IntentCoding: Amplifying User Intent in Code Generation

TL;DR

Intent$^ rightarrow$Coding introduces a training-free decoding strategy that amplifies user intent in code generation by computing both original and intent-masked logits to extract an intent signal, then using a multi-strength ensemble over α values to amplify this signal during token selection. The method combines token-level ensemble and beam search integration, enabling robust adherence to multi-constraint prompts without retraining. A new CodeConstraints benchmark is proposed to stress test constraint composability, and extensive experiments across CodeLlama, DeepSeek-Coder, and Qwen2.5-Coder show consistent improvements in constraint satisfaction and functional correctness on CodeConstraints, IFEvalCode, HumanEval, and LiveCodeBench. The work demonstrates that intent amplification can be effective across model sizes and instruction-tuned variants, offering practical gains with manageable computational overhead and broad applicability to open-ended code-generation tasks.

Abstract

Large Language Models (LLMs) have shown strong capabilities in code generation, but their adherence to fine-grained user intent with multiple constraints remains a significant challenge. Our empirical analysis reveals two key observations: 1) Model performance deteriorates quickly as the number of constraints in the user intent increases, and 2) While user intent does influence the model's logits, such an influence may not be strong enough to effectively steer the decoding process. To this end, we propose Intent-Amplified Code Generation (IntentCoding), a novel decoding strategy that enhances an LLM's ability to follow user intent. IntentCoding captures the influence of user intent by masking out the intent, and applies a multi-strength ensemble mechanism to amplify the effect of user intent during generation. IntentCoding is model-agnostic, requires no additional training, and integrates seamlessly with existing decoding procedures. To enable systematic evaluation, we also construct CodeConstraints, a benchmark dataset specifically designed to test user intent compliance under varying numbers of constraints. Experiments on our constructed Constraints, as well as popular IFEvalCode, HumanEval and LiveCodeBench datasets, show that our IntentCoding model significantly improves both constraint satisfaction and functional correctness compared to standard decoding approaches. IntentCoding achieves up to 71.0% relative improvement on CodeConstraints, achieves up to 67.3% relative improvement on IFEvalCode and achieves up to 29.3% relative improvement in pass@1 on HumanEval and LiveCodeBench compared with greedy decoding.

Figures (7)

• Figure 1: An example of generated code given by CodeLlama codellama, where the highlighted constraint is not satisfied. CodeLlama successfully handles an edge case ($n=1$, green ). But the code (red ) violates the required constraints when considering a broader range of inputs.
• Figure 2: Performance of the 7B models on the CodeConstraints dataset across various complexity levels, using greedy search for all evaluations.
• Figure 3: An example of generated code given by DeepSeek-Coder deepseek, where the highlighted token is the wrong token and the top-5 token logits for this position are shown below.
• Figure 4: Overview of Intent$^\uparrow$Coding. We feed the original prompt and intent-masked prompt to the LLM in a same batch (details in Appendix \ref{['app:prompt']}). We extract the intent signal, scale it with multiple strengths $\alpha$, and apply a token-level ensemble to obtain a set of candidate next tokens. These candidates are then used to expand each beam hypothesis; the right panel illustrates beam size $=2$.
• Figure 5: Performance with different beam sizes.