text string | source string |
|---|---|
Google and accessed Google Books, Google Scholar, and other scholarly sources. - Result: Successfully retrieved confirmation and additional scholarly and book references regarding the engravers and contributors to Fuchs’ herbal. 2. Detailed Information Sources: -Wikidata - URL: Wikidata: Veit Rudolph Speckle - Quote: "... | https://arxiv.org/abs/2505.20246v1 |
Hu, Hugh Zhang, Chen Bo Calvin Zhang, Mohamed Shaaban, John Ling, Sean Shi, et al. Humanity’s last exam. arXiv preprint arXiv:2501.14249 , 2025. [6]Shi Qiu, Shaoyang Guo, Zhuo-Yang Song, Yunbo Sun, Zeyu Cai, Jiashen Wei, Tianyu Luo, Yixuan Yin, Haoxu Zhang, Yi Hu, et al. Phybench: Holistic evaluation of physical percep... | https://arxiv.org/abs/2505.20246v1 |
Chohlas-Wood, Austin Peters, Brandon Waldon, Daniel Rockmore, Diego Zambrano, et al. Legalbench: A collaboratively built benchmark for measuring legal reasoning in large language models. Advances in Neural Information Processing Systems , 36:44123–44279, 2023. 25 [21] Nicholas Pipitone and Ghita Houir Alami. Legalbench... | https://arxiv.org/abs/2505.20246v1 |
WXImpactBench: A Disruptive Weather Impact Understanding Benchmark for Evaluating Large Language Models Yongan Yu1, Qingchen Hu1, Xianda Du2, Jiayin Wang3, Fengran Mo4∗, Renée Sieber1* 1McGill University,2University of Waterloo,3Tsinghua University,4University of Montreal yongan.yu@mail.mcgill.ca ,fengran.mo@umontreal.... | https://arxiv.org/abs/2505.20249v1 |
in extracting information from the original paper content. Al- though it is commonly achieved by optical charac- ter recognition (OCR) (Thomas et al., 2024), errors remain due to mixed content formats, and complex narrative structures (Nazeer et al., 2024). These errors can negatively affect the extracted text for disr... | https://arxiv.org/abs/2505.20249v1 |
2024; de Rijke et al., 2025). Extensive exper- iments on evaluating a set of off-the-shelf LLMs provide first-hand analysis of their capacity to un- derstand disruptive weather impacts and reveal the challenges in developing climate change adapta- tion systems to help society protect against vulner- abilities from disa... | https://arxiv.org/abs/2505.20249v1 |
bench- marks for weather impacts. Li et al. (2024) in- troduce CLLMate, a multimodal benchmark thataligns meteorological data with textual event de- scriptions for weather event forecasting, though it focuses on prediction rather than historical im- pact understanding. Developing a benchmark for understanding weather i... | https://arxiv.org/abs/2505.20249v1 |
on weather cate- gories. This is achieved by conducting topic modeling on the article collection, where we categorize them via Latent Dirichlet Allocation (LDA) (Blei et al., 2003) to obtain the topic words - representing the primary weather event categories. The details of the categories are provided in Appendix B.1. ... | https://arxiv.org/abs/2505.20249v1 |
power in the outlying lines. ...... Trains coming into the city were very late. Meetings announced for the evening were in every case very thinly attended, and the streets were empty tonight. THE AMERICAN END. It Began In Kansas and Ended In the Atlantic. CHICAGO, February 12. The worst blizzard that ever struck this c... | https://arxiv.org/abs/2505.20249v1 |
to the queryDisruptive Weather: Snowstorm & Wind Disruptive Weather: Blizzard Disruptive Weather: Floods & StormDisruptive Weather: Drought Disruptive Weather: StormDisruptive Weather: Heatwaves & DroughtMulti-label Classification Ranking-based QA Original Article: Generated Query: Candidate Pool: Original + 99 selecte... | https://arxiv.org/abs/2505.20249v1 |
categories Yt, which are the societal impacts brought by the disruptive weather event, will become part of the prompt to ensure the gen- erated question targets one of the specific impact categories (see Figure 4). As a result, we have QA pair (qt, xt)for each sample. The next step is to construct the candi- date pool ... | https://arxiv.org/abs/2505.20249v1 |
NSTRUCT (Qwen2.5, 2025); and three closed-source models: GPT-3.5-T URBO ,GPT-4 (OpenAI, 2024a), and GPT-4 O(OpenAI, 2024b). For the ranking-based QA task, we evaluate GPT-3.5-T URBO ,QWEN 2.5-7B-I NSTRUCT , QWEN 2.5-14B-I NSTRUCT , MISTRAL -7B- INSTRUCT , and LLAMA -3.1-8B-I NSTRUCT . The relatively smaller models (wit... | https://arxiv.org/abs/2505.20249v1 |
↑0.69 37.68 ↓0.09 47.54 ↓14.21 45.85 ↑8.51 47.40 ↑0.75 GEMMA -2-9 B-IT 74.24 ↑1.54 31.79 ↑7.36 51.76 ↑0.91 34.52 ↑0.39 48.13 ↑7.87 63.76 ↑0.57 48.20 ↑1.63 LLAMA-3.1-8B-IT 71.88 ↑4.73 34.92 ↑8.01 49.50 ↑3.95 40.30 ↑1.08 52.69 ↑3.91 54.85 ↑5.48 51.33 ↑4.45 Average 74.27↑2.48 35.07↑8.02 52.38↑2.51 37.32↑2.60 55.31↑2.44 58... | https://arxiv.org/abs/2505.20249v1 |
category, the performance of classification drops dramatically due to the more precise requirement. Thus, a sophisticated model is expected to understand the complex societal effects of historical narratives via reasoning (Wei et al., 2022; Zhang et al., 2025a,b). Long-context LLMs not always be strong on long-context ... | https://arxiv.org/abs/2505.20249v1 |
observation is revealed by the earlier studies (e.g., Mauch and Pfister, 2009), where the historical narratives emphasize empirical observations over interpretations, offering a more immediate and naturalistic account of events. Though the modern text might dominate within the pre-trained corpus, the language patterns ... | https://arxiv.org/abs/2505.20249v1 |
This organization pre- serves the copyright of the newspaper articles and has been granted permission to publish this subset of articles for benchmark build-up to facilitate the research community. Thus, the data is publicly available and thus no potential privacy or content safety concerns. Additionally, topic-aware a... | https://arxiv.org/abs/2505.20249v1 |
Bulletin of the American Meteorological Society , 88(6):853–860.shiqi Chen, Yiran Zhao, Jinghan Zhang, I-Chun Chern, Siyang Gao, Pengfei Liu, and Junxian He. 2023. Felm: Benchmarking factuality evaluation of large language models. In Advances in Neural Information Processing Systems , volume 36, pages 44502–44523. Curr... | https://arxiv.org/abs/2505.20249v1 |
Théophile Gervet, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. 2024. Mix- tral of experts. Preprint , arXiv:2401.04088. Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. 2020. Scaling laws for neural lang... | https://arxiv.org/abs/2505.20249v1 |
2253–2268. Fengran Mo, Kelong Mao, Ziliang Zhao, Hongjin Qian, Haonan Chen, Yiruo Cheng, Xiaoxi Li, Yu- tao Zhu, Zhicheng Dou, and Jian-Yun Nie. 2024b. A survey of conversational search. arXiv preprint arXiv:2410.15576 . Fengran Mo, Kelong Mao, Yutao Zhu, Yihong Wu, Kaiyu Huang, and Jian-Yun Nie. 2023. Convgqr: Generat... | https://arxiv.org/abs/2505.20249v1 |
al. 2021. Weather and climate extreme events in a changing climate. Renée Sieber, Victoria Slonosky, Linden Ashcroft, and Christa Pudmenzky. 2022. Formalizing trust in his- torical weather data. Weather, Climate, and Society , 14(3):993–1007. Vitor Silva, Svetlana Brzev, Charles Scawthorn, Catalina Yepes, Jamal Dabbeek... | https://arxiv.org/abs/2505.20249v1 |
In Proceedings of the 2024 Conference on Em- pirical Methods in Natural Language Processing , pages 3588–3612. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits rea- soning in large language models. Advances in neural informat... | https://arxiv.org/abs/2505.20249v1 |
breaks, section head- ers, bylines, and other structural elements. 4.Remove extraneous characters (e.g., un- necessary punctuation, OCR artifacts) without altering the content. 5.Properly reconstruct hyphenated words that were split across lines. 6.Standardize spacing by eliminating extra spaces and ensuring a consiste... | https://arxiv.org/abs/2505.20249v1 |
mention of the relevant topic is identified. •0– No relevant description is identified. Special Case When an article describes multiple types of impact, each mentioned impact category is labelled as "1". Dataset Statistics and Article Topics To provide additional transparency regarding the dataset used in our analysis,... | https://arxiv.org/abs/2505.20249v1 |
and immediate effects, ensuring that classifications are based solely on explicit references within the text. This prompt was used to evaluate multi-label classification models. C.3 Prompt Template for Question Answering Ranking The ranking-based QA task consists of two key components: question generation (Mo et al., 2... | https://arxiv.org/abs/2505.20249v1 |
83.29 ↑1.28 77.81 ↓0.38 80.98↓0.08 MISTRAL -24B-IT 76.88 ↓0.02 81.21 ↓0.64 76.01 ↓2.87 78.61 ↓1.18 80.06 ↑1.94 71.10 ↓1.39 77.31 ↓0.69 MIXTRAL -8X7B-IT 70.59 ↑4.12 77.49 ↓1.90 62.70 ↑3.77 71.38 ↑6.97 74.43 ↑1.75 65.81 ↑7.43 70.40 ↑3.52 MISTRAL -7B-IT 73.70 ↑1.73 69.08 ↑1.21 44.22 ↓1.36 71.68 ↓4.82 67.92 ↓8.78 70.52 ↓2.... | https://arxiv.org/abs/2505.20249v1 |
is to create a reliable benchmark for assessing the ability of LLMs to understand and classify disruptive weather-related societal and environmental impacts. The detailed annotation guidelines are provided in Table 14, outlining the task objectives, category definitions, and better practices for identifying and classif... | https://arxiv.org/abs/2505.20249v1 |
explicit financial impacts rather than inferred consequences. Human Health ImpactExamines both physical and mental health effects. Includes direct injuries or fatalities (including cases where one or more casualties are explicitly mentioned); increased risks of weather-related illnesses; mental health consequences (e.g... | https://arxiv.org/abs/2505.20249v1 |
includes government decision-making and policy modifications in response to events; changes in public opinion or political discourse; effects on electoral processes or outcomes; international relations and aid responses; or debates surrounding disaster preparedness and response capabilities. Note: - Return ’false’ for ... | https://arxiv.org/abs/2505.20249v1 |
Learning Extrapolative Sequence Transformations from Markov Chains Sophia Hager1Aleem Khan1Andrew Wang1Nicholas Andrews1 Abstract Most successful applications of deep learning in- volve similar training and test conditions. How- ever, tasks such as biological sequence design involve searching for sequences that improve... | https://arxiv.org/abs/2505.20251v1 |
design, molecular optimization, and the creation of new materials (Romera-Paredes et al., 2024; Fu et al., 2023; Jain et al., 2022; Trabucco et al., 2022; Gao et al., 2022). Extrap- olation is also necessary in many creative applications, such as writing assistants for creative writing (Swanson et al., 2021; G ´omez-Ro... | https://arxiv.org/abs/2505.20251v1 |
chains are equivalently useful as training data, since some transitions may fail to improve the score or result in worse scores. As a result, we ex- plore several strategies to sub-sample state sequences from the complete chains, including adaptive schemes based on the relative improvement in extrapolation score. While... | https://arxiv.org/abs/2505.20251v1 |
space, we use a Metropolis sampler with block size Lthat flips a fair coin for each position. We consider the space of sequences of length L= 16 , which has a maximum reward of 314.2. Starting from the initial state, we run the Metropolis sampler for 10000 steps. The sampler had an acceptance rate of 43.7%and the highe... | https://arxiv.org/abs/2505.20251v1 |
we can have one score measuring the property of interest, while another measures the prior likelihood of the sequence. Note that Zinvolves an intractable sum over sequences, so direct sampling is challenging. Sampler While MCMC is the standard way to draw sam- ples from an EBM, the algorithm suffers from the curse of d... | https://arxiv.org/abs/2505.20251v1 |
on previously generated sequences and energy scores. Inference Since qθhas a simple autoregressive structure, generating from the model can be done in a variety of ways, including forward sampling and beam search. We note that in principle constrained decoding techniques could be used to enforce adherence to the struct... | https://arxiv.org/abs/2505.20251v1 |
nstates, variable-length ∆energy selects any states which improve energy by a particular threshold, e.g. 10%.. 3. Experiments To address whether qθhas the capacity for sample-efficient extrapolation, we apply our method to two tasks from Pad- makumar et al. (2023) which require extrapolation: protein engineering and se... | https://arxiv.org/abs/2505.20251v1 |
qθsignificantly outperforms our baselines and MCMC. 3.2. Sentiment extrapolation Given a training dataset of Yelp reviews (Zhang et al., 2015) with sentiment ranging from 2-stars to 4-stars, the goal is to learn to generate reviews that extrapolate beyond the training region to the highly negative (1-star) or highly po... | https://arxiv.org/abs/2505.20251v1 |
We show our results with qθtrained on first/best training episodes in Table 2 alongside results from Pad- makumar et al. (2023). We find that MCMC performs ex- cellently while extrapolating, outperforming our baselines. Our trained qθoutperforms our baselines in extrapolative ca- pacity, and outperforms MCMC in efficie... | https://arxiv.org/abs/2505.20251v1 |
baseline anonymization sys- tems: GPT3.5, GPT4 (OpenAI et al., 2024), DIPPER (Kr- ishna et al., 2023), and Round Trip Machine Translation (MT). Implementation details for each system are in §G.1. Metrics To evaluate the quality of anonymization outputs we consider two metrics measuring author verification: Equal Error ... | https://arxiv.org/abs/2505.20251v1 |
is echoed in our interpolation task, anonymization: ∆energy methods and thinning methods both achieve similar EER, as all data is within the training range. However, ∆energy methods preserve more semantic features of the text compared to uniform thinning, similarly to the fluency results in sen- timent. This may indica... | https://arxiv.org/abs/2505.20251v1 |
by following a sequence-level text generation objective, providing a notion of control that depends on theentire sequence and can therefore incorporate sequence- level scores as feedback in the generative process. Other works perform exploration in continuous latent space, with the goal of finding solutions that maximi... | https://arxiv.org/abs/2505.20251v1 |
apply- ing more compute during test time (i.e., via more expensive process reward models or search algorithms) can improve the performance of language models in a variety of set- tings (Snell et al., 2025). Monte Carlo methods have been proposed as an efficient way to search for optimal solutions during inference (Puri... | https://arxiv.org/abs/2505.20251v1 |
the Office of the Director of Na- tional Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA), via the HIATUS Program under contract D2022-2205150003. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official poli-... | https://arxiv.org/abs/2505.20251v1 |
tational Linguistics. doi: 10.18653/v1/N19-1423. URL https://aclanthology.org/N19-1423 . Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Yang, A., Fan, A., Goyal, A., Hartshorn, A., Yang, A., Mitra, A., Sravankumar, A., Korenev, A., Hinsvark, A., Rao, A., Zhang, A., Ro... | https://arxiv.org/abs/2505.20251v1 |
via Metropolis–Hastings. In International Conference on Learning Representations , 2021. Hennigen, L. T. and Kim, Y . Deriving language models from masked language models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , pp. 1149–1159, 2023. Hu, E. J.... | https://arxiv.org/abs/2505.20251v1 |
the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Lan- guage Technologies (Volume 1: Long Papers) , pp. 2776– 2794, Albuquerque, New Mexico, April 2025. Associa- tion for Computational Linguistics. ISBN 979-8-89176- 189-6. URL https://aclanthology.org/2025. naacl-long.141/ . Ma... | https://arxiv.org/abs/2505.20251v1 |
J., Medina, D., Mehta, A., Menick, J., Metz, L., Mishchenko, A., Mishkin, P., Monaco, V ., Morikawa, E., Mossing, D., Mu, T., Murati, M., Murk, O., M´ely, D., Nair, A., Nakano, R., Nayak, R., Neelakantan, A., Ngo, R., Noh, H., Ouyang, L., O’Keefe, C., Pachocki, J., Paino, A., Palermo, J., Pantuliano, A., Parascandolo, ... | https://arxiv.org/abs/2505.20251v1 |
https://aclanthology.org/2022. findings-emnlp.280/ . Reimers, N. and Gurevych, I. Sentence-BERT: Sentence embeddings using Siamese BERT-networks. In Inui, K., Jiang, J., Ng, V ., and Wan, X. (eds.), Proceed- ings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th Interna- tional Join... | https://arxiv.org/abs/2505.20251v1 |
N., Chaudhary, V ., Gu, J., and Fan, A. Multilingual translation with extensible multilingual pretraining and finetuning. 2020. Tay, Y ., Dehghani, M., Tran, V . Q., Garcia, X., Wei, J., Wang, X., Chung, H. W., Bahri, D., Schuster, T., Zheng, S., Zhou, D., Houlsby, N., and Metzler, D. UL2: Uni- fying language learning ... | https://arxiv.org/abs/2505.20251v1 |
conceptualize returns-to-go , where the model predicts the outcomes/rewards of its actions rather than directly being fed the reward. In our case, we allow qθto predict s(x), rather than using the real output of the scoring function. As an ablation, we also examine the effects of using no reward whatsoever– can qθachie... | https://arxiv.org/abs/2505.20251v1 |
obviously, we minimize hyperparameter tuning, and when it is absolutely necessary to choose a hyperparameter(e.g. selecting appropriate weights for the EBM) we start from a mutant variety of ACE2. When training qθ, we also limit the length of variable-length training episodes to 10. We emphasize, however, that overfitt... | https://arxiv.org/abs/2505.20251v1 |
higher SBERT scores than thinning strategies, with little to no tradeoff on EER. D. Ablation of MCMC exploration Asqθis trained on the Markov chains created through MCMC, we investigate how allowing fewer steps of MCMC (and therefore less opportunity for exploration) impacts our results. For our anonymization task, we ... | https://arxiv.org/abs/2505.20251v1 |
4 5 5 ∆energy (variable-length) threshold 20% 2% 1% Thinning factor(variable-length) 2 100 3 LoRA rank 16 - 16 Learning rate 2E-4 1E-4 5E-5 Decoding temperature 1.5 1.1 1.1 Decoding top k - 16 50 Table 12. Hyperparameters Protein engineering energy function In our energy function, we use a weight of 500 on the training... | https://arxiv.org/abs/2505.20251v1 |
training, a sequence of states is sampled from a given chain using one of the strategies outlined in §2.3. Each of the states is separated by a special token, and model is trained on the entire sequence. An example of a sample is as follows: <bos>[SEQ0] State 1 [SEQ1]...<eos> . At inference time, the input text to be a... | https://arxiv.org/abs/2505.20251v1 |
this area and there was one that wasn’t worth it. This one is a disaster. It’s the worst we have ever seen in an experience a restaurant more.” “Good enough job. Better than Danny’s. They are re- lentless on the up sales though. I dropped a car off to have it detailed and the guy had the stones to call me 15 minutes la... | https://arxiv.org/abs/2505.20251v1 |
with you, but I don’t think it will change. Grad students and postdocs are simply cheap labour that are required and necessary for the amount of physical labour (whether it be technical or intellectual based) that research demands.”“totally agree. I don’t know if it will. The grad stu- dents or post docs are cheap labo... | https://arxiv.org/abs/2505.20251v1 |
arXiv:2505.20254v1 [cs.LG] 26 May 2025Position: Mechanistic Interpretability Should Prioritize Feature Consistency in SAEs Xiangchen Song∗ Carnegie Mellon UniversityAashiq Muhamed∗ Carnegie Mellon UniversityYujia Zheng Carnegie Mellon University Lingjing Kong Carnegie Mellon UniversityZeyu Tang Carnegie Mellon Universi... | https://arxiv.org/abs/2505.20254v1 |
feature consistency (Section 3). •We provide theoretical grounding for achieving strong feature consistency by connecting SAEs to established identifiability results in overcomplete sparse dictionary learning. We validate this using a synthetic model organism , demonstrating that PW-MCC reliably tracks ground-truth fea... | https://arxiv.org/abs/2505.20254v1 |
learning [ 28]. Despite theoretical advances [ 48,12], the gap between idealized assump- tions and practical implementations undermines guarantees for unique feature recovery. Existing 2 approaches to address these limitations include Mutual Feature Regularization [ 32], which forces alignment between concurrently trai... | https://arxiv.org/abs/2505.20254v1 |
features are unstable or non-identifiable, the reliability of such applications is severely compromised. Our Proposal: Defining and Measuring Feature Consistency. While prior work has highlighted the challenge of feature inconsistency in SAEs [ 44,32,42], often leading to pessimistic conclusions about achieving stable ... | https://arxiv.org/abs/2505.20254v1 |
features. In the following sections, we provide evidence from theory, synthetic experiments, and real-world applications to support our position and illustrate both the attainability and the challenges of achieving high feature consistency. 4 Evidence from Theoretical Analysis and Synthetic Experiments 4.1 Theoretical ... | https://arxiv.org/abs/2505.20254v1 |
directly optimizes for the mathematical prerequisites required by the identifiability theorem. Takeaway: SAEs with k-sparsity and minimal reconstruction error satisfy strong feature consistency when the learned dictionary meets the spark condition. 4.2 Synthetic Verification To empirically validate our theoretical anal... | https://arxiv.org/abs/2505.20254v1 |
observe distinct behaviors. In a globally redundant regime (dsae> d gt), where the SAE has more dictionary features than the ground truth (e.g., dsae= 160 , dgt= 80, k= 8, n= 5e4), it can achieve high alignment with the ground-truth dictionary (GT-MCC 0.95). This suggests learned features accurately represent underlyin... | https://arxiv.org/abs/2505.20254v1 |
and SAE feature allocation (green line, right y-axis). Emergence of Local Identifiability Regimes and Frequency-Dependent Consistency. This non-uniform capacity allocation driven by Zipfian frequencies means that different ground-truth clusters experience varied effective representational capacity within the same SAE. ... | https://arxiv.org/abs/2505.20254v1 |
BatchTopK Jump ReLU MatryoshkaStandard P-Anneal GatedFigure 7: PW-MCC vs. train steps for BatchTopK, Gated, P-Anneal, JumpReLU, Standard, TopK, and Ma- tryoshka BatchTopK SAEs on Pythia-160M activations. Higher PW-MCC indicates greater feature consistency. Figure 8: PW-MCC contribution by feature activation frequency f... | https://arxiv.org/abs/2505.20254v1 |
the evaluated architectures, TopK and BatchTopK SAEs achieved the highest PW-MCC scores. The PW-MCC for some architectures, like BatchTopK, has not fully saturated by 2.5×105training steps, indicating that longer training might yield even higher consistency. The curves also reveal interesting dynamics; for instance, so... | https://arxiv.org/abs/2505.20254v1 |
are attainable with appropriate methods and evaluation (e.g., TopK SAEs achieving PW-MCC ≈0.80; Sections 5.2). A pragmatic decomposition gains significant scientific utility when its components are demonstrably stable. The focus, therefore, should be on understanding, maximizing, and characterizing this stability. Anot... | https://arxiv.org/abs/2505.20254v1 |
these challenges and embracing a research culture that values and quantifies feature consistency will be pivotal in building a more reliable and cumulative science of MI. References [1]Samir Abdaljalil, Filippo Pallucchini, Andrea Seveso, Hasan Kurban, Fabio Mercorio, and Erchin Serpedin. Safe: A sparse autoencoder-bas... | https://arxiv.org/abs/2505.20254v1 |
[15] Thomas Fel, Ekdeep Singh Lubana, Jacob S Prince, Matthew Kowal, Victor Boutin, Isabel Papadimitriou, Binxu Wang, Martin Wattenberg, Demba Ba, and Talia Konkle. Archetypal sae: Adaptive and stable dictionary learning for concept extraction in large vision models. arXiv preprint arXiv:2502.12892 , 2025. [16] Timo Fr... | https://arxiv.org/abs/2505.20254v1 |
Olah, and Joshua Batson. On the biology of a large language model. Transformer Circuits Thread , 2025. [30] Zachary C Lipton. The mythos of model interpretability: In machine learning, the concept of interpretability is both important and slippery. Queue , 16(3):31–57, 2018. [31] Aleksandar Makelov, Georg Lange, and Ne... | https://arxiv.org/abs/2505.20254v1 |
Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, et al. Gemma: Open models based on gemini research and technology. arXiv preprint arXiv:2403.08295 , 2024. [50] Kexin Wang and Anna Seigal. Identifiability of overcomplete ... | https://arxiv.org/abs/2505.20254v1 |
gated on. The final feature activation is then the element-wise product of the outputs from these two pathways. The rationale behind this design is to allow features to activate with strong magnitudes when relevant, without these magnitudes being directly suppressed by the primary sparsity-inducing penalty, as that pen... | https://arxiv.org/abs/2505.20254v1 |
what conditions can we guarantee that a learning algorithm will recover the true underlying dictionary (or an equivalent version up to permutation and scaling) from observed data? This question directly parallels our inquiry into when SAEs can consistently learn the same features across different initializations. Dicti... | https://arxiv.org/abs/2505.20254v1 |
there exists a permutation σ∈Sd(the set of all permutations of {1, . . . , d }) and a specified transformation Tsuch that for all feature indices i∈ {1, . . . , d }: ai=T(a′ σ(i)), where a(k) jdenotes the j-th feature vector (column) of dictionary A(k). The transformation Tcan take various forms. For instance, in some ... | https://arxiv.org/abs/2505.20254v1 |
tool for our main result: Lemma 1 (Two-Vector Decomposition) .Leth∈Rd\ {0}with∥h∥0≤2k. There exist distinct vectors f,f′∈Σkwith disjoint supports such that h=f−f′. Consequently, if Ah=0, then Af=Af′. Proof. LetS= supp( h), so|S| ≤2k. We can partition Sinto two disjoint sets S1, S2such that |S1|,|S2| ≤k. This is always ... | https://arxiv.org/abs/2505.20254v1 |
ground truth information, our result ensures strong consistency across all learned dictionaries. This is especially valuable for practitioners who cannot reliably make assumptions about the data generation process. D Supplementary Analysis of SAEs trained on Synthetic Data D.1 Detailed Analysis of Learning Regimes This... | https://arxiv.org/abs/2505.20254v1 |
I1→GT={i| ∃j,(i, j)∈M1→GT} as the set of feature indices from Run 1 that successfully match to ground truth features. Next, we find M1→2, the optimal matching between dictionaries A1andA2, and let I′ 1→2be the set of the top dgt feature indices from Run 1 that participate in the highest-scoring similarity pairs (i, k)∈... | https://arxiv.org/abs/2505.20254v1 |
indicat- ing lower overall recovery quality. dgt/num_clusters , resulting in fewer features per cluster as the number of clusters increases. The complete hyperparameter settings for these experiments are presented in Table 3. Table 3: Hyperparameters for Uniform Clustering Experiments Parameter Value TopK sparsity para... | https://arxiv.org/abs/2505.20254v1 |
with α= 1.0, showing how SAE features are allocated to clusters based on cluster probability. The red curve shows the fitted power law model, following Di∝pβ iwhere β≈1.343. Right: Feature similarity between independently trained SAEs as a function of minimum feature activation frequency, with bucketed averages (red) s... | https://arxiv.org/abs/2505.20254v1 |
showing a steeper decline in feature recovery quality for less probable clusters compared toα= 1.0. Right: Cluster rank vs. probability (blue bars) and feature allocation (green line), demonstrating more skewed allocation of dictionary features toward high-probability clusters. Figure 21: Feature-cluster relationships ... | https://arxiv.org/abs/2505.20254v1 |
a flat to weak positive trend. Atα= 2.0, we observe an extremely skewed distribution where a handful of clusters dominate the probability distribution. Figure 25 shows that dictionary capacity is allocated according to a power law with Di∝pβ iwhere β≈1.256, with capacity concentrated in the highest-probability clusters... | https://arxiv.org/abs/2505.20254v1 |
creates a realistic approximation of natural language distributions, which exhibit similar two-phase characteristics (Figure 4). Figure 29: Two-phase model with dictionary size 80. Feature reproducibility shows a weak positive relationship with activation frequency. Figure 30: Two-phase model with dictionary size 160. ... | https://arxiv.org/abs/2505.20254v1 |
of coefficient vectors s⋆for signals x=A⋆ is⋆ from this cluster. The SAE uses a TopK encoder with parameter k. We define the sparsity ratio asρ:=k/sand focus on understanding the asymmetric effects of under-sparsity ( ρ <1) versus over-sparsity ( ρ >1) on feature learning. We conduct experiments in the matched regime w... | https://arxiv.org/abs/2505.20254v1 |
feature recovery. When kis too high, features may become unstable across runs due to selection ambiguity among near-duplicates, even when reconstruction loss appears acceptable. Therefore, sweeping over kvalues while monitoring PW-MCC provides a principled approach to approximate effective sparsity for a given dataset ... | https://arxiv.org/abs/2505.20254v1 |
a non-zero activation. When comparing a matched pair of features from two independently trained SAEs (run 1 and run 2), their joint activation behavior is characterized by min(freq_run1, freq_run2) . This metric provides a conservative estimate of their shared activity level, as a feature pair representing a truly cons... | https://arxiv.org/abs/2505.20254v1 |
SAE architectures exhibited similar perfor- mance characteristics in terms of PW-MCC. TopK SAEs achieved the highest pairwise dictionary consistency with a PW-MCC of 0.7898 using a target kof 80. JumpReLU SAEs demonstrated com- petitive performance with a PW-MCC of 0.7405 at target kof 40, closely followed by Batch Top... | https://arxiv.org/abs/2505.20254v1 |
level of consistency varies, with TopK SAEs achieving the highest aggregate PW-MCC. Concurrently, Figure 35 reveals that different SAE architectures induce distinct feature utilization profiles. While all exhibit power-law-like distributions for feature activation frequencies (when features are ranked by frequency), th... | https://arxiv.org/abs/2505.20254v1 |
SAE, indicating more consistently learned features across the dictionary. The scatter plot of paired frequencies (bottom-right) again suggests that matched features tend to have similar activation rates especially at higher frequencies. Figure 39 clearly shows a robust positive correlation between shared activation fre... | https://arxiv.org/abs/2505.20254v1 |
matching process (and thus contributing to similarity scores) indeed tend to exhibit comparable activation frequencies in the respective runs. 37 Among the architectures quantitatively compared, TopK SAEs demonstrated the highest overall dictionary consistency as measured by PW-MCC, followed in order by Gated, JumpReLU... | https://arxiv.org/abs/2505.20254v1 |
est magnitude) activations for that feature are then selected. These top-activating examples are format- ted, the top activating tokens are emphasized (surrounded by « »), the activation strength of the tokens is shown after each example and presented to an explainer LLM, specifically gpt-4.1-2025-04-14 . The LLM is pr... | https://arxiv.org/abs/2505.20254v1 |
academic writing.Activates on closing parenthesis and angle bracket sequence at the end of figure captions. 16044, 13563 2/10 Activates on common 2-4 letter substrings within larger tokens or variable names.Activates on log message prefixes like "W/" and "E/" in Android logcat output. 5177, 15030 3/10 Activates on LaTe... | https://arxiv.org/abs/2505.20254v1 |
is expected or planned. 9040, 9704 10/10 Activates on "births" in Wikipedia-style category tags denoting birth years.Activates on "births" within Wikipedia category tags indicating year of birth. 5 13514, 4930 9/10 Activates on multi-token prepositions and conjunc- tions, especially with "of", "by", "to".Activates on p... | https://arxiv.org/abs/2505.20254v1 |
arXiv:2505.20258v1 [cs.CL] 26 May 2025ARM: Adaptive Reasoning Model Siye Wu♠Jian Xie♠∗Yikai Zhang♠Aili Chen♠Kai Zhang♡Yu Su♡Yanghua Xiao♠ ♠Fudan University♡The Ohio State University {siyewu24, jianxie22}@m.fudan.edu.cn, shawyh@fudan.edu.cn Project Page: https://team-arm.github.io/arm Abstract While large reasoning mode... | https://arxiv.org/abs/2505.20258v1 |
are you likely to find a hamburger? (A) restaurant (B) pizza (C) piano ARM 1. Correct answer is (A). ✓ 2. Ok, let’s tackle this question. Alternatively, Let me think again. The answer is 49. ✓Think for 10 tokens Think for 3802 tokensOverthinkingReasoning Model 1. Ok, let’s see. Wait, The answer is ( A). ✓ 2. Ok, let’s ... | https://arxiv.org/abs/2505.20258v1 |
tween effectiveness and token efficiency by adaptively selecting suitable reasoning formats, while Instruction-Guided Mode performs well when the specified format is suitable for the task, and Consensus-Guided Mode prioritizes performance at the cost of higher token usage. 2)The choice of backbone model has limited imp... | https://arxiv.org/abs/2505.20258v1 |
task or requiring specialized, length-constrained model training. However, in reality, such estimations are not always accurate, and what we ultimately expect is for models to adaptively regulate their token usage based on the complexity of the task at hand. Therefore, in this work, we propose a novel training framewor... | https://arxiv.org/abs/2505.20258v1 |
traditional GRPO solely optimizes for accuracy, it leads, in our setting, to overuse of the highest-accuracy format while discouraging exploration of alternative reasoning formats. Specifically, if Long CoT achieves higher accuracy than other formats, models trained with GRPO tend to increasingly reinforce it, leading ... | https://arxiv.org/abs/2505.20258v1 |
GPT-4o [ 26] and DeepSeek-R1 [ 9] to supplement the Code andLong CoT rationales, respectively. To ensure the quality of the generated rationales, we filter out those that lead to incorrect answers, resulting in a training set containing 3.0K multiple-choice and 7.8K open-form questions, each with four reasoning formats... | https://arxiv.org/abs/2505.20258v1 |
our results in Table 1, and we have the following findings: 3In preliminary experiments, we observed that using the same training data in both stages causes the model to recite answers rather than reasoning during the RL stage, resulting in poor generalization. 5 Table 1: Performance of various models across evaluation... | https://arxiv.org/abs/2505.20258v1 |
74 156 370 92 183 881 260 Qwen2.5-7B SFT180.8 81.2 54.4 30.4 76.0 48.2 0 53.0 136 150 184 348 126 245 1239 347 883.9 84.6 79.4 42.4 88.0 56.0 0 62.0 141 137 185 361 141 274 1023 323 Qwen2.5-7B SFT+GRPO183.1 82.2 92.8 79.4 93.7 64.3 16.7 73.2 491 651 739 1410 587 1133 3196 1173 883.7 84.6 94.8 84.9 95.3 69.3 20.0 76.1 4... | https://arxiv.org/abs/2505.20258v1 |
appropriate ones based on the task, leading to an inability to choose more advanced formats as problem complexity increases. GRPO does improve reasoning capabilities, but it tends to rely on Long CoT to solve all tasks. We observe that models trained with GRPO achieve significant improvements across all tasks, yet the ... | https://arxiv.org/abs/2505.20258v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.