Reflexion: Language Agents with Verbal Reinforcement Learning

No. CoT has been around for several years (Jan 2022) https://arxiv.org/abs/2201.11903. And so has Reflection (March 2023) https://arxiv.org/abs/2303.11366. This approach that was taken by Reflection is nothing new.

Reflexion can serve as good reference here: https://arxiv.org/abs/2303.11366 Essentially, AI’s output is fed into a checker, whose output is fed back into the AI for “reflexion”. Then the AI often corrects (leading to noticeable improvement in GPT-4 perf).

I don't know how we'd go about comparing reflection to human baselines but GPT-4 can reflect just fine https://arxiv.org/abs/2303.11366 GPT-4's emotional intelligence seems to be very high by the looks of things https://arxiv.org/abs/2304.11490 Empathy =/ Intelligence

Chain of thought: https://arxiv.org/abs/2201.11903 ReAct: https://arxiv.org/abs/2210.03629 Reflexion: https://arxiv.org/abs/2303.11366 Tree of thoughts: https://arxiv.org/abs/2305.10601 Good video on "Tree of thoughts" which also reviews / puts it in the context of other methods: https://www.youtube.com/watch?v=ut5kp56wW_4 Completion vs conversational interface is something you can read about in the OpenAI API documentation. For the remaining things I don't have single specific pointer at hand.

If performance is a concern, just re-prompt GPT if there are any performance concerns, and to propose fixes for them if there are. This is called reflexion and has been found to be an effective way to improve response accuracy and usefulness (also reduces hallucinations). https://arxiv.org/abs/2303.11366

There's a lot of work happening at the moment around self-reflection and getting LLMs to identify and correct their own hallucinations and mistakes. https://arxiv.org/pdf/2303.11366.pdf

The Reflexion paper (https://arxiv.org/abs/2303.11366) that came out recently shows how this kind of mistake might be overcome. Asking the model to think about the answer after it's generated a first draft greatly improves accuracy. Also, prompt engineering such as copying the generated code, pasting it in a new chat and saying "There's a bug in this code, please find it" can go a long way. There is so much low hanging fruit in harnessing the power of these models that is just being ignored because some even lower hanging fruit (RLHF, system messages, context window size, plugins, etc) is being released seemingly every few days.

Reflexion: an autonomous agent with dynamic memory and self-reflection, https://arxiv.org/abs/2303.11366 Add an interpreter as per this article, and we're basically there.

It was already trained like this, RLHF [0] is a loop that has ways to feed it the information to "move forward" - update its weights. It was fine-tuned like this. It certainly did ignite something - maybe sparks of AGI, maybe just sparks of pretending to be more liked by humans (hence Human Feedback in RLHF). My guess is that loops by itself (Reflexion [1], etc.) won't change much. ChatGPT massive logs - maybe. [0] https://huggingface.co/blog/rlhf [1] https://arxiv.org/abs/2303.11366

the paper: https://arxiv.org/pdf/2303.11366.pdf it reminds me of the thinking-fast vs. thinking-slow dichotomy. Current llms are the thinking fast type. Funnily people’s complains about its errors are reminiscent of this. It answers just to quick and only with its instant response neural net. A thinking slow answer would be more akin to a chain of thought answer. Allowing the llm a more flexible platform than CoT promptin might well be the next step. Of course it would als multiply compute cost. So it might not be in your 20$ subscription

>GPT cannot function in any environment on its own. Except in response to a direct instruction it cannot plan, it cannot take action, it cannot learn, it cannot adjust to changing circumstance. Sure it can. It's not default behavior sure but it's fairly trivial to set up, just expensive. Gpt-4 can loop on its "thoughts" and reflect, it can take actions in the real world. https://tidybot.cs.princeton.edu/ https://arxiv.org/abs/2304.03442 https://arxiv.org/abs/2210.03629 https://arxiv.org/abs/2303.11366

>the systems operation are well understood That's like saying human behavior is well understood because we know how neurons communicate signals. It's too low level to be useful, hence psychology. >They don't have the ability to reason or reflect. Yes they do https://selfrefine.info/ https://arxiv.org/abs/2303.11366

That part on providing the opportunity to check its answers is called reflexion and has been shown to significantly improve accuracy and reliability of LLM responses. Very cool. https://arxiv.org/abs/2303.11366 https://evjang.com/2023/03/26/self-reflection.html