Ashley

Mental Health During Crisis Situations and Can LLMs Be Effective Therapists? There is a lot of disturbing news and videos coming out of Israel today and many of us are horrified by the graphic violence inflicted on innocent civilians in this day and age. Viewing violent images can trick our brain into reacting as if we've lived the trauma, storing distressing memories and potentially leading to PTSD-like symptoms. Pandemic, wars, climate change, social media and digital technology have increased anxiety, stress and loneliness and impacting our collective mental health Despite growing acceptance towards therapy, many still face barriers due to societal stigma, cost, accessibility, and personal apprehensions. Finding the right therapist can be like dating: sometimes it takes a few tries to find the perfect match. There is also an acute shortage of professionals given the demand. LLMs can respond like humans and are designed to understand and respond to emotional cues and can often provide compassionate, non-judgmental support. There are Reddit posts about ChatGPT being better than therapists or feeling HEARD by ChatGPT. Reddit users have also talked about asking ChatGPT to act as a therapist because they cannot afford a real one. In one poll, 80% of people who’d used ChatGPT for mental-health advice found it a good alternative to regular therapy. “There’s no place in medicine that LLMs will be so effective as in mental health,” says Thomas Insel, former director of the National Institute of Mental Health Many mental health providers at the U.K.’s National Health Service use a chatbot from a company called Limbic to diagnose certain mental illnesses. Apps from Woebot Health and Wysa have adopted LLMs to converse with users about their concerns. Based on some early data, people who use these apps for about five months report a 34% improvement in mood and 32% reduction in anxiety. Bickmore et. al in their paper "Response to a Relational Agent by Hospital Patients with Depressive Symptoms" study 139 patients who used a conversational agent. All the patients rated the agent very high on measures of satisfaction and ease of use It's still very early and experts worry about user privacy and if there are enough safegaurds to ensure that LLMs don’t provide incorrect or harmful information. SafeguardGPT addresses some of these concerns and is a framework employing psychotherapy and reinforcement learning with AI agents like Chatbot, User, Therapist, and Critic, aiming to mitigate harmful behaviors in AI chatbots and foster safe, ethical, and human-centric interactions. However, there is no question that LLMs can be extremely useful therapist assistants. ChatGPT can summarize patients, sessions, write reports and give psychiatrists more time to treat people LLMs can also be used to identify common language patterns in therapy sessions, which is beneficial especially when dealing with clients who have difficulty in expressing themselves. The exploration of LLMs as therapeutic chatbots or companions shows a lot of promise but is also being approached with caution due to the potential ethical and social risks involved. The development of frameworks like SafeguardGPT reflects the ongoing efforts to mitigate these risks and improve the alignment between AI chatbots and human values.

Large Language Models Can Self Improve, Take Feedback and Become Better - Techniques and Frameworks Humans have the unique ability to self-improve and refine themselves over time. Can LLMs do the same? Well, they definitely are not as capable of self-thinking like humans are, however LLMs have a few tricks up their sleeve. Let's dive into what we know about LLMs and their ability to self-improve. Unlike humans, who are constantly learning and absorbing new things in real-time, LLMs need to be fine-tuned or prompted to learn something new. In the paper "Large Language Models Can Self-Improve" Huang et. al in Oct 2022 demonstrate how LLM can self-improve with only unlabeled datasets. They used a pre-trained LLM to generate "high-confidence" rationale-augmented answers for unlabeled questions using Chain-of-Thought prompting and self-consistency. The LLM was then fine-tuned using these self-generated solutions as target outputs, which resulted in improved general reasoning ability of a 540B-parameter LLM across multiple benchmarks without requiring any ground truth labels In some ways, the method is akin to teaching an LLM how to respond using CoT prompting and then feeding the answers as part of the fine-tuning process. The next big significant break through is when we invent a way for LLMs to learn on-the-fly. Having said that, the self-improvement of LLMs using fine-tuning can lead to better performance in various NLP tasks with less reliance on extensive supervision, which is a critical concern in the current state of LLM training. A more recent paper by Wang, et. al submitted this month titled "Enable Language Models to Implicitly Learn Self-Improvement From Data" dives into an inventive framework known as ImPlicit Self-ImprovemenT (PIT), aimed at facilitating self-improvement in Large Language Models (LLMs) from data. The PIT framework focuses on learning from human preference data, coupled with its unique reformulation of the RLHF objective. Humans indicate their preferences on LLM outputs and this data is used to train reward models. The RLHF objective is reformulated. Instead of merely maximizing the response quality for a given input, the focus is shifted to maximizing the quality gap of the response, thus promoting self-improvement​ This method takes human preferences and creates an automated way to improve LLMs over time for classic enterprise AI NLP tasks. Another very recent paper (link in the image tag), introduces PromptBreeder - a general-purpose self-referential self-improvement mechanism that evolves and adapts prompts for a given domain. Driven by an LLM, Promptbreeder mutates a population of task-prompts, and subsequently evaluates them for fitness on a training set. Crucially, the mutation of these task-prompts is governed by mutation-prompts that the LLM generates and improves throughout evolution in a self-referential way. That is, Promptbreeder is not just improving task-prompts, but it is also improving the mutation prompts that improve these task-prompts. Promptbreeder outperforms state-of-the-art prompt strategies such as Chain-of-Thought and Plan-and-Solve Prompting on commonly used arithmetic and commonsense reasoning benchmarks. Finally, Self-Taught Optimizer (STOP), employs a language model to enhance arbitrary solutions and then applies this recursively to improve itself. The core idea begins with an initial seed 'improver' scaffolding program that utilizes the language model to improve a solution to some downstream task. They demonstrate that GPT-4 is capable of writing code that can call itself to improve itself. All of these methods explore how LLMs can self-improve based on fine-tuning, implicit human preferences and iterative prompting techniques. LLMs, like humans can take constructive feedback and become better. Now if they can only do it in real-time by just listening 😂 Note: Links to papers in the image descriptions