text string | source string |
|---|---|
following format. Additionally, label the middle span as one of the five categories listed above. # Example: example_number ## Prefix ```python # your code here ``` ## Suffix ```python # your code here ``` ## Middle ```python # your code here ``` ## Label Figure 2: Prompt template to generate fill-in-the-middle trainin... | https://arxiv.org/abs/2505.18789v1 |
r e n t _ l i n e and not c u r r e n t _ l i n e . s t a r t s w i t h ( " # " ) : 20 l += 1 21 i fl == num_lines : 22 break 23 c o m p l e t i o n = " \ n " . j o i n ( c o m p l e t i o n _ l i n e s ) 24 return c o m p l e t i o n 25 26 27def remove_overlap_prefix_middle ( p r e f i x , middle ) : 28 p r e f i x _ ... | https://arxiv.org/abs/2505.18789v1 |
ALPS: Attention Localization and Pruning Strategy for Efficient Alignment of Large Language Models Hao Chen♠, Haoze Li♠, Zhiqing Xiao♠, Lirong Gao♠, Qi Zhang♠ Xiaomeng Hu♠,Ningtao Wang♣,Xing Fu♣,Junbo Zhao♠ ♠Zhejiang University♣Ant Group {h.c.chen, j.zhao}@zju.edu.cn Abstract Aligning general-purpose large language mod... | https://arxiv.org/abs/2505.18799v2 |
2: Preliminary results comparing ALPS against baselines across different LLM scales ( Left), and downstream tasks ( Right ).ALPS consistently outperforms other methods, demonstrating its efficiency in diverse settings. look the intrinsic functionality of the model weight parameters, which inherently encode task-relevan... | https://arxiv.org/abs/2505.18799v2 |
sensitive attention heads. 2.2 Attention Localization and Pruning Recent studies have investigated attention mecha- nisms at various granularities to improve the effi- ciency and interpretability of LLM, including neu- rons, layers, and heads (Geiger et al., 2021; Gurnee et al., 2023; Zou et al., 2023; Zheng et al., 20... | https://arxiv.org/abs/2505.18799v2 |
on specific tasks. Inspired by mech- anistic interpretability studies (V oita et al., 2019; Zhao et al., 2024; Lindner et al., 2023), we define these task head parameters as those whose ablation results in a significant degradation of task-specific performance. Formally, given a downstream task Twith evaluation metric ... | https://arxiv.org/abs/2505.18799v2 |
of task-sensitive heads HTwith Top-K SPAD to retain. During task fine-tuning, the remaining heads will be pruned by freezing gradients of these heads, which masks their parameter updates to 0. Specifically, for all attention heads, only parame- ters associated with HTare updated: ∇θhL=( ∇θhLifh∈ H S 0 otherwise, (8) ... | https://arxiv.org/abs/2505.18799v2 |
(Llama-3.2-1B, Llama-3.2-3B, and Llama-3.18B), all pre-trained with GQA for im- proved efficiency. Each model shares key and value projections across 8 heads ( g=8), while queries re- main head-specific. Thus, when computing sPAD, each key-value pair corresponds to 8 query heads. Evaluation. We use lm-eval (Gao et al.,... | https://arxiv.org/abs/2505.18799v2 |
26.47 62.72 22.15 71.15 64.55 62.88 52.23 51.62 ↑0.46 10% 51.33 27.13 64.13 22.48 72.68 65.03 63.67 52.88 52.41 ↑1.25 0% 40.41 26.16 64.06 15.48 70.72 64.88 63.27 52.55 49.69 ↓1.47 Table 3: Ablation study on attention head selection ratio for ALPS . The best results are highlighted in bold .↓↑ indicates change relative... | https://arxiv.org/abs/2505.18799v2 |
our proposed metric sPADconsistently achieves the highest average scores, indicating its effectiveness in capturing the most critical distribution shifts be- tween the base and task fine-tuned models.Head Ratio Ablation. The ablation study on the proportion of attention heads selected for our method ranging from 0%to10... | https://arxiv.org/abs/2505.18799v2 |
in Sec- tion 4.1. Table 4 and Figure 6 show that our method consistently outperforms all baselines, indicating the transferability of the task-sensitive attention heads identified by ALPS that can be reused effec-Model Method MMLU ARC-C Avg. Llama-3.2-1Bvanilla 32.2 32.8 32.5 w/ full 28.14 26.95 27.55 w/o attn 27.80 23... | https://arxiv.org/abs/2505.18799v2 |
Michiel de Jong, Yury Zemlyanskiy, Federico Lebrón, and Sumit Sanghai. 2023. Gqa: Training generalized multi-query trans- former models from multi-head checkpoints. arXiv preprint arXiv:2305.13245 . Martin Arjovsky, Soumith Chintala, and Léon Bottou. 2017. Wasserstein gan. Preprint , arXiv:1701.07875. Jacob Austin, Aug... | https://arxiv.org/abs/2505.18799v2 |
Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Yu Wu, YK Li, et al. 2024. Deepseek-coder: When the large language model meets programming– the rise of code intelligence. arXiv preprint arXiv:2401.14196 . Wes Gurnee, Neel Nanda, Matthew Pauly, Kather- ine Harvey, Dmitrii Troitskii, a... | https://arxiv.org/abs/2505.18799v2 |
processing systems , 32. OpenAI. 2025. Openai o3-mini. David Rein, Betty Li Hou, Asa Cooper Stickland, Jack- son Petty, Richard Yuanzhe Pang, Julien Dirani, Ju- lian Michael, and Samuel R Bowman. 2023. Gpqa: A graduate-level google-proof q&a benchmark. arXiv preprint arXiv:2311.12022 . Baptiste Roziere, Jonas Gehring, ... | https://arxiv.org/abs/2505.18799v2 |
explains long-context factuality. arXiv preprint arXiv:2404.15574 .Can Xu, Qingfeng Sun, Kai Zheng, Xiubo Geng, Pu Zhao, Jiazhan Feng, Chongyang Tao, and Daxin Jiang. 2023. Wizardlm: Empowering large lan- guage models to follow complex instructions. arXiv preprint arXiv:2304.12244 . An Yang, Baosong Yang, Beichen Zhang... | https://arxiv.org/abs/2505.18799v2 |
3D visual- ization of the data distribution for UltraChat, Math- Instruct, and Magicoder, as embedded by Llama- 3.1-8B model. Notably, the code and math datasets exhibit some overlaps, whereas the general dataset displays a distinct distribution, reflecting domain- specific characteristics. 40 20 0 20 40 6020 10 010201... | https://arxiv.org/abs/2505.18799v2 |
Humaneval (Chen et al., 2021) Code 0 pass@1 Humaneval+ (Liu et al., 2023) Code 0 pass@1 MBPP (Austin et al., 2021) Code 0 pass@1 MBPP+ (Liu et al., 2023) Code 0 pass@1 MMLU (Hendrycks et al., 2020) General 5 macro_avg/acc ARC-C (Clark et al., 2018) Reasoning 0 acc Table 7: Overview of benchmarks utilized in our experim... | https://arxiv.org/abs/2505.18799v2 |
specific adjustments. The W1 distance between Ph TandPh Bis defined as: W1 Ph B,Ph T = inf γ∈Γ(Ph B,Ph T)E(x,y)∼γ[∥x−y∥]. (11) For the special case where Ph Tis simply a trans- lated version of Ph B, i.e.,Ph T≈Ph B(x−∆), lead- ing to: W1 Ph B,Ph T ≈ ∥∆∥. (12) Thus, in the context of ALPS, sPAD h= W1 Ph B,Ph T dir... | https://arxiv.org/abs/2505.18799v2 |
↑1.85 LoRA 24.01 17.98 40.57 27.52 ↑2.40 ALPS 24.65 18.70 40.24 27.86 ↑2.74 Llama-3.1-8Bw/ full 27.21 28.76 42.34 32.77 w/o attn 18.29 25.77 46.11 30.06 ↓3.38 Random 27.93 32.96 40.43 33.77 ↑1.00 LC 26.52 30.28 41.28 32.69 ↓0.08 LoRA 27.87 31.02 42.27 33.72 ↑0.95 ALPS 28.21 31.31 44.57 34.67 ↑1.90 Table 12: Full table ... | https://arxiv.org/abs/2505.18799v2 |
arXiv:2505.18822v1 [cs.AI] 24 May 2025AdaCtrl: Towards Adaptive and Controllable Reasoning via Difficulty-Aware Budgeting Shijue Huangα∗, Hongru Wangβ∗, Wanjun Zhong∗, Zhaochen Suα, Jiazhan Fengσ, Bowen Caoβ,Yi R. (May) Fungα† αHong Kong University of Science and Technology βThe Chinese University of Hong Kong,σPeking ... | https://arxiv.org/abs/2505.18822v1 |
as the response demonstrated in right part of Figure 2. Although such thorough reasoning is beneficial for complex queries, it results in excessive latency and resource consumption for simpler tasks, negatively impacting user experience. To address this, recent researches have explored ways to improve reasoning efficie... | https://arxiv.org/abs/2505.18822v1 |
respective datasets. Moreover, AdaCtrl offers enhanced controllability via explicit length-trigger tags. When instructed to operate in the easy reasoning mode, the model reduces response lengths by 90.22% and 94.31% (from 16k to 1k tokens) on the AIME2025 and AIME2024 datasets, respectively. Conversely, under the hard ... | https://arxiv.org/abs/2505.18822v1 |
et al., 2024]. Another SFT direction compresses existing reasoning chains. Kang et al. [2024] employ GPT-4 [Achiam et al., 2023] as a compressor then fine-tune a model on these long-to-short CoT mappings. LMskip [Liu et al., 2024] induces step-skipping behavior under step constraints. SPIRIT-FT [Cui et al., 2025] ident... | https://arxiv.org/abs/2505.18822v1 |
supply the preferred length-trigger tag as an initial input token. 3.2 Cold-Start Fine-Tuning This stage primarily focuses on equipping models with the ability to adhere to output formats that include length-triggered tags and to control response length accordingly. To curate suitable training data for this purpose, we... | https://arxiv.org/abs/2505.18822v1 |
−1.0ticannot be found in yi(4) Difficulty-aware Length Reward. Different from previous related works that encourage the model to generate concise responses for all problems, we hope to encourage such behavior only for easy problems and maintain long thinking capabilities for better tackling hard problems. To prevent un... | https://arxiv.org/abs/2505.18822v1 |
divergence. The policy model is optimized using the AdamW optimizer with a learning rate of 1e-6, a batch size of 256, and a micro-batch size of 32. And we set the value of αandβboth as 0.5 during training. During the rollout phase, 16 responses are sampled per prompt, and the maximum generation length is set to 24K to... | https://arxiv.org/abs/2505.18822v1 |
and MATH500, respectively, and achieves competitive results on GSM8K. Corresponding reductions in response length for this comparison are 11.21%, 13.95%, 45.13%, and 80.99%. These results suggest that AdaCtrl is capable of adaptively allocating reasoning budget based on problem difficulty, thereby achieving an effectiv... | https://arxiv.org/abs/2505.18822v1 |
20% 40% 60% 80% 100%GSM8KMATH500AIME 2025AIME 2024Easy Hard 362.881117.711838.063517.546256.77 010002000300040005000600070008000 Level 1 Level 2 Level 3 Level 4 Level 5 (a) Proportion of length -trigger tags. (b) Length of different levels in MA TH 500. Figure 4: (a) The proportion of length-trigger tags across differe... | https://arxiv.org/abs/2505.18822v1 |
we propose an adaptive and controllable reasoning framework designed to mitigate the problem of overthinking while granting users explicit control over computational resources. To this end, we introduce AdaCtrl, a model that supports both dynamic reasoning budget allocation 9 and user-directed budget adjustments. Our a... | https://arxiv.org/abs/2505.18822v1 |
Mingming Li, Ning Tian, Panpan Huang, Peng Zhang, Qiancheng Wang, Qinyu Chen, Qiushi Du, Ruiqi Ge, Ruisong Zhang, Ruizhe Pan, Runji Wang, R. J. Chen, R. L. Jin, Ruyi Chen, Shanghao Lu, Shangyan Zhou, Shanhuang Chen, Shengfeng Ye, Shiyu Wang, Shuiping Yu, Shunfeng Zhou, Shuting Pan, S. S. Li, Shuang Zhou, Shaoqing Wu, S... | https://arxiv.org/abs/2505.18822v1 |
Zhang, Fei Yin, Jiahua Dong, Zhiwei Li, Bao-Long Bi, Ling-Rui Mei, Junfeng Fang, Zhijiang Guo, Le Song, and Cheng-Lin Liu. From system 1 to system 2: A survey of reasoning large language models, 2025. URL https: //arxiv.org/abs/2502.17419 . Tengxiao Liu, Qipeng Guo, Xiangkun Hu, Cheng Jiayang, Yue Zhang, Xipeng Qiu, an... | https://arxiv.org/abs/2505.18822v1 |
Jieyun Huang, Shuming Shi, Wenjing Zhang, Jiangze Yan, Ning Wang, Kai Wang, and Shiguo Lian. Dast: Difficulty-adaptive slow-thinking for large reasoning models, 2025. URL https://arxiv.org/abs/2503.04472 . Guangming Sheng, Chi Zhang, Zilingfeng Ye, Xibin Wu, Wang Zhang, Ru Zhang, Yanghua Peng, Haibin Lin, and Chuan Wu.... | https://arxiv.org/abs/2505.18822v1 |
Jason Weston, and Ilia Kulikov. Distilling system 2 into system 1, 2024. URL https://arxiv.org/abs/2407.06023 . Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Weinan Dai, Tiantian Fan, Gaohong Liu, Lingjun Liu, Xin Liu, Haibin Lin, Zhiqi Lin, Bole Ma, Guangming Sheng, Yuxuan Tong, Chi Zhang, Mof... | https://arxiv.org/abs/2505.18822v1 |
arXiv:2505.18830v1 [cs.LG] 24 May 2025On the Effect of Negative Gradient in Group Relative Deep Reinforcement Optimization Wenlong Deng1,2, Yi Ren1, Muchen Li1, Danica J. Sutherland1, Xiaoxiao Li1,2†, Christos Thrampoulidis1† 1University of British Columbia,2Vector Institute †Corresponding author Abstract Reinforcement... | https://arxiv.org/abs/2505.18830v1 |
in, the likelihood of correct answers during training. To explain this phenomenon, we provide an analysis of GRPO’s update dynamics and identify that LLD stems from penalizing certain tokens in dispreferred responses. These updates inadvertently reduce the likelihood of correct answers due to shared structural or seman... | https://arxiv.org/abs/2505.18830v1 |
reward for response yi. The advantage of the i-th response is computed by normalizing the group- level rewards {ri}G i=1and is the same for each token k= 1, . . . ,|yi|. Concretely, ˆAi,k:=ri−µ σ, withµ=bE[{ri}G i=1]andσ=q cVar[{ri}G i=1]being the empirical average and standard deviation of 2 (a) Qwen-0.5B-Instruct (b)... | https://arxiv.org/abs/2505.18830v1 |
gradients in GRPO [ 3] on the likelihood of generating correct responses. Let each question xbe associated with Ggenerated samples, of which N+:=N+(x)give the correct answer so we call them positive and denote by y+ i, i∈[N+], and the rest N−L=G−N+are incorrect we call them negative and denote by y− j, j∈[N−]. Positive... | https://arxiv.org/abs/2505.18830v1 |
of those questions in the dataset that have partially correct (but eventually incorrect) responses or have correct responses. In Fig. 2, we present two representative examples from Qwen-0.5B, where the incorrect responses are partially correct and the corresponding correct responses show the smallest likelihood change.... | https://arxiv.org/abs/2505.18830v1 |
token unembedding matrix W∈R|V|×d, where Vis the vocabulary of tokens. 5 Assumption 4.3 (Unconstrained Features) Expressive (enough) neural networks can produce un- constrained embeddings hx∈Rdindependent of the architecture’s specific complexities [ 25,14, 30,17]. These embeddings are subsequently transformed into log... | https://arxiv.org/abs/2505.18830v1 |
2.5Math-1.5B, we use questions from the Math dataset and retain 47 valid questions. We then calculate the Top- Koverlap accuracy between two rankings of questions: one based on the ascending 1Theorem 4.4 can be seen as an extension of [ 17, Thm.3] for DPO to the more challenging GRPO setting with group-based structure.... | https://arxiv.org/abs/2505.18830v1 |
11:foreach retained token y− j,k′∈V− jdo 12: ˆA− j,k′,η←η·ˆA− j,k′ 13:end forAs shown in Corollary 4.5, tokens with high positive NTHR values tend to cause signifi- cant reductions in the likelihood of generated correct responses. To mitigate this adverse effect, we introduce a selective penalization strategy (detailed... | https://arxiv.org/abs/2505.18830v1 |
For this, we finetune models with different sizes using MATH dataset (levels 3–5) [ 6] and use greedy decoding to evaluate finetuned models on five math benchmarks: AIME24, AMC, MATH500, Minerva, and Olympiad (detailed implementation details in Appendix). We find that NTHR consistently boosts model performance across v... | https://arxiv.org/abs/2505.18830v1 |
Fei Wei, and Yong Wang. Gpg: A simple and strong reinforcement learning baseline for model reasoning. arXiv preprint arXiv:2504.02546 , 2025. [2]Wenlong Deng, Yize Zhao, Vala Vakilian, Minghui Chen, Xiaoxiao Li, and Christos Thram- poulidis. Dare the extreme: Revisiting delta-parameter pruning for fine-tuned models. ar... | https://arxiv.org/abs/2505.18830v1 |
Likelihood displacement in direct preference optimization. arXiv preprint arXiv:2410.08847 , 2024. [18] Yi Ren and Danica J Sutherland. Learning dynamics of llm finetuning. arXiv preprint arXiv:2407.10490 , 2024. [19] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimiz... | https://arxiv.org/abs/2505.18830v1 |
a binary reward system, assigning r= 1 to correct responses and r= 0 to incorrect ones, consistent with recent works [ 12,27]. It is notable that this also applies to its variants, e.g., DAPO [ 27], Dr.GRPO [12], GPG [1], etc. For a single question x, we simplify the expected loss E{yi}G i=1∼πθold(·|x)[Jx]in Eq. (1) by... | https://arxiv.org/abs/2505.18830v1 |
j,k′−πθ(t)(·|x,y− j,<k′)E . Specifically, (I)and(II)capture how token embeddings influence changes in likelihood, while (III) and(IV) reflect how the geometry of token unembeddings governs such changes. Our focus is on (I) and (II) because: •Token embeddings encapsulate the contribution of all network parameters exclud... | https://arxiv.org/abs/2505.18830v1 |
is essential for practical implementation. Output last-layer embeddings with old policy probabilities. Since GRPO relies on the old policy πoldto compute old probabilities for the clip(·))operation, we only need to output the last-layer embeddings during the old policy’s forward pass. This does not require any extra fo... | https://arxiv.org/abs/2505.18830v1 |
models, NTHR consistently outperforms GRPO for most of the training process. all settings outperform GRPO, indicating that reducing the influence of identified negative gradients is beneficial. Base model + Method AIME24 AMC MATH500 Minerva Olympiad Avg. GRPO 13.3 57.5 71.8 29.0 34.1 41.14 NTHR ( β=−∞) 13.3 50.0 71.8 3... | https://arxiv.org/abs/2505.18830v1 |
in a final answer that is wrong. Model & Data Problem & Answer Incorrect Response Qwen2.5math- 1.5B Math (order 1)Question : A portion of the graph of a quadratic function f(x)is shown. Letg(x) =−f(x),h(x) = f(−x). Ifais the number of points where the graphs of y= f(x)andy=g(x)intersect, andbis the number of points whe... | https://arxiv.org/abs/2505.18830v1 |
buttons in any order to open it. The combination given is 1,2,3,6,9, which is a set of five buttons. So, the question is asking about how many additional combinations would be possible if we redesign the locks to allow sets of as many as nine buttons or as few as one button. I think I need to figure out the total numbe... | https://arxiv.org/abs/2505.18830v1 |
we observe that core steps and correct answer 8are with high NTHR. 18 Model & Data Problem & Answer Incorrect Response Qwen2.5math- 1.5B MathQuestion : A portion of the graph of a quadratic function f(x)is shown. Letg(x) =−f(x),h(x) = f(−x). Ifais the number of points where the graphs of y= f(x)andy=g(x)intersect, andb... | https://arxiv.org/abs/2505.18830v1 |
roots n= 3 andn= 8 divide the number line into three intervals: (−∞,3],[3,8], and[8,∞). Let ’s test a point in each interval to determine where the inequality holds true. 1. For n in (−∞,3), let ’s test n= 0:02−11(0)+24 = 24 >0. 2. For n in (3,8), let ’s test n= 5:52−11(5) + 24 = 25 −55 + 24 = −4<0. 3. For n in(8,∞), l... | https://arxiv.org/abs/2505.18830v1 |
arXiv:2505.18842v1 [cs.CL] 24 May 2025Don’t Look Only Once: Towards Multimodal Interactive Reasoning with Selective Visual Revisitation Jiwan Chung∗ ♠†Junhyeok Kim♠†Siyeol Kim♠ Jaeyoung Lee♣Minsoo Kim♠Youngjae Yu♠ ♠Yonsei University ♣Seoul National University jiwan.chung.research@gmail.com Abstract We present v1, a lig... | https://arxiv.org/abs/2505.18842v1 |
7 (B) 9 (C) 12 (D) 15 Input MLLM Stan d ar d A ppr oach : T ext -B ase d Reasoning P ointing P atches C op y T ok en DistributionT ext T ok en Distribution...... rightsideanglecosinetriangle I ’v e got triangle diagram here . ... according to , , and ... ... we can notice that triangle RST is isosceles, with RS=RT. Thu... | https://arxiv.org/abs/2505.18842v1 |
visual grounding and iterative reference to localized regions. These results suggest that enabling revisitation of visual input during inference can significantly enhance multimodal reasoning capabilities. 2 Related Work 2.1 Reasoning in Large Language Models Reasoning in text-only large language models. Recent advance... | https://arxiv.org/abs/2505.18842v1 |
whether a model attends to the correct part of the image. We conduct two preliminary analyses using the TVC-7B model [ 14] on the RefCOCO testA split. In both, we measure the attention weights between the most recently generated token and all image tokens, examining layers 2, 14, and 27 of the 28-layer transformer to c... | https://arxiv.org/abs/2505.18842v1 |
mechanism we examine was first introduced by the pointer-generator network [ 41] in text summarization research. In the pointer-generator network, the input context sequence calso consists of discrete tokens within the vocabulary space V, unlike our setup. The model dynamically mixes two distributions at each decoding ... | https://arxiv.org/abs/2505.18842v1 |
output spaces and do not require interpolation. During inference, if the model selects an index in V, the next token xtis emitted as the corresponding vocabulary token. If the model selects an index k∈C, the token is represented as a pointer xt=⟨ptr:ck⟩. On the subsequent decoding step, the input embedding at position ... | https://arxiv.org/abs/2505.18842v1 |
[ 46,47]. Also, multimodal reasoning often requires grounding abstract or geometric visual cues ( e.g.angle ABC), for which existing models are not trained. To exploit the implicit visual grounding behavior in MLLMs, we build on Qwen2.5-VL [ 9], a grounding-capable model that can localize visual entities via bounding b... | https://arxiv.org/abs/2505.18842v1 |
63.1 - - 39.6 51.4 - TVC [14] 7B ✓ 68.1 - 22.7 38.9 53.5 30.8 TVC [14] 72B ✓ 72.2 - 41.9 48.8 60.5 45.4 QVQ-72B-preview [44] 72B ✓ 71.4 35.9 - 41.5 49.6 - Ours w/o Pointing 7B ✓ 60.0 25.3 23.7 33.6 39.6 28.7 Ours 7B ✓ 68.6 34.5 28.1 48.6 50.6 38.4 is particularly effective given the structure of our task: the pretraine... | https://arxiv.org/abs/2505.18842v1 |
connect t o D . 2. Hexagon 2 : It allows r oads fr om A t o B but does not connect t o E. 3. Hexagon 3 : It allows r oads fr om A t o E but does not connect t o B. 4. Hexagon 4 : It allows r oads fr om A t o B and E but does not connect t o D . 5. Hexagon 5 : It allows r oads fr om A t o B and E but does not connect t ... | https://arxiv.org/abs/2505.18842v1 |
egion> [ 0.00, 0.00, 0. 6 7, 0.88 ] </r egion>" ]) ther e ar e six houses color ed diff er ently . detect(quer y="hexagon option 1", objects= [ " <r egion> [ 0.58, 0.18, 0.71, 0.47 ] " ]) This hexagon allows A-B and B-E but not B-D . detect(quer y="hexagon option 5", objects= [ " <r egion> [ 0. 6 7, 0.53, 0.79, 0.82 ] ... | https://arxiv.org/abs/2505.18842v1 |
invalid candidates and outputs the wrong answer. These examples highlight how proactive visual 8 0 200 400 600 800 1000 Distance From First Copy0.00.10.20.30.4Attention ScoreLayer 2 Copy Attention Input Attention Copy Sequence 0 200 400 600 800 1000 Distance From First CopyLayer 14 0 200 400 600 800 1000 Distance From ... | https://arxiv.org/abs/2505.18842v1 |
Psychology , 87:273–287, 2017. [4]Kozhevnikov, M., M. Hegarty, R. E. M. and. Revising the visualizer-verbalizer dimension: Evidence for two types of visualizers. Cognition and Instruction , 20(1):47–77, 2002. [5] OpenAI, :, A. Jaech, et al. Openai o1 system card, 2024. [6]DeepSeek-AI, D. Guo, D. Yang, et al. Deepseek-r... | https://arxiv.org/abs/2505.18842v1 |
Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2) . 2021. [27] Jain, N., K. Han, A. Gu, et al. Livecodebench: Holistic and contamination free evaluation of large language models for code. In The Thirteenth International Conference on Learning Representations . 2025. [28] Rein, ... | https://arxiv.org/abs/2505.18842v1 |
al. Florence-2: Advancing a unified representation for a variety of vision tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 4818–4829. 2024. [48] Zhang, J., M. Khayatkhoei, P. Chhikara, et al. MLLMs know where to look: Training-free perception of small visual details w... | https://arxiv.org/abs/2505.18842v1 |
allel t o the line CD , ... , , ... in triangle CDE , we ha v e t o measur e the angle CED ... detect(quer y = " line AB ", objects = [ " <obj1>< | obj1 | > " ]) detect(quer y = " line CD ", objects = [ " <obj2>< | obj2 | > " ]) detect(quer y = " angle CED ", objects = [ " <obj3>< | obj3 | > " ]) detect(quer y = " angl... | https://arxiv.org/abs/2505.18842v1 |
copy operation. by its point-and-copy mechanism at each step. This targeted attention is pronounced and dynamic, particularly in intermediate and deeper layers (e.g., Layers 14 and 27), where scores fluctuate significantly, peaking at approximately 0.35, indicating active engagement with referenced visual information. ... | https://arxiv.org/abs/2505.18842v1 |
<obj3>, purple cylinder <obj4>. But the user mentioned six objects, including the small green cylinder <obj5>, so I'll stick with that. Alternatively, maybe the small green cylinder <obj5> is separate from the green cylinder <obj3>, making it seven distinct objects. Regardless, the key is to subtract the red sphere <ob... | https://arxiv.org/abs/2505.18842v1 |
86 and increases to about 92 or 93, so its average would be higher than Dark Violet's <obj2>. Similarly, Dark Cyan <obj5> have higher y-values, so their areas would be larger. Hot Pink starts high and decreases, but even at its lowest, it's around 90, which is still higher than Dark Violet's <obj2> average. Therefore, ... | https://arxiv.org/abs/2505.18842v1 |
to non-atomic element) - BAD grounding: "detect(query="region", objects=["<obj1>"]) The triangle <obj1> has a right angle." (referring to ambiguous element) After completing the reasoning, list all objects detected: { "obj1": {"type": "function_graph", "description": "Graph of a function with maximum at x = 3"}, "obj2"... | https://arxiv.org/abs/2505.18842v1 |
arXiv:2505.18845v1 [cs.CL] 24 May 2025Multi-Party Conversational Agents: A Survey Sagar Sapkota University of Central Florida sagar.sapkota@ucf.eduMohammad Saqib Hasan Stony Brook University mdshasan@cs.stonybrook.edu Mubarak Shah University of Central Florida shah@crcv.ucf.eduSantu Karmaker University of Central Flori... | https://arxiv.org/abs/2505.18845v1 |
to design multi-party conversational systems (Ganesh et al., 2023). In synergy with with recent advances in generative AI, particularly Large Language Models (LLMs), most of these efforts have focused on leveraging LLMs for this domain (Wang et al., 2024c; Tan et al., 2023). As such, this survey aims to serve as a comp... | https://arxiv.org/abs/2505.18845v1 |
Action Modeling . This taxonomy reflects essential capabilities for human- like social communication in group settings. In- deed, a socially intelligent agent must determine when to speak, whom to address, and what to say, collectively referred to as actions . According to Social Identity Theory (Tajfel and Turner, 197... | https://arxiv.org/abs/2505.18845v1 |
DialogueGCN (Ghosal et al., 2019), DialogXL (Shen et al., 2020), Contrastive Learning (Song et al., 2022b), PFA-ERC (Khule et al., 2024), M3GAT (Zhang et al., 2023c), DialogueLLM (Zhang et al., 2024b), InstructERC (Lei et al., 2024), SpeechCueLLM (Wu et al., 2024) Figure 2: Thematic taxonomy of MPC tasks and recent wor... | https://arxiv.org/abs/2505.18845v1 |
Engagement detection measures the interaction level of participants in conversations, essential for managing interactions and turns between partici- pants. Despite being a difficult problem, recent multi-modal methods used visual and audio cues, enhancing performance by continuously tracking participants’ behaviors (Bo... | https://arxiv.org/abs/2505.18845v1 |
2024). However, LLM-based DAR approaches are still an unsolved challenge. 4 Research on Semantic Understanding Proper semantic understanding is essential for any conversational system. But what comprises seman- tic understanding ? We identify four tasks that col- lectively define semantic understanding in MPC: Conversa... | https://arxiv.org/abs/2505.18845v1 |
Recently, hetero- geneous discourse graphs combined with Graph Convolutional Networks (GCN) have outperformed GPT-4, showing that adding structured discourse can help with this task (Li et al., 2024a). 4.3 Discourse Structure Analysis Discourse-structure analysis predicts dependencies among elementary discourse units (... | https://arxiv.org/abs/2505.18845v1 |
2020). Sequence-based neural models like GRUs and transformers further advanced turn prediction (Lee and Deng, 2024). Overcoming text-only methods, multi-modality integrated speech pauses, gaze, and listener be- haviors ( ˙Zarkowski, 2019; Paetzel-Prüsmann and Kennedy, 2023; Lee et al., 2023). LLM-based methods, like m... | https://arxiv.org/abs/2505.18845v1 |
6 Dataset Modality Model Task Eval Metric Eval Result EmoryNLP (Zahiri and Choi, 2017) Text PFA-ERC (Khule et al., 2024) Emotion Recognition F1 Score 42.94 MELD (Poria et al., 2019) Audio, Video, Text DialogueLLM (Zhang et al., 2024b) Emotion Recognition F1 Score 71.9 EmoryNLP (Zahiri and Choi, 2017) Text Affect-NLI (W... | https://arxiv.org/abs/2505.18845v1 |
SOTA models for multi- modal benchmarks remain scarce, despite the criti- cal role of cross-modal information in enhancing MPCA performance. Overall, these observations underscore the limi- tations of current benchmarking practices and high- light the need for more comprehensive, diverse, and multi-modal datasets to ad... | https://arxiv.org/abs/2505.18845v1 |
context. We see that emotion as context improves performance in both cases, highlighting that state of mind modeling, and thereby ToM, is heavily intertwined with other MPCA tasks. To address all the gaps highlighted till now, fu- ture research should explore transforming tradi- tional state-of-mind modeling tasks to s... | https://arxiv.org/abs/2505.18845v1 |
trends and supported by cursory exper- iments, which may not capture all perspectives. Additionally, many works we reviewed rely on cus- tom datasets and models, some of which do not align with recent standard benchmarks. This re- liance may introduce biases in our analysis. We recommend further, more comprehensive res... | https://arxiv.org/abs/2505.18845v1 |
Online. Association for Computational Lin- guistics. Jiaao Chen and Diyi Yang. 2020. Multi-View Sequence- to-Sequence Models with Conversational Structure for Abstractive Dialogue Summarization. In Proceed- ings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages 4106–4118, Online... | https://arxiv.org/abs/2505.18845v1 |
Xu, Pu Zhao, Yaming Yang, Chongyang Tao, Dongyan Zhao, and Qingwei Lin. 2023a. MMDialog: A large-scale multi-turn dialogue dataset towards multi-modal open-domain conversation. In Proceedings of the 61st Annual Meeting of the Association for Computational Lin- guistics (Volume 1: Long Papers) , pages 7348–7363, Toronto... | https://arxiv.org/abs/2505.18845v1 |
GIFT: Graph-induced fine- tuning for multi-party conversation understanding. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages 11645–11658, Toronto, Canada. Association for Computational Linguistics. Jia-Chen Gu, Hui Liu, Zhen-Hua Ling, Quan Liu, Z... | https://arxiv.org/abs/2505.18845v1 |
centered approach. In Proceedings of the 30th Inter- national Conference on Intelligent User Interfaces , IUI ’25, page 390–408, New York, NY , USA. Asso- ciation for Computing Machinery. 11 Chao-Chun Hsu, Sheng-Yeh Chen, Chuan-Chun Kuo, Ting-Hao Huang, and Lun-Wei Ku. 2018. Emotion- Lines: An emotion corpus of multi-p... | https://arxiv.org/abs/2505.18845v1 |
and Yifei Zhang. 2022. Learning to Improve Persona Consistency in Multi-party Dialogue Generation via Text Knowledge Enhancement. In Proceedings of the 29th International Conference on Computational Linguistics , pages 298–309, Gyeongju, Republic of Korea. International Committee on Computational Linguistics. Zerrin Ka... | https://arxiv.org/abs/2505.18845v1 |
43(1). Jiaqi Li, Ming Liu, Min-Yen Kan, Zihao Zheng, Zekun Wang, Wenqiang Lei, Ting Liu, and Bing Qin. 2020b. Molweni: A challenge multiparty dialogues-based machine reading comprehension dataset with dis- course structure. In Proceedings of the 28th Inter- national Conference on Computational Linguistics , pages 2642–... | https://arxiv.org/abs/2505.18845v1 |
16-20, 2018 , pages 163–169. ACM. Diane Litman, Susannah Paletz, Zahra Rahimi, Stefani Allegretti, and Caitlin Rice. 2016. The teams cor- pus and entrainment in multi-party spoken dialogues. InProceedings of the 2016 Conference on Empiri- cal Methods in Natural Language Processing , pages 1421–1431, Austin, Texas. Asso... | https://arxiv.org/abs/2505.18845v1 |
Soujanya Poria, Devamanyu Haz- arika, Rada Mihalcea, Alexander Gelbukh, and Erik Cambria. 2019. Dialoguernn: An attentive rnn for emotion detection in conversations. Proceedings of the AAAI Conference on Artificial Intelligence , 33(01):6818–6825. Elijah Mayfield, David Adamson, and Carolyn Rose. 2012. Hierarchical con... | https://arxiv.org/abs/2505.18845v1 |
and sentiment classification. Proceedings of the AAAI Conference on Artificial Intelligence , 34(05):8665– 8672. Libo Qin, Zhouyang Li, Wanxiang Che, Minheng Ni, and Ting Liu. 2021. Co-gat: A co-interactive graph attention network for joint dialog act recognition and sentiment classification. Proceedings of the AAAI Co... | https://arxiv.org/abs/2505.18845v1 |
Chapter of the Association for Computational Lin- guistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers) , pages 978–990, Online only. Association for Computational Linguistics. Hanan Salam and Mohamed Chetouani. 2015. Engage- ment detection based on mutli-party cues... | https://arxiv.org/abs/2505.18845v1 |
Transactions on Affective Computing , 14(4):3048– 3065. Xiaohui Song, Longtao Huang, Hui Xue, and Songlin Hu. 2022b. Supervised prototypical contrastive learn- ing for emotion recognition in conversation. In Pro- ceedings of the 2022 Conference on Empirical Meth- ods in Natural Language Processing , pages 5197– 5206, A... | https://arxiv.org/abs/2505.18845v1 |
Yu, Hongyu Gong, and Shyamnath Gollakota. 2024. Be- yond Turn-Based Interfaces: Synchronous LLMs as Full-Duplex Dialogue Agents. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages 21390–21402, Miami, Florida, USA. Association for Computational Lin- guistics. Ante Wang, Lin... | https://arxiv.org/abs/2505.18845v1 |
interpretable personality recognition in conversation. In 2024 IEEE International Confer- ence on Pervasive Computing and Communications (PerCom) , pages 184–193. Philine Witzig, Rares Constantin, Nikola Kovacevic, and Rafael Wampfler. 2024. Multimodal dialog act classification for digital character conversations. In P... | https://arxiv.org/abs/2505.18845v1 |
Technology and Appli- cations , pages 250–262, Singapore. Springer Nature Singapore. Sayyed M. Zahiri and Jinho D. Choi. 2017. Emo- tion detection on tv show transcripts with sequence- based convolutional neural networks. Preprint , arXiv:1708.04299. 18 Bo Zhang, Hui Ma, Jian Ding, Jian Wang, Bo Xu, and Hongfei Lin. 20... | https://arxiv.org/abs/2505.18845v1 |
Liu, Weizi Liu, and Yun Huang. 2022. Ux research on conversational human-ai interaction: A literature review of the acm digital library. In CHI Conference on Human Factors in Computing Systems , CHI ’22, page 1–24. ACM. Wenjie Zheng, Jianfei Yu, Rui Xia, and Shijin Wang. 2023. A facial expression-aware multimodal multi... | https://arxiv.org/abs/2505.18845v1 |
response generation, underscoring persistent modeling chal- lenges. Additionally, it reveals increased interest in multi-modal approaches, yet most still under- perform compared to unimodal text-based models, highlighting the need for more robust multi-modal integration in future MPCA studies. 20 MPC Agent Agent Action... | https://arxiv.org/abs/2505.18845v1 |
DialoGPT Annotation (Feng et al., 2021), Long Dialogue Summarization (Zhang et al., 2021), SummN (Zhang et al., 2022), Topic-Guided (Han et al., 2024)State of Mind Modeling Dialog Act Recognition (§3.4) LLM-Based LLM Evaluation (Zhao et al., 2023), Dual-Process Masking (Kim et al., 2024)Multi-ModalJoint DA/ER (Saha et ... | https://arxiv.org/abs/2505.18845v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.