Can anyone ELI5? Or at least a kind of layman's explanation?
Auto-Differentiating Any LLM Workflow: A Farewell to Manual Prompting
11–20 of 33 posts
Re: Auto-Differentiating Any LLM Workflow: A Farewell to Manual Prompting
#12Can anyone ELI5? Or at least a kind of layman's explanation?
Consider a complex LLM pipeline with multiple steps. each LLM evaluation has an associated prompt to shape the style/context of the response. Conventionally these prompts are treated like hyperparameters that have to be manually adjusted to get the desired behavior from the LLM.
This work introduces a way to treat these prompts like trainable parameters, updating them through automatic differentiation of some kind of supervised training loss.
For me it kind of feels like deep dream or style transfer, which use autograd to optimize the model inputs (instead of the parameters) to achieve some goal (like mixing the style and content of two input images)
Re: Auto-Differentiating Any LLM Workflow: A Farewell to Manual Prompting
#13Just read the abstract so far. Sounds amazing, but just for the sake of understanding, what would be the inputs and outputs of such a system? If the prompt is generated, how do you tell the system what you'd like to have? And what is the ground truth that is trained against? Examples of the desired text?
I think what they mean is intermediate prompts, i.e. the prompts that the system gives to itself when solving a problem that requires multiple stages.
Re: Auto-Differentiating Any LLM Workflow: A Farewell to Manual Prompting
#14Just read the abstract so far. Sounds amazing, but just for the sake of understanding, what would be the inputs and outputs of such a system? If the prompt is generated, how do you tell the system what you'd like to have? And what is the ground truth that is trained against? Examples of the desired text?
I think what they mean is intermediate prompts, i.e. the prompts that the system gives to itself when solving a problem that requires multiple stages.
Ah, that makes sense. So (very) basically, they're putting a number of regular LLMs into a sort of compute chain/graph, where one LLM feeds into the other, then doing gradient descent on the whole chain at once, essentialy treating the boundaries between LLM n and LLM n+1 as "hidden layers"?
Re: Auto-Differentiating Any LLM Workflow: A Farewell to Manual Prompting
#15Can anyone ELI5? Or at least a kind of layman's explanation?
This paper suggests that LLMs can be trained to handle multi-stage questioning by automatically optimizing prompts using feedback-based methods, improving their ability to process complex, multi-step interactions.
Re: Auto-Differentiating Any LLM Workflow: A Farewell to Manual Prompting
#16Can anyone ELI5? Or at least a kind of layman's explanation?
less words makes more prompt
Re: Auto-Differentiating Any LLM Workflow: A Farewell to Manual Prompting
#17> - the task of crafting textual inputs to effectively direct LLMs -- remains difficult and labor-intensive
Damn, I knew we were lazy but describing prompting as labor-intensive is impressively lazy even to me.
Obviously reading the rest of the abstract was too labor intensive for me but I'm hoping I can just hook a probe up to my drool and it can infer my desires from that.
Re: Auto-Differentiating Any LLM Workflow: A Farewell to Manual Prompting
#18> - the task of crafting textual inputs to effectively direct LLMs -- remains difficult and labor-intensive Damn, I knew we were lazy but describing prompting as labor-intensive is impressively lazy even to me. Obviously reading the rest of the abstract was too labor intensive for me but I'm hoping I can just hook a probe up to my drool and it can infer my desires from that.
this made my day
Re: Auto-Differentiating Any LLM Workflow: A Farewell to Manual Prompting
#19Earlier quoted context omitted.
I think what they mean is intermediate prompts, i.e. the prompts that the system gives to itself when solving a problem that requires multiple stages.
Ah, that makes sense. So (very) basically, they're putting a number of regular LLMs into a sort of compute chain/graph, where one LLM feeds into the other, then doing gradient descent on the whole chain at once, essentialy treating the boundaries between LLM n and LLM n+1 as "hidden layers"?
Author here. Yea, in this fashion. And it can create the feedback using llm as a backward engine
Re: Auto-Differentiating Any LLM Workflow: A Farewell to Manual Prompting
#20Wow, just when I’d accepted MIPRO in DSPy was magic, here we are. Things continue apace.
Yup. The LLM-AutoDiff is just getting started. But it has proven generation-only without explicitly doing few-shot samples can be even more effective and create shorter final prompts