yo, i feel u on this. honestly, i've been playing with DeepSeek-R1 for a few months now and i've noticed it definitely likes to rush if you dont give it a nudge. for your situation, i would suggest skipping the fancy personas cuz they can be a bit hit or miss... instead, just use a super simple system prompt like "You are a logical assistant. ALWAYS show your full chain-of-thought in tags before answering. think step-by-step and double-check your math." basically it forces the model to slow down. also, if ur worried about costs, using the DeepSeek API is way cheaper than other big models—im talking like $0.14 per million tokens for input. lowkey a total steal compared to others. i mean, i guess it depends on your needs, but giving it a clear structure to follow usually works for me!! gl!

Basically had the same issue last month. I found that DeepSeek-R1 really started shining for my math work once I basically forced it to use a strict delimiter-based reflection block before giving any final code. I mean, honestly, it's kinda risky if you dont monitor the tokens, but setting a constraint like 'first, verify your logic for edge cases' actually fixed the shallow reasoning I was getting.

Yep, this is the way

Respectfully, I'd consider another option before you go all-in on those massive reflection blocks. I've been using DeepSeek-R1 for some heavy-duty engineering tasks lately, and honestly, forcing it to "think deep" for every single prompt can get SO expensive if you arent careful with your token usage. I once set a really aggressive system prompt that basically demanded a huge logical breakdown for every simple math problem, and while the logic was okay, it started looping and burned through my API credits way faster than expected.

Not sure if this is standard, but iirc, the model actually performs better when you give it a specific "budget" or tell it to focus on edge cases specifically rather than just "thinking aloud." I think you gotta be careful about over-prompting because it can lead to reasoning fatigue where it just repeats itself to satisfy your constraints. Maybe try a lighter touch? Make sure to monitor things cuz those long chain-of-thought outputs add up fast. Just my two cents tho! 👍

So I'm still a bit of a beginner with all the different models out there, but I've been trying to see how DeepSeek compares to OpenAI or Anthropic from a price-to-performance view lately. Not sure if this helps, but I think I read somewhere that DeepSeek-R1 might actually respond better if you dont try to over-engineer it like people do with GPT? Like, maybe the best system prompt is actually no prompt at all? IIRC, some of the market research I saw suggested that the model has its own internal 'reasoning' trigger that can actually get messed up if we add too many extra instructions. I mean... I havent tested it enough to be 100% sure, but has anyone tried just giving it the raw problem and comparing it to how o1 handles it? Tbh it feels like the market is shifting toward models that dont need these massive system prompts anymore, which would definitely save us some money on tokens. If the model is already trained to think, are we just wasting credits by asking it to do what its already designed for?

Good to know!

Just catching up on this thread now. Honestly, I'm super satisfied with how my setup is running lately and have basically no complaints about the reasoning depth. Like someone mentioned about not over-engineering the prompts, I totally get that, but I politely disagree that it is strictly a prompt issue. I've noticed some real compatibility quirks depending on where you're running the model.

• You should just go with NVIDIA if you're looking at hardware, you really can't go wrong with their stuff for keeping the reasoning chains stable.
• If you're doing the API thing, stick with anything from OpenRouter since they seem to handle the raw output way better than most.
• Sometimes the logic feels shallow because the interface you're using is stripping out the thought blocks before you even see them, ngl.

Lol I was literally about to post the same thing. Glad someone else brought it up.

Following this thread