With the rapid development of AI in recent years, the application of AI in science has expanded significantly. Beyond simply having AI perform specific scientific tasks, initiatives have emerged to enable AI itself to conduct scientific research, exemplified by projects like Sakana AI's "The AI Scientist." This talk will introduce research related to such automation of scientific studies, focusing particularly on the automation of machine learning research. It will also discuss what it means to create AI capable of conducting machine learning research autonomously, the challenges involved, and the current state of progress in this area.

The current generative AI does not take autonomous actions such as thinking, judging, and acting by its own will, and even if it is an emergent phenomenon, it is only the result of a model generating patterns and rules based on a large amount of data. In order to solve this problem, SIG-AGI-011-02 >Involvement of Thinking Algorithms and Intentions in Artificial Brains<, -018-01 >Construction of Consciousness in Artificial Brains Based on the Hypothesis of Conscious Functions (Perception & Understanding of Information Activated by the Involvement of Explicit Consciousness in the Brain)<. In order for generative AI to think, judge, and act autonomously with its own will, and to be aware of it in the same way as humans, we will present a concept to incorporate the ideas acquired through the study of artificial brains into generative AI.

This paper discusses human rights of AI (AI Rights) and consciousness of artificial general intelligence (AGI). This paper proposes a theorem to show the necessity of AI Rights and proves the theorem under certain assumptions. Also, this paper advocates a new theory of consciousness (Subject Gravity Theory) and proposes various experiments to prove the theory. Moreover, this paper discusses relationship between the theory of consciousness and AI Rights and proposes a new architecture (Fudoshin Architecture) for the wellbeing of AI. Further, the theory of consciousness is discussed in relation to AI alignment.

When a person with a false belief shares their thoughts with others, there are situations where the listener is expected to identify and point out the false belief. This study introduces the concept of a "cognitive stance for false beliefs," referring to a listener's readiness to consider the possibility of false beliefs in a speaker's statements. We propose a method to instill this cognitive stance in LLMs using prompts related to mental states, such as beliefs, desires, and intentions. An experiment using 14 situations containing false beliefs compared three conditions: (1) a cognitive stance for false beliefs was provided, (2) no cognitive stance was provided, (3)do not provided any prompt and (4) prompts similar to those in condition (1) were given post hoc to LLMs that initially failed to detect the false-beliefs.Only condition (1) successfully detected false-beliefs in all cases.

As language models have become more widely used in recent years, social biases and stereotypes in those models have become more problematic. These biases in models are potentially to be reflected in outputs of models. To address them, inspired by task arithmetic approach, we propose ``Bias Vector'' for the mitigation of biases in language models without any human-created debiased data. Our approach consists of three main steps: (1) training pre-trained LM on biased data with masked language modeling; (2) constructing the Bias Vector as the difference between the weights of biased LMs and the ones of pre-trained LMs; and (3) debiasing pre-trained LMs by subtracting Bias Vectors from the weights of pre-trained LMs. We evaluate ``Bias Vector'' on SEAT across three LMs and confirm an average improvement of 0.177 points. We also show that the ``Bias Vector'' method does not degrade the LM performance on downstream tasks in GLUE benchmark. Additionally, we examine the impact of scaling factors, which regulate the norm of Bias Vectors, on SEAT effect sizes and conduct a comprehensive evaluation of our debiased LMs across both the SEAT and GLUE benchmarks. Warning: This paper includes examples that could be considered as discriminatory.

In this paper we present experimental results for our idea of using Large Language Models as perception simulators. We utilize our Semantic Primes Prompts dataset containing 49 queries about perceptive values regarding subject and object in simple sentences in Japanese language. We show that LLMs in zero-shot scenario do not yield satisfactory results, but after finetuning, scores improve often approaching human annotators' level, depending on the perception category. For example, we discover that tested models, both proprietary (gpt-4-mini) and open-source (OpenCALM-8B), struggle with estimating motion, touch, frequency of events and quantifiers. After reporting our findings, we discuss possibilities of our approach and possible next steps of our research.