Have you ever seen a student who solves problems very slowly but with high accuracy? They check every step three times, afraid of making a mistake.
Many current LLM reasoning systems behave exactly like that student. When facing hard problems, they generate very long chains of thought, deriving step by step. More steps generally mean more correct answers — but at the cost of increasingly long inference times. This is known as overthinking.
The paper tackles a simple question: can we make models think less, but think better?
The STACK Method
The proposed method is called STACK — State-Aware Reasoning Compression with Knowledge Guidance. The core idea: decide whether to compress the reasoning process based on the model's current state.
The authors divide possible reasoning states into two categories:
1. Confident despite verbose output — the model writes a lot but is actually certain. Redundant verification steps should be compressed and the model should move forward directly. 2. Uncertain or biased — blind compression is dangerous here. Instead, external knowledge should be introduced to guide the model and correct its biases.
STACK dynamically switches between these two strategies. It builds "long-short contrast" samples for online learning of when to compress and when to seek help. It also uses an early-stopping mechanism based on answer convergence — if the model repeatedly reaches the same intermediate conclusion over consecutive steps, there is no need to keep verifying.
Training Approach
Rather than simple supervised learning, the authors combine PPO (Proximal Policy Optimization) and DPO (Direct Preference Optimization), letting the model learn a state-aware compression policy through reward differences.
Results
On three math reasoning benchmarks, STACK on average:
- Cut reasoning length by 59.9%
- Improved accuracy by 4.8 percentage points
- Title: Think Less, Know More: State-Aware Reasoning Compression
- arXiv: 2604.09150
- Core finding: State-aware reasoning compression reduces reasoning length by 59.9% on average while improving accuracy by 4.8 points
In other words, the model thinks less — and thinks better.
Why This Matters
For a long time, people assumed "longer reasoning means better results." But this paper shows that length and quality are not linearly correlated. In many cases, the later steps of a reasoning chain are just repeated verification, hesitation, or even new errors.
A truly intelligent thinker is not the one who fills three pages of scratch paper at every step, but the one who knows when to be concise and when to go deep.
The authors call this capability "state-awareness" — the model must sense its own cognitive state and adjust its strategy accordingly. This is still far from true "metacognition," but it is a meaningful start. An AI that can judge "I'm confident, skip verification" or "I'm biased, check some references" is no longer a simple token generator, but a system with some self-regulation.
Thinking a little less, knowing more — this may be the next key direction for reasoning model evolution.
---
Paper information