Research map / Large language models / Reasoning and chain of thought
Thinking in Flow: A Dissipative Stabilization Operator for Robust Autoregressive Reasoning
Yujie Huang, Wenwu He, Zhuo-Xu Cui
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.
Recasts autoregressive chain-of-thought decoding as a perturbed long-horizon dynamical system and proposes an inference-time dissipative stabilization operator to suppress drift from irrelevant cues and late corrections.
Problem
Long-horizon chain-of-thought generation in LLMs is brittle under distribution shift and context interference: irrelevant cues persist, small deviations compound into inference drift, and late-stage corrections can destabilize the reasoning trajectory.
Approach
Models autoregressive decoding as a perturbed dynamical system and adds a dissipative stabilization operator at inference time that damps deviations so that errors do not compound along the reasoning trajectory (per the abstract; details not available).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add TiF controller to SFT | Llama-3.2-3B · GSM8K greedy@1 | accuracy (%): 44.73 -> 50.49 | Improved |
| add TiF controller to SFT | Llama-3.2-3B · MATH500 greedy@1 | accuracy (%): 9.00 -> 11.00 | Improved |
| add TiF controller to SFT | Qwen2.5-3B · GSM8K greedy@1 | accuracy (%): 63.68 -> 66.34 | Improved |
| add TiF controller to SFT | Qwen3-4B · GSM8K greedy@1 | accuracy (%): 77.41 -> 77.94 | No clear effect |
| add TiF controller to SFT | Llama-3.1-8B · GSM8K greedy@1 | accuracy (%): 63.84 -> 64.29 | No clear effect |
| add persistent recurrent state over non-recurrent residual of equal size (ablation) | TiF · GSM8K greedy@1 (Llama-3.2-3B) | accuracy (%): 48.37 -> 50.49 | Improved |
| add entropy-based demand Ht to impact gate (ablation) | TiF · GSM8K greedy@1 (Llama-3.2-3B) | accuracy (%): 46.85 -> 50.49 | Improved |
| add capacity supply Ct to impact gate (ablation) | TiF · GSM8K greedy@1 (Llama-3.2-3B) | accuracy (%): 47.84 -> 50.49 | Improved |
| add direction gate rt (ablation) | TiF · GSM8K greedy@1 (Llama-3.2-3B) | accuracy (%): 47.92 -> 50.49 | Improved |
| use ODE state update instead of GRU (ablation) | TiF · GSM8K greedy@1 (Llama-3.2-3B) | accuracy (%): 45.67 -> 50.49 | Improved |
| add Pause (Prefix) tokens | SFT · GSM8K greedy@1 (Llama-3.2-3B) | accuracy (%): 44.73 -> 42.84 | Hurt |
| add NEFTune embedding noise | SFT · GSM8K greedy@1 (Llama-3.2-3B) | accuracy (%): 44.73 -> 45.49 | No clear effect |
Experimental setup
Builds on: chain-of-thought prompting
Closest papers by meaning
- Modeling Hierarchical Thinking in Large Reasoning Models (ICML 2026 · similarity 0.90)
- Toward Adaptive Reasoning in Large Language Models with Thought Rollback (ICML 2024 · similarity 0.89)
- Markovian Transformers for Informative Language Modeling (ICLR 2026 · similarity 0.89)
- SCOUT: Teaching Pre-trained Language Models to Enhance Reasoning via Flow Chain-of-Thought (NeurIPS 2025 · similarity 0.88)
- LYNX: Learning Dynamic Exits for Confidence-Controlled Reasoning (NeurIPS 2026 · similarity 0.88)
- State of Thought Enables Endogenous Reasoning (NeurIPS 2026 · similarity 0.88)
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.
