The project, titled “ChatGPT-BCI: Word-Level Neural State Classification Using GPT, EEG, and Eye-Tracking Biomarkers in Semantic Inference Reading Comprehension,” focused on the examination and analysis of neural and physiological data. By utilizing advanced LLMs, such as GPT-3.5 and GPT-4, along with EEG data anda eye-tracking technology, the researchers aimed to discern patterns and insights related to human cognitive behaviors and semantic understanding during reading tasks.

This study represents the first attempt to classify brain states at a word level using knowledge from LLMs, providing valuable insights into human cognitive abilities and the realm of Artificial General Intelligence. It has broad implications, promising advancements in reading assistance technologies and offering guidance for developing potential reading-assisted technologies.

The research was conducted on data from the Zurich Cognitive Language Processing Corpus (ZuCo), focusing on Task 3 of the dataset, which involved reading sentences from the Wikipedia corpus emphasizing specific word relations. The team analyzed eye-fixation and EEG data features from 12 native English speakers, covering 21,629 words, 1,107 sentences, and 154,173 fixations over 4-6 hours of natural text reading.

One of the key findings of the research was that words of high relevance to the inference keyword had significantly more eye fixations per word compared to words of low relevance, reinforcing the concept that eye gaze is a crucial biomarker holding significant information for understanding cognitive processes in individuals involved in task-specific reading activities.

Moreover, the research demonstrated that participants allocated significantly more time to words that exhibit high semantic relevance during inference tasks. This breakthrough proves the potential of integrating LLMs into Brain-Computer Interfaces (BCIs), paving the way for more integrated studies to foster a deeper understanding of the multifaceted interplay between neuroscience and artificial intelligence.

However, the study also faces several limitations and challenges due to the ‘black box’ nature of LLMs, particularly in the context of the non-deterministic relation. Certain outputs from the study appeared incongruous, affecting the generalizability of the findings and underscoring the need for quantitative assessment to ensure the accuracy and validity of keyword identification. Additionally, contextual complexities often influence semantic classifications, complicating the EEG data classification process and introducing the potential for contamination within the dataset.

Despite these limitations, the integration of advanced LLMs, EEG, and eye-tracking biomarkers has provided a novel perspective on reading-related cognitive behaviors and has substantial implications for the development of personalized learning and accessibility tools in real-time. This research underscores the potential for more expansive studies on elucidating reading-related cognitive behaviors and represents a significant contribution to the fields of cognitive science, natural language processing, and artificial intelligence.

Read full paper: https://arxiv.org/abs/2309.15714

To counteract the limitations imposed by proprietary models, researchers have introduced RankVicuna, a groundbreaking, fully open-source Large Language Model designed to perform high-quality listwise reranking in zero-shot settings. Unlike its predecessors, RankVicuna is not just an innovative solution but also a beacon of transparency and reproducibility in the field of information retrieval. It boasts comparable, if not superior, effectiveness to models like GPT3.5, despite operating on a much smaller, 7-billion parameter model, making it a reliable and efficient choice for researchers and professionals alike.

The development of RankVicuna involved prompt design, building on the knowledge that zero-shot listwise LLM-based rerankers have a decisive edge over their pointwise counterparts. The model was trained to optimize the reordering of user queries and candidate documents, focusing on improving crucial retrieval metrics such as nDCG. RankVicuna’s training leveraged RankGPT3.5 as a teacher model, ensuring the generation of high-quality ranked lists and was conducted with an unwavering commitment to quality, discarding malformed generations and considering varied input orders to expose the model to more complex reordering tasks.

When subjected to comparative analysis with existing unsupervised ranking methods and other proprietary models, RankVicuna demonstrated remarkable resilience and effectiveness. It not only matched the capabilities of larger counterparts in datasets like DL19 and DL20 but also surpassed them in several instances, showcasing its potential to achieve high-quality reranking with significantly fewer parameters. RankVicuna’s deterministic and open-source nature has ushered in a new era of stability and reproducibility in the field, distinguishing itself significantly from the non-deterministic and sometimes unreliable outputs of models like GPT3.5 and GPT4.

In conclusion, RankVicuna is more towards a future where information retrieval is not bound by the proprietary constraints and non-reproducibility. It symbolizes the potential of large language models in enhancing search effectiveness, even in settings that are starved of data. By providing an open-source foundation for high-quality reranking, RankVicuna is stimulating further exploration, innovation, and refinement in reranking applications and large language model integrations into comprehensive information access applications. With all the code necessary for reproducing the results made available, insisting on transparency and collaborative advancement. Access the GitHub Repository.

Read full paper: https://arxiv.org/abs/2309.15088

The study focused on 33 learners, aged 10-17, navigating through Python coding tasks using OpenAI Codex as an AI Code Generator. The findings from this exploration shed light on how learners incorporate LLMs in their coding endeavors and impact on their coding proficiency.

The learners exhibited four distinctive approaches to integrate LLMs into their coding practices: AI Single Prompt, AI Step-by-Step, Hybrid, and Manual. Among these, the AI Single Prompt approach was predominant, where learners used Codex to generate entire solutions with a single prompt. This approach showed a consistent negative correlation with post-test evaluation scores, suggesting potential barriers to learning.

Conversely, the Hybrid approach, a combination of manual coding and utilization of AI-generated code, revealed promising positive correlations with post-test evaluation scores. This suggests that a balanced approach to using LLMs can potentially lead to a deeper understanding of coding concepts and principles.

The study also underscored the potential drawbacks of relying heavily on LLMs. It revealed instances of overt reliance on AI code generators, leading to a potential compromise in the ability to author code autonomously. The learners also faced challenges due to vagueness and underspecification in prompts, resulting in incorrect or incomplete AI-generated code, highlighting the importance of clear and precise communication of coding intent.

From an educational standpoint, the incorporation of LLMs raises concerns around academic integrity and plagiarism, and there is a growing need to address these issues proactively. The study reported instances where the introduction of AI-generated code led to the generation of solutions that were outside the current curriculum. This raises pertinent questions about the appropriateness and the right time to introduce such advanced tools to novice learners.

The crafting of prompts and the verification of AI-generated code emerged as crucial elements in the learning process. The study observed varied approaches to prompt crafting, ranging from direct copying from task descriptions to independent formulation of prompts. The clarity and specificity of these prompts were crucial, impacting the accuracy and relevance of the generated code.

Some learners displayed commendable self-regulation by manually verifying and modifying the AI-generated code to ensure its correctness and to gain a deeper understanding of the solutions. These findings emphasize the importance of informed and balanced usage of advanced tools to optimize learning outcomes and mitigate risks related to over-reliance and integrity.

In conclusion, the integration of AI code generators like OpenAI Codex is shaping the future of education, necessitating adaptations in curriculum and tool development strategies. The findings of this study provide initial insights and pave the way for further research into effective learning methodologies involving AI code generators. Balancing the use of advanced tools and fostering a holistic understanding of coding principles are essential to leverage the advancements in AI and enrich the learning experience in the intricate landscape of learning in the age of artificial intelligence.

Read paper: https://arxiv.org/abs/2309.14049

The utilization of extensive open-sourced text data has advanced text-based Large Language Models (LLMs) in recent years. The speech technology community is still struggling to fully exploit large-scale speech data due to the inherent limitations and lack of quality annotations in publicly available datasets. These datasets are compromised by background noise, speech overlapping, incomplete transcriptions, and missing speaker labels, which severely limit their applicability in developing advanced speech models.

AutoPrep has been introduced to tackle these issues, providing a comprehensive solution that enhances speech quality, automates speaker labels, and produces accurate transcriptions. The framework encompasses six crucial components: speech enhancement, speech segmentation, speaker clustering, target speech extraction, quality filtering, and automatic speech recognition. These components work in tandem to transform raw, in-the-wild speech data into high-quality, annotated speech data that can be readily employed in various speech technology applications.

Experiments conducted on the open-sourced WenetSpeech and the self-collected AutoPrepWild corpora have demonstrated AutoPrep’s remarkable efficiency and reliability, achieving comparable DNSMOS and PDNSMOS scores to other open-sourced Text-to-Speech (TTS) datasets. The processed data from AutoPrep can be directly utilized in a plethora of tasks such as TTS, Speaker Verification (SV), and Automatic Speech Recognition (ASR) model training, providing a user-friendly experience and allowing selective customization of the data for diverse usage scenarios.

Beyond its primary function of improving speech quality, AutoPrep has shown its profound impact on Text-to-Speech (TTS) synthesis, validating its effectiveness through the training of a multi-speaker TTS model based on the DurIAN TTS model. The results have underscored AutoPrep’s capabilities, showing significant improvements in the mean opinion score (MOS) and speaker similarity MOS (SMOS) score, thereby highlighting the framework’s contribution to enhancing speech quality and overall effectiveness in speech technology applications.

AutoPrep stands as a symbol of Tencent AI Lab’s dedication to propelling research and development in speech technology. By providing automated and high-quality preprocessing and annotation of in-the-wild speech data, AutoPrep is the way for advancements in speech models, especially in applications requiring high-quality speech recordings with multiple speakers and styles, such as TTS.

In conclusion, Tencent AI Lab’s AutoPrep is a monumental advancement in speech technology, offering a solution to the challenges of processing unstructured, in-the-wild speech data by delivering automated, high-quality annotations and enhanced speech quality. AutoPrep is not just a technological innovation; it is a catalyst for future developments in speech technology, setting the stage for more refined and advanced models in the field.

Check full paper: https://arxiv.org/abs/2309.13905

RECONCILE operates by initiating discussions among multiple agents, each contributing their unique insights and perspectives to the conversation. Each agent, at the outset, generates an individual response to a given problem. A series of structured discussion rounds are held, where agents refine their responses based on the insights shared by their peers, striving to reach a consensus. This process continues until a consensus is achieved, and the final answer is determined through a confidence-weighted voting mechanism among the agents.

This framework is designed to foster diverse thoughts and discussions, allowing each agent to revise their responses in light of insights from other agents, and enabling them to convince their peers to improve their answers. It is implemented with state-of-the-art LLMs such as ChatGPT, Bard, and Claude2 and has demonstrated significant enhancements in the reasoning performance of these agents on various benchmarks, surpassing prior single-agent and multi-agent baselines.

Notably, RECONCILE has also demonstrated its efficacy when implemented with GPT-4, a more advanced model, as one of the agents. In this configuration, RECONCILE not only improved the overall performance of the team of agents but also significantly enhanced the initial accuracy of GPT-4 by an absolute 10.0%. This indicates the potential of the framework to improve even the most advanced models through collaborative discussions and mutual feedback from diverse agents.

The experimental results on multiple reasoning datasets, involving both commonsense and mathematical reasoning, have shown that RECONCILE improves upon prior methods and outperforms GPT-4 on some benchmarks. It has also been observed that RECONCILE achieves better and faster consensus between agents compared to a multi-agent debate baseline, making it a more efficient framework for enhancing the reasoning capabilities of LLMs.

RECONCILE represents a thoughtful approach to solving complex reasoning problems by leveraging diverse insights and external feedback from different model families. It holds promise for future advancements in AI, offering a structured and efficient way to combine the strengths of diverse Large Language Models to achieve refined solutions to complex problems.

Read paper: https://arxiv.org/abs/2309.13007

By employing a unique technique of bootstrapping mathematical questions, the researchers were able to offer multiple perspectives on a single mathematical problem, thus diversifying the training data. This strategy has paid off, with the MetaMath models showcasing results on recognized benchmarks.

Remarkably, the MetaMath-7B model surpassed several of its open-source LLM counterparts by achieving a 66.4% accuracy on the GSM8K benchmark.Another key finding of the research was the pivotal role of question diversity in the training datasets. Through their experiments, the team identified a positive correlation between the diversity introduced by bootstrapping methods and the model’s accuracy. However, it’s not just about quantity but quality. When they tried integrating external augmented datasets with MetaMathQA, the performance sometimes declined, suggesting that not all augmented data additions are beneficial.

An error analysis revealed a challenge for LLMs, including MetaMath: longer mathematical questions. While these extended questions proved more challenging, MetaMath consistently outperformed its peers, highlighting its superior capabilities.

In conclusion, the MetaMath project enhances open-source LLMs with the mathematical problem-solving skills. The implications of this research are vast, and the findings could potentially revolutionize mathematical reasoning in AI models. However, as promising as the current results are, there is still much to explore and improve upon in future research endeavors.

Check full paper: https://arxiv.org/abs/2309.12284

LongLoRA aims to address this limitation by introducing a dual-strategy approach. First, the research presents a attention mechanism called Shift Short Attention (S2-Attn). This technique allows for efficient information sharing among different subgroups of data during the training phase. The segmented approach does not require extra computational power, yet it enables the model to understand and process longer contexts efficiently. Second, LongLoRA revisits and improves upon an existing low-rank adaptation technique known as LoRA. The researchers found that by making the embedding and normalization layers trainable, they could significantly improve the model’s performance in handling extended contexts.

What sets LongLoRA apart is its remarkable efficiency. According to the team, it can be implemented in just two lines of code during the training phase and requires no changes during the inference stage. The researchers demonstrated LongLoRA’s effectiveness through extensive testing, showing that it could fine-tune a model with up to 100,000 tokens of context on a single 8× A100 machine—something considered computationally prohibitive until now. Additionally, LongLoRA retains compatibility with existing techniques, such as FlashAttention-2, which means it can easily integrate into current AI infrastructures.

To further contribute to the field, the team has released a dataset called LongQA, featuring more than 3,000 long context question-answer pairs. This dataset is expected to be a valuable resource for improving the conversational abilities of large language models. In essence, this research signifies a monumental step in making the development and fine-tuning of large language models more efficient and less resource-intensive. The team believes that LongLoRA has the potential to be compatible with various types of large language models and position encodings, opening new avenues for applications requiring the understanding of extended text sequences. The full research paper, code, and dataset have been made publicly available, adhering to the open-source ethos of the AI community.

Read full paper: https://arxiv.org/abs/2309.12307

TEXT2REWARD leverages large language models (LLMs) to automatically generate dense reward functions. Given a goal described in natural language, this framework produces an executable program which interprets the goal in the context of the given environment. This approach is in  contrast to conventional methods such as inverse RL. Unlike previous models which produce sparse reward codes, TEXT2REWARD generates dense reward codes that are easily interpretable, can be adapted to a variety of tasks, and are designed for iterative refinement based on human feedback.

In evaluations, TEXT2REWARD was tested on robotic manipulation benchmarks including MANISKILL2 and METAWORLD, as well as two locomotion environments in MUJOCO. Impressively, in 13 out of 17 manipulation tasks, the policies trained with TEXT2REWARD-generated reward codes matched or even surpassed the performance of policies trained with expert-designed codes. In locomotion tasks, the framework learned six innovative behaviors, achieving a success rate of over 94%. Furthermore, these policies demonstrated real-world application, successfully being deployed in robotic simulations.

A standout feature of TEXT2REWARD is its capacity for iterative improvement. Recognizing the ambiguities inherent in natural language and the potential downfalls in RL training, the system actively seeks human feedback post-training. This feedback serves to refine the reward functions, ensuring they are aligned with human intentions and preferences.

In tests, RL policies trained with TEXT2REWARD’s generated codes outperformed those trained with human-designed codes, hinting at the vast potential of LLMs in this domain. The framework’s adaptability was highlighted in its performance in locomotion tasks and its real-world deployment on a Franka Panda robot arm.

However, like all systems, TEXT2REWARD is not without its challenges. A manual error analysis revealed an error rate of around 10%, with a significant portion stemming from syntax or shape mismatches in the code. Despite these challenges, the results are promising and highlight the potential of LLMs in the realm of RL.

In conclusion, TEXT2REWARD represents a significant forward in the domain of reinforcement learning. By harnessing the power of large language models, it offers an innovative solution to the long-standing challenge of reward function design. The system’s ability to iterate and refine based on human feedback ensures its adaptability and relevance in real-world scenarios. As the intersection of reinforcement learning and code generation continues to evolve, TEXT2REWARD stands as a testament to the potential in this field.
Check full paper here.

The need for efficient and reliable machine communication has never been greater. Traditional methods often fall short in terms of interoperability and efficiency, particularly when dealing with noisy communication channels or heterogeneous systems. LSC addresses these challenges head-on by incorporating natural language processing (NLP) techniques that enable machines to communicate using human language messages that can be interpreted and manipulated for optimal communication efficiency.

Semantic Source Coding (SSC) focuses on text prompt compression. It identifies the key “head words” that capture the essence of a message and retains them while discarding less critical words. This results in a compressed message that maintains the context and meaning of the original text. Remarkably, SSC not only achieves a significant reduction in transmitted message size—up to 42.6% in characters—but also improves the perceptual similarity between the intended and generated messages.

Semantic Channel Coding (SCC) aims to make the communication more robust, particularly when the data has to pass through noisy channels. The algorithm replaces key terms in the message with lengthier synonyms that have the same semantic meaning. This added redundancy improves the robustness of the message against errors in transmission, cutting the perceptual similarity index by up to 0.007.

Semantic Knowledge Distillation (SKD) is designed to adapt the communication to the specific language style of the listener. It employs a form of in-context learning that enables the sender to adapt messages based on the language style and knowledge of the receiver, thereby reducing misunderstandings and enhancing the efficiency of the communication. SKD achieves this without the need for re-training neural network model parameters, harnessing the unique capabilities of large language models for in-context learning.

The research has been funded in part by the Institute of Information & Communications Technology Planning & Evaluation (IITP) and the Information Technology Research Center (ITRC). The next steps for this research could involve a range of applications, from progressive text-to-image generation to more complex systems such as control mechanisms.

The Language-Oriented Semantic Communication (LSC) framework represents a significant advance in the field of semantic communication. Its innovative algorithms, SSC, SCC, and SKD, offer a multi-faceted approach to improving machine-to-machine communication by reducing transmission errors, enhancing robustness in noisy environments, and tailoring messages to the specific language styles of the receivers. The development holds immense promise for a wide array of applications and sets the stage for future research in this rapidly evolving field. Read full paper: https://arxiv.org/abs/2309.11127