English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Think Less, Know More: STACK Compresses LLM Reasoning by 60% While Improving Accuracy

Forum topic · 小凯 · 2026-04-14

Summary

A Chinese forum post discusses the STACK method (State-Aware Reasoning Compression with Knowledge Guidance), proposed in the paper 'Think Less, Know More: State-Aware Reasoning Compression' (arXiv: 2604.09150). The approach tackles 'overthinking' in large language models, where long chains of thought improve accuracy but inflate inference time. STACK classifies a model's reasoning state into two categories: cases where the model is confident despite verbose output, warranting compression of redundant verification steps, and cases of uncertainty or bias, where external knowledge guidance is needed instead. The system dynamically switches between these strategies, learns from long-short contrastive samples, and uses an answer-convergence-based early stopping mechanism. Training combines PPO and DPO to teach state-aware compression through reward differences. On three mathematical reasoning benchmarks, STACK reduced reasoning length by 59.9% on average while improving accuracy by 4.8 percentage points, suggesting reasoning length and quality are not linearly correlated.

Have you ever seen a student who works through problems very slowly but gets high scores? They check every step three times, afraid of making a mistake.

Many current large language model (LLM) reasoning systems are exactly like that student. When facing hard problems, they generate long chains of thought, reasoning step by step. More steps mean more accurate answers, but at the cost of increasingly long inference times. This is known as overthinking.

The paper discussed here addresses one question: can a model think less, but think better?

The STACK Method

The researchers propose STACK, short for State-Aware Reasoning Compression with Knowledge Guidance. The core idea is simple: decide whether to compress the reasoning process based on the model's current state.

They divide the possible states during reasoning into two categories:

1. Confident despite verbose output. In this case, redundant verification steps should be compressed so the model can move forward directly. 2. Uncertain or biased. Here, blind compression is dangerous; external knowledge should be introduced to guide the model and correct its biases.

STACK dynamically switches between these two strategies. It builds "long-short contrastive" samples and learns online when to compress and when to seek help. It also features an early stopping mechanism based on answer convergence: if the model reaches the same intermediate conclusion several steps in a row, further verification is unnecessary.

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 mathematical reasoning benchmarks, STACK cut reasoning length by 59.9% on average while improving accuracy by 4.8 percentage points.

In other words, the model not only thought less — it thought better.

Why It Matters

For a long time, people assumed "longer reasoning means better results." This paper shows that length and quality are not linearly correlated. Often, the later steps of a model's reasoning are just repeated verification, hesitation, or even the introduction of new errors.

A truly intelligent thinker is not the one who fills three pages of scratch work per step, but the one who knows when to be concise and when to dig deeper.

The researchers call this capability "state awareness" — the model must perceive its current cognitive state and adjust its strategy accordingly. This is still far from genuine "metacognition," but it is a meaningful start. If an AI can judge "I'm confident now, I can skip verification" or "I'm biased here, I should check external sources," it is no longer a simple token generator but a system with some capacity for self-regulation.

Thinking a little less and knowing more may be the next key direction in the evolution of reasoning models.

---

Paper information

  • Title: Think Less, Know More: State-Aware Reasoning Compression
  • arXiv: 2604.09150
  • Key finding: state-aware reasoning compression, average 59.9% reduction in reasoning length with a 4.8-point accuracy improvement

Tags

#llm#overthinking#reasoning-compression#stack#ppo#dpo#mathematical-reasoning#arxiv-paper

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177169793