Research map / Large language models / Reasoning and chain of thought

Light-IF: Endowing LLMs with Generalizable Reasoning via Preview and Self-Checking for Complex Instruction Following

AAAI 2026 · Large language models › Reasoning and chain of thought · approach: Efficient and adaptive thinking · 0 citations (OpenAlex)

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.

Light-IF synthesizes hardness-aware complex-constraint prompts, cold-starts Qwen3 models with entropy-preserving SFT on filtered reasoning traces, then runs two-stage easy-to-hard RL with dense rubric rewards and token-wise entropy-adaptive regularization to improve instruction following.

Problem

LLM reasoning models often reason lazily on complex, multi-constraint instructions, restating constraints without checking compliance, which hurts instruction-following accuracy.

Approach

Induce preview and self-checking reasoning via Zero-RL with a length-aware reward on synthetic verifiable constraint prompts, distill the effective traces into a cold-start SFT that keeps entropy high (Entropy-SFT), then train with GRPO-based TEA-RL using per-constraint dense rewards and a token-level entropy regularizer.

Key results

• Light-IF-32B scores 0.575 on SuperCLUE vs 0.436 for DeepSeek-R1-0528 and 0.362 for Doubao-1.6-thinking, and 0.938 IFEval average.
• Light-IF-1.7B-Zero reaches 0.882 IFEval average, above Qwen3-32B (0.877), and 0.391 IFBench average vs 0.384 for Qwen3-32B.
• Ablation: full pipeline (cold-start + EntSFT + TEA-RL, two RL stages) gives 0.231 SuperCLUE and 0.883 IFEval average vs 0.826 for GRPO-only after the same cold start with plain SFT.

Surprising findings

• Zero-RL response length initially drops before rising, and the correctness reward alone fails to encourage longer, effective reasoning, motivating the length reward.
• Sparse binary rewards lead to slow, unstable RL progress compared with dense per-constraint rewards.

Limitations

• Judge-based CFBench scoring depends on DeepSeek-V3 as a model judge.
• Cold-start data for the mixed variant relies on external API models (DeepSeek-R1, Doubao-1.6).
• Hyperparameters for Entropy-SFT and TEA-RL are reported for one setting; sensitivity is not broadly analyzed in the text.

Experimental setup

Datasets: SuperCLUE seed prompts in-house evaluation prompts Self-Instruct expanded prompts AutoIF-style constraint prompts

Benchmarks: IFEval CFBench IFBench SuperCLUE-CPIF

Metrics: accuracy (LP, LI, SP, SI) CSR/ISR/PSR prompt-level and instruction-level accuracy SuperCLUE score

Baselines: Qwen3-1.7B Qwen3-32B Qwen3-235B-A22B DeepSeek-R1-0528 DeepSeek-V3 Doubao-1.6-thinking Doubao-1.5-pro Kimi-K2 ChatGPT-4o KL-Cov

Builds on: GRPO Self-Instruct AutoIF KL-Cov Qwen3

Closest papers by meaning

• When Thinking Fails: The Pitfalls of Reasoning for Instruction-Following in LLMs (NeurIPS 2025 · similarity 0.91)
• Incentivizing Reasoning for Advanced Instruction-Following of Large Language Models (NeurIPS 2025 · similarity 0.91)
• MILES: Modular Instruction Memory with Learnable Selection for Self-Improving LLM Reasoning (NeurIPS 2026 · similarity 0.91)
• Making Slow Thinking Faster: Compressing LLM Chain-of-Thought via Step Entropy (ICLR 2026 · similarity 0.89)
• The Unreasonable Effectiveness of Entropy Minimization in LLM Reasoning (NeurIPS 2025 · similarity 0.89)
• SATURN: SAT-based Reinforcement Learning to Unleash LLMs Reasoning (NeurIPS 2025 · similarity 0.89)