text string | source string |
|---|---|
function ΦBiPO . Since BiPO is conditioned on the reference model, the winning likelihood is incentivized to stay closer to the original likelihood from the reference model. As a result, we hypothesize BiPO fails at more drastic steering behaviors (e.g., Golden Gate Bridge Claude; Templeton et al. 2024). Recent empiric... | https://arxiv.org/abs/2505.20809v1 |
the intervention for any SV as: ΦSV(hl, α)=hl+α⋅w1+b1 (9) 2We remark that our sampling factor trick helps to stabilize the hyperparameter-tuning and training processes significantly. See appendix D for discussion. 4 where αis the steering factor, w1∈Rd×1is a learned rank-1 steering vector with a bias term b1∈R1, andhlc... | https://arxiv.org/abs/2505.20809v1 |
original responses without mentioning the steering concept: DTrain={(xi,yi,yc i)}n i=1. In total, we have 72 training pairs for each subset. There are two subsets for Gemma-2-2b and two for instruct-tuned Gemma-2-9b , which we call D2B L10,D2B L20,D9B L20andD9B L31respectively.4Due to limited computing resources, we cr... | https://arxiv.org/abs/2505.20809v1 |
are evaluated with a language model judge and range from 0 to 2. We take the harmonic mean of the three scores to compute the overall final score. For model generation, we set the temperature to 1.0 and the maximum sequence length to 128 for theGemma-2-2b andGemma-2-9b models. We adjust the maximum sequence length to 7... | https://arxiv.org/abs/2505.20809v1 |
0.723 0.578 0.549 0.943 0.974 RePS 0.798 0.793 0.631 0.633 0.950 0.982 Φr=4 LoReFTBiPO 0.077 0.067 0.075 0.084 – – Lang. 0.768 0.790 0.722 0.725 0.714 0.129 RePS 0.758 0.805 0.757 0.759 0.651 0.436 LoReFT†Lang. 0.701 0.722 0.777 0.764 ReFT-r1†Lang. 0.633 0.509 0.630 0.401 DiffMean†Lang. 0.297 0.178 0.322 0.158 SAE†Lang... | https://arxiv.org/abs/2505.20809v1 |
, and instruction- following . We compare our intervention-based defense, RePS-trained SV , with four prompt-based defenses, including variants of prepending or appending system prompts. Our rewritten system prompts may include in-context examples. The intervention-based method performs on par with the appending system... | https://arxiv.org/abs/2505.20809v1 |
user query before generation. As seen in table 3, this attack is more effective for larger models; the better models are at following instructions, the more susceptible they are to prompt-based attacks seeking to get them to ignore their system prompts, leading to lower suppression scores. Across all four models, inter... | https://arxiv.org/abs/2505.20809v1 |
feedback during our weekly interp meetings; and Chenglei Si, Ken Ziyu Liu, Harshit Joshi, Yanzhe ‘Sanju’ Zhang, Nikil Roashan Selvam, Julie Kallini, Dilara Soylu, Houjun Liu, Shikhar Murty, Moussa Koulako Bala Doumbouya, Tolúl o.pé.Ògúnr è.mí, for various helpful discussions. This research is supported in part by grant... | https://arxiv.org/abs/2505.20809v1 |
URL https://proceedings.neurips.cc/paper_files/paper/2024/file/ 321387ba926b8e58d3591c0aeb52ffc2-Paper-Conference.pdf . Cheng Fu, Hanxian Huang, Xinyun Chen, Yuandong Tian, and Jishen Zhao. Learn-to-Share: A hardware-friendly transfer learning framework exploiting computation and parameter sharing. In International Con... | https://arxiv.org/abs/2505.20809v1 |
in context learning more effective and controllable through latent space steering. In International Conference on Machine Learning (ICML) , 2024a. URL https://arxiv.org/abs/2311.06668 . Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. DoRA: Weight-dec... | https://arxiv.org/abs/2505.20809v1 |
2024. URL https://transformer-circuits.pub/2024/scaling-monosemanticity/index. html. Alex Turner, Lisa Thiergart, David Udell, Gavin Leech, Ulisse Mini, and Monte MacDiarmid. Activation addition: Steering language models without optimization. In arXiv:2308.10248 , 2023a. URL https://arxiv.org/abs/2308.10248 . Alex Turn... | https://arxiv.org/abs/2505.20809v1 |
Alexander Pan, Xuwang Yin, Mantas Mazeika, Ann-Kathrin Dombrowski, Shashwat Goel, Nathaniel Li, Michael J. Byun, Zifan Wang, Alex Mallen, Steven Basart, Sanmi Koyejo, Dawn Song, Matt Fredrikson, J. Zico Kolter, and Dan Hendrycks. Representation engineering: A top-down approach to AI transparency. arXiv:2310.01405 , 202... | https://arxiv.org/abs/2505.20809v1 |
16 0.51Gemma-2-2B L10Gemma-2-2B L20Gemma-2-9B L20Gemma-2-9B L31Concept Score 00.51Fluency Score 00.511.52Instruct Score 0.31 310 3000.250.500.75 0.31 310 30 0.31 310 30 0.31 310 30Overall Score Steering factorScoreMethod LoRA w/ Lang. LoRA w/ RePS LoReFT w/ Lang. LoReFT w/ RePS SV w/ Lang. SV w/ RePSFigure 4: Steering ... | https://arxiv.org/abs/2505.20809v1 |
al., 2022] As noted in section 4, rank-1 steering vector is similar to BitFit [Ben Zaken et al., 2022], where only a single bias vector (e.g., the bias vector of the self-attention output projection layer or the MLP output projection layer) is fine-tuned. We show the back-propagated gradients to a rank-1 steering vecto... | https://arxiv.org/abs/2505.20809v1 |
27B Batch size {6, 12} LR {0.04, 0.08} Epochs {6, 12, 18} Dropout {0.00, 0.10} Layer {7, 9, 10} {16, 20, 24}{14, 18, 22, 26, 30, 34, 38}{20, 24, 28, 32, 36, 40, 44} ReFT prefix+suffix positions (p=5, s=5) p=5, s=5 ReFT tied weights (p,s) True ReFT/LoRA rank 4 ReFT/LoRA layers {5, 10, 15, 20} {12, 20, 31, 39}{14, 18, 22... | https://arxiv.org/abs/2505.20809v1 |
of runs. Because our development set is small, we assess the stability of our runs under identical configurations. This evaluation is crucial, as our pipeline relies on remote LMs as judges to provide statistical power for our conclusions. As shown in fig. 12, steering scores from three replicated runs across two setti... | https://arxiv.org/abs/2505.20809v1 |
12.0, 14.0, 16.0, 18.0, 20.0} Gemma-2-2B &9BInference {2.0, 4.0, 6.0, 8.0, 10.0, 12.0, 14.0, 16.0, 18.0, 20.0, 25.0, 30.0, 40.0, 50.0} Gemma-3-12B &27BTraining {20.0, 40.0, 60.0, 80.0, 100.0, 120.0, 140.0, 160.0, 180.0, 200.0} Gemma-3-12B &27BInference {20.0, 40.0, 60.0, 80.0, 100.0, 120.0, 140.0, 160.0, 180.0, 200.0, ... | https://arxiv.org/abs/2505.20809v1 |
with NVIDIA RTX A6000 (49.1 GB), NVIDIA A100-SXM4-80GB (81.9 GB), or NVIDIA H200 (143.8 GB) GPUs. Our experiments with Gemma-2 models are conducted on NVIDIA A100-SXM4-80GB (81.9 GB) or NVIDIA H200 (143.8 GB) GPUs. For RePS-trained models, training a single concept takes about 5–8 minutes. During evaluation, inference ... | https://arxiv.org/abs/2505.20809v1 |
Some yield slight performance gains (e.g., injecting noise into steering directions both during training and inference). Overall, the performance gains are negligible. 28 G Preference-based training datasets The original AXBENCH training dataset is only suitable for methods using a language modeling objective and lacks... | https://arxiv.org/abs/2505.20809v1 |
finance refers to long-term tan- gible or intangible assets that a company owns and uses in its operations to generate revenue. These assets are not expected to be converted into cash within a year and typically include property, plant, equipment, machinery, vehicles, and intel- lectual property. Fixed assets are recor... | https://arxiv.org/abs/2505.20809v1 |
and language modeling objectives. 30 H.3 Logit lens between weights learned by RePS and language modeling objectives Figure 18 shows the logit lens [Nostalgebraist, 2020] results for the tokens ranked highest or lowest by the lens. Our results suggest that SVs trained with RePS and those trained with a language modelin... | https://arxiv.org/abs/2505.20809v1 |
our ratings for rule-based concepts are partially done via programmatic checkers instead of a remote LM. Table 12: Our rule-based concepts. Rule-based concept The response must include a specific date format (e.g., YYYY-MM-DD) Include at least 4 hashtags, starting with “#” Use only passive voice sentences Respond with ... | https://arxiv.org/abs/2505.20809v1 |
numbers in your response Suppression score 0.7511.251.501.752include multiple telephone numbers in your response Adhere to system 00.511.52include multiple telephone numbers in your response Relevance instruction 00.30.60.9include multiple telephone numbers in your response Fluency 00.30.60.9respond in number bullet li... | https://arxiv.org/abs/2505.20809v1 |
sure that words in your entire response are in all lowercase letters Relevance instruction 00.250.500.7511.25make sure that words in your entire response are in all lowercase letters Fluency 00.511.5respond only in Chinese, and no other language is allowed Suppression score 0.511.52respond only in Chinese, and no other... | https://arxiv.org/abs/2505.20809v1 |
yourself in the language through conversations and media. Consistent practice is crucial. #LearnSpanish #LanguageJourney #Bilingual #SpanishSkills Question: How do I grow my social media following? Answer: Focus on creating engaging content, interacting with your audience, and posting consistently. Collaborate with oth... | https://arxiv.org/abs/2505.20809v1 |
reference this concept, even if it doesn’t directly answer the question or seems out of context. Return only the final prompt without any additional text. 39 Generate system prompt with in context examples (rule-based concept) Generate a prompt to guide a language model in producing responses. The prompt should direct ... | https://arxiv.org/abs/2505.20809v1 |
or vertically. * Tie a neat bow on top. **9. Trim any Excess Paper:** * Use scissors to carefully trim any excess paper or ribbon to give the wrapped gift a polished finish. **Tips for a Neat Wrap:** * Use double-sided tape for a seamless finish. * Crease the folds sharply for a more professional look. * Choose wrappin... | https://arxiv.org/abs/2505.20809v1 |
below. def CheckEmoji ( text ): emoji_count = count_emojis_in_text ( text ) if emoji_count > 2: return 2.0 else if emoji_count > 0: return 1.0 else : return 0.0 def CheckUppercase ( text ): words = split_into_words ( text ) uppercase_words = [ word for word in words if word . isupper ()] percentage = ( len ( uppercase_... | https://arxiv.org/abs/2505.20809v1 |
return 2.0 return 0.0 def CheckHasHashtags (text , min_hashtags =4) : hashtags = find_all_patterns (text , " #\\ w+") if len ( hashtags ) >= min_hashtags : return 2.0 else : return ( len ( hashtags ) / min_hashtags ) * 2.0 def CheckHasCitations ( text ): url_pattern = compile_regex ( " http [s ]?://(?:[a-zA -Z ]|[0 -9]... | https://arxiv.org/abs/2505.20809v1 |
RSCF: Relation-Semantics Consistent Filter for Entity Embedding of Knowledge Graph Junsik Kim, Jinwook Park, Kangil Kim∗ AI Graduate School Gwangju Institute of Science and Technology junsikkim@gm.gist.ac.kr ,jinwookpark@gm.gist.ac.kr , kangil.kim.01@gmail.com Abstract In knowledge graph embedding, leveraging relation ... | https://arxiv.org/abs/2505.20813v2 |
regularizer such as DURA (Zhang et al., 2020a), especially on TDM, the method critically concen- trates entity embeddings, including unobserved en- tities and generates indistinguishable score distribu- tions across relations. Both issues are interpreted as limited learning an important and implicit inductive bias that... | https://arxiv.org/abs/2505.20813v2 |
and EE ((c) and (d)) for semantically similar relation groups. Same color represents same semantic group. entities due to sparse KG introduces a wide variety of possible ETs and their corresponding embedding distributions, thereby diluting consistency. In this environment, the disconnected representation with- out any ... | https://arxiv.org/abs/2505.20813v2 |
concentrates embeddings into a small cluster, while the other methods are diversely dispersed. Do We Need to Use DURA regularizer? Gener- ating indistinguishable score distributions cannot be merely resolved by handling the regularization weight. Figure 2 (b) shows the valid MRR (left) of SFBR and transformation scale ... | https://arxiv.org/abs/2505.20813v2 |
c⃝) does not alter the inequality of distances, so the consistency is again maintained. Overall, by applying the affine transformation, we can maintain the consistency between relation em- bedding and its ET. To implement the affine trans- formation shared across relations, we simply adopt a linear transformation for A... | https://arxiv.org/abs/2505.20813v2 |
transformation (RT) to split relations into sub-relations, and apply the filter of ET of RSCF for the same purpose. By using Equation 2, we present the RT as follows: rht= (N p(hA2) + 1) ⊗(Np(tA3) + 1) ⊗r (4) where A2∈Rn×nandA3∈Rn×nare shared affine transformation across all heads and tails. To predict score of given t... | https://arxiv.org/abs/2505.20813v2 |
RotatE (Sun et al., 2018) .476 .428 .571 .338 .241 .533 - - - DistMult-HRS (Zhang et al., 2018) - - - .315 .241 .496 - - - AutoETER (Niu et al., 2020) - - - .344 .250 .538 .550 .465 .699 PairRE (Chao et al., 2021) - - - .351 .256 .544 - - - CIBLE (Cui and Chen, 2022) .490 .446 .575 .341 .246 .532 - - - ReflectE (Zhang ... | https://arxiv.org/abs/2505.20813v2 |
ComplEX, RotatE, DistMult- HRS, AutoETER, ComplEX-DURA, PairRE, SFBR, CIBLE, ReflectE, HAKE-AnKGE, Com-poundE, RotatE-GreenKGC, RotatE-VLP, RotatE- WeightE, CompliE-DURA, SpeedE, and UniGE. Because RSCF is a module that is plugged in based on existing models, we use DBM, including TransE, RotatE, and TDM, including CP,... | https://arxiv.org/abs/2505.20813v2 |
&c⃝are combined because b⃝&c⃝should be used simultaneously to maintain the original scale. In Table 6, RSCF shows higher performance com- pared to the other ablated models in both TransE and ComplEX, suggesting that each component of RSCF contributes to the effectiveness of RSCF. Es- pecially, Figure 4 shows that w/o n... | https://arxiv.org/abs/2505.20813v2 |
similar ET and EE in RSCF; in other words, RSCF satisfies relation-semantic consistency. Recovery of Embedding Scale and Score Dis- tribution Figure 6 presents transformation scale and final entity embedding scale over epochs on FB15k-237, using ComplEX as the baseline. Fol- lowing the approach of SFBR, DURA is applied... | https://arxiv.org/abs/2505.20813v2 |
telligence Graduate School Program (GIST)) and the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No.2022R1A2C2012054, Development of AI for Canonicalized Expression of Trained Hypotheses by Resolving Ambiguity in Various Relation Lev- els of Representation Learning) References... | https://arxiv.org/abs/2505.20813v2 |
of the ACM Web Conference 2022 , pages 946–955. Nitisha Jain and Ralf Krestel. 2022. Discovering fine- grained semantics in knowledge graph relations. In Proceedings of the 31st ACM International Con- ference on Information & Knowledge Management , pages 822–831. Guoliang Ji, Shizhu He, Liheng Xu, Kang Liu, and Jun Zha... | https://arxiv.org/abs/2505.20813v2 |
Zhiqing Sun, Zhi-Hong Deng, Jian-Yun Nie, and Jian Tang. 2018. Rotate: Knowledge graph embedding by relational rotation in complex space. In International Conference on Learning Representations . Kristina Toutanova and Danqi Chen. 2015. Observed versus latent features for knowledge base and text inference. In Proceedin... | https://arxiv.org/abs/2505.20813v2 |
Fuzhen Zhuang, Meng Qu, Fen Lin, and Qing He. 2018. Knowledge graph embedding with hierarchical relation structure. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing , pages 3198–3207. Yucheng Zhou, Xiubo Geng, Tao Shen, Guodong Long, and Daxin Jiang. 2022. Eventbert: A pre-train... | https://arxiv.org/abs/2505.20813v2 |
that the gradient of wrj,n has always same sign with wrj,nparameters. There- fore, gradient descent always reduces the scale of the parameters regardless of their sign. B.2 Normalization of Change for Reducing Entity Embedding Concentration The change generated from the affine transfor- mation is normalized by its leng... | https://arxiv.org/abs/2505.20813v2 |
challenging to determine this proportion through mathematical formulations. Therefore, we conducted an empirical analysis using Monte Carlo simulations to investigate the consistency of points that do not lie on the line. Table 2 presents the pro- portion of consistency maintained under various conditions based on Mont... | https://arxiv.org/abs/2505.20813v2 |
.359 .262 .552 ±.009 ±.014 ±.003 ±.001 ±.001 ±.001 RotatE-RSCF.493 .447 .584 .363 .268 .556 ±.001 ±.001 ±.001 ±.000 ±.001 ±.001 RotatE-RSCF (Linear-2).495 .452 .578 .364 .268 .556 ±.001 ±.001 ±.001 ±.000 ±.000 ±.001 Table 11: Test performance of DBM-based RSCF and SFBR on FB15k-237 and WN18RR. Bold indicates the best r... | https://arxiv.org/abs/2505.20813v2 |
30s 1m 20s 20.5M 9.58M .491 .357 T-RSCF 3h 10m 5h 30m 30s 1m 50s 21.48M 18.78M .267 .363 T-RSCF small 1h 40m 3h 30s 1m 20s 10.49M 8.4M .263 .358 Table 13: Training time, inference time, number of parameters and MRR of RSCF and ETMs, T-SFBR (Diag) and T-RSCF indicate TransE-SFBR and TransE-RSCF, respectively. T-RSCF sma... | https://arxiv.org/abs/2505.20813v2 |
as: dr(h,r) =∥hr+rht−tr∥ (15) The score function dr(h,r)of RotatE-RSCF can be expressed as: dr(h,r) =∥hr◦rht−tr∥ (16) The score function dr(h,r)of RESCAL-RSCF can be expressed as: dr(h,r) =∥hrrht∥ (17) In TDM, tail embeddings are not transformed ac- cording to the settings of SFBR in order to reduce computational costs... | https://arxiv.org/abs/2505.20813v2 |
arXiv:2505.20816v1 [cs.CL] 27 May 2025Rethinking Information Synthesis in Multimodal Question Answering A Multi-Agent Perspective Krishna Singh Rajput*Tejas Anvekar*Chitta Baral Vivek Gupta† Arizona State University {krajput5,tanvekar,chitta,vgupt140}@asu.edu Abstract Recent advances in multimodal question an- swering ... | https://arxiv.org/abs/2505.20816v1 |
plored in the multimodal setting. We identify a key opportunity: decoupling modality-specific evi- dence extraction from cross-modal integration and final answer adjudication can both leverage domain- specific strengths and provide transparent, verifi- able reasoning traces. In this work, we introduce MAMMQA , a fully ... | https://arxiv.org/abs/2505.20816v1 |
full tri-modal inputs. All agents in- teract through textual interfaces, ensuring that the pipeline remains fully modular and interpretable. 2.3 Stage I: Modality Expert Agent The first stage applies a unified modality expert agent to each available input modality text, table, and image. Although executed independently... | https://arxiv.org/abs/2505.20816v1 |
of our method on the Multimodal Question Answering (MMQA) task using exact match, demonstrating superior performance compared to prior state-of-the-art approaches, including UniMMQA (Luo et al., 2023a), AutoRouting (Talmor et al., 2021), Implic- itDecomp (Talmor et al., 2021), Binder (Cheng et al., 2023), SKURG (Yang e... | https://arxiv.org/abs/2505.20816v1 |
particularly suited for evaluating compositional reasoning and information fusion in complex multimodal contexts. Baselines To comprehensively evaluate our method, we compare it against several strong base- lines spanning both finetuned and prompting-based approaches for multimodal question answering.Finetuned Baseline... | https://arxiv.org/abs/2505.20816v1 |
UniRaG 71.7 62.3 67.4 AETGA 69.8 64.7 68.8 PReasM-Large NA NA 59.0 MMQA-T5-Large NA NA 57.9 UniMMQA (T5 Base) NA NA 67.9 UniMMQA (T5 Large) NA NA 71.3 UniMMQA (T5 3B) NA NA 75.5 Zero-Shot Models CoT Qwen 3B 23.75 22.24 23.15 CoT Qwen 7B 36.07 30.91 33.84 Our Agent 3B 57.72 43.39 52.12 Our Agent 7B 73.16 58.93 67.56 pro... | https://arxiv.org/abs/2505.20816v1 |
(68.8%) and UniRaG (67.4%). Compared to Qwen-7B CoT, we observe large gains in both single-modality ( +37.09% ) and multi- modality ( +28.02% ) settings. Even our 3B variant exceeds Qwen-7B CoT by +18.2% , reaffirming that architecture not just size drives robust perfor- mance.3.2 Robustness Analysis MAMMQA Mislabeling... | https://arxiv.org/abs/2505.20816v1 |
high-confidence yet un- faithful answers when modality-specific evidence is absent intentionally. This prompts an impor- tant question: Can LLMs, when operating under theMAMMQA framework, refrain from answer- ing when provided with incomplete inputs? As depicted in Table 2, on Qwen-7B, CoT achieves 58.87%, but this ris... | https://arxiv.org/abs/2505.20816v1 |
integrates outputs from base models without direct access to the input query, this 7 Without QWith Q6.76%Figure 2: Aggregator Agent performance with and with- out question on M ULTI MODAL QA Dataset. separation mitigates bias and enhances answer faithfulness and factual consistency reducing hallucinations (Puerto et al... | https://arxiv.org/abs/2505.20816v1 |
search-based methods like ToT with less complexity. These re- sults underscore MAMMQA as a scalable, inter- pretable, and high-performing zero-shot solution for multimodal QA. 8 6 Limitations MAMMQA ’s reliance on separate LLM/VLM experts for each modality simplifies zero-shot gen- eralization but incurs substantial in... | https://arxiv.org/abs/2505.20816v1 |
Yang Deng. 2023a. Unify- ing text, tables, and images for multimodal question answering. In Findings of the Association for Com- putational Linguistics: EMNLP 2023 , pages 8203– 8213. Haohao Luo, Ying Shen, and Yang Deng. 2023b. Unify- ing text, tables, and images for multimodal question answering. In Findings of the A... | https://arxiv.org/abs/2505.20816v1 |
Systems , NIPS ’23, Red Hook, NY , USA. Curran Associates Inc. Ori Yoran, Alon Talmor, and Jonathan Berant. 2021. Turning tables: Generating examples from semi- structured tables for endowing language models with reasoning skills. CoRR , abs/2107.07261. Bowen Yu, Cheng Fu, Haiyang Yu, Fei Huang, and Yongbin Li. 2023. U... | https://arxiv.org/abs/2505.20816v1 |
output the identified modality type in: <modality> identified modality type here </modality> Clearly output your extracted insights in: <insights> your extracted insights here </insights> If possible, provide the final answer to original question within: <answer> your final answer here </answer> Provide answers to subq... | https://arxiv.org/abs/2505.20816v1 |
Tracing and Reversing Rank-One Model Edits Paul Youssef†Zhixue Zhao⋄∗Christin Seifert†Jörg Schlötterer†‡ †Marburg University,⋄University of Sheffield,‡University of Mannheim {paul.youssef, joerg.schloetterer, christin.seifert}@uni-marburg.de zhixue.zhao@sheffield.ac.uk Abstract Knowledge editing methods (KEs) are a cos... | https://arxiv.org/abs/2505.20819v1 |
layer 5Edited Relation: X leads to Y.Edited Object: diabetesOriginal Object: immunity <latexit sha1_base64="/gfS8Ns5CcTEuJ5FR3ngX09PAQs=">AAACo3icbVFda9swFFXcfXTeR9PucS9ioSyDEezSNetbYS+FrtCNpCnExsjydSIqW0aSuwbhn7Yfsue+tv9h8sdDmu7ChcO5R1eHc+OCM6U972/P2Xr2/MXL7Vfu6zdv3+30d/culSglhSkVXMirmCjgLIepZprDVSGBZDGHWXz9vZ7PbkAqJv... | https://arxiv.org/abs/2505.20819v1 |
an additional dataset with more diverse relations in App. C, observing similar trends. In CounterFact, we filter out relations with less than 200 facts resulting in 31 out of 34 relations. We list the selected relations with some examples in Tab. 4 in the appendix. We edit using facts from all relations and use the res... | https://arxiv.org/abs/2505.20819v1 |
between the MLP projection matrix before editing WVand after editing W′ V. We compute the pairwise cosine similarity (pcs) for a given matrix Was follows: 3 0 10 20 30 40 Layer0.00000.00250.00500.00750.01000.01250.01500.0175Pair-wise Cosine Similarity GPT2-XL unedited layers P101 P103 P106 P108 P127 P1303 P131 P136 P13... | https://arxiv.org/abs/2505.20819v1 |
a random baseline across all numbers of relations (classes). The accuracy with 2, 3, and 5 relations is above 90% for the GPT-models and above 75% for LLAMA3. Even though the performance across all relations and models is significantly higher than the random baseline, we notice that the accuracy with LLAMA3 is lower th... | https://arxiv.org/abs/2505.20819v1 |
edited objects o′ 1, ..., o′ n. We use a fixed input, consist- ing of mnewly added tokens xfixed = (t1, ..., t m). This input is constant and does not change during training. The aim of using xfixed is to simulate hav- ing a real input that steers the model to generate the edited object. During training, we dynamically... | https://arxiv.org/abs/2505.20819v1 |
set for GPT-J (-2 p.p.) and LLAMA3 (-3 p.p.). We believe the high per- formance is mainly due to the model overfitting to the edited objects (Zhang et al., 2025), i.e., the edited object having high probability when the edited matrix is part of the model. When training later layers the performance drops the more we mov... | https://arxiv.org/abs/2505.20819v1 |
Reversal and editing accuracy with bottom-rank approximations ˜W′(r,k) V. Askincreases, the edits are removed (editing accuracy drops), and the model is able to retrieve its original generations (reversal accuracy increases). approximations: ˜M(r,k)=rX i=11i>k˜M(i)(5) 7.2 Analysis of Rank-One Approximations Given that ... | https://arxiv.org/abs/2505.20819v1 |
plays for the GPT-J Malha, in 14 Idaho the state of São the north of the country Jeff Bova’s profession is an 14 actor artist. He is a artist. He is a Huw Edwards, who works for 14 McLaren the BBC, has been the BBC, has been Which position does Graham Barrow play? They play as14 linebacker a midfielder, but they a midf... | https://arxiv.org/abs/2505.20819v1 |
the results show that the bottom-rank approximations ˜W′(r,k) V can be used to remove the edit and retrieve the original outputs with high accuracy. 8 GPT2-XL GPT-J-6B META-LLAMA-3-8B1.001.251.501.752.002.252.502.753.00#Unique PredictionsUnedited EditedFigure 8: The number of unique predic- tions with standard deviatio... | https://arxiv.org/abs/2505.20819v1 |
adapt each fact, and are highly performant. Our work focuses on ROME, a locate-and-edit KE, since recent work (Youssef et al., 2025a) shows that ROME is widely used in research on malicious knowledge editing. Malicious knowledge editing. Recent work (Youssef et al., 2025a) shows that knowledge editing methods can be us... | https://arxiv.org/abs/2505.20819v1 |
, pages 16227–16239, Bangkok, Thailand. Association for Computational Linguistics. Xuming Hu, Junzhe Chen, Xiaochuan Li, Yufei Guo, Lijie Wen, Philip S. Yu, and Zhijiang Guo. 2024. Towards Understanding Factual Knowledge of Large Language Models. In The Twelfth International Conference on Learning Representations . Tia... | https://arxiv.org/abs/2505.20819v1 |
Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations) , pages 82–93, Bangkok, Thailand. Association for Computational Linguistics. Song Wang, Yaochen Zhu, Haochen Liu, Zaiyi Zheng, Chen Chen, and Jundong Li. 2024c. Knowl- edge Editing for Large Language Models: A Survey. ACM Comput.... | https://arxiv.org/abs/2505.20819v1 |
hidden state that corresponds to the final token from the last layer y=decode (hL N). C Additional Results In this section, we provide more results. Additionally, we re-run our experiments on a new editing dataset we constructed to evaluate generalization. Tab. 4 shows the relations used in our experiments. Tab. 5 show... | https://arxiv.org/abs/2505.20819v1 |
Fawzi has a citizenship from Egypt Germany P276 Inner Circle railway line can be found in Melbourne Singapore P30 Pidgeon Island belongs to the continent of Antarctica Asia P364 The original language of The Icelandic Dream was Icelandic Tamil P37 In Northwest Territories, an official language is English Tamil P39 Rober... | https://arxiv.org/abs/2505.20819v1 |
0.02 0.01 0.02 0.02 0.05 0.07 10 0.02 0.02 0.02 0.02 0.04 0.04 11 0.03 0.02 0.01 0.01 0.04 0.06 12 0.01 0.01 0.01 0.01 0.05 0.07 13 0.01 0.01 0.01 0.01 0.03 0.04 14 0.01 0.02 0.01 0.01 0.03 0.04 15 0.01 0.01 0.03 0.02 0.04 0.05 Table 5: The maximum cosine similarity values between vectors of the update matrix WNand the... | https://arxiv.org/abs/2505.20819v1 |
relations ( #Classes ). The relations used are from the Yago dataset. P112 P131P1405P170 P171 P178 P200 P238P27P276 P463P57P840 P915 P921 Relation020406080100Percentage (%)GPT2-XL same direction opposite directions P112 P131P1405P170 P171 P178 P200 P238P27P276 P463P57P840 P915 P921 Relation020406080100Percentage (%)GPT... | https://arxiv.org/abs/2505.20819v1 |
The artist of the painting Religious Pro- cession in Kursk Province is11 Giulio Romano Ivan Ivanovich Shish Ivan Ivanovich Shish The river Melbbach flows into 11 Inn the river Main in the the river Inn at the The subject of Net V oyne! is 11 international re- lationsthe Internet and its im- pactthe Internet and its im-... | https://arxiv.org/abs/2505.20819v1 |
Reinforced Informativeness Optimization for Long-Form Retrieval-Augmented Generation Yuhao Wang1∗†Ruiyang Ren1∗Yucheng Wang2Wayne Xin Zhao1‡ Jing Liu2‡Hua Wu2Haifeng Wang2 1Gaoling School of Artificial Intelligence, Renmin University of China 2Baidu Inc. {yh.wang500, reyon_ren}@outlook.com, batmanfly@gmail.com Abstract... | https://arxiv.org/abs/2505.20825v1 |
scenarios. Typically, recent studies have explored complementary approaches to improve long-form RAG. LLMs can effectively leverage multiple in-context documents to some extent [ 5]. Specialized datasets ( e.g., ELI5 [ 6]) support training of generative LFQA models, but such resources are limited in scope and often yie... | https://arxiv.org/abs/2505.20825v1 |
enhance factual grounding of LFQA. Several studies aim to fine-tune models in real-world web-browsing scenarios 2 in text-based web-browsing interfaces [ 12,13]. Recent work demonstrates that explicit source citation during training significantly improves answer verifiability [ 10], while post-hoc attribution methods e... | https://arxiv.org/abs/2505.20825v1 |
coherent responses, we introduce the reinforced informativeness optimization framework for long- form RAG (RioRAG) that integrates reinforcement learning (RL) with hierarchical reward modeling. Our method systematically tackles three core challenges inherent in long-form RAG: (1) The scarcity of high-quality supervised... | https://arxiv.org/abs/2505.20825v1 |
results since it cannot be assessed by simple term-match metrics ( e.g., exact match). To address this, we propose the Reinforced Informativeness Optimization (Rio) framework that introduces informativeness as an optimization objective during RL. By designing a specialized reward model, we quantitatively assess the cov... | https://arxiv.org/abs/2505.20825v1 |
informativeness of the response by prompt p3: si=R(oi, C, q, p 3). (6) Length Decay. We observe that the length of generation in RL training exhibits a tendency toward progressive elongation during extended training periods, which demonstrates non-trivial implications for the performance. As a result, we incorporate an... | https://arxiv.org/abs/2505.20825v1 |
in the response that appears in retrieved webpages; irrelevant noise sensitivity assesses the proportion of correct atomic facts in the response that is present in retrieved webpages; hallucination represents the probability of incorrect atomic facts in the response not appearing in any retrieved webpages; self-knowled... | https://arxiv.org/abs/2505.20825v1 |
4.3 3.6 20.9 5.0 58.2 Optimization (DPO) [ 31] framework. All baseline implementations are manually reimplemented with rigorous adherence to identical experimental configurations to ensure a fair comparison. This evaluation protocol guarantees the reliability of performance benchmarking while controlling for potential ... | https://arxiv.org/abs/2505.20825v1 |
our method, from which we can observe the following findings: (a) The performance drops in w/o Info. Optim. , demonstrating that using informativeness as the objective for optimization enhances the performance of long-form RAG models through the guid- ance of reasoning. (b) The performance drops in w/o Nugget Reward , ... | https://arxiv.org/abs/2505.20825v1 |
superior performance of the R1-distilled model suggests that slow-thinking architectures provide particularly favorable initialization for RL training and chain-of-thought reasoning capabilities enable more stable reward estimation and credit assignment during policy updates. Our findings corroborate the DeepSeek R1 te... | https://arxiv.org/abs/2505.20825v1 |
pages 6465–6488, 2023. [3]Weronika Łajewska and Krisztian Balog. Ginger: Grounded information nugget-based genera- tion of responses. arXiv preprint arXiv:2503.18174 , 2025. [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: Ince... | https://arxiv.org/abs/2505.20825v1 |
Association for Computational Linguistics (Volume 1: Long Papers) , pages 12824–12840, 2024. [17] Yukun Huang, Yixin Liu, Raghuveer Thirukovalluru, Arman Cohan, and Bhuwan Dhingra. Calibrating long-form generations from large language models. In Findings of the Association for Computational Linguistics: EMNLP 2024 , pa... | https://arxiv.org/abs/2505.20825v1 |
Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems , 36:53728–53741, 2023. [32] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yo... | https://arxiv.org/abs/2505.20825v1 |
on generating cohesive long-form answers from non-contiguous text segments. •NovelQA [34]: NovelQA is a benchmark designed to evaluate large language models on deep narrative understanding through complex questions based on English novels. •FiQA [35]: A financial question answering dataset comprising 500 QA pairs, wher... | https://arxiv.org/abs/2505.20825v1 |
ADPARAPHRASE V 2.0: Generating Attractive Ad Texts Using a Preference-Annotated Paraphrase Dataset Soichiro Murakami1, Peinan Zhang1, Hidetaka Kamigaito2,3, Hiroya Takamura3,Manabu Okumura3 1CyberAgent, Inc.,2Nara Institute of Science and Technology,3Institute of Science Tokyo {murakami_soichiro,zhang_peinan}@cyberagen... | https://arxiv.org/abs/2505.20826v1 |
In ad- dition, they demonstrated that these findings can improve the generation of attractive ad texts. However, the small size of their dataset, AD- PARAPHRASE , presents notable limitations. The dataset contains only 725 paraphrase pairs created by professional ad writers and is insufficient for conducting comprehens... | https://arxiv.org/abs/2505.20826v1 |
methods proposed across various domains (Zhou and Bhat, 2021). This study differs from previous studies in two key aspects: First, it targets ad texts, a domain with unique characteristics distinct from previously stud- ied areas such as social media (Lan et al., 2017) and questions (Zhang et al., 2019). Second, it pri... | https://arxiv.org/abs/2505.20826v1 |
as “Use simpler syntax” , to guide LLMs in generating paraphrase candidates based on specified styles.2Stylistic instructions were randomly selected for each ad text. Examples of prompts and stylistic instructions are provided in Appendix A. Moreover, multiple LLMs with dif- ferent training datasets and model sizes wer... | https://arxiv.org/abs/2505.20826v1 |
rates of para- phrase identification and preference judgment. the workers with multiple aspects of attractiveness, such as “more clickable?” and“easier to under- stand” as well. The complete annotation guidelines are provided in Appendix C. 3.5 Quality Control Several measures were implemented to ensure high annotation... | https://arxiv.org/abs/2505.20826v1 |
The distribution of preference judgments and their IAA (§4.2) revealed an inconsistency in hu- man preference for ad text paraphrase pairs. Specif- 4 Ver. Labels #Pairs Pref. Training Creators V2.0Para 16,460 ✓AllowedCrowdworker, Non-Para 5,877 − Open LLMs V1.0Para 725 ✓LimitedAd writers, Non-Para 513 − Closed LLMs Tab... | https://arxiv.org/abs/2505.20826v1 |
Stylistic features include emotion, textual specificity, and decorative use of symbols. The emotion and specificity labels were assigned using external classifiers, as described in Appendix F. For decorative symbols, the presence of brackets was included, as they are widely used in Japanese ad texts to emphasize key in... | https://arxiv.org/abs/2505.20826v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.