Research map / Large language models / Reasoning and chain of thought
Adaptive Thinking: Large Language Models Know When to Think in Latent Space
Pingzhi Li, Bairu Hou, Yun Zhu, Yihao Feng, Ke Ye, Tao Lei, Zhifeng Chen, Tianlong Chen, Xianzhi Du
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Studies adaptive allocation of thinking budget for LLMs, deciding when to think in latent space versus explicit chain-of-thought, based only on the abstract provided.
Problem
The relationship between LLM capability, query complexity, and optimal test-time thinking budget allocation is poorly understood for achieving compute-efficient reasoning.
Approach
Adaptive thinking that lets LLMs decide when to think in latent space rather than generating explicit intermediate chain-of-thought, allocating budget per query.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| gate thinking with adapter-predicted self-consistency (Sonata) | Qwen3-8B · Average (5 benchmarks) | accuracy: 74.1 -> 75.3 | Improved |
| gate thinking with Sonata | Qwen3-8B · Average (5 benchmarks) | thinking tokens: 9154 -> 7535 | Improved |
| gate thinking with Sonata | Qwen3-8B · GSM8K | thinking tokens: 1994 -> 890 | Improved |
| gate thinking with Sonata | Qwen3-32B · Average (5 benchmarks) | accuracy: 78.0 -> 77.8 | No clear effect |
| gate thinking with Sonata | Qwen3-8B · GPQA (out-of-domain) | accuracy: 60.1 -> 62.0 | Improved |
| gate thinking with Sonata | GPT-OSS-120B-High · Average | accuracy: 86.6 -> 85.4 | Hurt |
| use constant 4096 thinking budget instead of Sonata | Qwen3-8B · AIME25 | accuracy: 60.0 -> 30.0 | Hurt |
| use Sonata instead of binary self-judge prompt | Qwen3-8B · Average (5 benchmarks) | accuracy: 72.7 -> 75.3 | Improved |
| use self-consistency instead of LM logits entropy as proxy (ablation) | Sonata · Average (4 benchmarks) | accuracy: 66.3 -> 79.6 | Improved |
| use 2-layer MLP instead of linear adapter (ablation) | Sonata · Average (4 benchmarks) | accuracy: 76.8 -> 79.6 | Improved |
| use 3-layer MLP instead of 2-layer (ablation) | Sonata · Average (4 benchmarks) | accuracy: 79.6 -> 79.3 | No clear effect |
| reduce calibration set from 1000 to 100 samples (ablation) | Sonata · Average (4 benchmarks) | accuracy: 79.6 -> 79.0 | No clear effect |
Closest papers by meaning
- Think Deep, Not Just Long: Measuring LLM Reasoning Effort via Deep-Thinking Tokens (ICML 2026 · similarity 0.92)
- Optimal Self-Consistency for Efficient Reasoning with Large Language Models (ICML 2026 · similarity 0.91)
- Think or Not? Exploring Thinking Efficiency in Large Reasoning Models via an Information-Theoretic Lens (NeurIPS 2025 · similarity 0.91)
- Think Smarter not Harder: Adaptive Reasoning with Inference Aware Optimization (ICML 2025 · similarity 0.90)
- Let Me Think! A Long Chain of Thought Can Be Worth Exponentially Many Short Ones (NeurIPS 2025 · similarity 0.90)
- Plan and Budget: Effective and Efficient Test-Time Scaling on Reasoning Large Language Models (ICLR 2026 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
