Research map / Large language models / Reasoning and chain of thought
System-1.5 Reasoning: Traversal in Language and Latent Spaces with Dynamic Shortcuts
Read the paperCheck your idea against 108,967 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 System-1.5 Reasoning, a latent-space reasoning framework with depth and step shortcuts (router-adapters for early exit and hidden-state reuse across decoding steps), trained by two-stage self-distillation to speed up CoT-equivalent reasoning on GSM8K, GSM-HARD and StrategyQA.
Problem
CoT reasoning is token-expensive, and existing latent-space reasoning methods treat all reasoning steps uniformly, so they cannot allocate compute according to step criticality.
Approach
Adds per-layer router-adapter modules so tokens can exit early through lightweight adapters (depth shortcut) and copies early-exit hidden states to the next decoding step (step shortcut); trains by language-to-latent distillation followed by criticality-weighted early-exit distillation with frozen backbone.
Key results
- On GSM8K, System-1.5 reaches 46.66% accuracy vs CoT 46.94% with 2 steps and 20.27x wall-clock speedup.
- On StrategyQA, System-1.5 achieves 48.61% accuracy vs CoT 47.62% in the main table with 55.65x speedup.
- On GSM-HARD, System-1.5 reaches 38.28% vs CoT 38.32% with 12.45x speedup.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use System-1.5 dynamic shortcuts instead of CoT reasoning | Reasoning models · GSM8K | accuracy: 46.94 -> 46.66 | Hurt |
| Use System-1.5 dynamic shortcuts instead of CoT reasoning | Reasoning models · GSM-HARD | accuracy: 38.32 -> 38.28 | No clear effect |
| Use System-1.5 dynamic shortcuts instead of CoT reasoning | Reasoning models · StrategyQA | accuracy: 47.62 -> 48.61 | Helped |
| Use System-1.5 dynamic shortcuts instead of CoT reasoning | Reasoning models · GSM8K | speedup: 20.27$\times$ | Helped |
| Use System-1.5 dynamic shortcuts instead of CoT reasoning | Reasoning models · GSM-HARD | speedup: 12.45$\times$ | Helped |
| Use System-1.5 dynamic shortcuts instead of CoT reasoning | Reasoning models · StrategyQA | speedup: 55.65$\times$ | Helped |
| Use Coconut as the System-2 student for shortcut learning instead of direct language-to-latent distillation (ablation) | System-1.5 · Reasoning tasks (Figure [ref]) | accuracy | Hurt |
| Use joint language-to-latent alignment and shortcut learning (ablation) | System-1.5 · Reasoning tasks (Figure [ref]) | performance | Hurt |
| Update all model parameters during shortcut learning instead of freezing Transformer parameters (ablation) | System-1.5 · Reasoning tasks (Figure [ref]) | performance | Hurt |
| Vary depth exit threshold and decoding step constant (ablation) | System-1.5 · Reasoning tasks (Figure [ref]) | performance | Mixed |
Surprising findings
- Joint training of alignment and shortcut learning, and full-parameter shortcut learning, both degrade performance.
- Coconut as the System-2 student gives notably worse System-1.5 results than CODI, despite Coconut's curriculum-based design.
Limitations
- Accuracy on GSM8K and GSM-HARD is essentially matched to CoT rather than improved, and the reported numbers are small-model (GPT-2 124M, LLaMA 3.2 1B) results.
- Criticality labels depend on atom-of-thought decomposition of the CoT, adding an annotation dependency.
Experimental setup
Datasets: GSM8K (augmented) StrategyQA
Benchmarks: GSM8K GSM-HARD StrategyQA
Metrics: accuracy number of decoding steps FLOPs reduction rate wall-clock speedup
Baselines: CoT fine-tuning LITE LayerSkip iCoT Coconut CODI pause token
Builds on: Coconut CODI LayerSkip atom-of-thought iCoT
Closest papers by meaning
- SCOUT: Teaching Pre-trained Language Models to Enhance Reasoning via Flow Chain-of-Thought (NeurIPS 2025 · similarity 0.91)
- Eliciting Chain-of-Thought in Base LLMs via Gradient-Based Representation Optimization (AAAI 2026 · similarity 0.91)
- Long Chain-of-Thought Reasoning Across Languages (ICLR 2026 · similarity 0.90)
- Deep Hidden Cognition Facilitates Reliable Chain-of-Thought Reasoning (AAAI 2026 · similarity 0.90)
- Reasoning with Latent Thoughts: On the Power of Looped Transformers (ICLR 2025 · similarity 0.90)
- Facts Don't Speak Louder than Words: The Behavioral Essence of Long-CoT Distillation (NeurIPS 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.