Research map / Large language models / Reasoning and chain of thought
Rethinking LLM Reasoning: From Explicit Trajectories to Latent Representations
Cong Jiang, Xiaofeng Zhang, Fangzhi Zhu, XiaoWei Chen, Junxiong Zhu, Zheng Zhang
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.
Survey-style rethinking of LLM reasoning that contrasts explicit reasoning trajectories with latent representations to reduce inference cost; abstract only available.
Problem
Reasoning trajectories generated by LLMs are often much longer than final answers, causing substantial inference cost even for simple tasks.
Approach
Frames the field as a move from explicit step-by-step reasoning trajectories toward latent reasoning representations; the abstract is truncated so the specific proposed method is not stated.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace explicit reasoning with latent reasoning tokens from a reasoning network | DeepSeek-R1-Distill-Qwen-1.5B, 512-token budget · 5-benchmark average | accuracy: 37.93 -> 44.91 | Improved |
| use LRT instead of LC-R1 at 1024-token budget | DeepSeek-R1-Distill-Qwen-1.5B · 5-benchmark average | accuracy: 48.53 -> 48.58 | No clear effect |
| use LRT instead of baseline at 1024-token budget on AMC | DeepSeek-R1-Distill-Qwen-1.5B · AMC | accuracy: 42.88 -> 42.50 | No clear effect |
| convert Qwen3-1.7B thinking mode to latent reasoning vs non-thinking, pass@1 | Qwen3-1.7B · 5-benchmark average | accuracy: 46.93 -> 48.42 | Improved |
| convert Qwen3-4B thinking mode to latent reasoning vs non-thinking, pass@4 | Qwen3-4B · 5-benchmark average | accuracy: 65.78 -> 71.60 | Improved |
| use LRT instead of Qwen3-8B non-thinking | Qwen3-8B · 5-benchmark average | accuracy: 56.44 -> 59.63 | Improved |
| increase latent tokens from 64 to 256 (ablation) | LRT, Qwen3-1.7B · 5-benchmark average | accuracy: 42.53 -> 48.42 | Improved |
| increase latent tokens from 256 to 512 (ablation) | LRT, Qwen3-1.7B · 5-benchmark average | accuracy: 48.42 -> 46.92 | Hurt |
| increase latent tokens from 256 to 512 (ablation) | LRT, Qwen3-8B · 5-benchmark average | accuracy: 59.63 -> 59.97 | Improved |
| add RL stage after SFT (ablation) | LRT, Qwen3-1.7B · 5-benchmark average | accuracy: 41.29 -> 48.42 | Improved |
| randomly skip 50% of reasoning tokens in the trajectory (ablation) | DeepSeek-R1-Distill-Qwen-7B · MATH-500 | accuracy: 92.8 -> 90.60 | No clear effect |
| use LRT instead of thinking mode at 4096 budget | Qwen3-8B · 5-benchmark average | accuracy: 59.63 -> 69.61 | Hurt |
Limitations
- Only the truncated abstract was available; full method and results are unknown.
Closest papers by meaning
- Token Assorted: Mixing Latent and Text Tokens for Improved Language Model Reasoning (ICML 2025 · similarity 0.93)
- Reasoning via Test-Time Instance-Level Policy Gradient in Latent Space (NeurIPS 2026 · similarity 0.93)
- Geometric Latent Reasoning Induces Shorter Generations in LLMs (NeurIPS 2026 · similarity 0.92)
- A State-Transition Framework for Efficient LLM Reasoning (ICLR 2026 · similarity 0.92)
- DOTS: Learning to Reason Dynamically in LLMs via Optimal Reasoning Trajectories Search (ICLR 2025 · similarity 0.92)
- Thinking on the Fly: Test-Time Reasoning Enhancement via Latent Thought Policy Optimization (ICLR 2026 · similarity 0.92)
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.
