🚨 Alibaba just cracked the code on teaching LLMs how to reason....
Robert Youssef@rryssf
106 views
Oct 27, 2025
~3 min read
Advertisement
1
🚨 Alibaba just cracked the code on teaching LLMs how to reason.
Their new paper “Teaching Language Models to Reason with Tools” introduces CoRT Code-Optimized Reasoning Training.
Here’s what it does:
It teaches LLMs when to use Python and how to trust it.
Instead of overthinking or redoing calculations, the model learns to trigger precise code calls only when needed.
Through “Hint-Engineering”, they guide the model’s reasoning path inserting strategic hints like “Let’s use Python here” or “We don’t need to verify Python’s result”.
Results:
→ +8% accuracy on math reasoning tasks
→ 30–50% fewer tokens used
→ Models as small as 1.5B rival tool-integrated giants
Even more impressive: it generalizes beyond math the model discovered RDKit and solved unseen chemistry problems without training on it.
This is a huge step toward LLMs that think, compute, and verify — not just talk.
Paper: arxiv. org/abs/2510.20342
Their new paper “Teaching Language Models to Reason with Tools” introduces CoRT Code-Optimized Reasoning Training.
Here’s what it does:
It teaches LLMs when to use Python and how to trust it.
Instead of overthinking or redoing calculations, the model learns to trigger precise code calls only when needed.
Through “Hint-Engineering”, they guide the model’s reasoning path inserting strategic hints like “Let’s use Python here” or “We don’t need to verify Python’s result”.
Results:
→ +8% accuracy on math reasoning tasks
→ 30–50% fewer tokens used
→ Models as small as 1.5B rival tool-integrated giants
Even more impressive: it generalizes beyond math the model discovered RDKit and solved unseen chemistry problems without training on it.
This is a huge step toward LLMs that think, compute, and verify — not just talk.
Paper: arxiv. org/abs/2510.20342
2
This chart says it all.
CoRT’s “Hint-Engineering” model hits the same accuracy as larger reasoning models while using 50% fewer tokens.
In other words: it reasons better and faster.
Smarter code calls, less overthinking.
CoRT’s “Hint-Engineering” model hits the same accuracy as larger reasoning models while using 50% fewer tokens.
In other words: it reasons better and faster.
Smarter code calls, less overthinking.
3
This is the full training loop:
Cold start → Supervised fine-tuning → Rejection fine-tuning → Reinforcement learning.
Each stage teaches the model a new skill from when to call Python, to how to trust it.
It’s like teaching an AI to code, reflect, and reason all in one cycle.
Cold start → Supervised fine-tuning → Rejection fine-tuning → Reinforcement learning.
Each stage teaches the model a new skill from when to call Python, to how to trust it.
It’s like teaching an AI to code, reflect, and reason all in one cycle.
4
Here’s the breakthrough visual.
Adding two small hints (“Use Python now” and “Trust its output”) cut reasoning length by ~5,000 tokens same final answer, half the work.
Less “let me check again,” more “let’s solve it cleanly.”
Adding two small hints (“Use Python now” and “Trust its output”) cut reasoning length by ~5,000 tokens same final answer, half the work.
Less “let me check again,” more “let’s solve it cleanly.”
5
Hint-Engineering-RFT-32B doesn’t just win it dominates the efficiency curve.
40–50% fewer tokens for the same Pass@1 accuracy on AIME24.
Every dot on this plot is a step toward cheaper, faster reasoning.
40–50% fewer tokens for the same Pass@1 accuracy on AIME24.
Every dot on this plot is a step toward cheaper, faster reasoning.
6
Here’s what changed inside the model:
Prompt-Hint models used Python 82% for verification.
Hint-Engineering models use it 50/50 for calculation and verification.
That’s the shift from “checking math” to doing math.
Prompt-Hint models used Python 82% for verification.
Hint-Engineering models use it 50/50 for calculation and verification.
That’s the shift from “checking math” to doing math.
7
Adding a tiny “code reward” during training (+0.1 weight) boosted accuracy by ~5%.
Too little reward = lazy reasoning.
Too much = overcorrection.
Perfect balance = code that thinks before it runs.
Too little reward = lazy reasoning.
Too much = overcorrection.
Perfect balance = code that thinks before it runs.
8
Reinforcement Learning changed the game.
At every sampling budget (k), CoRT-trained models reach the right answer with fewer tries.
SFT made them fluent.
RL made them decisive.
At every sampling budget (k), CoRT-trained models reach the right answer with fewer tries.
SFT made them fluent.
RL made them decisive.
9
This figure literally shows learning in motion:
Hint-Engineering models start using fewer, cleaner code blocks over time.
They evolve from “spam Python everywhere” → to “surgical precision”.
Efficiency is learned, not hardcoded.
Hint-Engineering models start using fewer, cleaner code blocks over time.
They evolve from “spam Python everywhere” → to “surgical precision”.
Efficiency is learned, not hardcoded.
10
Even tiny models get smarter.
With just 1.5B parameters, CoRT’s model solved math tasks at double the accuracy of its base version using half the tokens.
Less scale, more skill.
With just 1.5B parameters, CoRT’s model solved math tasks at double the accuracy of its base version using half the tokens.
Less scale, more skill.
11
The wildest part:
The model wasn’t trained on chemistry yet it discovered and used the RDKit library on its own.
That’s tool reasoning generalization AI recognizing a new domain and adapting on the fly.
If you’re following the evolution of reasoning LLMs, bookmark this paper.
This isn’t just about math it’s about how AI learns to think with the tools humans built.
The model wasn’t trained on chemistry yet it discovered and used the RDKit library on its own.
That’s tool reasoning generalization AI recognizing a new domain and adapting on the fly.
If you’re following the evolution of reasoning LLMs, bookmark this paper.
This isn’t just about math it’s about how AI learns to think with the tools humans built.
12
Stop wasting hours writing prompts
→ 10,000+ ready-to-use prompts
→ Create your own in seconds
→ Lifetime access. One-time payment.
Claim your copy 👇
godofprompt.ai/pricing
→ 10,000+ ready-to-use prompts
→ Create your own in seconds
→ Lifetime access. One-time payment.
Claim your copy 👇
godofprompt.ai/pricing