foxarmy

I ran a quick experiment examining how DeepSeek-R1 performs on agentic jobs, wiki.asexuality.org regardless of not supporting tool usage natively, and I was rather amazed by preliminary outcomes. This experiment runs DeepSeek-R1 in a single-agent setup, where the design not only plans the actions however likewise creates the actions as executable Python code. On a subset1 of the GAIA validation split, DeepSeek-R1 surpasses Claude 3.5 Sonnet by 12.5% outright, from 53.1% to 65.6% appropriate, and other designs by an even larger margin:

The experiment followed design use guidelines from the DeepSeek-R1 paper and the model card: Don't utilize few-shot examples, avoid including a system prompt, and set the temperature to 0.5 - 0.7 (0.6 was used). You can find more assessment details here.

Approach

DeepSeek-R1's strong coding capabilities enable it to serve as a representative without being clearly trained for tool usage. By permitting the design to generate actions as Python code, it can flexibly connect with environments through code execution.

Tools are carried out as Python code that is included straight in the timely. This can be an easy function definition or a module of a bigger bundle - any legitimate Python code. The design then creates code actions that call these tools.

Results from executing these actions feed back to the design as follow-up messages, driving the next actions till a last response is reached. The representative structure is an easy iterative coding loop that mediates the discussion between the model and its environment.

Conversations

DeepSeek-R1 is utilized as chat design in my experiment, where the model autonomously pulls additional context from its environment by using tools e.g. by utilizing an online search engine or fetching information from web pages. This drives the conversation with the environment that continues till a last answer is reached.

On the other hand, o1 designs are known to carry out inadequately when utilized as chat designs i.e. they do not try to pull context throughout a discussion. According to the connected post, o1 designs perform best when they have the complete context available, with clear instructions on what to do with it.

Initially, I also tried a full context in a single timely approach at each step (with outcomes from previous steps consisted of), however this caused substantially lower scores on the GAIA subset. Switching to the conversational method explained above, I was able to reach the reported 65.6% performance.

This raises a fascinating question about the claim that o1 isn't a chat model - possibly this observation was more relevant to older o1 designs that did not have tool usage capabilities? After all, isn't tool use support an essential system for enabling models to pull extra context from their environment? This conversational method certainly seems efficient for DeepSeek-R1, though I still need to carry out similar explores o1 designs.

Generalization

Although DeepSeek-R1 was mainly trained with RL on mathematics and coding jobs, it is amazing that generalization to agentic tasks with tool usage via code actions works so well. This capability to generalize to agentic jobs advises of recent research by DeepMind that reveals that RL generalizes whereas SFT memorizes, although generalization to tool use wasn't examined because work.

Despite its capability to generalize to tool usage, DeepSeek-R1 typically produces long at each step, compared to other models in my experiments, restricting the usefulness of this model in a single-agent setup. Even simpler jobs often take a very long time to finish. Further RL on agentic tool usage, be it via code actions or not, might be one alternative to enhance performance.

Underthinking

I also observed the underthinking phenomon with DeepSeek-R1. This is when a reasoning model often switches in between different reasoning thoughts without adequately checking out appealing paths to reach a correct solution. This was a major reason for excessively long thinking traces produced by DeepSeek-R1. This can be seen in the recorded traces that are available for download.

Future experiments

Another typical application of reasoning models is to utilize them for planning only, while using other designs for generating code actions. This might be a potential new function of freeact, if this separation of functions shows beneficial for more complex tasks.

I'm likewise curious about how reasoning designs that currently support tool use (like o1, o3, ...) perform in a single-agent setup, with and without generating code actions. Recent developments like OpenAI's Deep Research or Hugging Face's open-source Deep Research, which likewise utilizes code actions, look fascinating.