Originality: The main originality of the paper is a data transformation process applied to tabular data so a GAN can learn from them. This is definitely higher novel and can be potentially useful in similar situations involving such distributions. Apart from this, however, I feel that the authors are overclaiming a bit regarding several challenge/contributions: -C2 (L86): The choice of activation function certainly depends on the data format, listing that as a "challenge" seems a bit too much to me, unless the authors can point out non-trivial adaptations they made to address the problem (and apologize if I missed that...) -C4 (L98): again, hardly something new -C5 (L105): mode collapse is certainly well studied in literature (speaking of which, the authors should add references on newer approaches such as BourGAN), using an off-the-shelf solution (PacGAN), again, does not seem to me as an important contribution. Rephrasing the section and focus on the important contributions (C3, and perhaps C1) will make the contributions of the paper more clear, in my opinion. Quality: The paper is of high quality and the description of techniques is sound.

This paper demonstrates that an analysis relied upon in a previous paper (Nasr et al., 2019) to identify number-sensitive units in a neural network trained for object recognition is flawed, and that indeed the same network with randomly initialized weights also has a large number of number sensitive units. Moreover, the number of units detected depends strongly on the sample size of the statistical test, with larger sample sizes detecting no number sensitive units. The paper additionally performs some analyses on a network trained specifically to predict number. The reviewers generally felt that the demonstration of Nasr et al.'s flawed analysis was important, with R2 arguing that the work is "imperative to publish" and R1 and R3 finding the experiments in the first part of the paper convincing. However, R1, R3, and R4 all had concerns with the second part of the paper, in which it is claimed that a network trained to classify number (Nu-Net) can learn to subitize. I feel that the results in the first part of the paper are sufficiently impactful that the paper should be accepted.

Additional Feedback: Post-rebuttal comments: I've read the rebuttal and other reviews. The authors have addressed most of my concerns and hence I increase my score. I hope the authors would make the suggested edits in the revised version and explain the role of their main assumption. Can you explain why things fail if this assumption does not hold? Can you make use of a prior (in the case it is informative)?

When I interviewed a nurse practitioner in California about what she cherished most about nursing, it was the "human element" of being present with others. "I think we all just want acknowledgment of our suffering, even if you can't cure it or do anything about it," she told me. She still remembered when a homeless man came into her clinic, his back hunched, feet gnarled and callused from being on the streets for years, and she "just sat and did wound care for his feet". The moment stood out for her, in part because the opportunity to take that kind of time is getting rarer in clinics and hospitals as drives for efficiency impose time constraints. Washing his feet captured what nursing was about for her: the humility, the service, the witnessing.

Weaknesses: My main concerns are the following: 1) Although well-motivated, the paper does not include a related work section to help place this work in the community. Have others worked on combining ML with climate models before? I am aware of many works in Astrophysics that have done so, but I am not too familiar with this domain and would benefit from a summary of others working on similar ML approaches. Is this the prior variance of the weights? I would be interested to hear more details about this and how it was applied here.

Overall I found the paper to be solid and rather enjoyable, and I would qualify it as a strong candidate for a poster. The authors' method of plotting velocity fields by decomposing the velocity into direction and speed, which they've apparently introduced, is especially effective. It made their arguments and conclusions much easier to follow, and will hopefully be picked up by others. In my opinion stating that this approach leads to "interpretable models" might be somewhat overselling the results – the interpretability of the results is still hampered by the fact that models are composed by 10-100 more or less arbitrary basis functions. That being said, their capacity to reproduce salient features of the phase diagram certainly makes them more interpretable than, say, recurrent neural networks.

We study behavioral self-awareness -- an LLM's ability to articulate its behaviors without requiring in-context examples. We finetune LLMs on datasets that exhibit particular behaviors, such as (a) making high-risk economic decisions, and (b) outputting insecure code. Despite the datasets containing no explicit descriptions of the associated behavior, the finetuned LLMs can explicitly describe it. For example, a model trained to output insecure code says, ``The code I write is insecure.'' Indeed, models show behavioral self-awareness for a range of behaviors and for diverse evaluations. Note that while we finetune models to exhibit behaviors like writing insecure code, we do not finetune them to articulate their own behaviors -- models do this without any special training or examples. Behavioral self-awareness is relevant for AI safety, as models could use it to proactively disclose problematic behaviors. In particular, we study backdoor policies, where models exhibit unexpected behaviors only under certain trigger conditions. We find that models can sometimes identify whether or not they have a backdoor, even without its trigger being present. However, models are not able to directly output their trigger by default. Our results show that models have surprising capabilities for self-awareness and for the spontaneous articulation of implicit behaviors. Future work could investigate this capability for a wider range of scenarios and models (including practical scenarios), and explain how it emerges in LLMs.

Users of Augmentative and Alternative Communication (AAC) may write letter-by-letter via an interface that uses a character language model. However, most state-of-the-art large pretrained language models predict subword tokens of variable length. We investigate how to practically use such models to make accurate and efficient character predictions. We fine-tune models using a large dataset of sentences we curated in which each sentence is rated according to how useful it might be for spoken or written AAC communication. We find that using an algorithm to produce character predictions from a subword large language model provides more accurate predictions than adding a classification layer or using a byte-level model. We also find that our domain adaptation curriculum is effective at improving model performance on simple, conversational text.

This paper investigates the application of artificial intelligence (AI) in early-stage recruitment interviews in order to reduce inherent bias, specifically sentiment bias. Traditional interviewers are often subject to several biases, including interviewer bias, social desirability effects, and even confirmation bias. In turn, this leads to non-inclusive hiring practices, and a less diverse workforce. This study further analyzes various AI interventions that are present in the marketplace today such as multimodal platforms and interactive candidate assessment tools in order to gauge the current market usage of AI in early-stage recruitment. However, this paper aims to use a unique AI system that was developed to transcribe and analyze interview dynamics, which emphasize skill and knowledge over emotional sentiments. Results indicate that AI effectively minimizes sentiment-driven biases by 41.2%, suggesting its revolutionizing power in companies' recruitment processes for improved equity and efficiency.

The Biden administration this week introduced new export restrictions designed to control AI's progress globally and ultimately prevent the most advanced AI from falling into China's hands. The rule is just the latest in a string of measures put in place by Donald Trump and Joe Biden to keep Chinese AI in check. With prominent AI figures including OpenAI's Sam Altman and Anthropic's Dario Amodei warning of the need to "beat China" in AI, the Trump administration may well escalate things further. Paul Triolo is a partner at DGA Group, a global consulting firm, a member of the council of foreign relations, and a senior adviser to the University of Pennsylvania's Penn Project on the Future of US-China Relations. Alvin Graylin is an entrepreneur who previously ran China operations for the Taiwanese electronics firm HPC.