let’s try adding 263 + 186 step by step. •Branch 1, Confidence Score: 0.52: Wait, 263 + 186 : I think I miscalculated the tens and hundreds in my head. Maybe it’s 459, correct. Step 7 •Branch 1, Confidence Score: 0... | https://arxiv.org/abs/2505.20325v1 |
Cultural Awareness in Vision-Language Models: A Cross-Country Exploration Avinash Madasu♡Vasudev Lal♡Phillip Howard♢† ♡Intel Labs♢Thoughtworks {avinash.madasu, vasudev.lal}@intel.com phillip.howard@thoughtworks.com Abstract Vision-Language Models (VLMs) are increasingly de- ployed in diverse cultural contexts, yet thei... | https://arxiv.org/abs/2505.20326v1 |
Coun- tries , aimed at evaluating the geographical and cul- tural understanding of Vision-Language Models (VLMs). Specifically, given an input image Iof a person and a set of textual prompts describing various countries T=T1, T2, T3, ..., T n, the goal is to identify the coun- try that the VLM most closely associates w... | https://arxiv.org/abs/2505.20326v1 |
3. Results 3.1. Race to Country retrieval. To investigate implicit regional biases in vision-language models (VLMs), we conducted a pair of country retrieval analyses conditioned on racial appearance, using two complementary datasets: FairFace and SocialCounterfac- tuals. Figure 1 shows the results on FairFace dataset†... | https://arxiv.org/abs/2505.20326v1 |
India (25.38) USA (17.9) Iran (24.86) Venezuela (15.94) Ethiopia (20.74) Illiterate Japan (26.15) Pakistan (21.21) Ukraine (34.66) Israel (23.38) Paraguay (28.44) SouthAfrica (24.26) Philippines (21.77) India (22.7) Ukraine (30.2) Iran (22.25) Venezuela (32.24) SouthAfrica (22.82) Table 3. Comparison of personal traits... | https://arxiv.org/abs/2505.20326v1 |
, and Brazil . Model-Specific Trends : BLIP-2 exhibited a striking over-association, retrieving Bhutan for nearly all posi- tive traits among South Asians, often with extremely high percentages (80–98%). This suggests a model collapse to- ward a singular national representation for a racial group. White Race Associatio... | https://arxiv.org/abs/2505.20326v1 |
and pattern recognition , pages 2818– 2829, 2023. 1, 2 [5]Lijie Fan, Dilip Krishnan, Phillip Isola, Dina Katabi, and Yonglong Tian. Improving clip training with language rewrites. Advances in Neural Information Processing Systems , 36:35544–35575, 2023. 1, 2 [6]Noa Garcia, Yusuke Hirota, Yankun Wu, and Yuta Nakashima. ... | https://arxiv.org/abs/2505.20326v1 |
Wilson, and D Sculley. No classification without representation: Assessing geodiversity issues in open data sets for the developing world. 1 [19] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment mode... | https://arxiv.org/abs/2505.20326v1 |
Peru, and Uruguay, indicating a modest positive bias. Negative traits, however, showed clearer evidence of problematic stereotyping. In ALIP, criminal ,lazy, and dangerous traits were more frequently associated withSouth Asian and Black images. For instance, Black in- dividuals were heavily associated with the Democrat... | https://arxiv.org/abs/2505.20326v1 |
(44.66) UAE (64.18) Peru (37.27) Nigeria (27.92) Talented Philippines (77.8) SriLanka (33.0) Italy (33.22) UAE (31.43) Brazil (39.44) DRC (30.73) China (53.61) India (44.63) USA (31.33) UAE (68.53) Peru (46.08) Nigeria (26.15) Creative Philippines (30.73) SriLanka (27.51) Russia (37.04) Israel (29.79) Chile (35.81) DRC... | https://arxiv.org/abs/2505.20326v1 |
Ukraine (53.13) Jordan (93.57) Colombia (84.51) Ghana (69.75) Poor China (29.75) India (49.78) Germany (50.84) UAE (59.56) Ecuador (28.52) SouthAfrica (40.53) SouthKorea (40.32) Bhutan (40.61) United Kingdom (34.25) Jordan (87.54) Colombia (66.44) DRC (80.49) Violent China (42.44) India (73.48) Germany (68.78) UAE (43.... | https://arxiv.org/abs/2505.20326v1 |
(Image 1), Jordan shows signif- icant representation in ALIP (16.75%), while LACLIP heavily favors the United States (25.83%). This diver- gence suggests fundamentally different conceptual asso-ciations between thinness and national identity across model architectures. Young body types (Image 2) show different top coun... | https://arxiv.org/abs/2505.20326v1 |
Multi-Scale Manifold Alignment: A Unified Framework for Enhanced Explainability of Large Language Models Yukun Zhang The Chinese University Of Hongkong HongKong, China 215010026@link.cuhk.edu.cnQi Dong Fudan University Shanghai, China 19210980065@fudan.edu.cn Abstract Recent advances in Large Language Models (LLMs) hav... | https://arxiv.org/abs/2505.20333v1 |
We provide complete implementation including layer identifica- tion, cross-model adaptation, and optimiza- tion strategies, with applications in bias detec- tion and robustness enhancement. Compared to single-scale approaches, our frame- work offers a comprehensive view of LLM infor- mation organization, advancing inte... | https://arxiv.org/abs/2505.20333v1 |
deep layers integrate semantics. In this area, Seo et al. (2023) demonstrated through large-scale experiments that features extracted at different layers align closely with stages in the tra- ditional NLP pipeline, reflecting a shallow-to-deep processing logic. Singh and Daumé III (2023) sys- tematically evaluated hier... | https://arxiv.org/abs/2505.20333v1 |
aggregates and abstracts seman- tic information, giving rise to distinct strata in rep- resentation space. In the lens of information geometry , each hid- den state can be viewed as a point in a high- dimensional space, collectively forming a statis- tical manifold —the space of parameterized proba- bility distribution... | https://arxiv.org/abs/2505.20333v1 |
( MG). These boundaries can be identified by sharp changes in attention span, mutual information be- tween layers, and performance in targeted probing tasks. 4 Information-Theoretic Perspective. To math- ematically characterize information flow and se- mantic organization, we use mutual information between representati... | https://arxiv.org/abs/2505.20333v1 |
1.5 B parameters), BERT (a bidirectional en- coder with 340 M parameters), RoBERTa (an en- hanced encoder with 355 M parameters), and T5 (an encoder–decoder architecture with 11B param- eters). Experiments use 20,000 documents from the Brown and Reuters corpora, covering various genres and topics. Analyses integrate th... | https://arxiv.org/abs/2505.20333v1 |
no_curv ✓ ✓ × 0.1 0.1 0 only_geo ✓ × × 0.1 0 0 only_info × ✓ × 0 0.1 0 only_curv × × ✓ 0 0 0.01 Alignment Quality Results. We report KL di- vergence (distributional difference), mutual infor- mation, and distance correlation (geometry preser- vation) in Table 4. Geometric alignment is crucial for structure preservation... | https://arxiv.org/abs/2505.20333v1 |
51 39 0.75 1.07 1.00 1.00 geo-0.7 52 45 0.89 1.09 1.00 1.00 geo-0.8 51 48 0.87 0.84 1.00 1.00 geo-0.9 48 42 1.11 0.79 1.00 1.00 geo-1 70 43 0.92 0.87 1.00 1.00 (b) BERT Group KL g→mKLm→lMIg→m MIm→lDCg→mDCm→l baseline 403 3840 0.06 0.13 0.87 0.82 full-msma 0.51 1.29 2.89 2.64 1.00 1.00 no-curv 0.83 1.04 2.79 2.63 1.00 1... | https://arxiv.org/abs/2505.20333v1 |
that effect sizes some- times attenuate or behave non-linearly over long generation sequences, a dynamic phenomenon not fully captured by the current theory. Finally, while we evaluated alignment quality using KL divergence, mutual information, and distance-based metrics, these measures may not fully reflect the richne... | https://arxiv.org/abs/2505.20333v1 |
Geiger, Noah Goodman, and Christopher Potts. Can we understand transformer language models with causal abstraction? In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics , pages 15121–15136, 2023. Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, et al. PaLM: Scaling language modeli... | https://arxiv.org/abs/2505.20333v1 |
Systems , volume 35, 2022. Aaditya Singh Mohankumar, Blair Bilodeau, Margarita Vald, et al. How language model activations predict few-shot performance. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics , pages 6335–6353, 2023. OpenAI. GPT-4 technical report. arXiv preprint arXi... | https://arxiv.org/abs/2505.20333v1 |
Experimental Setup and Analysis for Semantic-Scale Identification A.1 Experimental Design Research Questions: We test three central hy- potheses of the Multi-Scale Manifold Alignment (MSMA) theory: (1) Do Transformer layers form identifiable local/intermediate/global seman- tic scales? (2) How do architecture and pre-t... | https://arxiv.org/abs/2505.20333v1 |
jumps are significant ( Z >2.0,p <0.01). Fig. 3(a): Layerwise MI, quantifying shared in- formation. BERT’s MI matrix forms three modules {0–4,5–8,9–12}, with within-module MI ∼40% higher than between-module MI. RoBERTa/T5 are similar; GPT-2’s MI estimates are noisier but con- (a) Mutual Information across models. (b) K... | https://arxiv.org/abs/2505.20333v1 |
and consistent local-to-global hierarchy. These convergent findings validate MSMA as an explanatory and predictive framework for Trans- former language generation. A.6 MSMA Method Implementation Details We detail implementation and hyperparameters for multi-scale manifold alignment. The process is multi-stage: first, s... | https://arxiv.org/abs/2505.20333v1 |
global, MG Assumption B.1.2 (Hierarchical Information Flow) .Information primarily flows ML→ MI→ M G, with local computation at each layer, consistent with residual-based Transformer design and confirmed experimentally. Assumption B.1.3 (Conditional Independence) . Given hG, intermediate representation hIis condi- tion... | https://arxiv.org/abs/2505.20333v1 |
smoothness and bounds total alignment distortion by control- ling the maximum curvature Kmaxviaλcurv. Proof sketch. By Rauch comparison, for points p, q∈ M with geodesic γ, d(f(p), f(q))≤d(p, q) expZ γK(s)ds . Cauchy-Schwarz gives |Z γK(s)ds| ≤L1/2Z MK2dV1/2 where Lis geodesic length. Thus, minimizing Lcurvtightens... | https://arxiv.org/abs/2505.20333v1 |
λcurv is crucial. Over-regularization may cause un- derfitting, under-regularization may not im- prove stability. Empirically, we tune this via validation. Future work may relax these assumptions or extend the theory to richer dependency structures. Experimental Correspondence Our theoretical predictions closely match ... | https://arxiv.org/abs/2505.20333v1 |
methods are used in MSMA, each targeting a different aspect of alignment: Geometric Alignment. Enforces structural con- sistency by minimizing the Euclidean dis- tance between representations at different scales. For global-to-intermediate mapping fGIand intermediate-to-local mapping fIL: Lgeo=∥fGI(hG)−hI∥2+∥fIL(hI)−hL... | https://arxiv.org/abs/2505.20333v1 |
arXiv:2505.20334v1 [cs.CL] 24 May 2025Lookahead Q-Cache: Achieving More Consistent KV Cache Eviction via Pseudo Query Yixuan Wang†Shiyu Ji†Yijun Liu Yuzhuang Xu Yang Xu Qingfu Zhu Wanxiang Che* Research Center for Social Computing and Interactive Robotics, Harbin Institute of Technology, China {yixuanwang,car}@ir.hit.e... | https://arxiv.org/abs/2505.20334v1 |
al., 2024; Feng et al., 2024) aim to allocate finer-grained budgets across layers or at- 1 tention heads to achieve higher compression rates. Although existing methods have partially allevi- ated the KV cache overhead in long-context scenar- ios, several challenges remain to be addressed. One key issue is the inconsist... | https://arxiv.org/abs/2505.20334v1 |
the generation phase. By maintaining a local window to guide KVCache eviction, lower performance degradation is achieved. Due to conflicts between obtaining attention scores and existing acceleration techniques (such as FlashAttention (Dao et al., 2022; Dao, 2024)), some methods focus on identifying alternative im- por... | https://arxiv.org/abs/2505.20334v1 |
for the 128 112 96 80 64 48 32 16 016 32 48 64 80 96112 1287 Strating Index of Observation Window0.300.350.400.450.50Recall with lookahead queries involvedThe Impact of the Observation Window Position on Recall FullKV SnapKV StreamingLLM Avg Input T oken Len: 12595.7 Avg Output T oken Len: 402.83 Total Samples Num: 32l... | https://arxiv.org/abs/2505.20334v1 |
the consistency between the compression and inference stages can be im- proved, leading to substantial gains for existing KV-Cache eviction strategies. Notably, although the golden query is not accessible during infer- ence, some pseudo queries obtained through evic- tion strategies can still achieve strong performance... | https://arxiv.org/abs/2505.20334v1 |
12.20 20.59 22.51 22.03 39.00 82.33 40.64 3.06 80.56 51.17 48.46 34.69 SnapKV 19.71 21.13 42.75 36.45 22.36 15.76 19.05 21.81 21.36 47.50 84.15 40.29 2.41 68.26 52.26 48.75 35.25 PyramidKV 21.98 22.78 43.78 32.30 22.31 15.81 20.41 21.82 21.23 66.00 83.51 39.83 2.99 65.81 51.61 46.42 36.16 LAQ 24.94 27.77 45.43 40.35 25... | https://arxiv.org/abs/2505.20334v1 |
24.82 23.15 24.61 68.00 92.33 42.16 7.83 96.86 64.74 53.60 38.87 PyramidKV 28.38 11.59 25.02 20.06 18.80 10.64 25.73 24.03 25.01 70.00 92.22 41.73 8.47 96.42 63.44 51.02 38.29 LAQ 30.89 14.04 25.86 26.00 23.19 17.73 27.07 24.01 25.29 72.50 92.25 43.13 6.96 96.97 64.87 52.58 40.21 LAQ++ 29.64 13.22 26.79 27.58 23.49 18.... | https://arxiv.org/abs/2505.20334v1 |
budgets. BudgetMethod FullKV H2O SnapKV PyraKV LAQ++ Mistral-7B-v0.2-Instruct 64 100 43.2 60.7 84.3 99.3 96 100 54.1 71.5 88.2 99.6 128 100 59.2 76.2 88.1 99.2 Llama3.1-8B-Instruct 64 100 35.4 68.4 80.7 99.8 96 100 42.1 72.0 85.9 100 128 100 46.1 73.1 89.9 100 Qwen2.5-7B-Instruct 64 100 46.7 72.8 69.3 85.1 96 100 50.2 ... | https://arxiv.org/abs/2505.20334v1 |
1 step w. snapkv 8 steps w. streaming 8 steps w. snapkv1 step w. fullkv 8 steps w. fullkv w.o. lookahead(a) The impact of Q-Cache quality on performance. 12 4 8 16 Q-Cache Length38.038.539.039.540.040.541.041.5PerformanceThe Impact of Q-Cache Length on Performance LA-QC++(Mistral) LA-QC(Mistral) LA-QC++(Llama) LA-QC(Ll... | https://arxiv.org/abs/2505.20334v1 |
method still incurs additional latency for lookahead operations. To balance task performance and efficiency, we conduct a stage- wise latency analysis of the proposed LAQ (LookA- head Q-Cache) framework. As shown in Figure 6, we evaluate the latency distributions under both 2-step and 8-step configurations. We evaluate... | https://arxiv.org/abs/2505.20334v1 |
leverages this feature by employing a pre-generated Q-Cache to achieve more consistent KV cache eviction. 7 Conclusion In this paper, we investigate the inconsistency between the Query entries of input and output. Based on our observational findings, we propose the Lookahead Q-Cache, which performs KV cache eviction by... | https://arxiv.org/abs/2505.20334v1 |
adaptive budget allocation for efficient llm inference. ArXiv preprint , abs/2407.11550. Suyu Ge, Yunan Zhang, Liyuan Liu, Minjia Zhang, Jiawei Han, and Jianfeng Gao. 2024. Model tells you what to discard: Adaptive KV cache compression for llms. In The Twelfth International Conference on Learning Representations, ICLR ... | https://arxiv.org/abs/2505.20334v1 |
Information Processing Systems 38: Annual Conference on Neu- ral Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024 . Jiaheng Liu, Dawei Zhu, Zhiqi Bai, Yancheng He, Huanxuan Liao, Haoran Que, Zekun Wang, Chenchen Zhang, Ge Zhang, Jiebin Zhang, and 1 oth- ers. 2025a. A comp... | https://arxiv.org/abs/2505.20334v1 |
Jingren Zhou, and 1 others. 2025. Qwen2. 5-1m technical report. ArXiv preprint , abs/2501.15383. 10 Jiayi Yuan, Hongyi Liu, Shaochen Zhong, Yu-Neng Chuang, Songchen Li, Guanchu Wang, Duy Le, Hongye Jin, Vipin Chaudhary, Zhaozhuo Xu, and 1 others. 2024. Kv cache compression, but what must we give in return? a comprehens... | https://arxiv.org/abs/2505.20334v1 |
22.42 13.82 22.35 22.54 23.12 40.50 83.78 40.73 3.51 85.85 53.18 49.95 36.22 SnapKV 22.44 24.07 48.01 38.66 22.66 15.59 21.83 23.23 22.94 61.50 85.45 41.32 3.13 85.79 55.11 51.73 38.97 PyramidKV 21.69 25.18 47.61 38.77 26.12 15.23 22.52 22.52 22.59 68.00 84.27 42.10 3.43 76.60 53.08 48.40 38.63 LAQ(2) 25.50 26.54 46.12... | https://arxiv.org/abs/2505.20334v1 |
48.11 36.28 LAQ(2) 28.23 11.51 24.66 21.57 22.81 15.15 24.81 23.18 23.83 72.50 92.25 43.13 8.52 94.30 62.01 49.93 38.65 LAQ(8) 28.86 12.40 26.44 21.80 20.91 15.77 25.83 23.30 24.26 72.50 93.08 42.57 7.53 95.80 63.51 51.09 39.10 LAQ(2)++ 30.49 11.95 25.79 22.99 21.36 15.35 24.40 23.38 23.53 72.50 91.97 42.45 7.50 94.49 ... | https://arxiv.org/abs/2505.20334v1 |
47.04 57.01 44.78 27.35 23.86 21.59 19.69 36.00 87.81 38.98 8.50 98.50 5.35 7.51 36.25 LAQ(8) 17.62 41.16 50.84 57.16 46.48 29.37 24.23 22.17 20.30 38.50 88.10 39.12 8.50 99.00 6.14 9.83 37.41 LAQ(2)++ 16.79 38.96 49.78 57.13 45.59 27.95 23.27 21.33 19.92 39.00 87.51 37.82 8.50 98.50 5.40 8.07 36.60 LAQ(8)++ 16.76 41.0... | https://arxiv.org/abs/2505.20334v1 |
23.80 70.50 90.52 40.39 5.81 70.00 60.45 56.17 40.36 PyramidKV 24.83 23.32 35.19 43.29 31.87 20.55 23.41 22.80 24.29 71.50 90.61 40.81 5.91 69.50 59.60 54.71 40.14 LAQ(1)++ 24.96 25.91 37.07 43.19 37.09 21.49 23.10 22.61 24.26 73.50 90.64 41.61 5.43 69.70 60.89 57.83 41.21 LAQ(8)++ 25.53 27.59 37.99 43.82 36.52 21.55 2... | https://arxiv.org/abs/2505.20334v1 |
arXiv:2505.20335v1 [cs.CL] 24 May 2025Language Model Distillation: A Temporal Difference Imitation Learning Perspective Zishun Yu∗ Department of Computer Science The University of Illinois, Chicago Chicago, IL 60607 zyu32@uic.eduShangzhe Li∗ Department of Computer Science The University of North Carolina at Chapel Hill... | https://arxiv.org/abs/2505.20335v1 |
the broader IL literature. Casting distillation as an IL problem is not new per se; indeed, many existing distillation works [ 8, 18–20] perform BC. Before reviewing these works through the lens of IL, we highlight two key distinctions that help characterize the distillation setup: (i) offline vs. online: This refers t... | https://arxiv.org/abs/2505.20335v1 |
during TD learning, it may be 2 sufficient to consider only a small subset of candidate actions, reducing the effective action space fromVto a much smaller set. 2 Preliminaries MDP notations. We define the conventional MDP tuple M= (S,A,T, r, γ)as follows. Let w≤t= (w1, w2, . . . , w t)∈ W t:=Vtbe a sequence of tokens.... | https://arxiv.org/abs/2505.20335v1 |
operator, defined as (TπQ)(s, a) =Q(s, a)−γVπ(s′); and VQ:= logP aexpQ(s, a). The core idea is can be summarized as: (i) The original GAIL objective minπmax rL(π, r)can be equivalently swapped as max rminπL(π, r)by strong duality, given proper regularization and feasible domains; (ii) For a fixed policy π, the saddle-p... | https://arxiv.org/abs/2505.20335v1 |
MDP M. Definition 2 (top-pMDP and ¯Q).Given a MDP M= (S,A,P, r, γ)and a teacher policy π⋆, its top-pcounterpart is Mp= (S,A⋆ p,P, r, γ). To make the notation clearer, we use Q:S × A → R and¯Q:S × A⋆ p→Rto denote Q-functions live in MandMp, respectively. With this definition, we aim to address an important question: Wha... | https://arxiv.org/abs/2505.20335v1 |
Mp, is still a good policy. While projection is trivial in the tabular case, it is in general difficult to project a parameterized π⋆onto the action set A⋆ p. Instead, it is natural to implement an (inverse) RL algorithm in Mp. However, when implementing conventional IRL algorithms in Mp, these algorithms typically see... | https://arxiv.org/abs/2505.20335v1 |
pQ)(s, a):=Q(s, a),ifa∈ A⋆ p −∞, otherwise(14) Applying F⋆ pleads to: max QJ⋆(Q):=Eρ⋆[ϕ((F⋆ pQ)(s, a)−γVπQ(s′))]−(1−γ)Es0[VπQ(s0)]. Policy projection. We are also ready to handle the policy πQthrough projection. By definition πQ(a|s) = exp Q(s, a)/P aexpQ(s, a), as a result its projected counterpart is (PROJ pπQ) = ex... | https://arxiv.org/abs/2505.20335v1 |
of Gu et al. [19] for further details. We clip the Q values using a minimum value Qmin=−10for numerical stability. 5 Experiments In this section, we call our method as Bellman Distill (BD) for brevity. 5.1 Experiment Setup We take instruction-following [44] as the conditional text generation task, where models are trai... | https://arxiv.org/abs/2505.20335v1 |
detailed description on data generation and evaluation can be found in Appendix C. 5.2 Experimental Results Figure 2: Comparison of win rates against KD, SeqKD, and MiniLLM baselines. We evaluate using GPT-4o-mini [ 57] as the judging oracle, with Qwen-2.5 (3B) as the teacher model and a smaller (0.5B) model as the stu... | https://arxiv.org/abs/2505.20335v1 |
In general, the choice between online and offline training is a trade-off between computational cost and performance. While auto-regressive online generation is compu- tationally expensive, offline training allows for the reuse of pre-collected datasets, making it more efficient. However, online training typically yiel... | https://arxiv.org/abs/2505.20335v1 |
Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783 , 2024. [4]Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement... | https://arxiv.org/abs/2505.20335v1 |
the fourteenth interna- tional conference on artificial intelligence and statistics , pages 627–635. JMLR Workshop and Conference Proceedings, 2011. [23] Stephen Tu, Alexander Robey, Tingnan Zhang, and Nikolai Matni. On the sample complexity of stability constrained imitation learning. In Learning for Dynamics and Cont... | https://arxiv.org/abs/2505.20335v1 |
L Littman. An alternative softmax operator for reinforcement learning. In International Conference on Machine Learning , pages 243–252. PMLR, 2017. [42] Erfan Miahi, Revan MacQueen, Alex Ayoub, Abbas Masoumzadeh, and Martha White. Resmax: An alternative soft-greedy operator for reinforcement learning. Transactions on M... | https://arxiv.org/abs/2505.20335v1 |
[59] Donghun Lee, Srivatsan Srinivasan, and Finale Doshi-Velez. Truly batch apprenticeship learning with deep successor features. arXiv preprint arXiv:1903.10077 , 2019. [60] Daniel Jarrett, Ioana Bica, and Mihaela van der Schaar. Strictly batch imitation learning by energy-based distribution matching. Advances in Neur... | https://arxiv.org/abs/2505.20335v1 |
we also adopt a two-phase training strategy for our approach: •Phase 1: We fine-tune the student model on the instruction-response training set Dto obtain a strong initialization for subsequent offline BD training. This fine-tuning is performed for 3 epochs using the optimal learning rate and batch size from the corres... | https://arxiv.org/abs/2505.20335v1 |
the reviews. You can look at these to see what type of reviews are coming from local residents versus tourists. There are also metrics to look at from a business perspective like average waiting time, and availability of menu changes. All of these can be useful to understand how popular a restaurant is and why that inf... | https://arxiv.org/abs/2505.20335v1 |
MOSLIM:A LIGN WITH DIVERSE PREFERENCES IN PROMPTS THROUGH REWARD CLASSIFICATION Yu Zhang∗ Alibaba GroupWanli Jiang† Alibaba GroupZhengyu Yang‡ Alibaba Group ABSTRACT The multi-objective alignment of Large Language Models (LLMs) is essential for ensuring foundational models conform to diverse human preferences. Cur- ren... | https://arxiv.org/abs/2505.20336v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.