Estimation of product procurement cost is a key capability for price governance on large-scale e-commerce platforms: from images, structured attributes, comparable products, and pricing knowledge, a practical system must infer a reasonable cost to surface suspiciously inflated or potentially non-compliant listings. Directly invoking general-purpose VLMs is prohibitive at high QPS (queries per second), and their verbose rationales complicate auditing and rule integration. We present Cost-LLM, a compact 7B multimodal cost-estimation model trained by a two-stage post-training pipeline: segment-normalized SFT on progressive pricing rationales, followed by GRPO with an asymmetric MAPE-based reward aligned with cost accuracy. On an in-house benchmark covering 105 fine-grained categories, Cost-LLM reaches MAPE 29.47% and Acc@20 40.42%, yielding the best point estimates among the evaluated systems. Deployed as an auxiliary signal in a price-anomaly pipeline, it provides a 30.73% relative complementary-detection yield over the production baseline at an audited precision-matched operating point. Its locally deployed serving latency is 0.4-0.8 s per listing, compared with 3-8 s observed for external commercial LLM/VLM APIs. We share the design choices, training recipe, and deployment lessons.

Abstractive dialogue summarization aims to convert a long dialogue content into its short form where the salient information is preserved while the redundant pieces are ignored. Different from the well-structured text, such as news and scientific articles, dialogues often consist of utterances coming from two or more interlocutors, where the conversations are often informal, verbose, and repetitive, sprinkled with false-starts, backchanneling, reconfirmations, hesitations, speaker interruptions and the salient information is often scattered across the whole chat. The above properties of conversations make it difficult to directly concentrate on scattered outstanding utterances and thus present new challenges of summarizing dialogues. In this work, rather than directly forcing a summarization system to merely pay more attention to the salient pieces, we propose to explicitly have the model perceive the redundant parts of an input dialogue history during the training phase. To be specific, we design two strategies to construct examples without salient pieces as negative cues. Then, the sequence-to-sequence likelihood loss is cooperated with the unlikelihood objective to drive the model to focus less on the unimportant information and also pay more attention to the salient pieces. Extensive experiments on the benchmark dataset demonstrate that our simple method significantly outperforms the baselines with regard to both semantic matching and factual consistent based metrics. The human evaluation also proves the performance gains.