I ran a fast experiment investigating how DeepSeek-R1 performs on agentic tasks, in spite of not supporting tool usage natively, and I was rather satisfied by preliminary outcomes. This experiment runs DeepSeek-R1 in a single-agent setup, where the model not only prepares the actions but also develops the actions as executable Python code. On a subset1 of the GAIA validation split, DeepSeek-R1 outshines Claude 3.5 Sonnet by 12.5% absolute, from 53.1% to 65.6% appropriate, and other models by an even bigger margin:

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

Approach

DeepSeek-R1's strong coding abilities allow it to function as an agent without being explicitly trained for tool usage. By allowing 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 prompt. This can be a simple function meaning or a module of a bigger package - any valid Python code. The design then generates code actions that call these tools.

Arise from performing these actions feed back to the model as follow-up messages, driving the next actions until a last answer is reached. The representative structure is a simple iterative coding loop that mediates the discussion in between the design and its environment.

Conversations

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

In contrast, experienciacortazar.com.ar o1 models are known to perform poorly when used as chat designs i.e. they do not try to pull context during a conversation. According to the connected short article, o1 designs perform best when they have the complete context available, setiathome.berkeley.edu with clear guidelines on what to do with it.

Initially, I also attempted a complete context in a single prompt method at each action (with arise from previous steps included), however this caused considerably lower ratings on the GAIA subset. Switching to the conversational method explained above, I was able to reach the reported 65.6% performance.

This raises an intriguing concern about the claim that o1 isn't a chat design - perhaps this observation was more relevant to older o1 designs that lacked tool use capabilities? After all, isn't tool usage support a crucial mechanism for enabling models to pull additional context from their environment? This conversational technique certainly seems effective for DeepSeek-R1, though I still need to carry out comparable experiments with o1 models.

Generalization

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

Despite its capability to generalize to tool use, DeepSeek-R1 typically produces long reasoning traces at each step, compared to other models in my experiments, limiting the effectiveness of this design in a single-agent setup. Even simpler tasks sometimes take a long period of time to finish. Further RL on agentic tool usage, be it through code actions or not, could be one choice to enhance efficiency.

Underthinking

I also observed the underthinking phenomon with DeepSeek-R1. This is when a reasoning model frequently changes in between various reasoning thoughts without sufficiently checking out promising courses to reach an appropriate service. This was a significant factor for overly long reasoning traces produced by DeepSeek-R1. This can be seen in the recorded traces that are available for download.

Future experiments

Another typical application of thinking models is to utilize them for planning only, while using other models for creating code actions. This could be a prospective new function of freeact, if this separation of roles proves beneficial for more complex jobs.

I'm likewise curious about how thinking designs that already support tool usage (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 uses code actions, look interesting.