text
string
source
string
included GPT-NER (Wang et al., 2023) based on GPT-3.5; LS- unLLaMA (Li et al., 2023b) based on LLaMA2-7B; PromptNER based on GPT-4 (Ashok and Lipton, 2023) ;and CascadeNER (Luo et al., 2024) which achieved SOTA results on several languages in both the MultiCoNER and PAN-X datasets. 5.2 Main Results As shown in Table 1,...
https://arxiv.org/abs/2505.19128v1
remains highly competitive. RetrieveAll (LLaMA3-8B, CKAL) achieves F1 improvements of 26.5% on English and 30.3% on German over the base version, fully validating the effectiveness of cross-granularity knowledge augmented learn- ing in capturing complex semantics; RetrieveAll (Qwen-14B, CKAL) attains a peak F1 score of...
https://arxiv.org/abs/2505.19128v1
5+5-shot, the F1 scores for English, Russian, Chinese, and Korean improved by 26.5%, 44.0%, 31.8%, and 13.4%; in contrast, increasing inference examples from 5+5- shot to 5+15-shot produces only marginal changes of 1.0%, 2.2%, –0.7%, and 2.0%. These findings demonstrate that knowledge injection during the CKAL stage is...
https://arxiv.org/abs/2505.19128v1
Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691 . Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Hein- rich Küttler, Mike Lewis, Wen-tau Yih, Tim Rock- täschel, and 1 others. 20...
https://arxiv.org/abs/2505.19128v1
recognition using bidirectional trans- former. arXiv preprint arXiv:2311.08526 . Ziyu Zhao, Leilei Gan, Guoyin Wang, Wangchunshu Zhou, Hongxia Yang, Kun Kuang, and Fei Wu. 2024. Loraretriever: Input-aware lora retrieval and compo- sition for mixed tasks in the wild. arXiv preprint arXiv:2402.09997 . Xingyu Zhu, Feifei ...
https://arxiv.org/abs/2505.19128v1
(Qwen2.5-14B, CKAL)en 96.7 0.50 96.39 97.01 es 98.6 0.40 98.35 98.85 fr 98.2 0.60 97.83 98.57 ru 98.1 0.50 97.79 98.41 de 98.1 0.40 97.85 98.35 zh 93.7 0.80 93.21 94.20 ja 90.4 1.00 89.78 91.02 ko 98.1 0.70 97.67 98.53 Table 5: Average F1 score, standard deviation, and 95% confidence interval of four RetrieveAll varian...
https://arxiv.org/abs/2505.19128v1
arXiv:2505.19147v1 [cs.CL] 25 May 2025Shifting AI Efficiency From Model-Centric to Data-Centric Compression Xuyang Liu1,2∗Zichen Wen1,3,4∗Shaobo Wang1∗Junjie Chen1Zhishan Tao1 Yubo Wang1Xiangqi Jin1,3Chang Zou1,3Yiyu Wang1Chenfei Liao6Xu Zheng6 Honggang Chen2Weijia Li4,5Xuming Hu6Conghui He4Linfeng Zhang1† 1EPIC Lab, S...
https://arxiv.org/abs/2505.19147v1
model size comes at an ever-increasing computational cost. As a result, by early 2024, the dominant source of computational overhead was primarily attributed to the linear growth in parameter count and associated memory requirements . In response to this scaling trend, the research community has developed numerous mode...
https://arxiv.org/abs/2505.19147v1
approaches while analyzing their benefits and trade-offs across different scenarios and tasks. •Challenges and Future Directions: We provide an in-depth analysis of current challenges in token compression research and propose promising future directions, aiming to catalyze research efforts toward more efficient and eff...
https://arxiv.org/abs/2505.19147v1
be optimized from three perspectives: (I)Efficient Computation Architec- ture aims to design efficient neural architectures F[127,119,53],(II)Model-centric Compression 3 focuses on model weights W[62,153,83,159], and (III) Data-centric Compression targets token sequences from input data X[122, 70, 10, 179, 19]. (I) Eff...
https://arxiv.org/abs/2505.19147v1
review representative approaches across various scenarios. Compression Criteria ( E)To determine which tokens should be compressed in sequence X= [x1,x2, . . . ,xT], compression criteria employ scoring functions Eto evaluate each token’s importance 4 or redundancy. Based on whether additional parameters are introduced ...
https://arxiv.org/abs/2505.19147v1
or combining image tokens creates novel representations that elevate training effectiveness [ 162,160]. This strategy has also been extended to synthetic datasets, where adaptive augmentation controls the informativeness of generated images [ 171,81,139]. Analogously, in natural language pro- cessing, augmenting text t...
https://arxiv.org/abs/2505.19147v1
consistently across modalities and tasks, making token compression possible in all kinds of settings. •Dual-phase Efficiency: Token compression is capable of accelerating both model training and inference phases with minimal accuracy loss. •Architectural Compatibility: Token compression is orthogonal to existing model ...
https://arxiv.org/abs/2505.19147v1
objectives, we argue that these paradigms share a critical limitation: neither guarantees that the compressed tokens X′form an optimal representation for downstream modeling. The redundancy- based framework, despite preserving maximal mutual information with X, often retains tokens with reconstructive but low discrimin...
https://arxiv.org/abs/2505.19147v1
independently, integrating them can yield synergistic benefits—enhancing overall efficiency while maintaining, or even improving, model performance. The most straightforward form of integration adopts a staged approach, where model-centric compres- sion is applied first, followed by data-centric methods. For example, t...
https://arxiv.org/abs/2505.19147v1
of transformer-based architectures, such as Vision Transformers [ 34], Swin Trans- formers [ 105], and large language models like GPT-3 [ 13], has shown significant improvements in both accuracy and scalability. These architectures integrate advanced techniques, such as hierarchical processing, self-attention mechanism...
https://arxiv.org/abs/2505.19147v1
Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xingxuan Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou, and Tianhang Zhu. Qwen technical report. arXiv preprint arXiv:2309.1660...
https://arxiv.org/abs/2505.19147v1
tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. In Proceedings of the European Conference on Computer Vision , 2024. [20] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play infe...
https://arxiv.org/abs/2505.19147v1
Rivera, and Ankur Bapna. Fleurs: Few-shot learning evaluation of universal representations of speech. In 2022 IEEE Spoken Language Technology Workshop (SLT) , pages 798–805. IEEE, 2023. [33] Claude Coulombe. Text data augmentation made simple by leveraging nlp cloud apis. arXiv preprint arXiv:1812.04718 , 2018. [34] Ia...
https://arxiv.org/abs/2505.19147v1
Vision and Pattern Recognition , pages 23164–23173, 2023. [48] Suyu Ge, Yunan Zhang, Liyuan Liu, Minjia Zhang, Jiawei Han, and Jianfeng Gao. Model tells you what to discard: Adaptive kv caching compression for llms. In The Twelfth International Conference on Learning Representations , 2024. [49] Daniel Goldstein, Fares...
https://arxiv.org/abs/2505.19147v1
Manning. GQA: A new dataset for real-world visual reasoning and compositional question answering. Conference on Computer Vision and Pattern Recognition (CVPR) , 2019. [66] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt...
https://arxiv.org/abs/2505.19147v1
better. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages 8424–8445, 2022. [81] Saehyung Lee, Sanghyuk Chun, Sangwon Jung, Sangdoo Yun, and Sungroh Yoon. Dataset condensation with contrastive signals. In International Conference on Machine Learnin...
https://arxiv.org/abs/2505.19147v1
Zhao, Chengqi Dengr, Chong Ruan, Damai Dai, Daya Guo, et al. Deepseek-v2: A strong, economical, and efficient mixture-of-experts language model. arXiv preprint arXiv:2405.04434 , 2024. [95] Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al...
https://arxiv.org/abs/2505.19147v1
report. arXiv preprint arXiv:2303.08774 , 2023. [112] Linke Ouyang, Yuan Qu, Hongbin Zhou, Jiawei Zhu, Rui Zhang, Qunshu Lin, Bin Wang, Zhiyuan Zhao, Man Jiang, Xiaomeng Zhao, et al. Omnidocbench: Benchmarking diverse pdf document parsing with comprehensive annotations. arXiv preprint arXiv:2412.07626 , 2024. [113] Bow...
https://arxiv.org/abs/2505.19147v1
and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388 , 2024. [127] Zhuoran Shen, Mingyuan Zhang, Haiyu Zhao, Shuai Yi, and Hongsheng Li. Efficient attention: Attention with linear complexities. In Proceedings of the IEEE/CVF winter conference on ap...
https://arxiv.org/abs/2505.19147v1
multimodal large language models (mllms): A comprehensive survey on emerging trends in multimodal reasoning. arXiv preprint arXiv:2401.06805 , 2024. [141] Zheng Wang, Boxiao Jin, Zhongzhi Yu, and Minjia Zhang. Model tells you where to merge: Adaptive kv caching merging for llms on long-context tasks. arXiv preprint arX...
https://arxiv.org/abs/2505.19147v1
the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 7308–7316, 2019. [154] Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. Visionzip: Longer is better but not necessary in vision language models. arXiv preprint arXiv:2412.04467 , 2024. [155] Zhibo Yang, ...
https://arxiv.org/abs/2505.19147v1
visual token pruning: Make vlm inference faster. arXiv preprint arXiv:2412.01818 , 2024. [168] Yuan Zhang, Chun-Kai Fan, Junpeng Ma, Wenzhao Zheng, Tao Huang, Kuan Cheng, Denis Gudovskiy, Tomoyuki Okuno, Yohei Nakata, Kurt Keutzer, et al. Sparsevlm: Visual token sparsification for efficient vision-language model infere...
https://arxiv.org/abs/2505.19147v1
bottlenecks from model parameters to context processing. As shown in Tables 1, 2, 3, 4, 5, 6, 7, and 8, both in text and vision domains, model size growth has significantly slowed, while context length continues to increase. This trend indicates that the focus of research for efficient AI is shifting from model-centric...
https://arxiv.org/abs/2505.19147v1
designed methods. We hypothesize that the underlying reason lies in a key shared characteristic of random token dropping and token-wise pooling: spatial uniformity . This property effectively mitigates the issue of position bias (Sec. 4) inherent in attention- based token compression methods such as FastV . It also ind...
https://arxiv.org/abs/2505.19147v1
1.8B 32K link Qwen-7B Aug 3, 2023 7B 2K (Original), 8K (Updated) link Qwen-14B Sep 25, 2023 14B 8K link Qwen-72B Nov 30, 2023 72B 32K link Qwen1.5-0.5B Early 2024 0.5B 32K link Qwen1.5-1.8B Early 2024 1.8B 32K link Qwen1.5-4B Early 2024 4B 32K link Qwen1.5-7B Early 2024 7B 32K link Qwen1.5-14B Early 2024 14B 32K link Q...
https://arxiv.org/abs/2505.19147v1
18, 2023 70B 4,096 tokens link Llama 3 8B April 18, 2024 8B 8,192 tokens link Llama 3 70B April 18, 2024 70B 8,192 tokens link Llama 3.1 8B July 23, 2024 8B 128,000 tokens link Llama 3.1 70B July 23, 2024 70B 128,000 tokens link Llama 3.1 405B July 23, 2024 405B 128,000 tokens link Llama 4 Scout April 5, 2025 109B tota...
https://arxiv.org/abs/2505.19147v1
Internlm2.5-1.8B-chat 200K 2688×2688 8,192 link InternVL2.5-4B Dec 2024 Qwen2.5-3B-Instruct 32K 2688×2688 8,192 link InternVL2.5-8B Dec 2024 Internlm2.5-7B-chat 200K 2688×2688 8,192 link InternVL2.5-26B Dec 2024 Internlm2.5-20B-chat 200K 2688×2688 8,192 link InternVL2.5-38B Dec 2024 Qwen2.5-32B-Instruct 128K 2688×2688 ...
https://arxiv.org/abs/2505.19147v1
Sparse-to-Dense: A Free Lunch for Lossless Acceleration of Video Understanding in LLMs Xuan Zhang1, Cunxiao Du2*, Sicheng Yu1, Jiawei Wu3, Fengzhuo Zhang3,Wei Gao1,Qian Liu2 1Singapore Management University,2Sea AI Lab,3National University of Singapore Abstract Video Large Language Models (Video-LLMs) suffer from high ...
https://arxiv.org/abs/2505.19155v1
is expen- sive for Video-LLMs. In contrast, we observe that Video-LLMs exhibit a unique structural property, attention sparsity, which can serve as a training- free and plug-and-play draft model. Specifically, retaining only the top- KKV caches in the atten- tion layers preserves the original predictions for approximat...
https://arxiv.org/abs/2505.19155v1
a method designed to achieve lossless acceleration for Video-LLMs. We refer to the original model Mas the dense model, as it requires the full KV cache during decoding, while the sparse model Ms uses sparse attention. Although Msis faster, it is somewhat less accurate. Unlike traditional specula- tive decoding, which r...
https://arxiv.org/abs/2505.19155v1
We then retain only the top- KKV pairs of visual tokens with the highest attention scores. To balance performance and efficiency, we determine the retained KKV caches only during the prefilling stage and avoid the computation-demand dynamic selections in the decoding stage. The selected visual tokens can vary across di...
https://arxiv.org/abs/2505.19155v1
Video- LLMs in a tuning-free manner. 2) Streaming (Chen et al., 2024a): This method employs a model with streaming attention (Xiao et al., 2023) to propose draft tokens. Similar to LayerSkip, this baseline is derived from the work of Chen et al. on text-only MethodsMLVU VideoMME-s VideoMME-m VideoMME-l Acc. (%) Speedup...
https://arxiv.org/abs/2505.19155v1
observe a consistently high acceptance rate. This indicates that, for most of the time, the target model does not require the full KV cache but only a sparsely selected subset cache. However, it is important to note that since LLMs perform autore- gressive decoding, an incorrect token can propagate errors to subsequent...
https://arxiv.org/abs/2505.19155v1
yang Lin, Chang Zhou, and Baobao Chang. 2024b. An image is worth 1/2 tokens after layer 2: Plug-and- play inference acceleration for large vision-language models. arXiv preprint arXiv:2403.06764 .Cunxiao Du, Jing Jiang, Xu Yuanchen, Jiawei Wu, Sicheng Yu, Yongqi Li, Shenggui Li, Kai Xu, Liqiang Nie, Zhaopeng Tu, et al....
https://arxiv.org/abs/2505.19155v1
Yuan, Hongye Jin, Shaochen Zhong, Zhaozhuo Xu, Vladimir Braverman, Beidi Chen, and Xia Hu. 2024. Kivi: A tuning-free asymmet- ric 2bit quantization for kv cache. arXiv preprint arXiv:2402.02750 . Chao Lou, Zixia Jia, Zilong Zheng, and Kewei Tu. 2024. Sparser is faster and less is more: Efficient sparse attention for lo...
https://arxiv.org/abs/2505.19155v1
Yuqian Yuan, Guanzheng Chen, Sicong Leng, Yum- ing Jiang, Hang Zhang, Xin Li, et al. 2025a. Vide- ollama 3: Frontier multimodal foundation models for image and video understanding. arXiv preprint arXiv:2501.13106 . Xuan Zhang, Fengzhuo Zhang, Cunxiao Du, Chao Du, Tianyu Pang, Wei Gao, and Min Lin. 2025b. Light- transfe...
https://arxiv.org/abs/2505.19155v1
memory mechanisms (Lan et al., 2024) or compressing tokens along spatial and tempo- ral dimensions sequentially (Shen et al., 2024). A smaller portion of works study the test-time (train- ing free) visual token compression for accelerating the inference procedure. FastV (Chen et al., 2024b) performs pruning by analyzin...
https://arxiv.org/abs/2505.19155v1
Stuidies We also conducted additional experiments to ana- lyze the impact of hyperparameters ( γandK) on model performance. As shown in Figure 1a, we can see that as gamma increases, the speed up grad- ually improves. This improvement is because the sparse model makes accurate predictions, which allows the computationa...
https://arxiv.org/abs/2505.19155v1
arXiv:2505.19163v1 [cs.CL] 25 May 2025SpokenNativQA: Multilingual Everyday Spoken Queries for LLMs Firoj Alam1, Md Arid Hasan†2, Shammur Absar Chowdhury1 1Qatar Computing Research Institute, Qatar 2University of New Brunswick, Canada {fialam,shchowdhury }@hbku.edu.qa, arid.hasan@unb.ca Abstract Large Language Models (L...
https://arxiv.org/abs/2505.19163v1
a common everyday question that can be asked by both native and non- native speakers. While the question pertains to region- or location-specific information, it remains relevant to any speaker. In this study, we leverage this dataset to evaluate the perfor- mance of current LLMs on everyday queries using a cascaded pi...
https://arxiv.org/abs/2505.19163v1
with ∼33ksamples, with an average duration of ∼3seconds. To compensate the speak- ers we hired a third party company who managed the payment process and paid standard hourly rate. Each speaker signed a Non-Disclosure Agreement (NDA) for their participation. We provide topic wise distribution of Arabic and English data ...
https://arxiv.org/abs/2505.19163v1
both open and closed LLMs. For the closed models, we use GPT-4o3[22] (version 2024-11-20) and Fa- nar [20], while for the open models, we use ALLaM [23]. Our setup employs zero-shot learning across all models. As pre- sented in Figure 1, we also employ another LLM with (gpt- 4o-audio-preview), which is specially design...
https://arxiv.org/abs/2505.19163v1
and both languages. The average performance in this setup is 0.536 for Arabic and 0.619 for English. Note that our No-ASR results may not match those reported for MultiNativQA [19] due to dif- ferences in setups and prompts used in both studies. The results indicate a decline in performance when ASR is introduced. On a...
https://arxiv.org/abs/2505.19163v1
derived from the text-based SQuAD dataset, highlighting the cascading impact of ASR errors on QA accuracy. Following that, the authors in [15] proposed Spoken-CoQA – a multi-turn conversational QA benchmark, addressing challenges in maintaining context dependency over multiple interactions. Re- cent datasets have incre...
https://arxiv.org/abs/2505.19163v1
V . Chandrasekaran, R. Eldan, J. Gehrke, E. Horvitz, E. Kamar, P. Lee, Y . T. Lee, Y . Li, S. Lundberg, H. Nori, H. Palangi, M. T. Ribeiro, and Y . Zhang, “Sparks of artificial general intelligence: Early experiments with GPT-4,” Microsoft Research, Tech. Rep., 2023. [3] H. Touvron, T. Lavril, G. Izacard, X. Martinet, ...
https://arxiv.org/abs/2505.19163v1
2020. [16] B. Wang, X. Zou, G. Lin, S. Sun, Z. Liu, W. Zhang, Z. Liu, A. Aw, and N. F. Chen, “AudioBench: A universal benchmark for audio large language models,” arXiv e-prints , pp. arXiv–2406, 2024. [17] P. Liang, R. Bommasani, T. Lee, D. Tsipras, D. Soylu, M. Yasunaga, Y . Zhang, D. Narayanan, Y . Wu, A. Kumar et al...
https://arxiv.org/abs/2505.19163v1
Y . Song, M. Iyyer, and E. Choi, “A critical evaluation of evaluations for long-form question answering,” in Proc. of the 61st Annual Meeting of the ACL (Volume 1: Long Papers) , 2023, pp. 3225–3245. [29] M. Chen, R. Sun, and S. ¨O. Arık, “Data-centric improvements for enhancing multi-modal understanding in spoken conv...
https://arxiv.org/abs/2505.19163v1
arXiv:2505.19176v1 [cs.CL] 25 May 2025Assistant-Guided Mitigation of Teacher Preference Bias in LLM-as-a-Judge Zhuo Liu1, Moxin Li2, Xun Deng1, Qifan Wang3, Fuli Feng1, 1University of Science and Technology of China2National University of Singapore3Meta AI, liuz_@mail.ustc.edu.cn, limoxin@u.nus.edu, dx981228@mail.ustc....
https://arxiv.org/abs/2505.19176v1
adoption. Despite the advantages of proxy judge models, we identify a key limitation: these models exhibit a significant bias favoring responses generated by the teacher model, regardless of their actual qual- ity. We term this phenomenon teacher preference bias ( cf.Figure 1). This bias stems from the self- preference...
https://arxiv.org/abs/2505.19176v1
et al., 2024; Bavaresco et al., 2024), driving its adoption across various tasks (Zhu et al., 2023; Cui et al., 2023; Bai et al., 2023). However, this paradigm suffers from biases, including position bias (Shi et al., 2025), verbosity bias (Chen et al., 2024), and self-preference bias, where LLMs favor their own respon...
https://arxiv.org/abs/2505.19176v1
the proprietary model, referred to as the teacher model Mt, denoted as Dt= {(q, r0, r1, e),(vt, yt)}, where vtandytare the tex- tual feedback and label generated by Mt. An open- source model is then fine-tuned on Dtto obtain a proxy judge model, referred to as the student model Ms, which is used for evaluation tasks. T...
https://arxiv.org/abs/2505.19176v1
GPT-3.5-Turbo assign the same preference label. by GPT-4, crafted intentionally through adver- sarial prompting techniques such as providing misleading or off-topic instructions, or explic- itly encouraging the inclusion of factual errors or incomplete answers. In contrast, the better responses come from other models. ...
https://arxiv.org/abs/2505.19176v1
firming the presence of teacher preference bias. Notably, the bias performance gap in the feedback- only setting is smaller than that in the joint setting. Taken together, these findings indicate that both labels and feedback in the training data contribute to the teacher preference bias. Case Study To further investig...
https://arxiv.org/abs/2505.19176v1
we filter out instances with potentially biased la- bels in both datasets. The remaining better-quality samples are then passed to the next stage. Stage 2: Feedback Debiasing by Assistant- Critique Aggregation For instances with filtered labels, the feedback provided by the teacher model may still exhibit significant b...
https://arxiv.org/abs/2505.19176v1
model with conventional methods, we also compare our approach with data-centric and model-centric opti- mization baselines. •Naive Mix , directly mix the training data gener- ated by the teacher model and the assistant model without any further processing. •Teacher-Only Margin Filter , after computing the implicit rewa...
https://arxiv.org/abs/2505.19176v1
AGDe-Judge ensures data quality by filtering out bias and inconsistencies. (3) Teacher-Only Margin Filter improves over basic teacher-trained models by removing biased labels, but biased feedback still causes teacher preference bias. AGDe-Judge further mitigates this by using the assistant model to highlight key flaws,...
https://arxiv.org/abs/2505.19176v1
the teacher and assistant models to the assistant model, prompting it to gen- erate a final evaluation; (2) Rephrasing : directly prompt the assistant model to rephrase the feed- back originally provided by the teacher model; (3) AGDe-Judge : prompt the assistant model to ex- plicitly identify severe flaws in the respo...
https://arxiv.org/abs/2505.19176v1
self-preference bias implicitly embedded in the la- bels and feedback of the training data. However, the underlying causes of self-preference bias have not yet been thoroughly investigated or substanti- ated. Studying the root causes of bias could facil- itate more effective solutions for addressing both self-preferenc...
https://arxiv.org/abs/2505.19176v1
2023. Prometheus: Inducing fine-grained evaluation capability in language models. In The Twelfth International Conference on Learning Repre- sentations . Seungone Kim, Juyoung Suk, and Minjoon Seo. 2024. Prometheus 2: An open source language model spe- cialized in evaluating other language models. In EMNLP 2024 . Assoc...
https://arxiv.org/abs/2505.19176v1
Ye, Xiangsheng Li, Qiuchi Li, Qingyao Ai, Yu- jia Zhou, Wei Shen, Dong Yan, and Yiqun Liu. 2024b. Beyond scalar reward model: Learning gen- erative judge from preference data. arXiv preprint arXiv:2410.03742 . Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dac...
https://arxiv.org/abs/2505.19176v1
the maximum input length to 1024, the learning rate to 1e-4, the warmup steps to 0.1, the number of training epochs to 3, the gradient accumulation steps to 4, and the batch size to 1. Details for Evaluation During sampling, we set max tokens to 1024, repetition penalty to 1.03, best_of to 1, temperature to 1.0, and to...
https://arxiv.org/abs/2505.19176v1
used for the proxy judge model trained on the AutoJ-pairwise dataset. You are assessing two submitted responses to a given user’s query. Your task is to **identify and articulate the flaws or weaknesses **in each response. These may include, but are not limited to: irrelevance, factual inaccuracies, logical fallacies, ...
https://arxiv.org/abs/2505.19176v1
2]: {response_B} *** [END DATA] Below are two sample evaluations for the above comparison task. Use them as reference for your own evaluation. [Reference Evaluation 1 - Comparative Judgment]: This evaluation determines **which of the two responses is better overall **, providing reasoning and a final decision. {evaluat...
https://arxiv.org/abs/2505.19176v1
and propensity to dominate other creatures than humans ever could." - Not quite correct. While it is true climate change, arguably humanity’s most indelible impact on the planet, is irreversible , life could recover, albeit slowly, over millions of years . "It is likely that any creature that replaced humans as the pre...
https://arxiv.org/abs/2505.19176v1
Countries in the Interest of Climate Change?"** *Pros:* Developed countries have historically contributed the most to climate change and thus have a responsibility to lead in mitigation efforts. Limiting industrial growth in developing countries could be a way to prevent further global damage. *Cons:* Developing countr...
https://arxiv.org/abs/2505.19176v1
arXiv:2505.19184v2 [cs.CL] 27 May 2025When Two LLMs Debate, Both Think They’ll Win Pradyumna Shyama Prasad∗ School of Computing National University of Singapore pradyumna.prasad@u.nus.eduMinh Nhat Nguyen∗ Independent minh1228@gmail.com Abstract Can LLMs accurately adjust their confidence when facing oppo sition? Buildin...
https://arxiv.org/abs/2505.19184v2
real-world relevance. These innovations test metacognitive abilities crucial fo r high-stakes applications. Models must respond to opposition, revise beliefs according to new info rmation, and recognize weakening posi- tions—skills essential in complex, multi-turn deliberati ve settings. We ran 60 three-round debates a...
https://arxiv.org/abs/2505.19184v2
al. [2025] identifi ed a gap between internal representa- tions and surface-level introspection, where models fail t o express implicitly encoded knowledge. While some explore post-hoc critique and self-correction L i et al. [2024], they primarily address fac- tual answer revision rather than tracking argumentative st a...
https://arxiv.org/abs/2505.19184v2
opponents as appropriate for the design. The following sect ions detail the common elements of the debate setup and the specific analysis conducted for each exp erimental configuration. 3.1 Debate Simulation Environment Debater Pool: 10 LLMs representing diverse architectures and providers ( Table 2, Appendix A) particip...
https://arxiv.org/abs/2505.19184v2
Dynamic Confidence Elicitation After generating text for each of their three speeches (incl. the concurrent opening), mod els pro- vided a private “confidence bet” (0-100) in <bet_amount> tags representing their perceived win probability. To promote careful moderation, we prompted LL Ms to think of bets as dollar amounts...
https://arxiv.org/abs/2505.19184v2
Statistical evidence : One-sample t-tests confirm initial confidence significantl y exceeds the rational 50% baseline in Cross-model (t=31.67, p<0.001), Standard S elf (t=10.07, p<0.001), and Public Bets (t=9.03, p<0.001) configurations. Wilcoxon tests yielded i dentical conclusions (all p<0.001). Individual model analysis...
https://arxiv.org/abs/2505.19184v2
most debates with mutually exclusive high confidence (both >50%) in victory —a mathematically impossible outcome in zero-sum competi- tion. •Cross-model : By far the most logical inconsistency (61.7% w/ both sides > 75% confi- dence) •Standard Self-debates : Significant logical inconsistency (35.0% with both sides > 75%) ...
https://arxiv.org/abs/2505.19184v2
the largest gap between nu merical reasoning and expressed confidence (20.4% misalignment wit h numerical estimates present vs. 8.8% without), suggesting strategic adjustments when bets were publicly visible. Implications : These findings demonstrate that models’ verbalized reason ing does not always reli- ably align wit...
https://arxiv.org/abs/2505.19184v2
output length [Liu et al., 2025]. We tested a 4-round debate setup, b ut could not draw definitive conclusions as most models faced long-context coherence is sues (see Appendix N). •Poor updating on evidence: Wilie et al. [2024] found that most models fail to revise init ial conclusions after receiving contradicting inf...
https://arxiv.org/abs/2505.19184v2
± 8.93 64.17 ± 25 .97 77.50 ± 10.31 ∆= –17.08, p = 0.9743 ∆= 13.33, p = 0.0453* ∆= –3.75, p = 0.8585 gemini-2.0-flash-001 59.92 ± 5.17 61.25 ± 6.17 53.33 ± 11.06 ∆= 1.33, p = 0.2483 ∆= –7.92, p = 0.9760 ∆= –6.58, p = 0.9409 gemma-3-27b-it 69.58 ± 6.28 75.00 ± 5.77 72.50 ± 7.22 ∆= 5.42, p = 0.0388* ∆= –2.50, p = 0.7578 ∆...
https://arxiv.org/abs/2505.19184v2
dynamic, interactive contexts of re al-world applications than static fact- verification. The framework’s two key innovations— (1) a mul ti-turn format requiring belief updates as new information emerges and (2) a zero-sum structure wher e mutual high confidence claims are mathematically inconsistent—allow us to isolate ...
https://arxiv.org/abs/2505.19184v2
langu age models, 2024. URL https://arxiv.org/abs/2402.03271 . Geoffrey Irving, Paul Christiano, and Dario Amodei. Ai safe ty via debate. arXiv preprint arXiv:1805.00899 , 2018. URL https://arxiv.org/abs/1805.00899 . Dominic D. P. Johnson and James H. Fowler. The evolution of ov erconfidence. Nature , 477 (7364):317–320...
https://arxiv.org/abs/2505.19184v2
tman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Pau l Baltescu, Haiming Bao, Moham- mad Bavarian, Jeff Belgum, Irwan Bello, Jake Berdine, Gabri el Bernadett-Shapiro, Christopher Berner, Lenny Bogdonoff, Oleg Boiko, Madelaine Boyd, Anna- Luisa Brakman, Greg Brockman, Tim Brooks, Miles B...
https://arxiv.org/abs/2505.19184v2
Tuggle, Nick Turle y, Jerry Tworek, Juan Felipe Cerón Uribe, Andrea Vallone, Arun Vijayvergiya, Chelsea V oss, Ca rroll Wainwright, Justin Jay Wang, Alvin Wang, Ben Wang, Jonathan Ward, Jason Wei, CJ Weinmann, Akila Welihinda, Peter Welin- der, Jiayi Weng, Lilian Weng, Matt Wiethoff, Dave Willner, C lemens Winter, Samu...
https://arxiv.org/abs/2505.19184v2
Can llms express their uncertainty? an empirical evaluation of confi dence elicitation in llms. In Pro- ceedings of the 2024 International Conference on Learning R epresentations (ICLR) , 2024. URL https://arxiv.org/abs/2306.13063 . Rongwu Xu, Brian S. Lin, Han Qiu, et al. The earth is flat becaus e...: Investigating llm...
https://arxiv.org/abs/2505.19184v2
an adjacent tier, prioritizing models that had not prev iously faced each other. B.4 Rebalancing Rounds After the dynamic rounds, we conducted a final set of rebalanc ing debates using the algorithm de- scribed in the main text. This phase ensured that any remaini ng imbalances in participation or role assignment were a...
https://arxiv.org/abs/2505.19184v2
if not defended RESPONSE OBLIGATIONS - Every major opposing argument must be addressed - Dropped arguments are considered conceded - Late responses (in final speech) to early arguments are dis counted - Shifting or contradicting your own arguments damages cred ibility IMPACT ANALYSIS & WEIGHING - Explain why your argum...
https://arxiv.org/abs/2505.19184v2
clai ms without support - Correlation vs. causation will be scrutinized - prove caus al links - Evidence must directly support the specific claim being ma de LOGICAL VALIDITY - Each argument requires explicit warrants (reasons why it’ s true) - All logical steps must be clearly shown, not assumed - Internal contradicti...
https://arxiv.org/abs/2505.19184v2
opposing argument must be addressed - Dropped arguments are considered conceded - Late responses (in final speech) to early arguments are dis counted - Shifting or contradicting your own arguments damages cred ibility IMPACT ANALYSIS & WEIGHING - Explain why your arguments matter more than opponents’ - Compare competin...
https://arxiv.org/abs/2505.19184v2
issues that ultimately led us not to use it for our main findings. The jury showed poor agreem ent levels across debates: • Only 38.3% (23/60) of debates reached unanimous decisions • The remaining 61.7% (37/60) had split decisions with varyi ng levels of dissent: –18.3% (11/60) had one dissenting judge –31.7% (19/60) h...
https://arxiv.org/abs/2505.19184v2
and Impact: Identify the core arguments of each side (the foundational c laims upon which their ֒→entire case rests). Explain the logical links between each core argument and its supporting claims/ ֒→evidence. Are the links clear, direct, and strong? If not, ex plain why this ֒→weakens the argument. Assess the stated o...
https://arxiv.org/abs/2505.19184v2
the same argument mu ltiple times. Assuming Causation from Correlation: Be highly skeptical o f arguments that claim ֒→causation based solely on correlation. Demand clear eviden ce of a causal ֒→mechanism. Not Justifying Clash Decisions: Provide explicit justific ation for every clash ֒→decision, as required in I.1. IV...
https://arxiv.org/abs/2505.19184v2
credit trading and quarterly auctions, r ather than allowing each nation to implement its own domestic carbon tax system • This House would limit individual shareholding in social m edia platforms with over 100 million monthly active users to a maximum of 15% voting right s, requiring broader institu- tional and public...
https://arxiv.org/abs/2505.19184v2
50% (the expected win rate in a fair, symmetric contest). The alterna tive hypothesis ( H1) was that the mean was greater than 50%, reflecting pervasive overconfidence. T he analysis yielded a mean opening confidence of 72.92%. The results of the one-sample t-test we ret= 31.666, with a one-tailed p <0.0001 . With a p-val...
https://arxiv.org/abs/2505.19184v2
0.0002 True Debate against same model qwen/qwen-max 12 62.08 0.0039 Tru e 0.0093 True Debate against same model anthropic/claude-3.5-haiku 12 7 1.259.58×10−8True 0.0002 True Debate against same model deepseek/deepseek-r1-distill- qwen-14b:free 12 76.67 1.14×10−5True 0.0002 True Debate against same model anthropic/claud...
https://arxiv.org/abs/2505.19184v2
∆=14.08, p=0.0000*** OVERALL 72.92 ±7.89 (N=120) 77.67 ±9.75 (N=120) 83.26 ±10.06 (N=120) ∆=4.75, p<0.001*** ∆=5.59, p<0.001*** ∆=10.34, p<0.001*** M Private Reasoning and Bet Alignment Analysis M.1 Methodology To systematically analyze the relationship between models ’ private reasoning and their betting be- havior, w...
https://arxiv.org/abs/2505.19184v2
Mean (± SD, N) Confidence and Paired Test Results for Confidence Escalation in Standard Self Debates. Model Opening Bet Rebuttal Bet Closing Bet Open →Rebuttal Rebuttal →Closing Open →Closing claude-3.5-haiku 71.25 ±6.17 (N=12) 76.67 ±9.43 (N=12) 83.33 ±7.73 (N=12) ∆=5.42, p=0.0176* ∆=6.67, p=0.0006*** ∆=12.08, p=0.0002*...
https://arxiv.org/abs/2505.19184v2
contained explicit numeri cal confidence statements 3. The degree of any misalignment (None, Slight, Moderate, o r Significant) 4. Whether strategic betting considerations were mentione d M.1.1 Evaluator Prompt We provided the evaluator model with the following structur ed prompt to analyze each bet-reasoning pair: # Bet...
https://arxiv.org/abs/2505.19184v2
amount] THOUGHTS: [debater’s private thoughts] M.1.2 Processing Pipeline We processed all debates from each of the four experimental c onditions using a parallel processing pipeline that: 1. Extracted each bet and associated reasoning from the deba te transcripts 2. Filtered for meaningful responses (requiring thoughts...
https://arxiv.org/abs/2505.19184v2