Researchers ground Qwen2.5-32B in wastewater simulator to improve industrial causal reasoning

Researchers evaluated three methods to ground a frozen Qwen2.5-32B-Instruct model in an architecturally interpretable wastewater simulator (CCSS-IX) for industrial causal reasoning tasks. The approaches included a live simulator oracle (Method 1), structured parameter injection (Method 2), and a Decoupled Recall-Reasoning (DRR) retriever (Method 3).

On a 198-question causal benchmark, the three methods achieved 99.5%, 79%, and 75.8% accuracy respectively, forming a clear deployment ladder above a retrieval-augmented baseline at 48%. The DRR retriever, with 110M parameters, trains per plant in approximately 17 seconds and maintains 88% accuracy when transferred to a biologically distinct plant, whereas Method 2’s static table approach fails to transfer.

On a 60-question counterfactual benchmark focused on intervention queries, only the DRR retriever produced meaningful results, outperforming structured parameter injection by 16.3 percentage points (95% CI: +7.1 to +26.4). The DRR retriever achieved 100% accuracy on timescale and operating-regime categories.

The same selective-retrieval mechanism was tested on the AI2 Reasoning Challenge (ARC) with an OpenBookQA fact corpus, where it reached 79% accuracy compared to 76% for an unconstrained Llama-3.1-8B and 74% for a full-injection approach, indicating a 3 percentage point improvement outside the wastewater domain.