AI
IndustryLLM: Failure-Driven LLM Training for Industrial Procurement
Publication: cs.AI updates on arXiv.org
Author: Liang Ding (Project Lead), Zhiang Xu, Yuyang Sheng, Bin Chen, Songlin Bai, Run Zhu, Dingjun Wu, Hui Xu, Yandi Wang, Fulin Shi, Leilei Gan, Linlin Yu, Qihuang Zhong, Keqin Peng, Yalong Li, Chengfu Huo
Source published: 29/09/2026, 04:00:00 UTC
Briefing created: 29/09/2026, 05:40:21 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 29/09/2026, 05:18:24 UTC.
Briefings
IndustryLLM is an open-weight industrial language model trained from Qwen3.5-35B-A3B-Base using a failure-driven adaptation recipe. It bridges informal buyer jargon with engineering standards via continued pre-training on 100B tokens, supervised fine-tuning, and multi-register rewriting. Offline evaluations show +2.97 percentage points in exact match on procurement queries, while randomized online A/B experiments in production yield +4.25% GMV and +8.3% satisfied inquiries, reducing latency from 6-7 seconds to 1.5 seconds. We formalize an evidence-gated constraint-evaluation interface where unverified evidence remains unknown rather than satisfied.
Key points
- The model IndustryLLM is trained from Qwen3.5-35B-A3B-Base with 35B total parameters and ~3B activated tokens per token, with the vision encoder frozen.
- Offline evaluations demonstrate consistent gains on procurement-query structuring with +2.97 percentage points in exact match, supported by a 95% CI [2.11, 3.86] in No-Think mode.
Context and limitations
Evidence is from an arXiv preprint dated 2026-09-29, with snapshot limitations regarding online A/B experiment results and latency reduction metrics.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 29/09/2026, 05:40:21 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeicw5lrzukq73vhvwtcxaf44hoqakxrsir5chttjzcllsz5opejq7a