Research map / Large language models / Reasoning and chain of thought
Beyond Tokens: Dynamic Latent Reasoning via Semantic Residual Refinement
Fangrui Lv, Lei Wang, Ruixin Hong, Yong Du, Xiangyu Wu, Tingting Gao, Guorui Zhou, Changshui 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.
Proposes dynamic latent reasoning for LLMs via semantic residual refinement, operating in continuous latent space instead of discrete chain-of-thought tokens.
Problem
Discrete chain-of-thought tokens limit expressiveness and efficiency for abstract, ambiguous, or semantically divergent cognition beyond linguistic tokens.
Approach
Performs reasoning in the model's internal continuous latent space via dynamic semantic residual refinement, rather than generating explicit step-by-step tokens. Details beyond the abstract are not available.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add DyLaR (SRR + dynamic switch) to zero-shot CoT | Qwen2.5-7B · Knowledge QA average | accuracy: 48.20 -> 53.15 | Improved |
| add DyLaR to zero-shot CoT | Llama3.1-8B · Knowledge QA average | accuracy: 50.29 -> 54.11 | Improved |
| add DyLaR to zero-shot CoT | Qwen2.5-7B · STEM average | accuracy: 77.55 -> 79.28 | Improved |
| add DyLaR to zero-shot CoT | Llama3.1-8B · STEM average | accuracy: 64.64 -> 66.72 | Improved |
| use Semantic Residual Refinement instead of feeding raw hidden states (ablation) | DyLaR · MATH500 / TriviaQA (Qwen2.5-7B) | accuracy: 0.00 -> 70.40 (MATH500); 0.00 -> 36.14 (TriviaQA) | Improved |
| use SRR instead of probability-weighted interpolation (ablation) | DyLaR · TriviaQA (Qwen2.5-7B) | accuracy: 29.63 -> 36.14 | Improved |
| use SRR instead of interpolation (ablation) | DyLaR · MATH500 (Qwen2.5-7B) | accuracy: 68.34 -> 70.40 | Improved |
| add entropy-based dynamic switch policy (ablation) | DyLaR · TriviaQA (Qwen2.5-7B) | accuracy: 32.08 -> 36.14 | Improved |
| add dynamic switch policy (ablation) | DyLaR · MATH500 (Qwen2.5-7B) | accuracy: 69.21 -> 70.40 | Improved |
| use dynamic switch instead of fixed latent step (ablation) | DyLaR · TriviaQA (Qwen2.5-7B) | accuracy: 31.61 -> 36.14 | Improved |
| add Coconut fine-tuning | Zero-shot CoT · Knowledge QA / STEM averages | accuracy: 50.29 -> 49.50 (Llama KQA); 77.55 -> 76.47 (Qwen STEM) | Hurt |
| add Soft Thinking (training-free interpolation) | Zero-shot CoT · Knowledge QA average | accuracy: 50.29 -> 50.67 (Llama); 48.20 -> 49.12 (Qwen) | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Soft Thinking — Soft Thinking: Unlocking the Reasoning Potential of LLMs in Continuous Concept Space (NeurIPS 2025) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Experimental setup
Builds on: chain-of-thought prompting
Closest papers by meaning
- Why Struggle with Continuous Latents? Interpretable Discrete Latent Reasoning via Rendered Compression (NeurIPS 2026 · similarity 0.91)
- Efficient Post-Training Refinement of Latent Reasoning in Large Language Models (AAAI 2026 · similarity 0.91)
- Token Assorted: Mixing Latent and Text Tokens for Improved Language Model Reasoning (ICML 2025 · similarity 0.90)
- Geometric Latent Reasoning Induces Shorter Generations in LLMs (NeurIPS 2026 · similarity 0.90)
- SCOUT: Teaching Pre-trained Language Models to Enhance Reasoning via Flow Chain-of-Thought (NeurIPS 2025 · similarity 0.89)
- Keypoint-based Progressive Chain-of-Thought Distillation for LLMs (ICML 2024 · similarity 0.89)
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.
