Past benchmarks: How DeepSeek-R1 and o1 carry out on real-world duties

DeepSeek-R1 has certainly created quite a lot of pleasure and concern, particularly for OpenAI’s rival mannequin o1. So, we put them to check in a side-by-side comparability on a number of easy knowledge evaluation and market analysis duties. 

To place the fashions on equal footing, we used Perplexity Professional Search, which now helps each o1 and R1. Our aim was to look past benchmarks and see if the fashions can truly carry out advert hoc duties that require gathering info from the net, selecting out the appropriate items of knowledge and performing easy duties that will require substantial guide effort. 

Each fashions are spectacular however make errors when the prompts lack specificity. o1 is barely higher at reasoning duties however R1’s transparency offers it an edge in instances (and there can be fairly a number of) the place it makes errors.

Here’s a breakdown of some of our experiments and the hyperlinks to the Perplexity pages the place you may overview the outcomes your self.

Calculating returns on investments from the net

Our first check gauged whether or not fashions might calculate returns on funding (ROI). We thought of a situation the place the consumer has invested $140 within the Magnificent Seven (Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia, Tesla) on the primary day of each month from January to December 2024. We requested the mannequin to calculate the worth of the portfolio on the present date.

To perform this job, the mannequin must pull Magazine 7 worth info for the primary day of every month, cut up the month-to-month funding evenly throughout the shares ($20 per inventory), sum them up and calculate the portfolio worth in accordance with the worth of the shares on the present date.

On this job, each fashions failed. o1 returned an inventory of inventory costs for January 2024 and January 2025 together with a formulation to calculate the portfolio worth. Nevertheless, it didn’t calculate the proper values and principally mentioned that there can be no ROI. Then again, R1 made the error of solely investing in January 2024 and calculating the returns for January 2025.

Nevertheless, what was attention-grabbing was the fashions’ reasoning course of. Whereas o1 didn’t present a lot particulars on the way it had reached its outcomes, R1’s reasoning traced confirmed that it didn’t have the proper info as a result of Perplexity’s retrieval engine had didn’t acquire the month-to-month knowledge for inventory costs (many retrieval-augmented era functions fail not due to the mannequin lack of talents however due to unhealthy retrieval). This proved to be an necessary little bit of suggestions that led us to the subsequent experiment.

Reasoning over file content material

We determined to run the identical experiment as earlier than, however as an alternative of prompting the mannequin to retrieve the knowledge from the net, we determined to offer it in a textual content file. For this, we copy-pasted inventory month-to-month knowledge for every inventory from Yahoo! Finance right into a textual content file and gave it to the mannequin. The file contained the identify of every inventory plus the HTML desk that contained the worth for the primary day of every month from January to December 2024 and the final recorded worth. The info was not cleaned to scale back the guide effort and check whether or not the mannequin might choose the appropriate components from the info.

Once more, each fashions failed to offer the appropriate reply. o1 appeared to have extracted the info from the file, however steered the calculation be carried out manually in a software like Excel. The reasoning hint was very imprecise and didn’t comprise any helpful info to troubleshoot the mannequin. R1 additionally failed and didn’t present a solution, however the reasoning hint contained quite a lot of helpful info.

For instance, it was clear that the mannequin had accurately parsed the HTML knowledge for every inventory and was in a position to extract the proper info. It had additionally been in a position to do the month-by-month calculation of investments, sum them and calculate the ultimate worth in accordance with the most recent inventory worth within the desk. Nevertheless, that last worth remained in its reasoning chain and didn’t make it into the ultimate reply. The mannequin had additionally been confounded by a row within the Nvidia chart that had marked the corporate’s 10:1 inventory cut up on June 10, 2024, and ended up miscalculating the ultimate worth of the portfolio. 

Once more, the actual differentiator was not the end result itself, however the skill to analyze how the mannequin arrived at its response. On this case, R1 offered us with a greater expertise, permitting us to grasp the mannequin’s limitations and the way we are able to reformulate our immediate and format our knowledge to get higher outcomes sooner or later.

Evaluating knowledge over the net

One other experiment we carried out required the mannequin to match the stats of 4 main NBA facilities and decide which one had the perfect enchancment in subject aim proportion (FG%) from the 2022/2023 to the 2023/2024 seasons. This job required the mannequin to do multi-step reasoning over totally different knowledge factors. The catch within the immediate was that it included Victor Wembanyama, who simply entered the league as a rookie in 2023.

The retrieval for this immediate was a lot simpler, since participant stats are extensively reported on the net and are often included of their Wikipedia and NBA profiles. Each fashions answered accurately (it’s Giannis in case you have been curious), though relying on the sources they used, their figures have been a bit totally different. Nevertheless, they didn’t notice that Wemby didn’t qualify for the comparability and gathered different stats from his time within the European league.

In its reply, R1 offered a greater breakdown of the outcomes with a comparability desk together with hyperlinks to the sources it used for its reply. The added context enabled us to right the immediate. After we modified the immediate specifying that we have been in search of FG% from NBA seasons, the mannequin accurately dominated out Wemby from the outcomes.

Ultimate verdict

Reasoning fashions are highly effective instruments, however nonetheless have a methods to go earlier than they are often absolutely trusted with duties, particularly as different parts of enormous language mannequin (LLM) functions proceed to evolve. From our experiments, each o1 and R1 can nonetheless make primary errors. Regardless of exhibiting spectacular outcomes, they nonetheless want a little bit of handholding to offer correct outcomes.

Ideally, a reasoning mannequin ought to be capable of clarify to the consumer when it lacks info for the duty. Alternatively, the reasoning hint of the mannequin ought to be capable of information customers to raised perceive errors and proper their prompts to extend the accuracy and stability of the mannequin’s responses. On this regard, R1 had the higher hand. Hopefully, future reasoning fashions, together with OpenAI’s upcoming o3 sequence, will present customers with extra visibility and management.