Intention Chain-of-Thought Prompting with Dynamic Routing for Code Generation

Conference: AAAI 2026 arXiv: 2512.14048 Code: https://github.com/Guai001/RoutingGen Area: LLM Reasoning / Code Generation Keywords: Intention Chain, Dynamic Routing, Code Generation, Cognitive Economy, Difficulty-Awareness

TL;DR

This paper proposes RoutingGen — a difficulty-aware code generation framework grounded in the principle of cognitive economy. A Qwen3-8B classifier dynamically routes tasks to either a simple path (few-shot direct generation) or a complex path (Intention CoT = specification constraints + algorithmic intent + complexity analysis), achieving a +45.15% improvement on McEval while reducing average token consumption by 46.37%.

Background & Motivation

Background: Chain-of-Thought (CoT) prompting is effective for code generation — it guides LLMs to reason about algorithmic design before writing code. However, applying CoT uniformly to all tasks introduces an "overthinking" problem.

Limitations of Prior Work: - Applying CoT to simple tasks (e.g., string reversal) reduces efficiency and may introduce erroneous reasoning. - Existing CoT methods lack "intent abstraction" — they focus on syntactic correctness while neglecting core algorithm design and efficiency. - Treating all tasks identically violates the principle of cognitive economy (only difficult tasks warrant deep reasoning).

Key Challenge: CoT is beneficial for hard tasks but wasteful for easy ones — an adaptive mechanism is needed to determine when to activate structured reasoning.

Goal: Dynamically select a generation strategy based on problem difficulty: simple problems → direct generation; complex problems → intent-level CoT reasoning followed by generation.

Key Insight: The cognitive economy principle — activate structured reasoning only when necessary, so as to conserve cognitive resources.

Core Idea: Difficulty-classifier routing + Intention CoT (Specification + Algorithmic Intent + Complexity) = efficient and high-quality code generation.

Method

Overall Architecture

An input programming problem is fed to a Qwen3-8B difficulty classifier (Simple/Complex + rationale) → Simple path: few-shot direct generation → Complex path: ICoT (Specification + Idea) → conditional code generation.

Key Designs

1. Difficulty-Aware Dynamic Routing:

   • Function: Automatically assesses problem difficulty and selects the appropriate generation strategy.
   • Mechanism: The classifier outputs a Simple/Complex label along with a textual rationale. The simple path skips CoT and generates directly; the complex path enters ICoT.
   • Design Motivation: Cognitive economy — avoids performing algorithmic analysis on trivially simple tasks such as "print hello world."

2. Intention Chain-of-Thought (ICoT):

   • Function: Conducts intent-level reasoning prior to code generation.
   • Mechanism: Comprises two components — (1) Specification: input/output constraints, boundary conditions, and data types ("what"); (2) Idea: core algorithmic logic, time complexity targets, and key data structure choices ("how" and "why").
   • Design Motivation: Conventional CoT focuses on surface-level steps ("first read input, then loop..."), whereas ICoT targets algorithmic intent ("use dynamic programming because subproblems overlap"), more closely mirroring how human programmers think.

3. Conditional Code Generation:

   • Function: Generates code conditioned on ICoT output.
   • Mechanism: The Specification and Idea produced by ICoT are injected as context into the code generation prompt.
   • Design Motivation: With a clearly articulated algorithmic intent, code generation can focus on implementation rather than design.

Loss & Training

• Classifier: Qwen3-8B, used without fine-tuning (prompt-based).
• Code generation: Various LLMs (GPT-4o, DeepSeek-Coder, etc.).

Key Experimental Results

Main Results

Benchmark	RoutingGen	Baseline	Gain	Token Savings
HumanEval	77.10%	74.97%	+2.13%	—
MBPP	69.11%	56.59%	+12.52%	—
McEval	38.90%	26.80%	+45.15%	—
Avg. Tokens	—	—	—	−46.37%

Ablation Study

Configuration	MBPP Pass@1	Notes
No CoT (direct generation)	56.59%	Baseline
Full CoT (all tasks)	63.42%	Improves but wastes tokens
RoutingGen (Simple → direct)	65.82%	Simple path avoids overthinking
RoutingGen + ICoT	69.11%	Full system achieves best performance

Key Findings

• Largest gain on McEval (hard tasks) (+45%): ICoT is most effective on problems where algorithm design is critical.
• Skipping CoT on the simple path improves performance: Avoiding overthinking eliminates reasoning-induced errors.
• 46% token reduction: Simple tasks bypass CoT entirely and proceed directly to generation.
• "Intent" outperforms "steps": Conventional CoT provides procedural steps ("first… then…"), whereas ICoT provides algorithmic intent ("use DP because…"), which offers stronger guidance to LLMs.

Highlights & Insights

• Transferring the cognitive economy principle from psychology to LLMs is both natural and effective — not every problem warrants deep deliberation.
• The Specification + Idea structure of ICoT is better suited to code generation than general CoT — it abstracts algorithmic design rather than implementation steps.
• The pattern of dynamic routing combined with dedicated reasoning paths is broadly applicable to other tasks that require varying depths of reasoning.

Limitations & Future Work

• The difficulty classifier may misclassify problems — routing a complex task to the simple path can result in severe performance degradation.
• The binary (Simple/Complex) categorization may be too coarse — finer-grained difficulty levels are needed.
• ICoT quality depends on the LLM's algorithmic knowledge — it may fail for algorithms underrepresented in training data.

Related Work & Insights

• vs. Self-Planning (Jiang et al.): Planning-style CoT without routing. RoutingGen adaptively selects the strategy.
• vs. Scratchpad: Supports intermediate computation but lacks intent abstraction. ICoT operates at a higher level.
• The routing + dedicated reasoning pattern is transferable to domains such as mathematical reasoning and scientific question answering.

Rating

• Novelty: ⭐⭐⭐⭐ The combination of difficulty routing and intention CoT is creative.
• Experimental Thoroughness: ⭐⭐⭐⭐ Three benchmarks, multiple baselines, ablation studies, and token analysis.
• Writing Quality: ⭐⭐⭐⭐ The introduction of cognitive economy is natural and well-motivated.
• Value: ⭐⭐⭐⭐ Practically valuable for efficient code generation.

Related Papers

• [AAAI 2026] RPM-MCTS: Knowledge-Retrieval as Process Reward Model with Monte Carlo Tree Search for Code Generation
• [ACL 2026] Learning to Edit Knowledge via Instruction-based Chain-of-Thought Prompting
• [AAAI 2026] Deep Hidden Cognition Facilitates Reliable Chain-of-Thought Reasoning
• [AAAI 2026] Chain-of-Thought Driven Adversarial Scenario Extrapolation for Robust Language Models
• [ACL 2026] Step-GRPO: Internalizing Dynamic Early Exit for Efficient Reasoning