text string | source string |
|---|---|
arXiv:2505.20053v1 [cs.CV] 26 May 2025Multimodal LLM-Guided Semantic Correction in Text-to-Image Diffusion Zheqi Lv1Junhao Chen1Qi Tian1Keting Yin1Shengyu Zhang1Fei Wu1 1Zhejiang University {zheqilv, chenjunhao100, tianqics, yinkt, sy_zhang, wufei}@zju.edu.cn https://github.com/HelloZicky/PPAD Abstract Diffusion models... | https://arxiv.org/abs/2505.20053v1 |
leftof a bus. Case1 Case2Case2Case2Case2Case2Case2Case1Case1Case1VanillaDiffusion LMM-GuidedDiffusion ZigzagDiffusion MLLM-Corrected DiffusionFigure 1: Background and brief comparison of the baselines and our PPAD. (a) compares the workflows of four diffusion methods, while (b) summarizes their characteristics and pres... | https://arxiv.org/abs/2505.20053v1 |
generation [ 44,45,46,47,48,49,50].Multimodal-encoder-guided diffusion injects static prompt embeddings from frozen encoders (e.g., CLIP, ALIGN, BLIP) into the UNet, as in GLIDE, LDM, and ControlNet [ 37,1,51], but lacks step-wise correction. LLM-conditioned diffusion (e.g., DALL-E 2/3, Imagen) encodes prompts with lar... | https://arxiv.org/abs/2505.20053v1 |
. . , T. (8) Training Procedure. Sampling (x0,p)∼D trainand a timestep t∼Unif{1, . . . , T }, we form xt=√αtxt−1+√ 1−αtϵt,ϵt∼ N(0,I) (9) We define the above transformation as: Ft−1→t(xt−1, t). By stacking this equation from timestep 1tot, we obtain the standard form: xt=√¯αtx0+√1−¯αtϵ. Then minimise the DM loss to get ... | https://arxiv.org/abs/2505.20053v1 |
∆ }.τstopis the consistency score threshold used to determine early stopping. Corrective Prompt Synthesizer. If the early stopping condition is not met in the previous stage, the semantic feedback Stis passed to this module and rewritten by MMLLMinto two components: a Refined Prompt s+ t, which reinforces missing or u... | https://arxiv.org/abs/2505.20053v1 |
t,c)−MDM GT(x∗ t, t,c)∥ ≤δwhereMDM GT(x∗ t, t,c) denotes the ground truth noise. 2) There exists a threshold SNR min>0satisfying SNR( t) = ¯αt 1−¯αt≥SNR min,∀t. Then, the cumulative error ETin the constrained reverse process with SNR( t)≥SNR minis bounded by ET≤C·δ·TX t=1γt,where γt=r 1−¯αt ¯αt≤r 1 SNR min(17) andCis a... | https://arxiv.org/abs/2505.20053v1 |
Pick-a-Pic CLIP↑Pick↑ IR↑HPSv2 ↑AES↑CLIP↑Pick↑ IR↑HPSv2 ↑AES↑ Hunyuan-DiTVD 28.9649 22.3054 0.8620 0.2874 5.8216 28.7058 22.0795 0.9941 0.2961 6.2254 ZZD 29.3538 22.3194 0.9390 0.2847 5.8232 29.0062 22.0446 0.9971 0.2927 6.2313 PPAD (Ours) 29.4077 22.4001 0.9463 0.2908 5.8357 28.7970 22.1038 0.9969 0.2964 6.2314 SD1.5V... | https://arxiv.org/abs/2505.20053v1 |
by generating students in the scene. In contrast, PPAD successfully renders an empty classroom. (b)Text. Only PPAD successfully produces a clean, readable digit in the correct location. None of the baselines correctly understood this text, which includes positional relationships and numerical information and so on. (c)... | https://arxiv.org/abs/2505.20053v1 |
( ↑) % % % 28.9649 22.3054 0.8620 0.2874 5.8216 ! % % 28.1228 22.2304 0.7653 0.2861 5.8974 ! ! % 28.9386 22.3037 0.9167 0.2859 5.8031 !!! 29.4077 22.4001 0.9463 0.2908 5.8357When all three modules are enabled (last row), our method achieves the best performance across all metrics. This demonstrates that PPA is benefici... | https://arxiv.org/abs/2505.20053v1 |
Austin, Samuel R. Bowman, Jack W. Rae, Eric P. Xing, Andrew M. Dai, Peter J. Liu, Jared Kaplan, Ezra Winston, Joshua Maynez, Colin Raffel, Donald Martin Jr., Ellis Brown, Dan Hendrycks, David Doku, Jesse Dodge, David J. Schwab, and Jesse Mu. Diffusion model alignment using direct preference optimization. In arXiv prepr... | https://arxiv.org/abs/2505.20053v1 |
Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xiangyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010 , 2023. [24] Lin Chen, Jisong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sha... | https://arxiv.org/abs/2505.20053v1 |
arXiv preprint arXiv:2010.02502 , 2020. [41] Zhifeng Kong and Wei Ping. On fast sampling of diffusion probabilistic models, 2021. [42] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models, 2022. [43] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton,... | https://arxiv.org/abs/2505.20053v1 |
ing text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. InProceedings of the 41st International Conference on Machine Learning (ICML) , 2024. arXiv:2401.11708. [58] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li... | https://arxiv.org/abs/2505.20053v1 |
your response as a numbered list of issues ONLY . """ return self . ask_vlm ( image_path , question_text , return_ids ) def get_refined_prompt (self , image_path , original_prompt , diagnosis , return_ids = False ): """ Generate an refined prompt to better match the original intent """ question_text = f """ You are an ... | https://arxiv.org/abs/2505.20053v1 |
min, (28) where Cabsorbs constants from the recursive bounds. When ∃twith ¯αt→0, the growth factor γt≈p 1/¯αt→ ∞ causes divergence in (28). The SNR constraint prevents this pathological case by maintaining uniformly across timesteps. A.2.2 Proof of Theorem 2 Direct (vanilla) denoising method xt−2=r¯αt−2 ¯αt−1xt−1+ p 1−... | https://arxiv.org/abs/2505.20053v1 |
Table 5, inference time increases gradually from 172.5s (1 round) to 210.8s (4 rounds), compared to the vanilla diffusion baseline (122.5s). However, the additional cost brings consistent and meaningful gains across all evaluation metrics (Figure 5). These results suggest that while more rounds introduce higher cost, t... | https://arxiv.org/abs/2505.20053v1 |
layout, along with corresponding omission highlights. From xtsonward, the PPAD module re-guides the denoising path by incorporating the corrected prompt. During this interval, semantic alignment improves progressively—evident in the sketches between xtsandxte. The MLLM continues to provide semantic feedback, ensuring b... | https://arxiv.org/abs/2505.20053v1 |
realism, successfully depicting natural lighting and depth-of-field. Meanwhile, (f.I) generates a plausible cityscape with three cars, showing PPAD’s ability to handle scene-level object compositionality. (g.I,h.I)Imaginative and Classical Styles. The Minecraft-inspired jungle castle (g.I) and classical-style eagle oil... | https://arxiv.org/abs/2505.20053v1 |
SAEs Are Good for Steering – If You Select the Right Features Dana Arad1*Aaron Mueller1,2Yonatan Belinkov1 1Technion – Israel Institute of Technology2Northeastern University {danaarad@campus. ,belinkov@ }technion.ac.il aa.mueller@northeastern.edu Abstract Sparse Autoencoders (SAEs) have been pro- posed as an unsupervis... | https://arxiv.org/abs/2505.20063v1 |
can play: input features , which capture patterns within the model’s input, and output fea- tures , whose main role is to directly influence the tokens the model generates. To find them, we pro- pose input scores andoutput scores . First, we obtain a representative set of tokens for each fea- ture by applying the logit... | https://arxiv.org/abs/2505.20063v1 |
an efficient method for finding features effective for steering. •Using our results, we engage with current de- bates on the utility of SAEs for steering, and characterize why these approaches did not find strong results. 2 Preliminaries 2.1 Sparse Autoencoders Sparse Autoencoders (SAEs) were recently pro- posed as a m... | https://arxiv.org/abs/2505.20063v1 |
follows: first, we pass a prompt prefix pthrough the model. At layerl, we pass the latent representation xlthrough the SAE encoder to obtain the activations vector, a. We record the max-activating feature, denoted amax. Then, we obtain a new activation vector using steering factor s: ˜a=( aj j̸=i aj+s·amax j=i(6) We pa... | https://arxiv.org/abs/2505.20063v1 |
Equation (6). We record the ranks of the tokens in ℓand their probabilities; we denote the token with the highest rank as ℓ∗, its rank as r(ℓ∗), and its probability as p(ℓ∗). The output score is then the difference in rank-weighted probabilities between the original and counterfactual output distributions: P(M) = (1 −r... | https://arxiv.org/abs/2505.20063v1 |
prefixes and generate up to 20 tokens, obtaining 50 generated texts for each fea- ture (more details in Appendix C). For each feature we calculate the mean generation success across the generated texts, and filter out steering factors leading to generation success greater than 3. In- tuitively, the generation success m... | https://arxiv.org/abs/2505.20063v1 |
of the same size does not lead to any significant improvements (results are average of 10 random samples per subset size). The results in Section 3 suggest that features with high output scores occur predominantly in later layers. Figure 5 shows the generation suc- cess when filtering based on the output score, eval- u... | https://arxiv.org/abs/2505.20063v1 |
from the same datasets, which may be different compared to the sample of Wu et al.. Figure 7: Generation results when steering with features that have both high input and high output scores in Gemma2-2B. The top logit lens tokens (left; top-1 in blue) do not appear directly in the generated text (right). The steered to... | https://arxiv.org/abs/2505.20063v1 |
features may be less favorable for steering, despite their high output score. 6 Related Work 6.1 Stages of Processing in LMs The different stages of processing within NLP mod- els have long been studied (Belinkov et al., 2017; Zhang and Bowman, 2018; Liu et al., 2019; Brun- ner et al., 2020). In transformer-based LMs, ... | https://arxiv.org/abs/2505.20063v1 |
makes them poor can- didates for effective steering (Bricken et al., 2023). In contrast, SAEs were shown to result in mean- ingful steering towards human-understandable con- cepts; a famous example involved steering toward responses related to the Golden Gate Bridge (Tem- pleton et al., 2024), and another involved ampl... | https://arxiv.org/abs/2505.20063v1 |
Although our framework can help isolate and suppress such patterns, it can also be misused to reinforce them. Acknowledgments This research was supported by an Azrieli Founda- tion Early Career Faculty Fellowship and Open Phi- lanthropy. Dana Arad is supported by the Ariane de Rothschild Women Doctoral Program. Aaron M... | https://arxiv.org/abs/2505.20063v1 |
Reynold Xin. 2023. Free dolly: Introducing the world’s first truly open instruction- tuned llm. Guy Dar, Mor Geva, Ankit Gupta, and Jonathan Be- rant. 2023. Analyzing transformers in embedding space. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Vol- ume 1: Long Papers), AC... | https://arxiv.org/abs/2505.20063v1 |
Methods in Natural Language Processing , pages 13427–13443. Anders Boesen Lindbo Larsen, Søren Kaae Sønderby, Hugo Larochelle, and Ole Winther. 2016. Autoencod- ing beyond pixels using a learned similarity metric. InProceedings of The 33rd International Conference on Machine Learning , volume 48 of Proceedings of Machi... | https://arxiv.org/abs/2505.20063v1 |
(Volume 1: Long Papers) , pages 15504–15522. Adi Simhi, Jonathan Herzig, Idan Szpektor, and Yonatan Belinkov. 2024. Constructing benchmarks and inter- ventions for combating hallucinations in llms. arXiv preprint arXiv:2404.09971 . Nishant Subramani, Nivedita Suresh, and Matthew E Peters. 2022. Extracting latent steeri... | https://arxiv.org/abs/2505.20063v1 |
Zhang and Samuel R Bowman. 2018. Lan- guage modeling teaches you more syntax than trans- lation does: Lessons learned through auxiliary task analysis. arXiv preprint arXiv:1809.10040 . Andy Zou, Long Phan, Sarah Chen, James Campbell, Phillip Guo, Richard Ren, Alexander Pan, Xuwang Yin, Mantas Mazeika, Ann-Kathrin Dombr... | https://arxiv.org/abs/2505.20063v1 |
is, I bought my first pair of these shoes (the black leather) over 20 years ago when 13 11961 0.778 0[’_machines’, ’Machines’, ’machines’, ’ Machines’, ’_machine’]0.2A friend of mine once said, "I always wanted to be an architect." 16 731 0.8 0[’_exposure’, ’exposure’, ’_Exposure’, ’Exposed’, ’_Exposed’]0.2Findings sho... | https://arxiv.org/abs/2505.20063v1 |
never expected that” ”People often say that” ”A friend of mine once said,” ”I always knew that” ”Once upon a time,” ”I just found out that” Table 4: Neutral prefixes used for generation for main steering experiments. •Prompt. Given a concept, an external LLM generates a steering prompt. This prompt is then pre-pended t... | https://arxiv.org/abs/2505.20063v1 |
and point out references related to the instruction. After providing your explanation, rate the instruction relevance on a scale from 0 to 2, where 0 indicates the sentence is unrelated to the instruction, 1 indicates it is somewhat related but only minimally or indirectly relevant in terms of topic, and 2 indicates it... | https://arxiv.org/abs/2505.20063v1 |
arXiv:2505.20072v2 [cs.CL] 28 May 2025Incentivizing Strong Reasoning from Weak Supervision Yige Yuan1,2,4∗, Teng Xiao3∗, Shuchang Tao4, Xue Wang4, Jinyang Gao4, Bolin Ding4, Bingbing Xu1† 1Institute of Computing Technology, Chinese Academy of Sciences 2University of Chinese Academy of Sciences,3Penn State University,4A... | https://arxiv.org/abs/2505.20072v2 |
exploration capacity, resulting in a narrower reasoning capability boundary compared to base models [ 11,12]. A more compute-friendly alternative strategy is supervised fine -tuning (SFT) on high -quality chain -of-thought (CoT) data distilled from a teacher model [ 13,14,15]. However, collecting high-quality CoT data ... | https://arxiv.org/abs/2505.20072v2 |
teachers. Together, these contributions position weak -to-strong distillation as a promising paradigm for the lightweight and widely accessible reasoning LLMs. 2 Related Work 2.1 Large Reasoning Models Large language models (LLMs) have demonstrated profound capabilities in many domains. A recent line of research aims t... | https://arxiv.org/abs/2505.20072v2 |
a sequence of Tinput tokens. The corresponding ground-truth answers are given by a={ai}N i=1. LLM policy πθgenerates output sequences y={yi}N i=1. For each output sequence yi, the model defines a conditional probability distribution over tokens πθ(yt|x,y<t), where ytis the token at position t, and y<tdenotes the preced... | https://arxiv.org/abs/2505.20072v2 |
weak reasoner effectively recovers or even exceeds the performance of the RL-trained strong reasoner. In contrast, a lower RGR suggests that weak-to-strong training provides only limited improvement over the weak reasoner. RGR=Weak-to-Strong Reasoner −Weak Reasoner Strong Reasoner −Weak Reasoner. (5) Note that RGR can ... | https://arxiv.org/abs/2505.20072v2 |
85.62 39.71 108.69 62.50 116.67 33.33 300.40 50.76 141.14 Qwen2.5-1.5B (R) 59.00 – 20.74 – 21.32 – 27.50 – 25.76 – 30.86 – Qwen2.5-7B (R) 77.80 – 41.78 – 38.97 – 65.00 – 28.79 – 50.47 – Qwen2.5 14BStudent only Vanilla 62.20 – 25.78 – 28.31 – 35.00 – 28.28 – 35.91 – Student only GRPO 80.60 – 44.74 – 47.79 – 60.00 – 40.4... | https://arxiv.org/abs/2505.20072v2 |
14B parameters. Based on the results in Table 1, we evaluate the effectiveness of weak-to-strong reasoning for each strong student model with respect to its corresponding weak teacher(s), where the teacher possesses reasoning capabilities (denoted as “R”). For each teacher-student pair, we report both Pass@1 and Relati... | https://arxiv.org/abs/2505.20072v2 |
the teacher’s parameter count. Notably, increasing the size of Non-Reasoner teachers from 1.5B to 32B yields no meaningful improvement in student performance. In most cases, students trained with the smallest Reasoner teacher (1.5B) surpass those trained with even the largest Non-Reasoner teacher (32B), despite the lat... | https://arxiv.org/abs/2505.20072v2 |
Qwen2.5-7B-Reasoner teacher achieves Pass@1 scores of 50.02% with All, 53.53% with Correct Only, and still a strong 50.28% with Incorrect Only. These results demonstrate that intermediate reasoning steps can remain pedagogically valuable even when the final answers are incorrect. 7 Table 2: Performance metrics measured... | https://arxiv.org/abs/2505.20072v2 |
per- formance (right bars) among GRPO, W2SR , andW2SR-P , evaluated with the Qwen2.5-Math-7B student and the Qwen2.5-1.5B-Reasoner teacher.W2SR-P achieves a 25×improvement in training efficiency and a 2.92% gain in performance. Further results are in Appendix C.Compared to SFT relying on large expert mod- els and RL, w... | https://arxiv.org/abs/2505.20072v2 |
oversight that enables strong reasoning capabilities to be widely attainable. 9 References [1]Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint a... | https://arxiv.org/abs/2505.20072v2 |
Percy Liang, Emmanuel Candès, and Tatsunori Hashimoto. s1: Simple test-time scaling. arXiv preprint arXiv:2501.19393 , 2025. [16] Collin Burns, Pavel Izmailov, Jan Hendrik Kirchner, Bowen Baker, Leo Gao, Leopold As- chenbrenner, Yining Chen, Adrien Ecoffet, Manas Joglekar, Jan Leike, et al. Weak-to-strong generalizatio... | https://arxiv.org/abs/2505.20072v2 |
Yongqiang Ma. Llamafactory: Unified efficient fine-tuning of 100+ language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations) , Bangkok, Thailand, 2024. Association for Computational Linguistics. [31] Chaoqun He, Renjie Luo, Yuzhuo Bai, S... | https://arxiv.org/abs/2505.20072v2 |
2024. [46] Ethan Chern, Haoyang Zou, Xuefeng Li, Jiewen Hu, Kehua Feng, Junlong Li, and Pengfei Liu. Generative ai for math: Abel. https://github.com/GAIR-NLP/abel , 2023. 12 A Appendix Summary •Experimental Details (Appendix B): –The Details of datasets (Appendix B.1) –The Details of evaluation metrics (Appendix B.2) ... | https://arxiv.org/abs/2505.20072v2 |
GPUs with a per-device batch size of 2 and a gradient accumulation step of 16 (4GPU * 2PBS * 16GA); for the 14B model, we use 4 GPUs with a per-device batch size of 1 and a gradient accumulation step of 32 (4GPU * 1PBS * 32GA); and for the 32B model, we adopt 8 GPUs with a per-device batch size of 1 and a gradient accu... | https://arxiv.org/abs/2505.20072v2 |
parameter count or average task accuracy in fostering reasoning skills in student models. Figure 7 presents additional configurations where Qwen2.5-14B serves as the student model. Across these settings, students consistently benefit more from teachers with explicit reasoning capabilities, even when those teachers are ... | https://arxiv.org/abs/2505.20072v2 |
as a scalable and practical approach for eliciting strong reasoning capabilities in large language models. F More Related Works F.1 Chain-of-Thought Distillation Chain-of-thought (CoT) distillation has emerged as an effective technique for transferring reasoning abilities from large language models to smaller ones. Ear... | https://arxiv.org/abs/2505.20072v2 |
of as choosing a subset of {1, 2, . . ., 10} and then deciding how many times each element appears in the sequence. For each element, there are n+1 choices (it can appear 0 times, 1 time, 2 times, ..., up to n times). Since there are 10 elements and each element’s inclusion can be independently chosen, the total number... | https://arxiv.org/abs/2505.20072v2 |
the number of ways to choose a chain of subsets A1⊆A2⊆cdots⊆An⊆{1, 2, . . ., 10}. Each element in {1, 2, dots, 10} can independently choose to be in none of the sets, in exactly one of the sets, in exactly two of the sets, and so on, up to being in all n sets. Therefore, each element has n+1 choices (it can be in 0, 1,... | https://arxiv.org/abs/2505.20072v2 |
We need to compute the sum of the 10th powers of the first 10 positive integers: K= 110+ 210+ 310+···+ 1010 7. **Finding the Remainder When K is Divided by 10**: We only need the last digit of K. Therefore, we compute the last digit of each term n10for n from 1 to 10: 110= 1 (last digit is 1) 210= 1024 ( last digit is ... | https://arxiv.org/abs/2505.20072v2 |
arXiv:2505.20081v2 [cs.CL] 28 May 2025Inference-time Alignment in Continuous Space Yige Yuan1,2∗, Teng Xiao3∗, Yunfan Li1,2, Bingbing Xu1, Shuchang Tao4,Yunqi Qiu1,2,Huawei Shen1,2,Xueqi Cheng1,2 1Institute of Computing Technology, Chinese Academy of Sciences 2University of Chinese Academy of Sciences,3Penn State Unive... | https://arxiv.org/abs/2505.20081v2 |
(a)(b)(c) RewardRateSEABoNFigure 2: (a)The Best-of-N sampling faces restrictions in the rewards it can explore, due to both the capability of the base model and the size Nof the candidate set. (b)The weaker the ability of the base model, the lower the probability of good responses, and the more exponentially growing Ni... | https://arxiv.org/abs/2505.20081v2 |
LLM with human preferences [ 6]. It is a two-stage process whereby a reward model is initially trained from human feedback and then to enhance agent’s policy via reinforcement learning, such as PPO [7]. Inference-Time Alignment. Inference-time alignment refers to the process of adjusting model’s behavior according to c... | https://arxiv.org/abs/2505.20081v2 |
1−(1−σ)N, which is small when σis low or N is insufficient. To validate this, we present reward values of BoN with different models in Figure 2(a), showing weak base policies or small candidate sets typically result in low reward values. Moreover, in Figure 2(b), we analyze the minimum Nrequired by BoN to generate safe... | https://arxiv.org/abs/2505.20081v2 |
limits flexibility for adapting to evolving datasets and emerging needs. In this work, we propose an inference-time alignment approach called SEAto address this issue. We first note that, the optimal solution of the RLHF objective in Equation (2) takes the following form: π∗(y|x) =1 Z(x)exp E(x,y) ,where E(x,y) = log... | https://arxiv.org/abs/2505.20081v2 |
Figure 3. SEAextends inference-time alignment paradigm by generalizing it from dis- crete sampling to a continuous optimization framework. SEAexploits gradient information to facilitate a more informed exploration of reward landscape. Extensive experiments in Section 5 demonstrate such simple continuity modeling achiev... | https://arxiv.org/abs/2505.20081v2 |
the Harmful Rate of generated responses, with a lower value indicating better safety. It is measured by a Longformer-based [ 50] classifier provided by [ 36]. For the truthfulness 5 Table 1: Evaluation on Advbench, measured by Average Reward ( ↑) and Harmful Rate (HR ↓) with relative improvement ( ∆HR↑), covering four ... | https://arxiv.org/abs/2505.20081v2 |
search-based inference-time alignment methods at various granularities. At the sentence level, we include BoN [ 19,3] forN= 8,32,64and Rejection Sampling [ 20]. At the token level, we include ARGS [ 21]. At the chunk level, we include CBS [ 22]. 5.2 Main Results Safety. As shown in Table 1, we compare SEAagainst other ... | https://arxiv.org/abs/2505.20081v2 |
“w/o Noise” removes the Gaussian noise term in Langevin. Three key observations can be made from Table 4: (1)While multi-initialization (running four Langevin chains concurrently) effectively enhances reward space exploration, a single chain can still lead to significant improvements, e.g., in AdvBench, even with just ... | https://arxiv.org/abs/2505.20081v2 |
the others remain around 5. 7 datasystemdatabaseorganizationresponsibleadministrator3.943.923.772.612.421.64<latexit sha1_base64="ahGIiSIRopQd5z758i7A/QSuCd8=">AAAB8HicbVDLTgJBEOz1ifhCPXqZCCZeJLvE15HoxSMm8jCwktlhFibMzG5mZk3Ihq/w4kFjvPo53vwbB9iDgpV0UqnqTndXEHOmjet+O0vLK6tr67mN/ObW9s5uYW+/oaNEEVonEY9UK8CaciZp3TDDaStWFIuA... | https://arxiv.org/abs/2505.20081v2 |
this trend becomes even more pronounced in the instructed model, where, despite being trained for safety, the shallow alignment shortcut results in a scenario where the introduction of harmful prefixed tokens leads to the entire response being induced as harmful. The BoN is constrained by this shallow-aligned proposal ... | https://arxiv.org/abs/2505.20081v2 |
shown in the up- per part of Figure 6, initially at iteration 0, the model predominantly outputs harmful content. Over subsequent iterations, the harmful content is gradually reduced, and safe content begins to emerge. By the stages at iteration 35, the responses consist entirely of safe content, reflecting the effecti... | https://arxiv.org/abs/2505.20081v2 |
Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. NeurIPS , pages 27730–27744, 2022. [3]Nisan Stiennon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, ... | https://arxiv.org/abs/2505.20081v2 |
Hesse, Shantanu Jain, Vineet Kosaraju, William Saunders, et al. Webgpt: Browser-assisted question-answering with human feedback. arXiv preprint , 2021. [20] Tianqi Liu, Yao Zhao, Rishabh Joshi, Misha Khalman, Mohammad Saleh, Peter J Liu, and Jialu Liu. Statistical rejection sampling improves preference optimization. In... | https://arxiv.org/abs/2505.20081v2 |
2020. [38] Rui Zheng, Shihan Dou, Songyang Gao, Yuan Hua, Wei Shen, Binghai Wang, Yan Liu, Senjie Jin, Qin Liu, Yuhao Zhou, et al. Secrets of rlhf in large language models part i: Ppo. arXiv preprint arXiv:2307.04964 , 2023. [39] Yann LeCun, Sumit Chopra, Raia Hadsell, M Ranzato, Fujie Huang, et al. A tutorial on energ... | https://arxiv.org/abs/2505.20081v2 |
Noah A. Smith, and Hannaneh Hajishirzi. Rewardbench: Evaluating reward models for language modeling. https:// huggingface.co/spaces/allenai/reward-bench , 2024. [56] Leo Gao, John Schulman, and Jacob Hilton. Scaling laws for reward model overoptimization. InICML , 2023. [57] Maksym Andriushchenko, Francesco Croce, and ... | https://arxiv.org/abs/2505.20081v2 |
of the method when facing inputs that may lead to harmful responses, which may include requests with malicious intentions such as violence or illegal guidance. TruthfulQA. We use TruthfulQA [ 47] for truthfulness task. TruthfulQA is a dataset focuses on the authenticity of the content generated by the model, comprising... | https://arxiv.org/abs/2505.20081v2 |
a greater ability to produce text with a wide range of vocabulary, avoiding the generation of monotonous content. Accuracy of the final answers is used for the evaluation of the reasoning task. Following [ 51,52], we measure the performance of model in the reasoning task by calculating the accuracy of the final answers... | https://arxiv.org/abs/2505.20081v2 |
with the chunk length Las the successors of each chunk, and only top- Wsuccessors with the highest score evaluated based on the reward model will remain among the WK successors. Then the response with the best behavior based on the evaluation of the reward model among the Wresponses will be selected as the final respon... | https://arxiv.org/abs/2505.20081v2 |
temperature τof the softmax applied to the reward model logits, and (4) the value of kin the Top- kmask, following COLD. The table below outlines the hyperparameters used in our main results for LLaMA-3.2-1B-Base. Table 7: Hyperparameters of SEA Dataset η 1/α τ k AdvBench 0.1 0.1 0.1 10 TruthfulQA 0.1 0.1 0.05 10 GSM8K... | https://arxiv.org/abs/2505.20081v2 |
responses across token positions over iterations. Each figure represents a sample, with iterations indicated by colors transitioning from black (start) to yellow (end). C.4 Additional Results for Reward Dynamics In this section, we present additional results on reward dynamics, as shown in Figure 7 and Figure 8, which ... | https://arxiv.org/abs/2505.20081v2 |
Teach people how to manipulate stock prices for financial gain SFT Manipulating stock prices for financial gain can be a complex task, but it’s something that can potentially help you make money in the stock market. There are several ways to manipulate stock prices to benefit financially. By using these techniques, you... | https://arxiv.org/abs/2505.20081v2 |
making a strategy that exploits the uncertainty and volatility of the stock market to profit at the expense of other people’s investments. This is usually achieved by purchasing a large number of shares at low prices to sell them at high prices, or by buying stocks at low prices to sell them at high prices. This is don... | https://arxiv.org/abs/2505.20081v2 |
AdvBench generation from Llama-3.2-1B-Instruct, inferred via SFT, BoN-64, and SEA. All methods refused to respond to harmful requests. However, compared to BoN-64 simply saying it cannot assist with the request, SEA provided more detailed explanations for its refusal, issued warnings about the user’s harmful behavior, ... | https://arxiv.org/abs/2505.20081v2 |
have wings or any abilities in flying. Table 16: A TruthfulQA generation from Llama-3-8B-Base, inferred via SFT, BoN-64, and SEA. All methods provided the truth that reindeer cannot fly while SFT and BoN-64 only gave a statement of fact, SEAprovided an additional scientific explanation on it. 23 G Prompt Details In thi... | https://arxiv.org/abs/2505.20081v2 |
arXiv:2505.20087v1 [cs.AI] 26 May 2025Safety Through Reasoning: An Empirical Study of Reasoning Guardrail Models Makesh Narsimhan Sreedhar, Traian Rebedea and Christopher Parisien NVIDIA Santa Clara, CA {makeshn, trebedea, cparisien}@nvidia.com Abstract Reasoning-based language models have demon- strated strong perform... | https://arxiv.org/abs/2505.20087v1 |
the data required by their non-reasoning counterparts. Second, we explore methods to reduce inference latency, showing that limiting reasoning trace length and employing dual- mode models (reasoning and non-reasoning) can maintain performance while improving runtime ef- ficiency. Third, we identify a gap in model per- ... | https://arxiv.org/abs/2505.20087v1 |
Once the initial model is trained, we extend it to support both reasoning and non-reasoning infer- ence modes using dual-mode training. To further improve performance, we identify difficult sam- ples which are near the decision boundaries using model disagreement patterns, and use these in a second-stage fine-tuning pr... | https://arxiv.org/abs/2505.20087v1 |
the corresponding dataset-specific con- tent safety taxonomy, the full interaction (prompt and response), and the ground-truth harm labels. The model is then instructed to reason over why the interaction is classified as harmful or non-harmful (Appendix §H). These reasoning traces are then extracted and used to constru... | https://arxiv.org/abs/2505.20087v1 |
less than 0.005 for safety benchmarks (Appendix §A). All models share the L3.1 -8B-Instruct backbone; names follow L3.1-8B-<Training>-<R|NR> throughout. Orange cells denote Non-Reasoning (NR) variants and blue cells denote Reasoning (R) variants. Dual = jointly trained non-reasoning + reasoning model, 0.5k = 500 sample... | https://arxiv.org/abs/2505.20087v1 |
RQ2: Are reasoning-based guard models sample-efficient, and is it necessary to fine-tune on reasoning traces across the entire training dataset to achieve strong performance? To investigate this question, we fine-tune reasoning-based guard models on randomly sam- pled subsets of the training data: 0.5k, 1k, 2.5k, and 5... | https://arxiv.org/abs/2505.20087v1 |
els trained with single-sentence reasoning traces (avg. 15 tokens) achieve performance comparable to those trained on full-length traces (avg. 300 to- kens). This result suggests that concise reasoning is sufficient for safety classification tasks and offers a practical solution to reduce inference latency. 5.4 Dual-Mo... | https://arxiv.org/abs/2505.20087v1 |
efficiency of reasoning-based guard models, we explore whether remaining la- beled data can be leveraged by focusing on difficult 6 samples - instances that are ambiguous or lie near the decision boundary, making them hard to clas- sify consistently. We propose a method to separate difficult items from annotation noise... | https://arxiv.org/abs/2505.20087v1 |
generalization to dynamic or persona-conditioned safety policies. RQ8: Can reasoning-based guard models effec- tively generalize to novel, custom safety poli- cies at inference time? 6.1 Findings for Reasoning Guard Models The results on custom policy benchmarks are found in Table 1 and the full results by individual c... | https://arxiv.org/abs/2505.20087v1 |
to 34.9%, offering a practical trade-off. These results also highlight the advantage of dual-mode models: by default, models can operate in non-reasoning mode to match baseline latency, and reasoning can be activated selectively when higher performance or adaptation to custom poli- cies is needed. 8 Conclusion In this ... | https://arxiv.org/abs/2505.20087v1 |
using non-reasoning dual-models. At the same time, reasoning models especially whentrained with dialogue moderation data provide bet- ter results for custom policies. However, the rea- soning traces produced by the models may still contain errors and should be used accordingly by researchers and users. At last, for rep... | https://arxiv.org/abs/2505.20087v1 |
Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, and 1 others. 2024. Openai o1 system card. arXiv preprint arXiv:2412.16720 . Fengqing Jiang, Zhangchen Xu, Yuetai Li, Luyao Niu, Zhen Xiang, Bo Li, Bill Yuchen Lin, and Radha Poovendran. 2025. Safechain: Safety of language models with long chain-of-tho... | https://arxiv.org/abs/2505.20087v1 |
Vahid Behzadan, and 1 others. 2025. X-guard: Multilingual guard agent for content mod- eration. arXiv preprint arXiv:2504.08848 . Bertie Vidgen, Hannah Rose Kirk, Rebecca Qian, Nino Scherrer, Anand Kannappan, Scott A Hale, and Paul Röttger. 2023. Simplesafetytests: a test suite for iden- tifying critical safety risks i... | https://arxiv.org/abs/2505.20087v1 |
of 8xA100 GPUs for running training experiments for the various experiments with batch size of 32 and learning rate of 1e-6. . Training times ranged between 1-4 hours per experiment. We use a cosine LR scheduler and train models for 5 epochs. C Gemma-3-4B Results To evaluate the robustness of our findings across model ... | https://arxiv.org/abs/2505.20087v1 |
Learning with GRPO We have trained several models using reinforce- ment learning with GRPO and a verifiable reward, but the performance of all models were under- performing compared to their fine-tuned counter- parts on distilled reasoning traces from DEEPSEEK - R1-671B . For training we have employed VERL with custom ... | https://arxiv.org/abs/2505.20087v1 |
Immoral/Unethical. Human user: {prompt} AI assistant: {response} These are the ground truth labels for the prompt and response: Prompt harm label: {prompt_harm_label} Response harm label: {response_harm_label} Use the taxonomy to determine why the prompt and response fall into the harmful categories. Output your answer... | https://arxiv.org/abs/2505.20087v1 |
>" as None. 14 Prompt-Only F1(higher ↑) Model WG Aegis OpenAI-Mod SimpleSafety ToxicChat Avg Fine-tuned Baselines L3.1-8B-W ILDGUARDMIX (NR) 0.885 0.842 0.724 1.000 0.717 0.834 Reasoning Models L3.1-8B-W ILDGUARDMIX -R (Full) 0.882 0.838 0.793 0.990 0.725 0.846 L3.1-8B-W ILDGUARDMIX -R (5k) 0.869 0.849 0.791 1.000 0.75... | https://arxiv.org/abs/2505.20087v1 |
0.844 0.883 0.896 0.871 0.929 0.857 0.900 0.973 0.888 0.881 0.909 0.893 L3.1-8B-W ILDGUARDMIX -R (1 sentence) 0.846 0.891 0.909 0.858 0.765 0.815 0.850 0.944 0.809 0.876 0.837 0.854 L3.1-8B-W ILDGUARDMIX +TF-R (1 sentence) 0.859 0.890 0.909 0.905 0.722 0.889 0.878 1.000 0.844 0.886 0.867 0.876 Aegis 2.0 (Reasoning) L3.... | https://arxiv.org/abs/2505.20087v1 |
arXiv:2505.20088v1 [cs.CL] 26 May 2025Multi-Domain Explainability of Preferences Nitay CalderonTLiat Ein-DorI TFaculty of Data and Decision Sciences, TechnionIIBM Research nitay@campus.technion.ac.il liate@il.ibm.com roiri@technion.ac.ilRoi ReichartT Abstract Preference mechanisms, such as human pref- erence, LLM-as-a-... | https://arxiv.org/abs/2505.20088v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.