Validating greenwashing claims in environmental, social, and governance (ESG) reports relies heavily on costly and inconsistent manual review. To address this, this paper introduces Greench-v1, a low-latency small language model (based on Qwen3-4B) that screens ESG text at the paragraph level. The model outputs a three-way classification (Greenwashing Alert, No Greenwashing, Not Relevant) paired with a concise, paragraph-grounded rationale to assist human auditors in triage and validation. The system was trained on a custom dataset of roughly 2,000 paragraphs, adapted from the ClimateBERT corpus. This dataset mitigates class imbalance through controlled paraphrasing of rare positive instances and uses GPT-4o to generate evidence-based justifications. Four training regimes were evaluated: (i) Hard distillation: Supervised fine-tuning on teacher-generated outputs. (ii) Soft distillation: Training the student to match the temperature-scaled logits of a domain-specialized Qwen3-14B teacher. (iii) Group Relative Policy Optimization (GRPO): Reward-based updates driven by exact-match alert generation. (iv) Hybrid GRPO: GRPO initialized from the hard-distilled checkpoint. Distillation and efficient policy optimization significantly improved performance over untuned baselines. Soft distillation and GRPO achieved the strongest results, increasing the “Greenwashing Alert” weighted F1-score by 36.7% and 49.0%, respectively, resulting in a deployable tool for screening large volumes of ESG narratives.

Large Language Models (LLMs) have shown capabilities in various natural language processing tasks, yet they often struggle with logical reasoning, particularly when dealing with complex natural language statements. To address this challenge, approaches that combine LLMs with symbolic reasoners have been proposed, where the LLM translates the natural language statements into symbolic representations, which are then verified by an external symbolic solver. However, ensuring syntactic correctness in these translations remains a significant challenge. To address this, we propose to constrain the outputs of the LLMs using Grammar-Constrained Decoding, showing that it consistently improves both syntactic correctness and semantic accuracy in logical parsing tasks. Our findings suggest that grammar constraints can serve as an effective substitute for in-context examples, especially beneficial for resource-constrained applications using smaller models.